Real-Time Concrete Workability Estimation in Transit via an IoT-Enabled Cyber-Physical System

Abstract

Concrete quality on construction sites depends on maintaining certain workability during transportation. However, traditional slump testing is manual and cannot assess concrete in transit. This study presents a cyber-physical system (CPS) integrating IoT sensors with machine learning models to estimate concrete workability in real time during delivery. The proposed approach first correlates traditional slump test parameters with sensor measurements, allowing for automated replication of established workability evaluation. The CPS continuously captures IoT sensor data, such as drum rotation, vibration, internal pressure, and temperature, processes these data in real time, and infers workability before unloading. The developed CPS was validated in a real-world case study, achieving high workability prediction accuracy with an average error of approximately 0.9 cm and timely, automated feedback of below 200 ms. These results enable continuous in-transit monitoring of concrete workability and lay a practical foundation for data-driven operational improvements in construction logistics.

1. Introduction and Related Work

Concrete’s strength and durability are fundamentally dependent on its workability, which reflects the material’s ability to flow and consolidate with minimal external effort [1,2]. Traditionally, workability is assessed through manual slump tests performed before loading and after delivery, typically at the batching plant or construction site [3]. However, such assessments are limited to discrete points and fail to capture workability changes during transportation. Manual testing is subject to operator variability, which increases uncertainty in quality control [4].

During transit, concrete experiences continuous agitation, vibration, and temperature fluctuations, which significantly affect workability [5]. Despite the known impact of transportation, no widely adopted method currently exists for continuous, in-transit workability monitoring. This gap leads to undetected workability loss, rejected batches, material waste, and increased environmental and economic costs [6].

Recent advances in digitalization and sensor technology offer new possibilities for addressing these challenges. The integration of industrial Internet of Things (IoT) sensors and CPS holds the promise of enabling continuous, automated monitoring of key physical parameters throughout transportation. In turn, the combination of real-time sensor data with machine learning analytics can facilitate estimating workability online, providing actionable feedback to stakeholders before unloading.

In this study, a CPS is introduced that uses IoT data acquisition and machine learning for real-time, in-transit estimation of concrete workability. This approach is validated using data from real-world concrete deliveries, demonstrating its potential to improve quality assurance and operational decision-making in ready-mix concrete logistics. Nevertheless, a review of existing approaches is essential to clarify the contributions and limitations of prior work. Efforts in AI-based workability estimations, IoT sensing, and CPSs reveal both the potential and current gaps in the state of the art.

Workability is traditionally evaluated using manual slump tests at the plant and on-site [7,8], typically via the Abrams cone procedure [9,10]. However, the manual, point-in-time nature of slump tests precludes continuous or in-transit assessment [11], and human involvement introduces uncertainty and variability [12].

To reduce reliance on manual slump tests, several AI-based approaches estimate concrete workability at the end of transit based on mix properties [13]. Vergas et al. [14] constructed an MLP-based pipeline that utilized concrete mix design components per cubic meter to predict the workability of concrete during its delivery. However, these methods do not account for in-transit environmental effects such as temperature, as noted in [15,16,17,18,19], nor do they provide real-time decision support during transport.

Pertaining to IoT-enabled concrete workability estimation, Yu et al. [20] used IoT-acquired images during the concrete mixing stage to evaluate the concrete’s current workability. In contrast, in [7], temperature data acquired during the early induction phase of concrete placement were used for real-time concrete strength evaluation, which is directly correlated to the concrete’s workability [21]. Similar studies estimate concrete workability during its mixing phase, as described in [22,23]. Nonetheless, no studies to date have leveraged IoT sensor data collected during transit for real-time, automated workability estimation.

In the context of predictive modeling for concrete workability, deep learning models have demonstrated potential in capturing large sequential historical patterns [20], typically needed in scenarios where both long-term historical contexts and short-term information is needed for accurate predictions [24]. Long Short-Term Memory (LSTM) networks are well suited, as they mitigate vanishing gradients and successfully retain both long-term historic trends and short-term fluctuations in data streams [25]. Hybrid architectures have emerged that augment the capabilities of LSTMs by extracting spatial (CNN-LSTM [26]) or temporal (TCN-LSTM [27]) features via convolutional or temporal layers before sequence modeling, improving predictive capacity. Similarly, Bidirectional LSTM (Bi-LSTM) models offer improved context capture by processing sequences in both forward and backward directions, resulting in superior convergence and richer temporal feature extraction than TCN-LSTM hybrids [28]. These architectures outperform Gated Recurrent Units (GRUs), which have fewer parameters and struggle with long-range dependencies, and Transformers, which can lose temporal order due to their attention mechanism and have high memory requirements [29,30].

