Synchronous Detection Method of Physical Quality for Korla Fragrant Pear with Different Damage Types During Storage

Abstract

Mechanical damage reduces the marketability of Korla fragrant pears, severely restricting industry development. To enhance the commercial value of pears, this study investigated the effects of impact, compressive, and combined impact-compressive damage types on the weight loss rate, L*, a*, and b* of pears, and constructed a multi-output prediction model for the weight loss rate, L*, a*, and b* of damaged pears during storage by integrating partial least squares regression (PLSR), support vector regression (SVR), and long short-term memory (LSTM), from which the optimal prediction model was selected to achieve synchronous detection of the physical quality of damaged pears during storage. The results indicated that during storage, the weight loss rate, a*, and b* of pears subjected to different damage types gradually increased with prolonged storage time, while L* gradually decreased. Under the same damage volume situation, pears subjected to impact-static pressure combined action exhibited the fastest storage quality change speed, followed by impact action, static pressure action. The SVR multi-output model demonstrated optimal performance in predicting the weight loss rate, L*, a*, and b* of damaged pears during storage, achieving mean coefficient of determination R², root mean square error (RMSE), and residual prediction deviation (RPD) values of 0.988, 0.513, and 10.072, respectively, for these four quality indicators. These results establish a theoretical foundation for the development of simultaneous monitoring techniques for fruit storage quality.

1. Introduction

Korla fragrant pear is a characteristic pear cultivar from Xinjiang, China [1]. Due to its unique geographical location and climatic conditions, the pear is characterized by its juiciness, thin skin, bright color, and rich nutritional value. It is acclaimed as a “treasure among pears” and possesses high market appeal [2,3]. To ensure the quality and commercial value of pears, industry practitioners typically store them after the concentrated harvest period and release them to the market at opportune times to extend supply [4,5,6]. Due to features such as thin skin and crisp flesh, pears are susceptible to different forms of mechanical damage during harvesting. Impact and compressive loads represent common damage forms encountered during fruit picking and transportation [7]. Pears are susceptible to mechanical damage during postharvest handling. Industry practitioners typically discard damaged fruit before storage. This results in an annual loss rate exceeding 30% due to such damage, which severely impedes the sustainable development of the Korla fragrant pear industry and causes significant economic losses [8]. However, Korla fragrant pears possess a protective peel and self-healing capacity. Those with minor mechanical damage can still be processed into popular commercial products such as pear syrup and fruit wine before quality deterioration occurs, thereby retaining commercial value [9]. Selling these damaged pears before their storage quality deteriorates to levels below industry standards can effectively reduce resource waste from complete discard and enhance the economic returns for practitioners [8,10]. Therefore, clarifying the pattern of quality changes in pears during storage can provide a reference for quality control and optimal sales timing, contributing to enhanced economic benefits for the pear industry.

During actual harvesting and processing, fruits are highly susceptible to different forms of mechanical damage, which affects their storage quality. Extensive research has been conducted on the changes in the storage quality of fruits subjected to static pressure and impact damage. For instance, Shao et al. [11] demonstrated that when Citrus Blanco fruits were subjected to static compression loads exceeding 12 mm, the fruit firmness, soluble solids content, and titratable acidity exhibited a downward trend with prolonged storage time. Pathare et al. [12] indicated that when Omani pomegranate fruit was subjected to impact loading, bruising at high impact levels significantly reduced firmness and geometric mean diameter. Impact bruising level and storage temperature decreased the L*, b*, and browning index of the fruit while increasing its a*. These studies have collectively demonstrated that mechanical damage exerts certain detrimental effects on the storage quality of fruits. However, these investigations primarily focused on the impacts of single-load forms (impact or compressive loads). Considering that fruits are frequently subjected to a combination of loads, such as impact and compression, during actual harvesting operations, it remains necessary to investigate the influence of Combined Load on the storage quality of pears.

