Research Progress on Control Algorithms for Grain Combine Harvesters

Abstract

Intelligent control algorithms are essential for enhancing combine harvester efficiency and minimizing losses, especially as global food demand rises and labor shortages become more severe. This paper provides a comprehensive overview of the evolutionary progression from single-subsystem control to the current core challenge of multi-system co-optimization. We examine the technological development of the cutter, threshing, scavenging, and motion control systems, highlighting persistent bottlenecks that impede global performance improvements due to parameter coupling and conflicting objectives. This review serves as a reference for future advancements in the field. Future research should focus on lightweight reinforcement learning, hybrid control strategies, multimodal perception, and dynamic optimization frameworks for digital twins to drive technological breakthroughs and practical applications.

1. Introduction

In today’s world, the population continues to rise [1], and the demand for food [2,3] is increasing. According to the Agricultural Outlook 2024–2033 report, jointly published by the Food and Agriculture Organization of the United Nations and the Organization for Economic Cooperation and Development, it is projected that halving global food loss and waste by 2030 could lead to a 4% reduction in agricultural greenhouse gas emissions [4] and a decrease of 153 million undernourished people [5]. However, the world will still fail to meet the Sustainable Development Goal (SDG) of eradicating hunger by 2030 [6].

According to the Food and Agriculture Organization of the United Nations (FAO), total global cereal production in 2023 is projected to reach 2.81 billion tons, with losses during the harvesting process estimated to be as high as 8% to 15%. In China, the average operational loss rate for rice and wheat combine harvesters ranges from 5% to 8%, which is significantly higher than that of leading international brands such as Japan’s Yanmar (3%) and Germany’s Claas (2.5%). The primary reason for this disparity is that the control algorithms of domestic combine harvesters lack sufficient intelligence, particularly in complex farmland environments, making it difficult to achieve synergistic optimization across multiple subsystems.

The shortage of agricultural labor and the aging population problem [7] further exacerbate the demand for intelligent agricultural machinery. According to statistics, the proportion of China’s agricultural labor force aged over 55 will exceed 34% in 2022. Meanwhile, countries such as Japan and Germany have already achieved unmanned operation of combine harvesters through high-precision control algorithms. Traditional algorithms face challenges such as insufficient adaptability and strong parameter coupling in complex farmland environments. Therefore, this paper systematically reviews the research progress of control algorithms for various components of grain combine harvesters to provide theoretical references for overcoming technical bottlenecks.

This challenge is accelerating the transformation of traditional agricultural production methods into modern, sustainable agricultural mechanization. With the advent of Agriculture 4.0 and even 5.0 [8], digital technologies such as the Internet of Things, Big Data, and Artificial Intelligence are profoundly reshaping the entire agricultural production process, shifting the traditional experience-driven farming model to a data-driven, predictive, and precise production paradigm [9].

A critical synthesis that systematically connects the theoretical potential of these algorithms with the complex (such as

Table 1. Control algorithm for combine harvester.

Control Algorithms	Characteristics	Applicable Scenarios	Limitations
model predictive control [10]	Strongly inclusive of complex constraints, adaptable to multivariate systems	Flat farmland path tracking, etc.	Dependent on accurate models, limited adaptability to system changes
Self-Adaptive control [11,12]	Self-adjustment of parameters, strong anti-interference	height adjustment of cutting deck, etc.	Stability and convergence are difficult to prove, and there is a lag in adapting to fast time-varying systems.
Classical control (PID)	Good robustness, low dependence on model accuracy	Cutting table control, etc.	Poorly adapted to complex environments
Neural networks [13]	Neural networks Highly nonlinear mapping	Complex terrain path planning	Difficult to guarantee generalization ability, possible control delay when real-time requirement is high
Reinforcement learning [14]	Autonomous learning, trial and error optimization	Dynamic environment Autonomous decision	making Limited model generalization, poor adaptability
Fuzzy Logic Control [15,16,17]	Rule base driven, no need for accurate models	Fuzzy adjustment of clear selection parameters	Insufficient modeling of deep nonlinear dynamics, insufficient adaptation to dynamic conditions, low optimization efficiency
Model + data fusion [18]	Combining model robustness and data flexibility	Multi-parameter coordination of threshing and sorting	Difficult to accurately describe complex time-varying characteristics
Fuzzy adaptive control [19]	Fuzzy rules optimize parameter adaptation	Coupled feed-velocity control	Difficult decoupling of multi-timescale dynamics, limited adaptability to complex time-varying systems
Digital twin control [20]	Virtual-reality interaction, real-time simulation	Full life cycle dynamic optimization	Deficiencies in system robustness and anomaly handling [21]

, etc.), real-world challenges of the agricultural environment is currently lacking. A significant gap exists in the literature: the absence of a comprehensive and systematic analysis explaining why the transition from optimizing individual subsystems to achieving robust, energy-efficient, synergistic control of the entire machine remains a formidable challenge.

To address this gap, this paper offers a new critical synthesis of the field. The primary contributions of this study are highlighted in the following four aspects:

First, this paper provides a comprehensive overview of the evolutionary progression from single-subsystem control to the current central challenge of multi-system co-optimization. We trace the technological development paths of the cutter, threshing, scavenging, and motion control systems, highlighting persistent bottlenecks that hinder overall performance improvement due to parameter coupling and conflicting objectives.

The second and most distinctive contribution of this paper is a critical assessment of the limitations of the dominant control paradigm, achieved by directly linking it to the often-overlooked first-principle constraints.

Third, we construct a comprehensive, multi-dimensional comparison framework that is based not only on performance metrics such as error and response time but also incorporates key real-world factors, including model dependence, computational cost, robustness to uncertainty, and theoretical stability guarantees. This approach provides a more pragmatic and application-oriented assessment of mainstream algorithms than traditional research.

Fourth, based on the above critical analysis, we outline a future-oriented research agenda focused on integrated system-level solutions. Rather than merely listing future trends, we propose a concrete pathway to overcome the cooperative control challenge. This pathway involves a deep convergence of emerging technologies, specifically (a) Digital Twin (DT) [22] frameworks for creating high-fidelity virtual-reality interaction environments that simulate complex multi-system dynamics and enable multi-objective optimization prior to real-world deployment (e.g., balancing loss rate, energy consumption, and operational efficiency using algorithms such as NSGA-III); (b) lightweight [23,24], edge-deployed AI for real-time, airborne intelligent decision-making in network-constrained environments, including lightweight reinforcement learning and compressive neural network models that can run on embedded controllers to adapt to immediate environmental changes; and (c) hybrid and physics-informed models that bridge the gap between purely data-driven and purely model-driven approaches, for example, by combining model predictive control with deep reinforcement learning or utilizing physics-informed neural networks to create more robust and generalizable models of agricultural processes.

Based on the critical analysis above, we have developed a future-oriented research agenda centered on integrated system-level solutions. Rather than merely listing future trends, we propose a concrete pathway to overcome the challenges of cooperative control.

2. Methodology

In this study, we employed a systematic review approach to analyze the existing literature on the characterization of control algorithms for combine harvesters.

We identified the core academic databases used for this review, including Scopus, Web of Science, and IEEE Xplore, which are the most representative and authoritative sources in the fields of engineering, technology, and agricultural sciences. The keyword combinations employed in the searches are detailed in both English and Chinese. These include not only general terms such as “combine harvester” and “control algorithm” but also more specific keywords for each subsystem, such as “cutter” and “control algorithm.” For example, “header height control” was combined with “model predictive control” and “fuzzy logic,” while “path planning” was paired with “reinforcement learning.”

The primary body of literature in this field has been organized over the past five years, with additional entries highlighting the development of control algorithms over time. The scientific search engine Google Scholar was utilized to ensure that the review reflects mainstream techniques and the latest advancements within this period.

Only rigorously peer-reviewed journal articles and high-level conference papers were included to ensure the originality and advanced quality of the content analyzed.

The focus was on grain combine harvesters (e.g., rice, wheat, corn), including research on non-grain harvesters only when the methodology employed is highly generalizable and innovative. For example, IPSO-Fuzzy-PID control research on cabbage harvesters was cited due to the broad applicability of its methodology.

3. Control Challenges in Complex Environments

The grain combine harvester achieves efficient harvesting through the coordinated operation of harvesting, threshing, and cleaning processes. The basic control requirements include the following: (1) Self-adaptive adjustment of the cutting platform height to accommodate variations in terrain and crop height requires real-time adjustments of 2% to 5% [25]; manual adjustments based on traditional hydraulic methods result in 8% to 12% variation. (2) Feed fluctuations must be matched to the threshing gap, as a fixed gap increases threshing losses by 2% to 5%. (3) Optimization of cleaning parameters, including sieve vibration frequency and wind speed, should be automatically adjusted according to crop moisture content.

The current mainstream rice combine harvester (Figure 1) utilizes multi-sensor fusion to achieve adaptive control of the cutting platform height, reducing the loss rate to 3%. In contrast, domestic models still rely on PID control, resulting in a loss rate of 5% to 8% on complex terrain.

Taking the hilly rice fields in the south as an example (

Table 2. Grain combine harvesters face three types of complex environmental challenges.