CPSs and digital twins can facilitate real-time decision-making and adaptation in domains such as manufacturing and logistics [31,32,33,34]. In the context of decision-making in the concrete transportation industry, Weerapupra V. et al., in [35], utilized digital twins fed with data on concrete mixing operation times and concrete mixing properties to evaluate abnormalities in concrete mixing and water dosing. While not directly estimating concrete workability, the CPS is used to recommend to engineers on potential approaches to adjust process parameters or halt concrete production before dispatch. On a similar note, in [36], a digital-twin-based CPS was constructed to facilitate decision-making during the delivery of ready-mix concrete. As denoted in [36], the CPS evaluates aspects such as mixing time, temperature history during transportation, and waiting time on site to evaluate the optimal delivery window of concrete. Lastly, similar approaches have been presented in [37,38]. Nevertheless, it is evident from the reviewed studies that CPSs have yet to be adopted to facilitate the decision-making process of operators involved in the transportation of concrete during transit.

Overall, while evident from the literature that concrete workability estimations during transit are crucial for improving decision-making in the context of aborting or redirecting transportation when workability deteriorates [7], current approaches are solely focused on either the concrete mixing phase or the on-site arrival phase of the concrete.

Hence, the present study targets a critical gap in current practice by introducing a CPS that enables continuous, real-time estimation of concrete workability during transit using in-drum IoT sensor data and machine learning. The proposed approach provides direct, automated workability assessment throughout the delivery process, supporting proactive quality assurance and operational decision-making that were previously unattainable with manual or endpoint-only measurements.

2. Methodology

2.1. Empirical Workability Estimation

Concrete workability in this study is referenced by the slump test, standardized in ASTM C143/C143M-20 [39] and EN 12350-2 [40]. This test provides the ground truth for model calibration and classification according to BS EN 206:2013+A2:2021 [41], as seen in

Table 1. Classes of workability.

Workability Measurement (cm)	Workability Classification
10–20	S1-a
21–30	S1-b
31–40	S1-c
40–90	S2
90–150	S3

.

Traditionally, workability is modeled as a linear function of mix design and environmental variables [7]. Equation (1) represents the empirical relationship of workability to the aforementioned variables:

$$W=k_1\ast {\displaystyle \frac{w}{c}}+k_2\ast A+k_3\ast T+k_4\ast H+\epsilon$$

(1)

where $W$ is workability (mm), ${\displaystyle \frac{w}{c}}$ is the water to cement ratio, $A$ is admixture content (% cement mass), $T$ is ambient temperature (°C), $H$ is relative humidity (%), $k_1$, $k_2$, $k_3$, $k_4$ are empirically determined coefficients, and $\epsilon$ is the error. According to Equation (1), a higher water-to-cement ratio, ${\displaystyle \frac{w}{c}}$, or the presence of plasticizing or other admixtures, $A$, indicates an increased fluidity of concrete, thus increased workability. In addition, higher ambient temperatures accelerate moisture loss and hydration resulting in reduced workability as time progresses [42]. While this captures baseline workability immediately after mixing, it fails to account for dynamic, in-transit changes due to agitation and temperature fluctuations.

2.2. Mathematical Modeling

The rheological behavior of fresh concrete can be described by the Bingham model [43], as denoted in Equation (2).

$$\tau ={ \tau }_y+{\mu }_p\ast \dot{\gamma }$$

(2)

where $\tau$ denotes the shear stress, ${\tau }_y$ the yield stress, ${\mu }_p$ the plastic viscosity, and $\dot{\gamma }$ the shear rate. In a rotating mixer, increased yield stress and viscosity are reflected in higher internal drum pressure and resistance to rotation [44]. Thus, in-drum sensor data can be mapped to rheological parameters and, by extension, to workability.