Developing an efficient method for detecting the physical quality of damaged pears during storage is essential for optimizing storage process parameters and reducing postharvest loss rates, thereby enhancing commodity circulation efficiency and economic benefits. During storage, moisture loss within pears occurs due to evaporation and respiration consumption, directly leading to changes in fruit weight and skin color [13]. Weight loss rate and color parameters (L*, a*, b*) are commonly used to evaluate storage quality. These are important physical qualities for fruits like pears, significantly impacting consumer purchasing decisions [14]. Machine learning and deep learning, characterized by their capabilities for automatic complex feature extraction and powerful pattern recognition, have been extensively applied in recent years for fruit quality detection and model construction [15]. Liu et al. [16] utilized Fourier transform near-infrared spectroscopy to develop a method for evaluating weight loss rate and color quality of nano-packaged pakchoi during storage. The results demonstrated that the standard normal variate transformation combined with the PLSR algorithm yielded the best prediction performance for weight loss rate (R² = 0.96, RMSEP = 1.432%), while the first derivative combined with PLSR provided the most accurate prediction for color, with an R² of 0.89 and an RMSEP of 3.25 mg/100 g for L*. Xia et al. [17] utilized near-infrared spectroscopy combined with machine learning to develop detection models for the color indices L, a, and b* of Korla fragrant pears. The results showed that the PLSR model, trained on data preprocessed and with feature bands extracted, achieved an R of 0.80 and an RMSE of 1.19 for predicting L; an R of 0.84 and an RMSE of 1.28 for predicting a; and an R of 0.84 and an RMSE of 1.25 for predicting b*. The constructed models for predicting the storage quality of damaged fruits have demonstrated promising results. However, current research on detecting the storage quality of damaged pears is primarily focused on the development and application of single-indicator detection models. In the assessment of agricultural product quality, comprehensive analysis often requires the integration of multiple key indicators, as detection of a single parameter cannot comprehensively reflect quality characteristics. To address this limitation of single-parameter detection models in practical applications, agricultural product multi-quality synchronous detection technology is widely explored. For instance, Ouyang et al. [18] developed a synchronous detection method for multiple chemical components of matcha. The results demonstrated that the PLS model optimized with bootstrapping soft shrinkage yielded R values of 0.8077, 0.7098, 0.7942, and 0.8473 for caffeine, tea polyphenols, free amino acids, and chlorophyll, respectively. Peng et al. [19] combined random forest and near-infrared spectroscopy to develop a synchronous detection method for protein, fat, moisture content, and acidity in goat milk powder, with R values of 0.9846, 0.9642, 0.9915, and 0.9819, respectively. Zhang et al. [20] constructed a Synchronous detection model for four main catechin contents in tea using deep learning near-infrared spectroscopy based on channel and spatial attention mechanisms, showing that this model outperformed existing methods in prediction accuracy and stability, achieving R² values greater than 0.90 for predicting epicatechin, epicatechin gallate, epigallocatechin, and epigallocatechin gallate in tea. Zhu et al. [21] proposed a synchronous detection method for multiple indicators based on a multi-output model combining a residual network and LSTM. This method quantifies nicotine, total sugar, reducing sugar, total nitrogen, potassium, chlorine, and pH in tobacco. The proposed model simultaneously predicts the content of seven chemical components and demonstrates superior predictive performance compared to other existing machine learning methods. The synchronous detection approach introduced in the aforementioned study enables the simultaneous quantification of multiple key indicators, enhancing detection efficiency and showing promising applicability in agricultural product quality inspection. Nevertheless, research on developing Synchronous detection methods for the physical quality of Korla fragrant pears with different damage types during storage is rarely reported.

This study used weight loss rate, L*, a*, and b* as evaluation indicators for the physical quality of pears during storage. It examined the effects of impact load, compressive load, combined impact-compressive load, and different damage levels on pear storage quality. Based on PLSR, SVR, and LSTM, a Synchronous detection method was developed for predicting the weight loss rate, L*, a*, and b* of damaged pears during storage. The optimal prediction model was subsequently identified. This work aims to provide theoretical guidance for predicting the storage quality of damaged fruits.

2. Materials and Methods

2.1. Sample Collection

The Korla fragrant pear samples used in this experiment were harvested on 22 September 2022, from a conventionally managed orchard located at Company 15 of Regiment 10 in Alar City, Xinjiang Production and Construction Corps (Alar, China). As of 2022, the trees were 15 years old. To avoid mechanical damage during collection, all samples were manually harvested with gloves, wrapped in foam nets, and immediately transported to the laboratory. Pears selected for the experiment were uniform in size (130 ± 5 g), smooth-surfaced, undamaged, and free of pest infestation.

2.2. Korla Fragrant Pear Damage Experiment

Prior to conducting the damage experiments on pears, preliminary tests were performed. Under impact load, when the pear surface showed no apparent damage but internal browning occurred in a small area after peeling, the measured damage volume was 900 mm³. When the pear skin formed shallow marks under mechanical load, and the internal flesh was extensively destroyed with partial tissue cell damage, the damage volume was 2400 mm³. When the pear surface exhibited obvious deformation, the flesh was substantially damaged, the tissue structure was disrupted, and juice exuded; the damage volume was 5700 mm³. Consequently, pears subjected to different forms of mechanical damage were categorized into three damage levels—900 mm³, 2400 mm³, and 5700 mm³—to investigate the effect of varying damage severity on the storage quality of the pears.