Type of Environment	Control Difficulties	Typical Impacts	Algorithm Requirements
Hilly terrain	Dynamic changes in cutting deck inclination	Uneven stubble height (error > 10 cm)	Multi-sensor fusion + RABL
High humidity crops	Grain adherence to threshing drum	3~7% increase in unthreshed rate	Humidity feedback + gap adaptation
Unevenly spaced fields	Dramatic fluctuation of feeding volume	Clearance inclusion rate over 5%	Feeding quantity prediction + fuzzy control

), the conventional PID control led to a frequency of cutter collisions which resulted in the standard deviation of stubble height being 8.2 cm; whereas the algorithm based on the MPC could reduce the number of collisions to 0, while the standard deviation of the stubble height was narrowed down from 8.2 cm to 2.1 cm.

Intelligent control of grain combine harvesters relies on a three-layer closed-loop architecture comprising “perception, decision, and execution.” This architecture (Figure 2) achieves efficient operation in complex farmland environments through the synergistic integration of multi-source data fusion technology, dynamic optimization algorithms, and precise actuators. The specific architecture is as follows:

The robustness of a control algorithm depends primarily on the reliability of the sensing system. Unlike industrial robots operating in structured factory environments, sensors in combine harvesters encounter numerous significant challenges that are not merely random noise but involve state-dependent, and sometimes catastrophic, signal degradation or failure.

Machines operate with high levels of vibration and thermal effects, which directly impact the performance of precision sensors such as inertial measurement units. Additionally, widespread dust, rain, and debris generated during crop harvesting in the field can interfere with optical sensors like LiDAR and vision cameras. For instance, LiDAR’s target detection accuracy drops from 80% to 30.9% in light rain and nearly fails in moderate rain. Furthermore, the accuracy of multi-sensor fusion algorithms decreases by more than 8% in dense fog or heavy rain [26].

Crop canopies can act as barriers that attenuate or completely block GNSS satellite signals, leading to a significant decrease in positioning accuracy or even total signal loss. Additionally, woods or buildings at the edge of a field can cause multipath effects, resulting in erroneous positioning data.

Intense daylight, rapidly changing shadows, and low-light conditions during night operations pose significant challenges to the image quality of vision cameras, often resulting in overexposure, underexposure, or loss of image features (

Table 3. Sensor challenges in agricultural environments.

Sensor	Major Challenges	Environmental Causes	Common Mitigation Strategies
GNSS [27]	Signal loss/attenuation, multipath effects	Crop Canopy Shade, Forest Strip Shade	Tight/loose coupling fusion (EKF/UKF) with IMUs/odometers to fill signal gaps using waypoint derivation
IMU [28]	Bias Drift, Noise Integration Error Accumulation	Vehicle vibration, engine heat effect	On-line estimation and compensation of drift (e.g., EKF state vectors contain bias terms), periodic correction using GNSS/visual odometry
LiDAR [29]	Sparse/degraded point clouds, noisy points, measurement failures	Dust, rain, fog, crop foliage shading, bright sunlight	Point cloud filtering algorithms (e.g., statistical filtering, radius filtering), intensity-based filtering, fusion with other sensors
Vision Camera [30]	image overexposure/underexposure, blurring, feature loss	Intense sunlight/shadow bursts, high speed movement, dust, rain	Utilizes HDR cameras, adaptive exposure control, image enhancement algorithms, multi-sensor fusion (e.g., VIO)

).

4. Challenges of Formal Modeling

Formal mathematical models are the cornerstone of modern control theory and are essential for performing robustness analysis and designing high-performance controllers. Currently, a notable feature of the combine control field is the lack of a widely accepted unified theoretical framework that can cover all operating conditions. Establishing a unified, accurate, first-principle-based mathematical model for the entire combine system is extremely difficult under the current technical conditions. This explains why the existing literature is full of hybrid methods, data-driven models and subsystem optimization strategies. The underlying reasons can be attributed to the following three major challenges:

(1)

Biological Variability [31]: Combine harvesters operate on biomass, whose physical characteristics (e.g., crop moisture, density, straw strength, maturity) are highly uncertain, non-uniform, and unpredictable within a field and even within a single trip. This inherent biological variability makes it extremely difficult to accurately and reliably parameterize any physical model [32], and once a model has been established, its validity can be quickly invalidated by changes in crop status.

(2)

Complex Dynamics: The combine harvester is a typical multiple-input multiple-output strongly coupled system. There are complex nonlinear couplings between subsystems such as feed rate, travel speed, drum speed, and fan speed (e.g., fluctuations in feed rate directly affect the load on the threshing drum and the load on the cleaning screen). In addition, there are significant and condition-dependent time delays in material transport within the machine (e.g., from the cutting table to the threshing drum), which presents a significant challenge for any control algorithm (e.g., MPC) that relies on accurate modeling.

(3)

Harsh Operating Environment: Harvesters operate in harsh environments filled with dust, vibration, and crop debris. These factors can severely contaminate sensor measurements (e.g., LIDAR, inertial measurement units), making accurate state estimation and system identification difficult, thus weakening the basis of control models that rely on accurate state observations [33].

5. Evolution and System Analysis of Control Algorithms for Combine Harvesters

5.1. Cutting Table Profiling and Height Control

The structure of the cutting table (Figure 3) consists of four functional units: the grain separator, the cutter, the pivot unit, and the conveyor. The cutting process proceeds as follows: The separator first distributes the crop and directs it to the cutting area. The pivot wheel and pivot plate holder then advance the crop to the cutting position. The cut crop is conveyed to the threshing mechanism via a screw conveyor [34].

As a key component, the cutter can be classified into two types based on its operating principle. The reciprocating design is primarily used in grain harvesters due to its high reliability. Additionally, the cutter body is categorized into horizontal and vertical configurations according to its spatial arrangement. The horizontal configuration is suitable for harvesting short-stalked crops, while the vertical configuration is mainly employed for tall-stalked crops.

Currently, cutting platform profiling and height control primarily employ three technological approaches: mechanical control, hydraulic control, and electro-hydraulic control. Height measurement of the cutting platform mainly relies on detection methods such as image processing, ultrasonic ranging, and multi-sensor fusion. In terms of control algorithms, relevant research includes not only the application of classical fuzzy-based control algorithms but also the exploratory implementation of novel control methods within the system [35].

The core of cutting table profiling control lies in the real-time sensing of terrain undulation and crop height, combined with the dynamic adjustment of the cutting table’s attitude to minimize harvesting losses. Early cutting platform height control systems utilized mechanical-hydraulic mechanisms, requiring operators to manually adjust the platform height through mechanical linkages or hydraulic valves based on their experience. These systems featured simple structures and low costs. However, they exhibited a response lag exceeding 0.5 s, very low accuracy, and a leakage rate of up to 12%.

A PID algorithm has been used for height control in combine harvesters since the 1970s, as demonstrated by a 1984 patent for an “Automatic Header Height Control” system (WO1984004652A1). PID control regulates the system output through error feedback, generating control signals based on height deviations detected by LiDAR or tilt sensors.

A seminal 2013 study by Xie et al. [36] on the fundamental limits of header height control identified an unavoidable time delay inherent in the electro-hydraulic actuator subsystem of up to 0.3 s. This delay is not attributable to a single factor but results from the cumulative duration of several sequential physical events: the time required to open the hydraulic valves, the time to stroke the load-sensing pump, the time needed to build sufficient pressure in the hydraulic lines, and the time to overcome the initial static friction in the cylinder seals and mechanical linkages. This inherent actuator delay imposes a hard physical limit on the achievable closed-loop bandwidth of the control system, which the study calculates to be below 0.53 Hz. Practically, this means the system is physically incapable of effectively responding to terrain variations that occur more rapidly than approximately twice per second.

After 2015, Model Predictive Control (MPC) was introduced into combine harvester cutter control systems. The following objective function was developed by integrating terrain elevation point cloud data, inclination measurements from the inertial measurement unit, and crop density distribution data:

$$\underset{\Delta H}{min}\sum _{k=0}^{N_p}(\parallel H_k-H_{ref}{\parallel }_Q^2+\parallel \Delta H_k{\parallel }_R^2)+\rho ·\parallel F_{\mathrm{v}}{\parallel }^2$$

(1)

where $H_{ref}$ is the desired height; Q, R is the weight matrix; $\Delta H_k$ is the rate of change of height; $F_{\mathrm{v}}$ is the cutter vibration penalty (

Table 4. Key variable naming and definition.

Symbolic	Define	Formula
$H_{ref}$	Desired cutter height	(1)
Q, R	Weighting matrix for balancing control objectives	(1)
$\Delta H_k$	Rate of change of cutter height	(1)
$H_k$	Actual height of the cutting table at moment k	(1)
$F_{\mathrm{v}}$	Cutter vibration penalty term	(1)
$N_p$	Prediction time domain, i.e., the number of steps the model predicts forward	(1)(7)
ρ	Control increment penalty factor or overrun penalty factor	(1)(7)
a	Chaotic sequence in the interval	(2)
d	Chaos parameter, which usually takes the value of 0.7	(2)
n	Population size	(2)
k	Number of iterations	(2)(3)(4)
b	Chaotic sequence that satisfies the range of values of the particle	(2)
$ul,ll$	Upper and lower boundaries of the particles	(2)
${\omega }_{\min }$, ${\omega }_{\max }$	Minimum and maximum value of the inertia weight	(3)
ω	Inertia weights at k iterations	(3)
$k_{max}$	Maximum number of iterations	(3)(4)
α, β	Steepness and location of the midpoint of the characterization curve	(3)
$c_1, c_{21}$	Learning factor	(4)
$c_{1max},$ $c_{1min}$	Maximum and minimum values of the local learning factor	(4)
$c_{2max},$ $c_{2min}$	Maximum and minimum values of the global learning factor	(4)
$x_m$	State vector of the reference model	(5)
$A_m,B_m$	System matrix and input matrix of the reference model	(5)
r	Reference input signal (desired trajectory)	(5)
y(k)	Output of the system at moment k	(6)
u(k)	Output of the controller at moment k	(6)
$N{N}^{-1}$	Inverse dynamics neural network model	(6)
$N_c$	Control time domain	(7)
${\hat{Y}}_{t+k|t}$	Predicted value of the output metric at moment t for moment t + k	(7)
$Y_{ref}$	Reference value of the output metric	(7)
$\Delta U_{t+k}$	Change in control inputs in neighboring moments	(7)
μ	Overrun penalty coefficients	(7)
$[U_{t+k}-U_{max}{]}_{+}^2$	Overlimit penalty term	(7)
$y_i^{*}$	Center value of each fuzzy set	(8)
${\mu }_{max}(y_i)$	Maximum degree of affiliation	(8)
$y^{**}$	Output value after defuzzification	(8)
$M$	Number of fuzzy sets	(8)
i	Summation index	(8)

).