The shear stress is directly related to the in-drum pressure. The rotating mass produces a dynamic pressure head that is proportional to the average shear stress at the drum wall. However, external vibrations, typically measured by the drum’s linear acceleration, temporarily lower the apparent yield stress and plastic viscosity. In this context, in-drum pressure is related to the instantaneous effective yield stress under vibration and the instantaneous effective plastic viscosity, as described in Equations (3) and (4), resulting in the inner-drum pressure being denoted by Equation (5).

$${\tau }_y^{eff}={\tau }_y-k_{\tau }\ast a_{RMS}$$

(3)

$${\mu }_P^{eff}={ \mu }_P-k_{\mu }\ast a_{RMS}$$

(4)

$$P=k_P\ast ({\tau }_y^{eff}+{\mu }_P^{eff}\ast {\displaystyle \frac{R}{\delta }}\ast \omega )$$

(5)

where ${\tau }_y^{eff}$ is the instantaneous effective yield stress, ${\mu }_P^{eff}$ is the instantaneous plastic viscosity, ${\tau }_y$ the static (no-vibration) yield stress, ${\mu }_P$ the static plastic viscosity, $a_{RMS}$ the root-mean-square drum-frame acceleration over the last revolution, $k_{\tau }$ a calibration factor linking vibration level to yield-stress reduction, $k_{\mu }$ a calibration factor linking vibration level to viscosity reduction, R the inner radius of an inclined drum, and $\delta$, where $\delta <R$ is the thickness of a thin annular layer where the dominant shear occurs.

Furthermore, by neglecting secondary slumping waves, the shear rate is approximated as denoted in Equation (6).

$$\dot{\gamma }={\displaystyle \frac{R}{\delta }}\ast \omega$$

(6)

Building on Equations (3) and (4), which account for the effective yield stress and plastic viscosity under vibration, the static Bingham parameters ${\tau }_y$ and ${\mu }_p$ can be estimated using paired (R, ω) samples. By collecting two non-collinear operating points $(P_1, {\omega }_1)$ and $(P_2, {\omega }_2)$ during trip segments, ${\tau }_y$ and ${\mu }_p$ are calculated based on Equations (7) and (8), respectively.

$${\tau }_y= {\displaystyle \frac{P_1}{k_P}}-{\mu }_p\ast {\displaystyle \frac{R}{\delta }}\ast \omega$$

(7)

$${\mu }_p={\displaystyle \frac{P_2-P_1}{k_P\ast \frac{R}{\delta }({\omega }_2-{\omega }_1)}}={\displaystyle \frac{\delta \ast (P_2-P_1)}{k_P\ast R\ast ({\omega }_2-{\omega }_1)}}$$

(8)

By leveraging Equations (2)–(8), it is evident that the behavior of fresh concrete is related to the inner-drum pressure, the angular velocity of the drum, and the drum’s linear acceleration. Thus, such aspects are mapped to an analogous to the Equation (1) model for estimating the concrete’s workability and are denoted in Equation (9).

$$\widehat{W}\left(t\right)=c_0+c_1\ast P_{drum}\left(t\right)+c_2\ast \omega \left(t\right)+c_3\ast \alpha \left(t\right)+c_4\ast T\left(t\right)+\in (t)$$

(9)

where $\widehat{W}\left(t\right)$ denotes the estimated workability at time t, $P_{drum}$ the internal drum pressure, $\omega \left(t\right)$ the angular velocity at the time t during transit, $\alpha \left(t\right)$ the drum linear acceleration, as a proxy for concrete mass agitation and road-induced vibration, $T\left(t\right)$ the ambient temperature, while $c_0\dots c_4$ are model coefficients, and $\in \left(t\right)$ is the corresponding modeling error.

Supporting Equations (3)–(9), an increase in yield stress, ${\tau }_y$, due to lower water to cement ratio, ${\displaystyle \frac{w}{c}}$, or reduced admixture, $A$, as concrete ages, will result in increased concrete resistance to drum rotation, ${\tau }_y$, raising internal drum pressure. Similarly, an increased plastic viscosity, ${\mu }_p$, pinpoints to increased incremental pressure during higher rotational speeds of the drum.

Concrete mass agitation results from the mixer drum’s movement and linear acceleration. Since agitation describes the degree of internal mixing within the drum, it can be inferred using drum acceleration as a proxy. Furthermore, the drum’s movement induces additional agitation [45], which affects workability, since sustained vibrations due to abrupt turns or bumps in the road can temporarily fluidize the concrete [46]. Lastly, temperature T in Equation (9) is a direct match to Equation (2) to account for thermal effects on workability loss during transit.