2.2.1. Impact Damage Experiment

This study employed a custom-built impact damage test bench for pears (Figure 1), which consists of a lifting device and an adsorption device. The specific procedure was as follows: The lifting device raised the vacuum chuck on the cantilever along a graduated linear guide rail to a predetermined height, and the adsorption device secured the pear sample to the chuck. Subsequently, the adsorption device was deactivated, causing the pear to fall freely onto the test platform to generate impact damage. In preliminary tests, pears dropped from heights of 276 mm, 398 mm, and 670 mm produced damage volumes of 900 mm³, 2400 mm³, and 5700 mm³, respectively. Therefore, impact heights of 276 mm, 398 mm, and 670 mm were set for the formal tests. Twenty replicate tests were performed at each height, yielding a total of 60 pears with varying damage volumes.

2.2.2. Static Pressure Damage Experiment

Compressive damage tests on pears were conducted using a universal testing machine (WD-D3-7, Shanghai Zhuoji Instrument Equipment Co., Ltd., Shanghai, China). The pear samples were placed horizontally on the lower platen of the universal testing machine. The upper platen was then driven downward by computer control to compress the samples. The compression rate was set to 5 mm/s. Compression was halted when the specified compression displacement was reached. Simultaneously, the upper platen returned to its starting position. Preliminary tests revealed that damage volumes of 900 mm³, 2400 mm³, and 5700 mm³ corresponded to compression displacements of 3.52 mm, 4.66 mm, and 7.16 mm. Therefore, the compression displacements for the compressive tests were set at 3.52 mm, 4.66 mm, and 7.16 mm. For each compression displacement, 20 repeated tests were performed, yielding a total of 60 pear samples with different damage volumes.

2.2.3. Combined Load Damage Experiment

In actual transportation environments, impact loads and compressive loads typically coexist. To investigate the quality changes in pears under combined impact and compressive loads during storage, a combined impact-compression damage experiment was designed. Based on three damage levels (900 mm³, 2400 mm³, and 5700 mm³), the impact heights for the combined load test were determined as 235 mm, 330 mm, and 570 mm, with compression amounts of 3 mm, 3.7 mm, and 4.4 mm, respectively. First, impact damage tests were conducted on pear samples at drop heights of 235 mm, 330 mm, and 570 mm. Subsequently, static compression tests were performed on the same damaged areas of impacted samples at compression levels of 3 mm, 3.7 mm, and 4.4 mm. Each test group was repeated 20 times, yielding a total of 60 pear samples with different damage volumes.

2.2.4. Pear Damaged Volume Calculation

Pears subjected to damage testing were positioned for 24 h to facilitate the identification and measurement of the damaged tissue volume. At the peel contact site, the damaged area exhibited distinct browning after peeling, with the surface region presenting an elliptical shape (Figure 2a). The pear was bisected equally along the stem-calyx axis to obtain a depth profile of the surface damage, revealing that the damage extended to a certain depth (Figure 2b). To quantify the damage volume, the entire damaged region was treated as a partial ellipsoid for calculation [22]. The major axis a and minor axis b of the damaged area were measured using an electronic vernier caliper. The depth d of the browning region in the pear was then measured from the damage profile. As shown in Equation (1), the damage volume was calculated according to the method described by Li et al. [23].

$$\mathrm{V}={\displaystyle \frac{\mathrm{\pi }\mathrm{d}}{24}}(3\mathrm{ab}+4{\mathrm{d}}^2)$$

(1)

where a represents the major axis of the fruit damage area, mm; b represents the minor axis of the damage area, mm; and d represents the depth of the fruit damage area, mm.

2.3. Pear Measurement of Weight Loss Rate

Fruit weight loss rate refers to the percentage of mass lost by fruit relative to its initial mass at harvest after a certain period of transportation or storage, resulting from its own physiological metabolism and external environmental factors [13]. The collected pears were weighed, labeled, and their initial mass was recorded. During storage, they were weighed every 10 days, with 20 fruits randomly selected as test samples each time. The weight loss rate was calculated from the mass difference between two measurements and then averaged. The weights from the two measurements were substituted into Equation (2) to calculate the weight loss rate.

$$\mathrm{Weight} \mathrm{loss} \mathrm{rate}={\displaystyle \frac{{\mathrm{M}}_0-{\mathrm{M}}_1}{{\mathrm{M}}_0}}\times 100\%$$

(2)

In the formula, M₀ represents the initial weight of the pear (g); M₁ represents the weight of the pear after storage (g).