The multi-sensor data fusion model predictive control (MSD-MPC) algorithm proposed by Chunjiang Zhao’s team [18] has been tested in multiple field trials at various operating speeds and header heights. The results indicate that the harvester’s τ values—which represent the proportion of actual stubble retained within the error tolerance—were 90%, 90%, 94%, and 92%, with a mean value of 91.5% and an SSR of 0, at operating speeds of 2, 3, 4, and 5 km/h, respectively. At header heights of 10, 15, 20, and 25 cm, the τ values were 90%, 94%, 91%, and 90%, respectively, demonstrating that the developed header contour control system was highly accurate and stable. Additionally, the coefficient of variation, used to evaluate the responsiveness and stability of the cutter simulation control system, was significantly lower for the proposed header height profiling control system across different operating speeds and header height profiles compared to existing studies. Field experiments further showed that the MSD-MPC algorithm reduced the cutter height control error from ±5.3 cm (conventional PID,

Table 5. Performance comparison of cutting height control algorithms.

Algorithm	Error (cm)	Response Time (ms)	Applicable Terrain	Validation Methods	Test Condition
PID	±5.3	639	Flat farmland	field trial	Grain Flat farmland speed 1.84 km/h
MSD-MPC [18]	±2.1	60–90	Hilly/sloping land	field trial	Rice, slope ≤ 8°, speed 2–5 km/h
Self-Adaptive control [11]	±1.8	54	Slightly undulating	Test beds + simulated field trials	rice–wheat Small obstacles and relatively flat working environment speed 1.62 km/h
Fuzzy control [19]	±1.0	50	Moderate undulations	field trial	Grain Significant topographic relief Speed 1.08 km/h
IDBO-PID [37]	±2.0	42	Complex terrain	simulation + field trials	Grain Complex terrain Speed 1.84 km/h

) to ±2.1 cm and decreased the number of cutter collisions from 2–3 to zero per acre on hilly terrain with slopes ≤ 8°.

The performance of MPC [38] is highly dependent on accurate system modelling. However, in agricultural environments characterized by variable soil and crop conditions, establishing and maintaining model accuracy is extremely challenging. Model mismatches can directly lead to a degradation in control performance. Additionally, the high computational load associated with online rolling optimization poses a significant obstacle for real-time applications. In contrast, fuzzy adaptive control does not rely on precise models and is highly adaptive, but its theoretical foundation has notable weaknesses. The design of fuzzy rules and membership functions depends heavily on expert experience and lacks a systematic design methodology. More importantly, for these nonlinear, heuristic controllers, it is very difficult to formally prove their stability and convergence.

In 2021, Long Z. et al. [11] proposed an indirect method for measuring cutter height based on body and cutter inclination. They validated the performance of a cutter height adaptive adjustment system, which utilizes inclination sensors, through tests assessing control accuracy, system response speed, and simulated field obstacle crossing. The results indicate that the maximum error in cutter height control is 18 mm; the average rising speed of the cutter is 0.22 m/s, while the falling speed is 0.17 m/s. In the simulated obstacle crossing test, under identical conditions, activating the system reduced the coefficient of variation in cutter height stability from 10.77% to 2.79%. This demonstrates that low-cost, real-time adaptive adjustment is achievable.

Fuzzy adaptive control has become mainstream in recent years. For example, Zhang F. et al. [37] addressed the issues of cutter height control accuracy and response speed in unmanned farm combine harvesters. They proposed a height detection compensation method based on dual inclination sensors and established a correlation model between cutter inclination and height using the least-squares method, achieving a correlation coefficient of 0.9958. Additionally, they introduced the Improved Dragonfly Algorithm (IDBO) to optimize the PID controller. The algorithm enhances optimization capability through Bernoulli chaos mapping. The system employs Particle Swarm Optimization (PSO) and t-distributed perturbation. Simulations demonstrate an overshoot of only 0.6% and a response time of 42 ms, significantly outperforming conventional PID controllers. Field tests show that the cutting table rises at 0.44 m/s and falls at 0.32 m/s, with altitude error controlled within 0.02 m. The system adapts effectively to fluctuations in feed volume, including the rise and fall of the cutting table. However, fuzzy adaptive control still faces challenges such as difficulty in stability verification, convergence that depends on initial parameters, and insufficient dynamic decoupling capability across multiple time scales.

In 2024, Zhang C. et al. [19] employed fuzzy adaptive finite impulse response Kalman filtering (FA-FIR-KF), which integrates GNSS and IMU data, to enhance terrain elevation detection accuracy to ±3 cm. This approach reduces the error by 42% compared to traditional extended Kalman filtering, enabling high-precision, stable, and reliable height control of the cutting deck. Several field tests conducted at different running speeds and cutting deck profiling heights demonstrated that the algorithm achieves an average cutting deck height control accuracy of approximately 91.5%, an average coefficient of variation of about 3.5%, and an average number of cutting deck landings of zero.

5.2. Threshing Control

The classification of grain harvesting and threshing technologies primarily depends on the spatial arrangement of the threshing components on the implement. This classification results in three major technological approaches: cross-threshing technology [39], longitudinal threshing technology [40,41], and vertical threshing technology [42,43]. Based on this framework, the current mainstream threshing methods can be summarized into four categories: longitudinal axial flow threshing, cross tangential threshing, vertical axial flow threshing, and pre-cutting and combing threshing.

The development of a loss monitoring and control method for the threshing device (Figure 4) is a critical technical challenge in achieving intelligent harvesting with combine harvesters. Given that a combine harvester is essentially a complex system with multiple components operating in unison, its dynamic working conditions involve significant uncertainties. This study primarily establishes an evaluation system based on two key indicators: crushing rate and entrainment loss rate. It identifies drum speed, concave plate screen opening, threshing gap, and deflector angle as the key control variables and integrates intelligent algorithms—such as neural network control and fuzzy control—to implement optimized control. This approach ensures high stability of threshing quality across different crop types and field conditions [44].

In the 1950s, the Allis-Chalmers G model employed purely mechanical adjustment and fixed gap control, relying on the operator’s expertise to manually adjust the threshing gap. After the 1980s, simple threshold controls were introduced on harvesters, such as the strain gauge torque sensor on the John Deere 9500. However, the accuracy of these adjustments remained limited by mechanical constraints, resulting in gap tolerances of ±3 mm. In the 1970s, the Chinese tractor thresher used manual bolts to adjust the gap, which led to an 8% failure rate due to fluctuating feed rates.

In the mid to late 20th century, PID algorithms were first applied to combine harvesters, reducing losses to 5–8%. However, the PID algorithm has limitations in adapting to nonlinear and time-varying systems. Currently, improvements to the PID algorithm primarily focus on optimizing traditional PID parameter tuning by incorporating adaptive mechanisms, fuzzy logic, neural networks, and other intelligent strategies to enhance control accuracy and robustness under complex operating conditions.

In 2024, Yu Y. et al. [45] proposed a fuzzy neural network PID control algorithm for the real-time adjustment of the threshing concave plate gap in a corn combine harvester. Simulink simulations verified that its response speed was 30% faster than that of a traditional PID controller. Field tests demonstrated that the system reduced the entrainment loss rate from 1.28% to 0.72%. However, the fuzzy neural network PID control algorithm faces several challenges: the monitoring signal is limited by the sensor installation position and the characteristics of the material flow field; there is a need to develop a multi-objective synergistic optimization model addressing entrainment loss, crushing rate, and impurity rate; it lacks dynamic adaptability under high-speed extreme working conditions; and it depends heavily on specific crop parameters. This dependence limits the model’s generalizability. Therefore, further research is required in multi-sensor fusion monitoring, multi-objective control strategies, and cross-condition robust optimization.

IPSO introduces three improvement strategies based on PSO: chaotic mapping, nonlinear inertia weights, and asymmetric learning factors. The specific formulas are as follows:

(1) Chaotic Mapping: To address the challenge of achieving a uniform distribution of randomly initialized populations within the boundary range of the problem dimension in the PSO algorithm, the introduction of chaotic sequences in place of random initialization can enhance the population diversity. The mathematical model is as follows:

$$a_{k+1}^n=\{\begin{array}{cc}{\displaystyle \frac{a_k^n}{d}}& a_k^n\in (0,d]\\ {\displaystyle \frac{1-a_k^n}{1-d}}& a_k^n\in (d,1]\end{array}\hspace{1em}{b}_j^h=a_k^n(ul-ll)+ll$$

(2)

where a is the chaotic sequence in the interval [0, 1]; d is the chaos parameter, which usually takes the value of 0.7; n denotes the population size; k represents the number of iterations; b is the chaotic sequence that satisfies the value range of the particle; ul and $ll$ are the upper and lower boundaries of the particle, respectively.