2.3. CPS for In-Transit Workability Estimation

Coefficients in Equation (9) are not constant during transit, limiting the equation’s applicability in a real-world scenario. To address this limitation, this study leverages the theoretical mapping of the drum’s internal pressure, its angular velocity, and its linear acceleration to concrete workability, and introduces a data-driven CPS for in-transit workability estimation using IoT-captured data to provide actionable recommendations to concrete mixer truck operators regarding trip termination or mixer truck operational parameter adjustment. A high-level representation of the proposed CPS is illustrated in Figure 1.

To capture time-varying and nonlinear relationships, the CPS adopts a Bi-Directional Long Short-Term Memory (Bi-LSTM) neural network with an attention mechanism that is trained to capture the complex, time-varying relationships between multivariate sensor time series and concrete workability [47]. This architecture models the dynamic influence of multivariate signals on workability, with internal weights functioning analogously to empirical coefficients. Model development, feature extraction, and evaluation procedures are detailed in subsequent paragraphs.

Sensor data, including internal drum pressure ($\mathrm{k}\mathrm{P}\mathrm{a}$), drum linear acceleration (acting as a proxy to concrete mass agitation) (${\displaystyle \frac{\mathrm{m}}{{\mathrm{s}}^2}}$), angular velocity (${\displaystyle \frac{\mathrm{r}\mathrm{a}\mathrm{d}}{\mathrm{s}}}$), and ambient temperature ($°\mathrm{C}$) were continuously recorded from the mixer truck during transit. Data streams were partitioned into increments corresponding to individual drum revolutions (~X seconds), ensuring that each input sequence captured the most recent evolution of concrete behavior. Within each increment, features were extracted, including descriptive statistics (mean, median, standard deviation, minimum, maximum), frequency-domain features (fast Fourier transform, wavelet energy), and statistical moments (skewness, kurtosis) in order to characterize both steady-state and transient behaviors relevant to workability.

Sequences were processed by the implemented Bi-LSTM neural network with an attention mechanism. The Bi-LSTM network models temporal dependencies in both past and future directions, while attention enables the model to focus on the most informative timesteps and sensor events. This approach allows the model to improve its capacity for generalization over the linear empirical model of Equation (9) in complex, time-evolving relationships observed during transport. Input sequences were zero-padded to a fixed length to accommodate variability in trip duration, with masking applied to exclude padded timesteps from gradient updates.

Model performance was evaluated using root mean squared error (RMSE), mean absolute error (MAE), mean squared error (MSE), and coefficient of determination (R²), using a describe split, e.g., leave-one-trip-out cross-validation strategy [48].

Workability estimations generated by the Bi-LSTM are provided with a rule-based mechanism that is triggered when the estimated workability’s class differs from the initial measured class. The CPS flags class changes and provides recommendations to truck operators. Recommendations are generated based on the rules presented in

Table 2. Recommendations generated by the CPS in transit.

Relative to Initial Class Detected Event	CPS Recommendation	Reasoning
No change in class is estimated	Resume current operations	Since no class change is detected, no need to adjust mixer parameters
One class drop	Increase concrete mixer rotations	Increased rotation (higher agitation), slows down concrete stiffening
Severe class drop (higher than two class drop)	Terminate delivery	Severe class drops signals errors during initial concrete mixing

.

3. Implementation

The Bi-LSTM model of the CPS was developed, trained, and validated on a Windows 10, version 22H2 (Microsoft Corporation, Redmond, WA, USA) workstation equipped with an Intel Core i9-10850K CPU (Intel Corporation, Santa Clara, CA, USA), 32 GB RAM, and an NVIDIA GeForce RTX 2070 SUPER GPU (NVIDIA Corporation, Santa Clara, CA, USA). The model training pipeline and evaluation were implemented in Python (version 3.10; Python Software Foundation, Wilmington, DE, USA), using NumPy (version 1.21.6) and pandas (version 2.2.3) for data manipulation, scikit-learn (version 1.6.1) for preprocessing and performance metrics, and TensorFlow (version 2.18.1; Google LLC, Mountain View, CA, USA) for Bi-LSTM model construction and optimization. Lastly, the hyperparameters used for model training can be seen in

Table 3. Hyperparameters of the Bi-LSTM model.