2.4. Pear Color Measurement

The color of Korla fragrant pear samples was measured using a precision colorimeter (SC-10, Shenzhen 3nh Technology Co., Ltd., Shenzhen, China). The color of fruits and vegetables is commonly expressed using L*, a*, and b* values. L* represents lightness, with higher values indicating a brighter surface and lower values indicating a darker surface. a* represents the red-green chromaticity difference, where positive values denote red and negative values denote green, with higher absolute values indicating a deeper red or green. b* represents the yellow-blue chromaticity difference, where positive values denote yellow and negative values denote blue, with higher absolute values indicating a deeper yellow or blue. Prior to testing, all pear samples were wiped clean. As shown in Figure 3, four equidistant measurement areas were selected and marked on the equatorial region of each pear for color index measurement. The L*, a*, and b* values of each pear sample were measured at the marked points using a precision colorimeter. Subsequently, the average values of the L*, a*, and b* values across the four measurement points were calculated to serve as the final color index value for each pear sample.

2.5. Korla Fragrant Pear Storage Experiment

To investigate the relationship between storage time and quality indicators (weight loss rate, color parameters [L*, a*, b*]) of pears under different damage types and severity levels, a storage experiment was conducted on damaged pears at the Key Laboratory of Modern Agricultural Engineering, Tarim University, Xinjiang. Undamaged pears and pears with different damage levels induced by impact load, compressive load, and impact-compressive combined load were stored under room temperature (the average temperature is 15 °C), with the average humidity maintained at 56%. The storage period was 60 days. Every 10 days, 20 samples from each damage type at different damage levels were randomly selected for measurement of color parameters and weight loss rate; the average values were calculated.

2.6. Construction of the Korla Fragrant Pear Multi-Output Model

2.6.1. PLSR

PLSR is a multivariate statistical analysis method, mainly used to solve regression problems between independent variables (predictor variables, usually denoted as matrix X) and dependent variables (response variables, usually denoted as matrix Y) [24]. PLSR can quantify the contribution of each independent variable to the model through the Variable Importance in Projection, screen key influencing factors [25]. This algorithm extracts latent variables from X and Y through an iterative way, has high predictive and explanatory abilities in complex data scenarios, and is applicable for high-dimensional data and collinear data [26].

2.6.2. SVR

SVR, a supervised learning algorithm based on statistical learning theory, is widely applied in machine learning for its proficiency in handling non-linear relationships and small-sample data [27]. This algorithm identifies an optimal hyperplane that minimizes the cumulative error of all sample points to the hyperplane. By utilizing kernel functions to map the original data into a high-dimensional space, it finds the optimal regression model for complex distributions [28].

2.6.3. LSTM

LSTM is a variant of recurrent neural networks, mainly used to solve the “long-term dependency” problem caused by gradient vanishing or explosion when traditional recurrent neural networks process long-sequence data, specifically for handling time-series data tasks. The core structure of LSTM includes a forget gate (${\mathrm{f}}_{\mathrm{t}}$), input gate (${\mathrm{i}}_{\mathrm{t}}$), and output gate (${\mathrm{o}}_{\mathrm{t}}$), which selectively memorize and forget input information through these three gating units [29].

2.6.4. Model Evaluation

This study employed R², RMSE, and RPD as evaluation criteria for model prediction performance to screen for the optimal prediction model of damaged Korla fragrant pear quality. The calculation formulas for R², RMSE, and RPD are as follows:

$${\mathrm{R}}^2 = 1-{\displaystyle \frac{\sum _{\mathrm{i}=1}^{\mathrm{n}}{\left({\widehat{\mathrm{y}}}_{\mathrm{i}}-{\mathrm{y}}_{\mathrm{i}}\right)}^2}{\sum _{\mathrm{i}=1}^{\mathrm{n}}{\left({\stackrel{\mathrm{-}}{\mathrm{y}}}_{\mathrm{i}}-{\mathrm{y}}_{\mathrm{i}}\right)}^2}}$$

(3)

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{\mathrm{n}}}{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{y}}_{\mathrm{i}}-{\widehat{\mathrm{y}}}_{\mathrm{i}}\right)}^2}$$

(4)

$$\mathrm{RPD}={\displaystyle \frac{\mathrm{SD}}{\mathrm{RMSE}}}$$

(5)

where $\mathrm{n}$ is the number of samples, ${\widehat{\mathrm{y}}}_{\mathrm{i}}$ represents the predicted value of the $\mathrm{i}$-th sample, ${\mathrm{y}}_{\mathrm{i}}$ represents the actual value of the $\mathrm{i}$-th sample, ${\stackrel{\mathrm{-}}{\mathrm{y}}}_{\mathrm{i}}$ represents the mean of the actual values, and SD is the standard deviation of the analytical samples. RPD is a widely adopted model performance evaluation metric in chemometrics, enabling a comprehensive assessment of prediction robustness and practical applicability in analytical scenarios [17]. An RPD > 2.5 indicates excellent model validity with high predictive accuracy, 2.0 < RPD ≤ 2.5 indicates acceptable model performance for preliminary screening, and RPD < 1.4 indicates insufficient model reliability, making it unsuitable for analytical purposes [30].