(2) Non-linear decreasing adjustment weights: In the standard PSO algorithm, if the inertia weights are fixed, the algorithm is prone to insufficient global search capability and slow convergence. In this context, a nonlinear decreasing inertia weight adjustment strategy is introduced to help improve the global convergence performance of the algorithm. The mathematical model is as follows:

$$\begin{array}{c}\omega ={\omega }_{\min }+{\displaystyle \frac{{\omega }_{\max }-{\omega }_{\min }}{1+e^{-[\alpha (\frac{k}{k_{\max }}-\beta )]}}}\end{array}$$

(3)

where ${\omega }_{min}$ and ${\omega }_{max}$ are the minimum and maximum values of the inertia weights; $k_{max}$ represents the maximum number of iterations; α and β are used to characterize the steepness of the curve and the location of the midpoint, respectively.

(3) Improved learning factors: The introduction of an asymmetric learning factor enhances the individual learning capability of the PSO algorithm during the early stages of the search. In contrast, during the later stages, it emphasizes strengthening the algorithm’s global search performance. This approach is mathematically modeled as follows:

$$\begin{array}{ccc}{c}_1(k+1)& =& c_{1max}-(c_{1max}-c_{1min}){\displaystyle \frac{k}{k_{max}}}\\ c_2(k+1)& =& c_{2max}-(c_{2max}-c_{2min}){\displaystyle \frac{k}{k_{max}}}\end{array}$$

(4)

where $c_{1max}$ and $c_{1min}$ are the maximum and minimum values of the local learning factor, respectively; $c_{2max}$ and $c_{2min}$ are the maximum and minimum values of the global learning factor, respectively.

In 2025, Zheng J. et al. [46] designed an adaptive control system based on an IPSO-fuzzy PID controller for a cabbage harvester cutter, which typically suffers from high damage due to parameter mismatching (Figure 5). They optimized the fuzzy PID parameters by introducing adaptive inertia weights and the helix position updating mechanism of the Whale Optimization Algorithm. Additionally, they developed a negative feedback control model for the transverse and longitudinal displacements of the clamping and root-cutting mechanisms. Simulations demonstrated that the system’s response time was only 0.146 s without overshoot. Indoor tests at speeds of 0.1–0.5 m/s achieved an average harvesting qualification rate of 97.19%, with displacement errors of 1.31 mm transversely and 0.92 mm longitudinally. Field tests on multiple cabbage cultivars at speeds below 0.4 m/s yielded a harvesting qualification rate exceeding 96.42%.

Fuzzy control algorithms have been applied to threshing control since 1990. The principle involves mapping inputs, such as drum speed error and load current, into fuzzy sets, constructing IF-THEN rules based on expert knowledge, and then converting the fuzzy outputs into precise control values using the center of gravity method or the maximum membership degree method.

Xu B. et al. [16] addressed the issues of loss rate and grain breakage in combine harvester threshing by considering uncertainties such as environmental noise and other disturbances. They designed a control strategy based on a Type-2 Fuzzy Logic Controller for the threshing separation component. Their approach involved conducting experiments to acquire basic data on the threshing system, establishing the control strategy, and evaluating its performance using the IT2-FLS MATLAB Toolbox 1.0.0.0. The Type-2 fuzzy logic controller was found to outperform the Type-1 controller in terms of integral squared error and integral absolute error.

In 2010, adaptive gap control was introduced for combine harvester threshing systems, dynamically coupling the feed rate and gap settings. A recursive least squares method was employed to update the model parameters online, allowing adaptation to changes in crop moisture content. A reference model was developed to represent the ideal threshing dynamics, and global convergence was ensured based on Lyapunov stability theory.

$${\dot{x}}_m=A_m{x}_m+B_{m}r$$

(5)

where ${\dot{x}}_m$ is the state vector of the reference model; Am, ${\mathit{B}}_{\mathit{m}}$ are the system matrix and input matrix of the reference model; $r$ is the reference input signal (desired trajectory).

The threshing gap is fixed, meaning it cannot be adjusted when the feed rate fluctuates. This limitation leads to reduced threshing efficiency, increased power consumption, and potentially higher rates of grain loss and breakage. Fan C. et al. [47] developed a threshing unit with an automatic threshing gap control system based on the feeding rate.

In 2000, Neural Network Inverse Control (NNIC) was applied to threshing control by using a backpropagation (BP) network to construct an inverse dynamic model of the threshing system. The network weights were updated online through a gradient descent algorithm to approximate the optimal control input.

$$u(k)=N{N}^{-1}(y(k+1),y(k),u(k-1))$$

(6)

where $\mathit{y}(\mathit{k})$ is the output of the system at moment k (e.g., threshing drum speed); $\mathit{u}(\mathit{k})$ is the output of the controller at moment k (e.g., motor control volume); $\mathit{N}{\mathit{N}}^{-\mathbf{1}}$ is the inverse dynamics neural network model.

Lan M. et al. [13] developed a BP neural network model (Figure 6) using threshing drum rotational speed, concave sieve threshing clearance concave sieve separating clearance, and feed quantity as input variables. The output variables were crushing rate, impurity rate, and loss rate, representing the nonlinear threshing process of a flexible threshing device. The model was validated using 25 test data sets, yielding a correlation coefficient of 0.980, a RMSE of 0.139, and mean absolute error values. Through training and validation, the regression coefficient of determination between predicted and experimental values was also obtained. Sensitivity analysis identified rotational speed (15.00%), threshing clearance (14.89%), and separating clearance (14.32%) as the key factors influencing threshing performance, while feed quantity had the least impact (11.65%). This data-driven modeling approach elucidates the influence of key parameters, facilitating performance optimization of the flexible threshing device. The study employs the backpropagation neural network algorithm to construct the predictive model, overcoming the limitations of traditional mechanism-based models and effectively capturing complex nonlinear relationships.

The threshing process of a combine harvester is a complex, time-varying, nonlinear system characterized by strong multivariate coupling. Its control accuracy directly impacts the grain loss rate and impurity rate. Digital twin technology enables closed-loop control through a “data-driven model prediction and real-time optimization” framework by constructing a high-fidelity virtual representation of the physical threshing components—namely, the drum, concave plate, and cleaning device—and by establishing objective functions alongside constraint processing.

$$J=\sum _{k=1}^{N_p}\parallel {\hat{Y}}_{t+k|t}-Y_{ref}{\parallel }_2^2+\rho \sum _{k=0}^{N_c-1}\parallel \Delta U_{t+k}{\parallel }_2^2+\mu \sum _{k=0}^{N_c-1}[U_{t+k}-U_{max}{]}_{+}^2$$

(7)

where $N_p$ is the prediction time domain; $N_c$ is the control time domain; ${\hat{Y}}_{t+k|t}$ is the predicted value of the output indicator at time t for time t + k; $Y_{ref}$ is the reference value of the output metrics; $\rho$ is the control incremental penalty factor; $\Delta U_{t+k}$ is the change in control inputs in adjacent moments; $\mu$ is the overrun penalty coefficient; ${[U}_{t+k}-U_{max}{]}_{+}^2$ is the overrun penalty term.

In 2025, Guo D. et al. [48] proposed a digital twin (DT)-based online optimization method for the highly nonlinear and uncertain threshing process (Figure 7). This method employs a DT to predict future grain breakage rate (GBR) trends and integrates genetic algorithms to dynamically optimize threshing parameters. It actively adjusts these parameters before the GBR exceeds a predefined threshold, thereby avoiding the lag associated with traditional feedback control. Experimental results demonstrate that, compared to manual and feedback control, the online optimization method based on digital twin (OOMDT) reduces GBR by 2.08% and 1.00%, respectively, and increases operating speed by 1.12 km/h and 1.47 km/h. This improvement corresponds to a reduction of 10–20 kg of grain breakage per ton, significantly enhancing economic efficiency. Notably, this method introduces the concept of predictive maintenance into agricultural machinery control for the first time, providing a new paradigm for real-time decision-making in complex agricultural systems.

Neural networks can effectively handle the complex nonlinear mapping relationships in the threshing process, eliminating the dependence on mechanistic models. However, its “black box [49]” characteristics make the control process difficult to interpret, and the stability and generalization capabilities of the model are challenging to guarantee theoretically. Optimization algorithms such as IPSO [50] can identify better controller parameters, but their theoretical guarantee of convergence to the global optimum is weak, and their performance is sensitive to hyperparameter settings. Although digital twin methods [51] show promise, they heavily depend on the high fidelity of the virtual model and the real-time accuracy of data interaction between the physical and virtual worlds. The high cost of model construction and calibration remains a significant challenge for practical application.

5.3. Motion Control

Path planning technology (Figure 8) in the field of agricultural equipment can be categorized into global path planning and local path planning based on the planning scale [52]. Global path planning considers the entire operation area as the optimization domain and enhances operational efficiency by generating global paths that maximize coverage and minimize operational costs. In contrast, local path planning focuses on dynamic environmental responses, ensuring that agricultural machinery performs tasks safely and efficiently under complex field conditions by adjusting the trajectory in real time.