LSTM Layer 1 Num. of Units	LSTM Layer 2 Num. of Units	Learning Rate	Weight Decay	Dropout	Batch Size	Loss Function
128	64	3 × 10−3	10−4	0.3	32	MSE

.

The CPS was deployed on an edge device, as shown in Figure 2. The IoT sensors were mounted on concrete mixer drums, continuously capturing data, transmitted via a 4G cellular connection to an onboard Raspberry Pi 4 model B (Raspberry Pi Ltd., Cambridge, UK), serving as an edge computing device. The data preprocessing pipeline was exposed via Docker containers (Docker, Inc., Palo Alto, CA, USA; Engine version 25.0.3) running a Dockerized FastAPI (version 0.70.0), and preprocessed data were sent via a RESTful POST request to the Bi-LSTM model. The FastAPI application containing the Bi-LSTM model also encapsulates the recommendation system of the CPS. Generated recommendations are communicated to a React-based user interface (React version 18.2.0; Meta Platforms, Inc., Menlo Park, CA, USA) via a WebSocket connection deployed on a tablet fitted directly in the cabin of the concrete mixer truck.

To address connectivity issues which may arise during transit, the IoT devices securely buffer all captured sensor data locally. Equipped with 32 GB of onboard storage and operating at an average data rate acquisition of approximately 500 MB/h, the sensors can retain up to 65 h of sensor data before reaching storage capacity. All data entries are timestamped to ensure accurate synchronization once transmission resumes. To prevent data loss, the IoT sensors issue a storage alert when usage exceeds 80% and is configured to overwrite the oldest data if the limit is reached. During field trials, no data loss occurred, as connectivity disruptions were not observed.

4. Use Case

Description

To validate the proposed CPS for real-time concrete workability estimation, a 4-week field deployment with a regional concrete manufacturer in Attica, Greece, was conducted. Over 68 ready-mix concrete deliveries (initial workability S1-b, 21–30 cm) were monitored using three sensor-equipped mixer trucks. Operational conditions are summarized in

Table 4. Experiment’s operational conditions.

Minimum Trip Duration (min.)	Maximum Trip Duration (min.)	Minimum Ambient Temperature (°C)	Maximum Ambient Temperature (°C)
11	81.5	9.7	24.7
Mean Drum Rotation During Transit (RPM)	Minimum Drum Angle (°)	Maximum Drum Angle (°)
2	15	20

, reflecting the practical range encountered during the pilot.

IoT devices were installed on three mixer trucks. The pressure sensor was mounted inside the drum, while the gyroscope and accelerometer were attached externally. Data frequencies are presented in

Table 5. Frequency of IoT data collection.

Sensors	Frequency	Unit
Accelerometer	~3500 Hz	g (m/s2)
Gyroscope	~7000	Degrees per second (°/s)
Pressure	~7000	Bar
Temperature	Measured once at the start and end of the delivery	Celsius

.

In addition, specific assumptions were made to facilitate the use case execution:

• Drum rotation speed was maintained as steady throughout each transit, consistent with standard agitation protocols.
• No significant variation in concrete mix design was assumed during the study period.
• Based on expert feedback, manual workability measurement error was estimated at 2 cm.
• To facilitate interpolation of temperature and workability evolution, both parameters were linearly interpolated between start and end values, generating regular data points.

5. Results

Across the 68 transits performed (average duration: 50 min), approximately 3.5 billion raw sensor data points were collected. Using this data, Equation (3) coefficients were approximated, based on 20% of the dataset. This subset was also used for Bi-LSTM model validation, ensuring fair comparison. Coefficient approximations can be seen in

Table 6. Coefficient approximations using real-world IoT data.

c0		c1		c2		c3		c4
Min	Max	Min	Max	Min	Max	Min	Max	Min	Max
24.6	25.8	−0.0043	−0.0022	−0.15	−0.09	0.04	0.06	−0.69	−0.8

.

5.1. Empirical Model Validation

To quantify the predictive performance of Equation (9) following the approximation of the coefficients, the RMSE, MSE, and $R^2$ error metrics were evaluated (

Table 7. Empirical model’s performance metrics.