3. Results and Analysis

3.1. The Change Rule of Weight Loss Rate in Korla Fragrant Pears

Figure 4 illustrates the variation in the weight loss rate of Korla fragrant pears under different damage types with storage time. Under varying damage severities, the weight loss rate of pears subjected to different damage types increased with prolonged storage. When the damage volumes were 900 mm³, 2400 mm³, and 5700 mm³, respectively, the average weight loss rate of damaged pears after 60 days of storage increased from an initial 0% to 8.92%, 9.15%, and 9.50%. At identical damage volumes, pears damaged by combined load exhibited the most rapid increase in weight loss rate and demonstrated the highest weight loss rate at any given storage time, followed by those damaged by impact load and compressive load; undamaged pears exhibited the lowest weight loss rate. This phenomenon may be attributed to secondary damage inflicted by the compressive force on impact-damaged pears under combined load. The pear cells were gradually compressed, with some reaching their yield limit and rupturing. This released intracellular water into the intercellular spaces, increasing internal moisture content and accelerating its dissipation rate. When pears were subjected to the same load type, those with a damage volume of 900 mm³ exhibited a slower increase in weight loss rate. Pears with a damage volume of 2400 mm³ showed a greater rate of weight loss increase under different load types, while those with a damage volume of 5700 mm³ displayed the most pronounced increase.

3.2. The Variation Pattern of Pear Color

3.2.1. Variation Pattern of L*

Figure 5 shows the variation pattern of L* in Korla fragrant pear samples with different damage types during storage. As storage time increased, the L* value of pear samples with different damage types gradually decreased. For pear samples with a damage volume of 900 mm³ (Figure 5a), the decline rate of L* was less pronounced compared to undamaged pears. The average L* of pears with different damage types decreased from an initial value of 64.76 to 61.27. Pear samples with a damage volume of 2400 mm³ (Figure 5b) exhibited a faster decline rate in L* than those with a damage volume of 900 mm³. The average L* of pears with different damage types decreased from 64.76 initially to 61.12. Pear samples with a damage volume of 5700 mm³ (Figure 5c) showed the fastest average L* decline rate, decreasing from 64.76 initially to 60.80.

When the storage time was constant, the lightness value (L*) of damaged pears decreased fastest under the combined load, followed by the loadimpact load and the compressive load. The L* of undamaged pear samples decreased the slowest. With increasing damage severity, the difference in the L* decline rate under different damage types became more pronounced.

3.2.2. Variation Pattern of a*

Figure 6 shows the variation pattern of the a* value of Korla fragrant pear samples subjected to different damage types with storage time. The results indicate that as storage time increased, the a* value of pear samples under different damage types gradually increased, and the green color of the pear peel gradually faded, which is consistent with the findings of Zhang et al. [10]. Under the same degree of damage, the a* value of pears subjected to a combined impact-compressive load increased the fastest with storage time, followed by those subjected to impact load and compressive load. The a* value of undamaged pears increased the slowest.

Under the same storage time, the increase rate of the a* value (red-green chromaticity) in pear samples with different damage types accelerated with increasing damage volume. Among these, pears with a damage volume of 900 mm³ (Figure 6a) exhibited the smallest increase, with their average a* value rising from an initial −12.50 to −1.80 after 60 days of storage. Pear samples with a damage volume of 2400 mm³ (Figure 6b) showed a larger increase in a* value compared to those with 900 mm³ damage; the average a* value for pears across different damage types rose from an initial −12.50 to −1.39. Pear samples with a damage volume of 5700 mm³ (Figure 6c) displayed the fastest increase in average a* value, rising from an initial −12.50 to −1.03, indicating that pears with higher damage levels developed a redder peel color with prolonged storage time.

3.2.3. Variation Pattern of b*

The variation patterns of the b* value in pears under different damage types with storage time are shown in Figure 7. The b* value of damaged pears subjected to various load forms increased with storage time across all three damage levels. Under the same load type, pears with a higher damage volume exhibited a faster increase in peel b* value. Specifically, after 60 days of storage, the average peel b* values for pears damaged by compressive load at damage volumes of 900 mm³, 2400 mm³, and 5700 mm³ were 49.95, 50.33, and 51.75, respectively. For pears damaged by impact load at the same volumes, the average peel b* values were 50.12, 51.24, and 52.35, respectively. For pears damaged by combined load (impact-compressive) at these volumes, the average peel b* values were 50.84, 51.75, and 52.36, respectively. The results indicate that under the same load type, the peel b* value is positively correlated with damage volume; higher damage volumes lead to a more pronounced yellowness in the pear peel.