Motion control of combine harvesters must achieve high-precision path tracking, dynamic matching of feed velocity, and multi-machine cooperative operation in complex farmland environments. The technical challenges include spatiotemporal asynchrony of heterogeneous sensor data, nonlinear dynamic modeling, and algorithm lightweighting under stringent real-time constraints. Precision agriculture demands point-to-point navigation [53].

In 2021, Li S. et al. [54] proposed a FA-FIR-KF algorithm (Figure 9) to improve navigation and positioning accuracy and stability, addressing the issue of low GNSS localization accuracy caused by noise in the automatic navigation of agricultural machines. Static fixed-point filtering and positioning experiments demonstrated that the algorithm can correct GNSS non-differential static positioning data to values close to the differential true value, significantly reducing positioning error. Tobias Peschke et al. [55], focusing on the propulsion system of a combine harvester, derived a nonlinear model to reduce noise and exhaust emissions while simplifying the operator’s task, and applied MPC to the combine harvester’s cruise control system, achieving performance significantly superior to traditional control schemes. Zhang Y. et al. [56] proposed an adaptive neural network control model (Figure 10) for autonomous combine harvesters and designed a weight update law based on Lyapunov stability theory to address input saturation and parameter uncertainty. Comparative tests conducted in wheat and rice fields showed that the lateral tracking error of this algorithm was 0.033 m, which is 28% lower than that of a traditional PID controller, with a marked improvement in terrain adaptability.

In 2022, He J. et al. [10] used the body of agricultural machinery operating in paddy fields as the control object and developed a linear MPC path-tracking controller based on the attitude kinematics model of the machinery. They jointly optimized lateral deviation, heading angle error, and speed fluctuation within the objective function. In field tests involving three-line linear path tracking, the average root-mean-square error was 0.043 m, and the maximum absolute error was 0.056 m, representing a 40% improvement in accuracy compared to traditional PID control. However, the algorithm is prone to accumulating state estimation errors when operating on ground with a low adhesion coefficient, and the linear weighting assumption in the multi-objective optimization fails to capture the nonlinear interactions among strongly coupled variables.

In 2023, Deng L. et al. [12] developed an adaptive algorithm for the coupled feed volume–travel speed system of rice harvesters operating in the hilly mountainous regions of southern China. The feed volume was calculated in real time by detecting the torque and rotational speed of the power shaft of the tilting conveyor, while the travel speed was dynamically adjusted based on a fuzzy rule base. Field tests demonstrated that when the feed volume increased from 6 kg/s to 10 kg/s, the travel speed decreased from 1.5 m/s to 0.9 m/s, the response time was ≤1.5 s, and the harvesting loss rate remained below 3.4%.

He Y. et al. [57] used the cone index to classify paddy ground conditions and constructed an eigenvector to characterize PGC based on the vibration acceleration of the harvester. A particle swarm optimization support vector machine (PSO-SVM) algorithm was employed to identify PGC categories. A steering control model, adapted to the three PGC categories, was developed using the orthogonal test method, and a new adaptive tracking control strategy was implemented to follow the target path. Field path tracking tests demonstrated that the standard deviation of the harvester’s lateral deviation was 0.053 m, 0.039 m, and 0.045 m, while the standard deviation of the heading deviation was 1.120°, 0.895°, and 0.877°, respectively, for each PGC category of paddy field. The algorithm enhances straight-line path tracking performance and enables the crawler combine harvester to operate stably in paddy fields with varying ground conditions, achieving high path tracking accuracy. However, the method relies on feature vectors derived from the cone index and vibration acceleration to characterize ground conditions, making its classification accuracy susceptible to sensor noise and the quality of feature engineering. Additionally, the SVM exhibits limited generalization ability to untrained ground categories, and control bias occurs in the fuzzy region near category boundaries.

Conventional GNSS/IMU fusion positioning typically exhibits a positioning error exceeding 20 cm due to multipath effects and IMU zero-bias instability [58]. In contrast, the FA-FIR-KF algorithm dynamically adjusts the process noise covariance matrix using fuzzy rules, reducing the root mean square error of non-differential positioning from 0.35 m to 0.21 m. However, the parameters of its fuzzy affiliation function still rely on grid search. Although MPC achieves a 15% reduction in fuel consumption through multi-computer collaboration, the communication delay in its distributed solver results in a 40% decrease in the Nash equilibrium convergence speed. DRL performs end-to-end path tracking optimization via policy gradient methods (Proximal Policy Optimization algorithm), but the discrepancy between simulation and real-world conditions leads to an overestimation of the Q-value function during field trials.

With its strong robustness, low model dependence, and excellent anti-interference capabilities, ADRC (Figure 11) has demonstrated significant advantages in managing complex, unstructured agricultural environments—such as soil variability, climate perturbations, and system nonlinearities. It has been successfully applied to the motion control of agricultural equipment, including steering control of unmanned ground vehicles, achieving a response time of less than 2 s and an accuracy improvement exceeding 50% [59]. Additionally, ADRC has been utilized for high-precision navigation and field tracking, achieving zero convergence of path tracking error under slip compensation. In agricultural production processes (Figure 12), it has reduced the steady-state error of weeding mechanisms to less than 6 mm and attained a sweet vine crushing pass rate of 89.41%. These applications significantly enhance operational efficiency and control stability. Further performance improvements are achieved through technology integration: linear ADRC simplifies parameterization via bandwidth tuning and, when combined with sliding mode control, enhances system robustness, resulting in a 37% improvement in position control accuracy of agricultural machines compared to single ADRC.

Automated agricultural ground vehicles (AGVs), including automated tractors and combine harvesters, face significant navigation challenges when operating in the unstructured environments typical of agricultural fields. These environments are characterized by heterogeneous crop distributions, uneven terrain, and the presence of transient obstacles. The dynamic and unstructured nature of such settings requires robust navigation systems that can adapt to variable conditions while simultaneously improving operational efficiency and minimizing energy consumption. Although conventional navigation methods have made progress through sensor fusion and algorithmic enhancements, their inherent limitations become increasingly evident in complex agricultural scenarios. In contrast, Deep Reinforcement Learning (Figure 13) offers a promising alternative. By enabling autonomous interaction with the environment and continuous policy optimization, DRL reduces reliance on predefined models and provides a novel approach to overcoming the constraints of traditional navigation techniques in unstructured agricultural terrains.

Wang et al. [60] proposed a nested and hybrid helicoidal scaling method, MRS-NM, to address the challenge of achieving full coverage operations for autonomous navigation path planning in agricultural machinery. Yu et al. [61] developed a high-performance obstacle avoidance control method based on reinforcement learning by embedding an improved neural network model into the Double DQN architecture and conducted field tests. The average results for the shortest distance during obstacle avoidance, the length of the obstacle avoidance trajectory, and the time required for obstacle avoidance were 2.37 m, 0.53 m, and 2.7 s, respectively. Zhang et al. [62] proposed a fuzzy speed regulation system based on IPSO-SVM to address the problem of combine harvester clogging caused by high feeding speeds during field operations. The system’s reliability was verified through field tests, demonstrating the ability to adjust speed within 0.5 to 2 s in response to slight clogging. Building on the DRL framework, Li H et al. [14] constructed a motion model for agricultural machinery and designed a reward function to jointly optimize path tracking precision and power consumption. Simulation and field test results show that the DRL controller achieves a lateral tracking error of 0.053 m with a standard deviation of σ = 0.012, representing a 42% improvement over the conventional Pure Pursuit algorithm, along with a 15% reduction in energy consumption.

Current research on path planning for unmanned agricultural systems shows a trend toward multi-dimensional optimization. Addressing the limitations of classical algorithms, Shen et al. [63] corrected the path point offset problem in the A* algorithm using tree row position information. Xu et al. [64] optimized the A* algorithm to reduce the number of global path loops and improve path smoothing. Feng et al. [65] introduced a global/critical region dual sampling strategy into the RRT algorithm to enhance search capability and path feasibility. Kong et al. [66] combined Bessel curves with kinematic constraints to achieve global path smoothing with a minimum turning radius and continuous curvature.

To address the local optimality defect of the artificial potential field method, Wu et al. [67] employ an annealing algorithm combined with an obstacle-exclusion potential field function to achieve effective circumvention. Meanwhile, Boryga et al. [68] utilize polynomial transition curves to optimize the steering trajectory, significantly reducing non-operating distance and improving energy efficiency. At the algorithmic convergence level, two paradigms emerge prominently: first, local planning techniques enhance the stability of tracked harvesters by reducing steering frequency [69]; second, the boundary between global and local planning becomes increasingly blurred [53]. This is evident in the development of improved global algorithms—such as multi-objective genetic algorithms [70]—that can be applied to local scenarios, as well as local optimization strategies that can be adapted for global planning. This convergence trend improves all-path coverage by integrating time or distance cost minimization, adaptation to external obstacles, and energy consumption constraints.