RMSE	MSE	${\mathit{R}}^2$
2.66	2.32	0.24

), resulting in a 95% prediction interval of $\pm 5 \mathrm{c}\mathrm{m}$ for workability predictions, based on assumed normality.

For the training of the deep learning model, all IoT signals were aggregated into 30-s increments. Figure 3 demonstrates significant variability in the norm of the accelerometer and gyroscope at this window, supporting the chosen increment.

Aggregated 30-s increment raw data’s dimensionality was reduced using the preprocessing approaches listed in Section 3. For each increment, features included mean, standard deviation, min, max, skewness, kurtosis, fast Fourier transform, and wavelet energy, capturing both central tendency and variability.

Building upon the linear relationship of IoT variables and the measured workability denoted in Equation (9), Pearson correlation [49] was evaluated for all extracted features against the concrete’s workability, whose results can be seen in

Table 8. Pearson correlation analysis results of experimental IoT data.

Features	Correlation Values	Features	Correlation Values	Features	Correlation Values
Temperature start 1	−0.50	Pressure skewness 1	0.42	Acceleration wavelet max	−0.18
Temperature end 1	−0.54	Pressure kurtosis 1	−0.51	Gyro norm mean 1	0.45
Pressure FFT std	−0.11	Acceleration norm mean 1	−0.52	Gyro norm std	0.14
Pressure FFT mean	−0.02	Acceleration norm std	−0.19	Gyro norm min	0.15
Pressure FFT min	−0.03	Acceleration norm min	−0.08	Gyro norm max	−0.04
Pressure FFT max	0.06	Acceleration norm max	−0.22	Gyro norm skewness 1	−0.45
Pressure wavelet mean	−0.18	Acceleration norm skewness	−0.05	Gyro norm kurtosis 1	−0.54
Pressure wavelet max	−0.18	Acceleration norm kurtosis	−0.10	Gyro FFT mean	−0.26
Pressure mean 1	−0.43	Acceleration FFT mean	−0.21	Gyro FFT std	0.16
Pressure std	0.19	Acceleration FFT std	0.05	Gyro FFT max	0.13
Pressure min	−0.36	Acceleration FFT max	0.12	Gyro wavelet mean	−0.09
Pressure max	−0.22	Acceleration wavelet mean 1	−0.45	Gyro wavelet max	−0.18

¹ Features demonstrating correlation higher than 0.4.

.

As depicted in Table 8, a number of features demonstrated a weak correlation, with the Pearson correlation coefficient being below $\left|r\right|<0.4$. Features with weak correlation to concrete workability were excluded from model training and inference in the real-world CPS application.

Following feature selection, the final dataset for deep learning modeling comprised approximately 60,000 points, evenly distributed among ten input features, all standardized via z-score normalization.

Table 9. Dataset Sample.

Pressure Min (Bar)	Pressure Skewness (Bar)	Pressure Kurtosis (Bar)	Acc. Norm Mean (m/s2)	Acc. Wavelet Mean (m/s2)	Gyro. Norm Mean (Dps°/s)	Gyro. Norm Skewness (Dps°/s)	Gyro. Norm Kurtosis (Dps°/s)	Starting Temperature (Celsius)	Workability (mm)
49.8049	7.4053	437.2824	1.01519	956.6658	22.4718	0.1858	1.4406	10.7294	23.82
49.8024	10.4955	877.7744	1.0156	1784.5491	22.5144	0.2033	1.2606	10.7588	23.77
49.8016	12.8636	1318.2666	1.0156	2563.1054	22.5346	0.2113	1.4163	10.7882	23.34
49.8012	14.8590	1758.7588	1.0162	3389.1709	22.5438	0.2353	1.4987	10.8176	23.31
49.8009	16.6165	2199.2511	1.0163	4195.8640	22.5477	0.2333	1.4993	10.8470	23.23

provides a representative data sample.

Eighty percent (80%) of this dataset was used to train the model, with the remaining 20% for validation. The performance of the Bi-LSTM model was evaluated using MAE, RMSE, MSE, and R². Nevertheless, to strengthen the selection of the Bi-LSTM model, additional deep learning algorithms suitable for capturing large sequential information were evaluated. These included single-layer unidirectional LSTM, multi-layer LSTM, a hybrid convolutional LSTM (CNN-LSTM), and a bidirectional LSTM (Bi-LSTM) without an attention mechanism. The performance of each model is depicted in

Table 10. Performance metrics of the evaluated deep learning models.