Under identical damage levels, the rate of change in the b* value of pears subjected to the combined load with storage time was the highest, followed by that under the impact load and then the compressive load. Undamaged pears exhibited the smallest rate of change in the b* value.

3.3. Pear Quality Prediction

This study utilized PLSR, SVR, and LSTM to construct multi-output prediction models for the weight loss rate, L*, a*, and b* of damaged Korla fragrant pears during storage. The damage type inflicted on the pears (undamaged, compressive load damage, loadimpact load damage, and combined load damage), the damaged volume of the pears, and the storage time were used as model inputs, while the four quality indicators of the pears (weight loss rate and color indices L*, a*, b*) served as simultaneous model outputs. Seventy percent of all test data were randomly selected as the model training set, with the remaining 30% serving as the test set.

During model training and prediction phases, the correlations between predicted and measured values of weight loss rate for Korla fragrant pears during storage by PLSR, SVR, and LSTM models are shown in Figure 8a, Figure 8b and Figure 8c, respectively. The R², RMSE, and RPD of each model in the training and test sets are listed in

Table 1. Prediction of weight loss rate, L*, a*, b* in Korla fragrant pears using PLSR, SVR, and LSTM models.

Quality Indicator	Model	Training Stage			Prediction Stage
		R2	RMSE	RPD	R2	RMSE	RPD
weight loss Rate (%)	PLSR	0.972	0.005	5.989	0.968	0.005	5.678
	SVR	0.987	0.004	8.766	0.984	0.003	8.025
	LSTM	0.983	0.004	7.769	0.974	0.004	6.256
L*	PLSR	0.973	0.198	6.145	0.965	0.221	5.442
	SVR	0.992	0.109	11.324	0.992	0.101	11.532
	LSTM	0.989	0.130	9.760	0.975	0.166	6.458
a*	PLSR	0.973	0.620	6.176	0.971	0.646	5.919
	SVR	0.995	0.272	14.189	0.994	0.295	12.791
	LSTM	0.995	0.265	14.947	0.988	0.374	9.398
b*	PLSR	0.989	1.368	9.743	0.964	2.561	5.370
	SVR	0.984	1.691	8.045	0.984	1.654	7.940
	LSTM	0.997	0.721	19.157	0.974	1.999	6.330

. For predicting weight loss rate, all models achieved R² > 0.95 in the test set, indicating that all trained models accurately predicted the weight loss rate of damaged pears during storage. Specifically, the PLSR model yielded R² = 0.968, RMSE = 0.005, and RPD = 5.678 on the test set. The SVR model achieved R² = 0.984, RMSE = 0.003, and RPD = 8.025. The LSTM model attained R² = 0.974, RMSE = 0.004, and RPD = 6.256. These results demonstrate that the SVR model performed best in predicting the weight loss rate of damaged pears during storage, with the highest R² and RPD values and the lowest RMSE.

The correlations between predicted and measured L* values for Korla fragrant pears during storage, as determined by the PLSR, SVR, and LSTM models during the training and prediction phases, are shown in Figure 9a, Figure 9b and Figure 9c, respectively. Results in Table 1 show that for predicting L* of pears, the PLSR model achieved an R² of 0.965, an RMSE of 0.221, and an RPD of 5.442 on the test set. The SVR model achieved an R² of 0.992, an RMSE of 0.101, and an RPD of 11.532 on the test set. The LSTM model achieved an R² of 0.975, an RMSE of 0.166, and an RPD of 6.458 on the test set. The SVR model achieved the highest R² and RPD, and the lowest RMSE on the test set, demonstrating its superior performance in predicting L* for damaged pears during storage, indicating its applicability for quality control.

Figure 10a, Figure 10b and Figure 10c display the correlation between predicted and measured a* values of Korla fragrant pears during storage for the PLSR, SVR, and LSTM models during the model training and prediction phases, respectively; the prediction results are presented in Table 1. For the prediction of pear a*, the PLSR model achieved an R² of 0.971, an RMSE of 0.646, and an RPD of 5.919 on the test set. The SVR model yielded an R² of 0.994, an RMSE of 0.295, and an RPD of 12.791 on the test set. The LSTM model attained an R² of 0.988, an RMSE of 0.374, and an RPD of 9.398 on the test set. The results indicate that the SVR model possesses the highest R² and RPD, along with the lowest RMSE on the test set, demonstrating that the SVR model provides the most accurate predictions for the a* value of damaged pears during storage.