Significant breakthroughs have also been made in the fields of random sampling and bio-inspired algorithms. The RRT family of algorithms enhances the efficiency of complex spatial exploration through mechanisms such as bidirectional scaling (RRT-Connect [71]) and safe region modeling (SRT [72]), while RRT* further approximates the optimal solution. In bio-inspired optimization, multi-objective decision-making techniques (AHP [73]), hybrid strategies (Voronoi diagram–ant colony [74]), as well as MOPSO [75] and non-dominated sequential genetic algorithms [76] are widely employed to synergistically optimize path length, safety, smoothness, and energy efficiency. Notably, dynamic environment adaptation has emerged as a new focus; for example, enhanced genetic algorithms [77] and dynamic MOPSO [78,79] have been applied to unstructured scenarios where water flows and obstacles coexist. This development indicates that path planning research is evolving toward multi-constraint coupling, cross-level fusion, and dynamic robustness.

MPC [80] performs well in path tracking; however, its linear modeling assumptions limit its ability to handle the strong nonlinear dynamics of the vehicle, especially on surfaces with low coefficients of adhesion (e.g., slippery paddies), which tend to cause the accumulation of state estimation errors. Although adaptive neural network control [81,82] can manage uncertainty, its stability proof relies on Lyapunov theory, and the computational effort required for online weight adjustment can be substantial. DRL [83] shows great potential as an end-to-end learning method, but its primary challenge is that a policy trained perfectly in a simulation environment may perform significantly worse in the real world, which is full of noise and uncertainty. Additionally, the DRL training process involves extensive trial and error, exhibits low sample efficiency, and lacks theoretical guarantees for stability.

5.4. Sorting Device Control

The grain cleaning system of a combine harvester achieves efficient material separation through the combined action of multiple physical fields (Figure 14). Its core components include a grain pan, a double-deck screen, and a directional airflow device [34]. The system operates based on differences in the physical properties of materials: (1) the aerodynamic separation stage utilizes the vertical airflow generated by the fan to perform primary sorting by exploiting the terminal velocity differences between grain and impurities [84]; (2) the mechanical screening stage relies on the synergistic effect of the oscillatory motion of the screen mesh and gravity, combined with geometric screening principles, to accomplish secondary separation based on particle size and shape differences [85]. A specially designed chaffer extension enhances tailings separation efficiency by prolonging material residence time, while the inclined conveyor mechanism facilitates the directional transport of clean grains. This composite separation system enables graded processing: low-density impurities are removed through pure pneumatic separation, whereas medium-density particles require pneumatic-mechanical coupling for effective separation. This multimodal separation mechanism significantly improves adaptability across different crop varieties and enhances cleaning efficiency.

Research on the structural optimization of combine harvester cleaning systems has revealed the influence of different fan configurations on separation efficiency [86]. It has been noted that the traditional single-outlet fan system, although technologically mature, suffers from non-uniform airflow distribution, leading to significant particle losses and higher impurity ratios. This issue stems from the difficulty of achieving laminar separation of materials within the turbulence field generated by a single duct [87,88]. To address this, the dual-outlet fan system regulates airflow momentum through graded air supply from upper and lower outlets: the upper airflow performs pre-separation for grain tray dropout, while the lower airflow enhances the fluidization effect on the screen surface. This design improves the discharge efficiency of lightweight MOG (Material Other than Grain) by 19% and has been adopted in mainstream models such as the Claas Lexion 770 (Manufactured by CLAAS in Hasewink, North Rhine-Westphalia, Germany). Further development of the multi-fan system, featuring a staggered-flow centrifugal fan combination [89], achieves long and short straw separation efficiencies of 92% and 85%, respectively, through multi-physical field coupling. Its technological advantages have been validated in models such as the Kubota Pro688Q (Manufactured by Kubota Corporation, produced in Utsunomiya, Japan). However, the gap effect caused by the juxtaposition of multiple fans results in a 34% reduction in airflow rate in localized areas of the screen surface [90] and a 22% increase in system power consumption, leading to diminishing marginal returns. Studies [91,92] have shown that the cleaning system configuration must establish a dynamic balance among airflow uniformity, separation accuracy, and energy efficiency, providing a theoretical foundation for subsequent optimization of the fan system based on computational fluid dynamics.

Before the 20th century, control systems were primarily traditional, relying on PID control to adjust parameters such as sieve amplitude and fan speed through sensor feedback. After the 20th century, fuzzy control was introduced into combine harvester scavenging systems. The control principle involves constructing a fuzzy rule base based on expert experience, fuzzifying sensor signals for impurity rate and loss rate into linguistic variables such as high, medium, and low and determining the control output through Mamdani or TSK inference methods.

The development of intelligent control technology for grain cleaning systems is primarily demonstrated through the advancement of closed-loop control strategies. Maertens [93] proposed a regulation method based on monitoring the upper sieve crop load threshold, which dynamically adjusts the upper sieve position to optimize material return by detecting load deviations. Additionally, he developed a sieve overload control algorithm to initiate the optimal control action.

Craessaerts et al. [94] designed a closed-loop fan speed control system that utilized grain loss sensor readings and workload data in combination with fuzzy logic [95] to process differential air pressure signals, significantly reducing grain losses compared to static fan settings. Omid et al. [15] extended closed-loop control to a threshing, separation, and cleaning system, relying solely on grain loss sensor data and applying fuzzy logic. The system characterized stem entrainment loss and upper sieve loss as input variables using fuzzy sets, achieving a significant reduction in grain loss compared to static fan settings. This characterization effectively reduced particle loss in both subsystems and demonstrated the efficacy of fuzzy logic in addressing nonlinear, distributed, parameter-uncertain, or model-unknown systems. Liang et al. developed a fuzzy logic controller by integrating linguistically based expert knowledge [96] to construct control rules and employed a monitoring device to obtain real-time information on sieve grain loss, thereby optimizing the cleaning system. Linde [97] proposed a load-responsive control strategy that calculates the fan drive load based on crop parameters and dynamically adjusts fan speed by comparing it with a predefined threshold. In 2020, Chai X. et al. [98] developed an adaptive system integrating grey prediction and fuzzy control to address the nonlinear, time-varying, and time-lag issues of cleaning loss in oilseed rape combine harvesters, achieving real-time optimization of cleaning loss through an integrated cleaning loss sensor and a louvered sieve opening adjustment mechanism.

Wold et al. [99] described a system integrating a residue separation device, an airflow-guided fan, a crop characteristic sensor, and a controller that adjusts wind speed in real time based on sensor signals for precise residue management. In recent developments, Li W. et al. [17] proposed an improved FLC algorithm by combining FLC theory with big data and fuzzy control algorithms for a cleaning system installed in a combine harvester. This approach enables real-time synergistic adjustment of multiple parameters of the cleaning sieve, significantly enhancing adaptability to complex field conditions.

$$y^{**}={\displaystyle \frac{{\sum }_{i=1}^M(y_i^{*}{\mu }_{max}(y_i))}{{\sum }_{i=1}^M{\mu }_{max}(y_i)}}$$

(8)

where ${\mathit{y}}_{\mathit{i}}^{*}$ is the center value of each fuzzy set; ${\mathit{\mu }}_{\mathit{m}\mathit{a}\mathit{x}}({\mathit{y}}_{\mathit{i}})$ is the maximum degree of affiliation.

In 2022, Li Y. et al. [100] employed a system identification method to model the threshing and cleaning system. They proposed a multi-parameter collaborative state-space model and utilized a particle swarm optimization wavelet neural network (PSO-WNN) for system identification. The model achieved a variance explained rate of 81.7% and a RMSE of ≤0.602, significantly outperforming the traditional ARX model.

In 2023, Wu J. et al. [101] developed a self-leveling cleaning screen device and a control system based on a fuzzy PID control algorithm (Figure 15), enabling a combine harvester to operate on gentle slopes of up to 10°. Indoor test results demonstrated that the system exhibited effective tracking performance when the inclination angle did not exceed 10°. Liu P. et al. [102] investigated the impact of scavenging parameters on scavenging quality through experimental tests and established a linear relationship between these parameters and evaluation indices. The experiments showed that, under fixed parameter control, the impurity rate increased from 1.8% to 4.5% as the feeding rate rose from 4 kg/s to 12 kg/s. However, by dynamically adjusting parameters f and v, the impurity rate was stabilized at 2.2% ± 0.3%.

The primary advantage of FLC [103] is that it does not require an exact mathematical model and can effectively leverage expert knowledge to address nonlinear problems. However, its main drawback arises from this reliance: ensuring the completeness and optimality of the rule base is challenging, and the system’s performance heavily depends on the quality of expert knowledge and the tuning process. Additionally, the number of fuzzy rules increases exponentially with the number of input variables, leading to the curse of dimensionality. Similar to other intelligent algorithms, FLC lacks rigorous mathematical proofs for stability and convergence. Although system identification-based methods can develop more accurate data-driven models, their accuracy depends on the quality and coverage of the data used for identification, and their generalization ability is limited when applied to conditions beyond the training data.

6. Transient Response Challenges for Electrohydraulic Systems

The performance of any control algorithm is ultimately constrained by the physical system it governs. In combine harvesters, the primary actuators responsible for functions such as adjusting the height of the cutting deck are predominantly electro-hydraulic systems [104]. Although these systems can deliver substantial actuation power, their transient response characteristics present considerable control challenges.

6.1. Inherent Time Delay

As noted in the seminal study by Xie et al., there is an unavoidable time delay—typically greater than 0.3 s—in electrohydraulic subsystems. This delay is not caused by a single factor but rather results from the accumulation of several physical processes, including valve opening time, stroke time of the load-sensitive pump, the time required to build sufficient pressure in the hydraulic lines, and the time needed to overcome the initial static friction in the cylinder seals and mechanical linkage. This inherent actuator delay imposes a hard physical upper limit on the bandwidth of closed-loop control, rendering the system physically incapable of responding efficiently to high-frequency terrain changes [105].