Deep Learning Models	MSE	RMSE	MAE	R2
Single-layer unidirectional LSTM	9.387	3.064	2.281	0.256
Multi-layer LSTM	6.783	2.604	1.964	0.494
Hybrid CNN-LSTM	5.509	2.347	1.789	0.560
Bi-LSTM without an attention mechanism	4.155	2.038	1.709	0.669
Bi-LSTM with an attention mechanism	0.971	0.985	0.739	0.897

.

As seen in Table 10, the Bi-LSTM model with an attention mechanism demonstrated the highest predictive accuracy, averaging an error of 0.98 cm and R² = 0.89, indicating strong generalization. When compared to the other deep learning algorithms included in the comparison, the selected model demonstrated an average improvement in $R^2$ by an average of 46%.

Based on expert feedback and prior literature [50], the error of manual slump tests resides in the $\pm 2 \mathrm{c}\mathrm{m}$ range of the measured slump. This ±2 cm range corresponds to the 95% confidence interval of the manual reference method. Assuming zero-mean Gaussian workability measurement noise, this range corresponds to a standard deviation of approximately $\sigma = 1 \mathrm{c}\mathrm{m}$. Any error metric computed against such ground truth therefore contains an irreducible noise floor of σ. The Bi-LSTM attains an RMSE of 0.98 cm and an MAE of 0.74 cm, resulting in an error of equal or lower than the noise floor. By treating the manual slump test error and the model error as independent, their variances add to (Equation (10)):

$${RMSE}_{obs}^2 = {\sigma }^2+{\sigma }_{model}^2$$

(10)

where ${RMSE}_{obs}$ is the root-mean-square difference between the Bi-LSTM predictions and the manual slump readings in the validation set, σ is the standard deviation of the random error inherent in the manual slump measurements, and ${\sigma }_{model}$ is the unknown true standard deviation of the Bi-LSTM’s error relative to the noise-free slump measurement.

Since ${RMSE}_{obs}^2 = {0.985}^2 = 0.97$, lower than the ${\sigma }^2,$ which is equal to 1, the unbiased estimate of ${\sigma }_{model}^2$ is close to zero, resulting to the residual scatter being mostly attributable to manual measurement uncertainty. Using the NIST one-sided procedure [51] with N equal to 14 (20% of transits used for model validation), yields a 95% upper confidence bound for the actual RMSE of the Bi-LSTM model of ${\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}}_{\mathrm{m}\mathrm{o}\mathrm{d}\mathrm{e}\mathrm{l}}\le 1.06$. Consequently, with 95% confidence, any individual Bi-LSTM estimate will differ the most from a noise-free slump value by approximately $\pm 1\mathrm{c}\mathrm{m}$, thereby providing workability predictions that are both more precise and free of manual-induced bias.

Model predictions on the evaluation set were also visualized on the 30-s increment steps, as illustrated in Figure 4. For improved readability, Figure 4 visualizes the seven transits of the validation set, each trip denoted with different color, demonstrating the highest error in estimated workability at the end of transportation. The initial 16 increments show convergence behavior due to input padding. After this, predictions stabilize, enabling actionable decisions.

5.2. CPS Decision Support in Operation

The CPS’s decision-making mechanism is triggered after the 16th increment has passed. Recommendations are generated by the decision-making mechanism when the estimated final workability is accompanied by a class change, as denoted in Section 3. To evaluate the bidirectional nature of the CPS, concrete mixer truck operators were presented with recommendations through the interface demonstrated in Figure 5.

The decision-making mechanism of the CPS was evaluated in the use case through a real-world deployment in the context of 1 week. During the evaluation period, a total of 14 transportation trips were conducted, and the CPS recommendations were generated based on Table 2. No initial slump measurements were conducted, while the workability at the end of each transit was manually evaluated using manual slump measurements to validate the system’s performance and decision-making capacity.