Figure 11a, Figure 11b and Figure 11c illustrate the correlation between predicted and measured b* values of Korla fragrant pears during storage for the PLSR, SVR, and LSTM models during the model training and prediction phases, respectively. The prediction results are presented in Table 1. For predicting pear b, the PLSR model achieved R², RMSE, and RPD values of 0.964, 2.561, and 5.370, respectively, on the test set. The SVR model yielded R², RMSE, and RPD values of 0.984, 1.654, and 7.940, respectively, on the test set. The LSTM model produced R², RMSE, and RPD values of 0.974, 1.999, and 6.330, respectively, on the test set. The SVR model demonstrated the highest R² and RPD, and the lowest RMSE on the test set, indicating that it most effectively predicted the b* values of damaged pears during storage.

The above results indicate that by inputting damage type, damage volume, and storage time into the optimized PLSR, SVR, and LSTM models, the Synchronous detection of physical quality indicators—including weight loss rate and color parameters (L*, a*, b*)—for damaged pears during storage can be achieved, yielding satisfactory prediction performance. Among these, the SVR multi-output model proved to be the optimal model for predicting the weight loss rate, L*, a*, and b* of damaged pears. In the training set, the average R², RMSE, and RPD for predicting pear weight loss rate and color L*, a*, b* were 0.990, 0.519, and 10.581, respectively. In the test set, the average R², RMSE, and RPD were 0.988, 0.513, and 10.072, respectively.

In previous related research, Tian et al. [31] established linear regression models for the compression deformation of citrus fruits versus damage rate and storage deterioration rate, with R² of 0.94 and 0.97, respectively. In comparison, this study investigated not only compressive damage but also the effects of impact damage and impact-compression combined damage on fruit storage quality. The prediction model for storage quality of damaged pears under different load types demonstrated superior predictive performance (with average R², RMSE, and RPD values of 0.988, 0.513, and 10.072, respectively). This improvement is likely attributable to the fact that mechanical damage disrupts pear cell structure, triggering abnormal physiological metabolism (e.g., accelerated respiration, enzymatic browning, moisture loss), leading to highly non-linear quality changes. The SVR model excels at handling such complex non-linear relationships and small-sample data [32]. Zhang et al. [33] constructed a comprehensive quality prediction model for winter jujube by integrating a BP neural network optimized with particle swarm optimization. This model incorporated a weighted combination of 22 physicochemical indicators (including color parameters) during storage, achieving an overall fitting rate exceeding 95%. Ali et al. [34] constructed a pineapple storage quality prediction model by combining infrared thermal imaging technology with PLSR. The constructed model achieved R² values exceeding 0.94 in predicting color, firmness, pH, total soluble solid content, and moisture content. In contrast, the storage quality prediction model for damaged Korla fragrant pears based on SVR developed in this study shows better predictive performance. This advantage may stem from the SVR multi-output model’s ability, compared to traditional models, to effectively capture the inherent correlations among various quality indicators and the highly non-linear mapping relationships between the complex physiological changes in damaged pears during storage and their quality indicators [35]. While the aforementioned studies constructed prediction models for fruit quality indicators during storage, their models could only analyze single components, making it difficult to comprehensively reflect quality characteristics. Furthermore, model construction and computational costs were high, and detection efficiency was low. This study, however, established a synchronous detection method based on an SVR multi-output model for predicting weight loss rate and color quality (L*, a*, b*) of damaged Korla fragrant pears during storage, enabling simultaneous prediction of the weight loss rate, L*, a*, and b* quality indicators.

4. Discussion

Weight loss rate and color (L*, a*, b*) are important indicators for evaluating the storage quality of fruits such as pears [14,36]. This study explored the effects of varying degrees of impact load, compressive load, and combined impact-compressive load on the storage quality of Korla fragrant pears, constructing a multi-output prediction model for weight loss rate, L*, a*, and b* during storage using damage type, damage volume, and storage time as inputs. The weight loss rate of pears under different damage types gradually increased with prolonged storage time, and larger damage volumes resulted in the most pronounced increase in weight loss rate. This is because mechanical damage stimulates metabolic activity within the fruit, accelerating the respiration rate of Korla fragrant pears and decomposing respiratory substrates such as sugars, starches, and organic acids, further promoting water loss and increasing weight loss rate [37]. For the color of Korla fragrant pears, the L* value decreased gradually with prolonged storage time under different damage types, showing a negative correlation with storage time, and larger damage volumes resulted in the most obvious decrease in pear weight loss rate. This aligns with findings by Spricigo et al. [38], indicating that compressive-damaged tomato samples exhibit deeper color changes than undamaged fruit with prolonged storage. Lawati et al. [39] demonstrated this when investigating mechanical damage sensitivity in zucchini storage quality. This occurs because mechanical damage accelerates enzymatic reactions within the fruit, such as hydrolase-mediated decomposition of cell wall components (cellulose, polysaccharides), damaging cell walls and membranes of Korla fragrant pear epidermis [40]. Higher mechanical damage severity caused more extensive internal tissue destruction; this disruption accelerates exosmosis and oxidation of intracellular substances, inducing browning and ultimately reducing skin chroma. The a* and b* values of pears were positively correlated with storage time. With prolonged storage, a* and b* values continuously increased in Korla fragrant pears under different damage types. This phenomenon arises because mechanical damage ruptures cell membranes, releasing vacuolar phenolic substances that contact cytoplasmic polyphenol oxidase, triggering enzymatic browning [41,42]. While chlorophyll content in the peel decreased sharply, anthocyanins accumulated gradually. Higher damage intensities accelerated anthocyanin accumulation, causing the peel to gradually transition toward red or yellow hues [43,44]. Additionally, under identical damage volumes, Korla fragrant pears subjected to combined impact-compressive loads exhibited the fastest storage quality deterioration, followed by impact load and compressive load, while undamaged pears showed the slowest changes. This occurs because Korla fragrant pears contain abundant phenolic compounds, and combined loads inflict secondary damage, enhancing enzymatic reactions. Under polyphenol oxidase catalysis, phenolics oxidize into brown quinones, causing browning [45]. Respiration in Korla fragrant pears intensified, increasing internal ethylene release and accelerating chlorophyll degradation [46].