6.2. Non-Minimum Phase Behavior

A More Subtle Challenge: A more subtle yet equally critical challenge in some hydraulic systems is non-minimum phase (NMP) behavior. NMP systems are characterized by an initial response to a control input that moves in the opposite direction of the final steady-state response [106]. For example, a command to raise a cutter may initially cause a slight, momentary dip in the cutter before it begins to rise. This phenomenon is notoriously difficult for standard controllers to manage. If the control algorithm fails to accurately model and compensate for this behavior, it can mislead the controller into making incorrect adjustments, potentially causing system oscillations or even instability [107]. Although this phenomenon is poorly documented in the harvester-specific literature, it is a recognized feature of complex hydraulic and mechanical systems and represents a crucial yet understudied aspect of harvester control.

Therefore, analyzing output response times and characterizing the transient behavior of control actions—especially in the context of NMP dynamics—is an important future direction for connecting theoretical algorithms with real-world field performance [108].

7. Discussion

Current research on control algorithms for grain combine harvesters has gradually shifted from optimizing individual subsystems to achieving multi-system cooperative control; however, significant limitations remain (

Table 6. Control algorithm comparison.

Algorithm Types	Real-Time	Model-Dependent	Multivariate Processing	Academic Hotspots
PID [112]	High	Low	Poor	Adaptive Gain Scheduling Expert system fusion
Fuzzy control [113]	Medium	Low	Medium	1. genetic algorithm optimization rule base 2. Deep learning enhancement 3. Multi-objective optimization
MPC [114]	Low	High	Excellent	Distributed solver acceleration
DRL [115]	Medium	None	Excellent	Lightweight networks and migration learning
Hybrid control algorithms [116]	Medium	Medium	Excellent	1. PID + fuzzy control 2. MPC + DRL 3. expert system + deep learning
Digital twin system [48]	Medium	High	Excellent	1. Lightweight model construction 2. Real-virtual interaction optimization 3. Edge computing deployment
Particle Swarm Optimization Algorithm PSO [117]	Low	Low	Excellent	Dynamic parameter tuning with hybrid strategy

). In cutting deck profiling and height adaptive control, MPC combined with multi-sensor fusion technology significantly enhances adaptability to hilly terrain, but reliance on accurate models restricts generalizability in dynamic farmland environments. In threshing control, algorithms based on dynamic adjustment of the feed gap reduce unthreshed grain and breakage rates; nevertheless, multi-objective co-optimization still faces challenges such as strong parameter coupling and increased energy consumption. For cleaning device control, the application of fuzzy logic and neural networks improves the stability of impurity rates, but insufficient modeling accuracy of multivariate nonlinear relationships limits control efficiency. Additionally, although path planning and motion control benefit from DRL and multi-sensor fusion to improve navigation accuracy, real-time responsiveness to complex terrain and variable crop distribution remains inadequate. Overall, existing research has yet to overcome the trade-offs between local optimization and global synergy, model robustness and data flexibility, as well as real-time performance and energy efficiency. Intelligent control is a critical core technology for enabling agricultural equipment to achieve autonomy [109]. Currently, more advanced path planning algorithms—such as improved path search methods based on the A* algorithm [110] and Dijkstra’s algorithm [111]—can be employed to realize high-precision path tracking for grain combine harvesters, thereby reducing harvesting omissions and overlaps and enhancing land utilization and harvesting efficiency.

Combine harvester operating environments are highly complex, presenting multiple simultaneous challenges such as nonlinearity (e.g., electro-hydraulic actuators), strong coupling (e.g., interaction between feed rate and travel speed), time-varying conditions (e.g., variations in crop moisture), and uncertainty (e.g., sudden changes in terrain). Typically, a single algorithm is effective in addressing only one or two of these issues.

Hybrid control integrates the strengths of various algorithms. For instance, it leverages the global search capabilities of intelligent optimization algorithms to address the challenging parameter tuning of traditional controllers. Additionally, it utilizes the robust predictive power of deep learning models to guide the real-time operation of classical controllers. Hybrid control is a crucial approach for resolving the paradox of global co-optimization across multiple subsystems. The global optimization capability of PSO [46], the nonlinear processing capability of fuzzy logic, and the stable execution capability of PID are mixed to achieve excellent control results, with a fast response of 0.146 s without overshooting in the simulation, and more than 96% of the harvesting pass rate in the field experiments.

An incremental PID controller can receive the dynamically changing predicted speed from the LSTM as its control target (setpoint) [118]. Its task is no longer to maintain a static value but to accurately and smoothly track this real-time, forward-looking speed command generated by the LSTM. This model effectively responds to sudden changes in feed volume and proactively adjusts the speed before issues such as clogging occur, resulting in a significant improvement in overall machine efficiency and stability.

By intelligently combining MPC and DRL, it is possible to develop an intelligent control system that dynamically balances performance and energy consumption. This is achieved by leveraging prior physical knowledge for safe planning and continuously optimizing through online learning. In the hierarchical or supervised architecture, the DRL agents operate at the upper layer, responsible for making high-level, tactical decisions—such as determining the optimal target traveling speed or feed rate based on real-time crop density. These decisions consider long-term energy efficiency and operational effectiveness. Subsequently, the lower-level MPC acts as the executor, receiving high-level commands from the DRL and utilizing its precise short-term prediction and optimization capabilities to compute specific, safe control sequences (e.g., steering wheel angle, throttle position) necessary to implement the commands while ensuring compliance with the vehicle’s dynamic and safety constraints. In the residual or calibrated architecture, the MPC controller provides a baseline, stable control output based on its internal model, which is optimized for energy efficiency. Simultaneously, a parallel DRL agent learns a “residual” or “correction” signal that is added to the MPC output. This residual signal compensates for discrepancies between the MPC model and the real physical environment (i.e., model mismatch). For example, when a harvester operates on soft ground, the actual tire slip rate may be significantly higher than assumed in the MPC model, leading to reduced engine efficiency. In this scenario, the DRL agent learns and outputs an additional drive torque correction signal to counteract the effects of this unmodeled dynamic, thereby restoring actual energy consumption to an optimal level. The hybrid MPC-DRL architecture achieves a dynamic balance between operational performance and energy consumption through this synergistic mechanism.

The current multi-system synergistic control capability is weak, as the control strategies for subsystems such as the cutting table, threshing, and clearing are mostly designed independently, lacking a global optimization framework. This results in parameter coupling and energy losses that are difficult to coordinate effectively. Additionally, there is insufficient adaptability and robustness to complex environments. Traditional model-driven methods rely on accurate mathematical models, which struggle to handle sudden changes in humidity and crop sparsity in farm environments. Although deep reinforcement learning offers autonomous learning capabilities, it has limited generalization and often fails to meet stringent real-time performance requirements. Furthermore, there are bottlenecks in the accuracy and reliability of sensor fusion. The asynchrony and noise interference among LiDAR, inertial measurement units, and spectral data significantly increase the complexity of fusion algorithms, and the stability of fuzzy adaptive Kalman filtering under extreme working conditions requires improvement.

Lightweight reinforcement learning [119] is a significant area of focus for the future, as it can substantially reduce model size and computational demands. This enables the deployment of complex deep learning and reinforcement learning models on edge devices with limited computational resources, such as embedded controllers on harvesters, drones, or field sensors. This approach addresses the challenge of making real-time intelligent decisions in remote farmland areas where network signals are unstable or unavailable, making it highly relevant. To achieve model lightweighting, techniques such as model pruning [120,121], weight quantization [122], knowledge distillation [123], and the use of efficient network architectures (e.g., MobileNets) are currently employed.

In the future, the integration of multimodal intelligent sensing and lightweight reinforcement learning can be employed to develop a heterogeneous data fusion framework for multimodal sensors—such as hyperspectral imaging, millimeter-wave radar, and haptic sensors [124,125,126]. This approach aims to achieve end-to-end optimization of environmental sensing and decision-making through lightweight reinforcement learning algorithms utilizing distributed near-end policy optimization. Low-latency inference of deep reinforcement learning models will be enabled by combining these algorithms with edge computing devices, while transfer learning techniques will address cross-crop and cross-region generalization challenges.

Digital twin technology [127] is also a prominent area of research (

Table 7. Comprehensive comparative analysis of mainstream control algorithm paradigms.

Algorithm Types	Cost	Core Limitations	Stability/Convergence
PID [112]	Low	Poorly adapted to strongly nonlinear, large time lag and parameter time-varying systems, with empirically dependent parameterization.	Strong for linear systems and complex to analyze for nonlinear systems.
Fuzzy control [113]	Medium	The design of the rule base and the affiliation function relies on expert knowledge, lacks systematicity, and is prone to local optimization; the number of rules grows exponentially with the input dimension.	Weak and difficult to analyze and prove formally
MPC [114]	High	Highly dependent on accurate system dynamics models, model mismatch seriously affects performance; online optimization is computationally intensive and real-time is challenging.	Stronger, stability can be guaranteed if the model is accurate and the optimization problem is solvable.
DRL [115]	High cost of training, medium to high cost of reasoning	“Black-box” characteristics, poor interpretability; requires a large amount of high-quality training data, generalization ability is the main challenge; stability is difficult to prove formally.	Extremely weak, a central challenge for current research
Particle Swarm Optimization Algorithm PSO [117]	Middle to high	Prone to local optimality (premature convergence), performance is sensitive to hyperparameters (e.g., inertia weights); weak theoretical guarantees of convergence to the global optimum.	Weak, convergence analysis is still an active research area
Hybrid control algorithms [116]	Middle to high	High design and commissioning complexity; interactions between components may lead to unintended behavior; theoretical stability analysis of the overall system is extremely difficult	Weak, the stability of the overall system is difficult to analyze formally.
Digital twin system [48]	High	The high fidelity required for the virtual model is extremely high; construction and calibration are costly; system robustness and anomaly handling are challenges.	Dependent on the underlying control algorithm and model fidelity.