In 10 of the 14 trips, CPS estimated that workability remained within the original S1-b and S2 classes, leading to no recommendations generated. In 3 out of the 14 trips, the CPS estimated a single-class drop (S1-b to S1-a), generating a recommendation to operators suggesting an increase in drum rotations. In all three cases, operators followed the CPS recommendations. During the arrival at the destination, the workability measurements resulted in 22.3, 23.8, and 22.8 cm, respectively, potentially indicating the CPS’s ability in facilitating decision-making. Lastly, in a single trip out of the 14 total, the CPS predicted a drop in workability from S2 to S1-a (two-class drop) and generated a trip termination recommendation. However, the truck’s operator disregarded the recommendation and continued the transportation, based on knowledge of the specific concrete’s mixing properties and drum real-time operational parameters displayed in its dashboard. Subsequently, upon delivery, workability was measured at 25.8 cm (class S2), pinpointing to a false-positive trip termination recommendation of the CPS.

Given the false-positive trip termination recommendation of the CPS, it is evident that the Bi-LSTM model of the system does not generalize adequately to unseen data from different classes of concrete other than the ones it has been trained upon, despite its high $R^2$. In addition, raw data from the false-positive trip were analyzed. The analysis indicated an elevated standard deviation of accelerometer-generated features compared to the original raw data used for model training, suggesting that a different agitation protocol was adopted during transit. Consecutively, it is evident that the model’s predictive uncertainty increases for out-of-distribution inputs and points to the need for data augmentation. Nonetheless, this also indicates the importance of human involvement in the decision-making process, signifying the need for the introduction of a feedback mechanism where operators can provide valuable feedback to the CPS for its continuous improvement.

The provision of human feedback will be enabled by the allowance of on-route appraisal from an expert driver who is aware of the initial workability and the target workability at the end of the transportation trip. Drawing on real-time observations of weather and road conditions, factors experienced drivers can recognize that accelerate slump loss, the operator will be able to flag each estimated workability as plausible or implausible. Driver feedback will be cross-validated with measured end-slump values. Only feedback that aligns with the final measured workability will be used in the next model update cycle, incorporating human expertise and protecting the CPS from catastrophic forgetting or parameter drift.

Lastly, the time between in-transit feature preprocessing and recommendation generation was evaluated. The average response times of the two RESTful APIs were calculated during the 1-week real-world deployment. The response time of the feature preprocessing API was calculated at an average of 58 ms, while the average response time of the Bi-LSTM and recommendation system API averaged 128 ms. In total, the average response time of the CPS accumulates to 186 ms, affirming its real-time capacity for workability estimations in transit.

6. Conclusions

This study proposes and experimentally validates a CPS for real-time estimation of concrete workability during transport, which is a challenge not addressed by traditional, manual slump testing methods. In particular, the proposed approach includes integrating high-frequency IoT sensor data from mixer drums with a Bi-LSTM neural network grounded in concrete rheology, for the CPS to continuously monitor and predict workability, thus enabling timely, automated recommendations to operators in transit.

Field deployment of over 68 deliveries demonstrated that the CPS achieves an average estimation error of ±0.9 mm, surpassing manual measurement accuracy and supporting actionable in-transit decision-making. However, the evaluation also revealed limitations. The Bi-LSTM model generated a false-positive recommendation for an unfamiliar mix, highlighting issues of model generalizability and the enduring importance of operator expertise. These originate from the limited diversity of concrete classes within the collected dataset. Lastly, the solution will also be evaluated in remote areas with limited 4G coverage to ensure no data packets between the IoT devices and the edge device are lost during transit.

Along with the above-mentioned next steps, to mitigate future false-positives, the CPS will be extended to incorporate a continual learning pipeline, where lower-level layers of the Bi-LSTM will be frozen and high-level layers of the model will be retrained on continuously collected datasets from new mixes. Such few-shot updates will be performed on the edge with elastic-weight-consolidation regularization to avoid catastrophic forgetting. Related IoT-enabled architectures for real-time monitoring have been demonstrated [52], which provide a suitable foundation that could be integrated with adaptive models to improve predictions during construction.

These findings highlight both the promise and the current limitations of data-driven CPS in construction logistics. For practical adoption, such systems should function as decision support tools, augmenting, rather than replacing, skilled human operators, and need to incorporate mechanisms for continual learning and operator feedback. Future work will focus on expanding model training to include a broader range of concrete classes and operational conditions to better ensure model robustness and generalizability, and on embedding adaptive feedback loops to further enhance system robustness and interpretability.