Compared with previous studies, this study further investigated the effect of the Combined Load (impact and static pressure acting together) on the storage quality of Korla fragrant pears. A synchronous detection method for physical quality, based on a multi-output model, was developed. Compared with single quality detection methods, this synchronous approach significantly improves detection efficiency by enabling the acquisition of multidimensional information reflecting the overall physical state of the fruit through a single measurement. This provides more efficient and reliable technical support for comprehensively evaluating changes in the storage quality of damaged pears and optimizing storage process parameters, serving as a key strategy to reduce postharvest losses and enhance commercial value. In previous studies, Panchbhai et al. [47] developed a multi-quality synchronous detection method for edible oils using multi-output least squares support vector regression. The method achieved root mean square errors of 0.4967%, 0.8400%, and 1.0199% for oleic and linoleic acids, with correlation coefficients of 0.8133, 0.9992, and 0.9981. Zhao et al. [48] introduced a synchronous detection method for tannin and protein content in sorghum using multi-output Gaussian process regression. For tannin and protein prediction sets, R² values were 0.9790 and 0.9500, RMSEP values were 0.0587 and 0.1699, while RPD values were 6.8928 and 4.4710. Comparatively, the SVR-based synchronous detection method for damaged pear storage quality demonstrated superior performance, achieving average R², RMSE, and RPD values of 0.988, 0.513, and 10.072 when predicting weight loss rate, L*, a*, and b*. The superior performance may be attributed to the multi-output SVR model’s enhanced capability in handling non-linear relationships and small datasets. By sharing feature representations and implicitly capturing inter-output correlations, the model yields more robust and accurate predictions on limited data [49]. However, current research still has certain limitations; for instance, it only preliminarily explored quality changes under the single combined load sequence of “impact followed by static compression,” without comparing effects from other combined load forms (e.g., static compression followed by impact, simultaneous application, cyclic impact followed by static compression). During actual fruit harvesting, transportation, or sales, in addition to impact damage, static compression damage, and impact-static compression combined damage, pears are often subjected to damage from other load types, such as vibration or multiple combined loads (vibration-impact, vibration-static compression combined loads) [50]. Additionally, storage experiments on damaged fragrant pears were conducted at room temperature, whereas in practical production, storage temperature significantly influences physical quality during storage [8]. In subsequent research, the quality change patterns of damaged pears under different storage conditions or different harvest periods will be further investigated. Investigating the effects of other load forms, particularly combined loads, on the storage quality of pears will further optimize the model’s practicality and generalizability. Future work will validate the prediction effectiveness of the synchronous detection method for Korla fragrant pear quality during storage on other fruit varieties and key indicators, providing a reference for establishing prediction methods for other quality indicators of Korla fragrant pears and for the determination of storage quality in other fruits.

5. Conclusions

For Korla fragrant pear samples subjected to different damage types, the weight loss rate, a*, and b* values increased progressively with prolonged storage time, while the L* value decreased continuously. Under the same damage type and load, the weight loss rate, a*, and b* values exhibited a positive correlation with the damage volume, whereas the L* value showed a negative correlation. At identical damage volumes, pears subjected to the combined load exhibited the fastest rate of change in storage quality, followed by those subjected to impact load and then compressive load; Undamaged pears demonstrated the slowest rate of change. The SVR multi-output model achieved optimal prediction performance for the weight loss rate and color indices (L*, a*, b*) of damaged pears, with average R², RMSE, and RPD values of 0.988, 0.513, and 10.072, respectively.