). It enables the construction of a digital twin system encompassing the machine—virtual environment—control algorithm to predict the dynamic behavior of cutting, threshing, and cleaning processes through real-time simulation of virtual-real interactions. This system achieves Pareto-optimal optimization based on multi-objective algorithms such as NSGA-III, targeting reductions in loss rate, impurity rate, and energy consumption. A time-graph neural network (e.g., T-GNN) is employed to model the coupling relationships among multiple subsystems, and the digital twin model is dynamically updated via an online learning mechanism to enhance prediction accuracy under complex working conditions. Guo et al. [48] developed a digital twin-based online optimization system for the threshing process of a harvester. Their system predicted future grain breakage rate (GBR) trends using a high-fidelity model and optimized threshing parameters in advance through a genetic algorithm. Field trials demonstrated a 2.08% reduction in grain breakage and a 1.12 km/h increase in operating speed compared to manual control, corresponding to a 10–20 kg reduction in losses per ton of grain. Ma et al. [128] proposed a lightweight network-based method for building a digital twin system for combine harvesters and conducted performance tests and fuel consumption prediction experiments. The system achieved a response time of ≤78 ms, a memory footprint of ≤331 MB, average CPU and GPU utilizations of 17% and 30%, respectively, and maintained a frame rate of 75.6 frames per second during high-intensity operation at a data update frequency of 20 Hz. This work provides a valuable reference for the application of digital twin technology in combine harvesters.

The control algorithm integrates GNSS positioning, IMU attitude data, visual imaging, spectral analysis, and soil sensor information [129] to provide a comprehensive and accurate perception of the harvester’s operating status and the farmland environment. For example, navigation control based on a fuzzy adaptive finite impulse response Kalman filter algorithm effectively fuses GNSS and IMU data, significantly improving positioning accuracy and stability. Multi-source data fusion also supports precision control by monitoring factors such as crop yield distribution, water content, and soil fertility changes, enabling precise management of fertilizer application, irrigation, and other operations to achieve precision agriculture. Addressing energy and environmental concerns, the harvester’s power system control strategy can be optimized through intelligent speed control and load-adaptive control technologies [130], which adjust engine power and rotational speed in real time according to operational load, thereby reducing energy consumption and emissions. Furthermore, the control algorithm can be integrated with precision agriculture technologies [131] to enable accurate application of fertilizers and pesticides, minimizing chemical use, reducing environmental impact, and promoting sustainable agricultural development. The currently proposed adaptive algorithm outperforms traditional PID methods and meets practical application requirements; however, improvements are needed in sensor stability and detection accuracy. Additionally, further research is required to evaluate vibration and noise reduction effects. Future work will involve comparing different control strategies and conducting both quantitative and qualitative studies. Reason: The revision improves clarity, flow, and technical precision by restructuring sentences, enhancing vocabulary, and correcting grammar and punctuation. It also clarifies the relationships between technologies and their benefits, ensuring the text is more readable and professionally presented.

Unmanned farm experiments [132] have demonstrated that a multifunctional integrated platform can increase land utilization efficiency by 12% and reduce manual intervention by 30%. An autonomous navigation system based on full-stack unmanned technology can be developed to integrate multifunctional modules—such as harvesting, fertilizing, and seeding—and automate the entire farm operation process through cloud-based collaborative control.

Realizing the advanced control described in this paper requires a comprehensive set of hardware and software systems. Current market prices indicate that a GPS/GNSS autopilot system with high-accuracy RTK capability ranges from $5000 to over $20,000. Complete autosteering kits from mainstream brands (e.g., Case IH, John Deere) are priced between $4000 and $12,000. Additionally, sensory equipment such as high-resolution drones and multispectral cameras can cost thousands to tens of thousands of dollars each. On the software side, modern farm management platforms typically use an annual subscription model, ranging from approximately $950 for a basic version to more than $15,000 for an enterprise version. Farmers generally expect new technologies to yield a return on investment of at least 3:1 before considering adoption [133]. While research has shown that precision agriculture technologies can deliver 10–15% yield gains and 15–20% input savings, such returns are not guaranteed. Benefits depend on a combination of factors, including farm size, crop type, soil heterogeneity, weather conditions, and manager skill level, making it difficult for small farmers to accurately predict and trust the economic benefits [134]. Because the technology itself is economically and operationally challenging to scale directly to smallholder farmers, innovation in business models is essential to achieving “service scalability” by purchasing services on an as-needed basis (e.g., per-acre UAV planting, hourly rentals for automated seeding services). This approach converts high fixed-asset inputs into flexible operational expenses, effectively addressing the economic scalability challenge for smallholder farmers.

8. Conclusions

• Currently, research on control algorithms for grain combine harvesters has shifted from optimizing individual subsystems to achieving multi-system synergy; however, numerous challenges remain. In cutter control, MPC enhances adaptability to hilly terrain through multi-sensor fusion, but its reliance on accurate models limits its generalizability in dynamic farmland environments. Although the threshing mechanism dynamically adjusts the gap based on feeding rate, multi-objective cooperative optimization faces conflicts between parameter coupling and increased energy consumption. The application of fuzzy logic and neural networks in the scavenging system improves impurity rate stability, yet the lack of accurate nonlinear modeling hinders significant improvements in control efficiency. Control strategies for the cutting table, threshing, and cleaning subsystems are typically designed independently without a global optimization framework, resulting in difficulties coordinating parameter coupling and causing high energy losses. While MPC performs well in controlling the cutting table, it lacks dynamic integration with the threshing gap. The existing algorithm lacks a multivariate coupled dynamic adjustment mechanism due to the independent design of the control strategies of the cutting platform, threshing, cleaning, and other subsystems. Whenever there is a sudden increase in the feeding volume, the threshing gap and the height of the cutting table are not adjusted in concert, resulting in an increase in the unthreshed rate, and the model-driven method in hilly terrain with a slope of >8°, the IMU tilting noise leads to the mismatch of the MPC model, and the error in the height of the cutting table expands to ±5 cm, while the data-driven method has an error of more than 50% in the tracking error of DRL paths under extreme working conditions and there is a bottleneck of real-time performance. The fusion reliability is reduced, and the energy consumption of the fixed-parameter scavenging system increases by more than 30% with the fluctuation of the feeding volume, and the Pareto-optimal loss rate and energy efficiency have not been realized, and there are difficulties in global synergy, environmental adaptability, data fusion reliability, and energy efficiency optimization that have not been solved yet.
• Regarding the cleaning system of the combine harvester, future efforts should focus on optimizing the fan blade and screen motion parameters. A coupled DEM-CFD model should be developed to replace costly experiments. Additionally, deep learning techniques should be introduced, and MPC should be enhanced to effectively manage the nonlinear system.
• Despite the excellent performance of ADRC in high-precision navigation and trajectory tracking applications for combine harvesters, most current research remains confined to simulation and laboratory settings, lacking sufficient field validation. Future efforts should focus on enhancing field environment validation, deepening integration with intelligent algorithms and deep learning to enable parameter self-tuning, and expanding the integration of actuator control with hyperspectral sensing applications. The continuous optimization and interdisciplinary integration of self-immune controller technology will provide core technical support for Agriculture 4.0, promoting the evolution of agricultural production toward intelligence, high precision, and sustainability. This advancement will effectively address global challenges such as food security, resource efficiency, and environmental pollution.
• Although the model-driven approach, exemplified by MPC, offers strong interpretability and a solid theoretical foundation, its performance heavily depends on the accuracy of the system model, which is often challenging to establish. In contrast, data-driven methods, such as those based on neural networks, can flexibly handle complex nonlinear relationships but suffer from the generalization ability, and theoretical challenges related to stability that are difficult to prove. These theoretical limitations—namely the characteristic, insufficient generalization, and stability issues—pose significant obstacles. Looking ahead, we believe that the most critical path to resolving this core contradiction and advancing the technology lies in the cross-fertilization of algorithms, which is an inevitable trend toward achieving higher levels of intelligence.
• The superior performance of many advanced control algorithms reported in the current literature is often achieved through fine-tuning under highly specific conditions. Since the physical characteristics (e.g., plant height, stem toughness, water content) and growing environments (e.g., soil type, topography) of different crops vary significantly, the performance of a model trained for one scenario may decline drastically in another. This often necessitates large-scale data re-collection and model re-training. In agriculture, collecting extensive labeled data that cover various crops, growth stages, weather, and soil conditions is extremely costly and challenging. Consequently, trained models are often overfitted to their limited training data, making it difficult to generalize to new and unseen situations. Achieving strong generalization capability in control algorithms remains a major research challenge in the field of smart agricultural machinery. Addressing this issue requires not only algorithmic innovations (e.g., transfer learning, domain adaptation techniques) but also concerted efforts in data collection and sharing to establish standardized and diverse benchmark datasets.