Intelligent prediction and monitoring system for mushroom drying process

Abstract

This invention relates to an intelligent prediction and monitoring system for the mushroom drying process, belonging to the technical fields of machine vision and edible fungi deep processing. The system includes: an image acquisition and region segmentation module, a visual feature screening module, a drying stage identification module, and a moisture content prediction and closed-loop control module. It acquires images of mushrooms during the drying process and extracts images of effective mushroom regions; extracts initial visual features, eliminates redundant features through correlation analysis, and selects the optimal subset of visual features using a dynamic feature importance ranking mechanism; constructs a drying stage identification model, performs mushroom segmentation and real-time identification of drying stages, and outputs the results; combines drying environment conditions and mushroom weight change data, predicts moisture content through a hyperparameter-optimized regression model, constructs a dynamic closed-loop control mechanism, and generates control commands. This system improves the accuracy of moisture content prediction and drying stage identification, optimizes the drying process, and enhances the quality and production efficiency of dried mushrooms.

Description

Technical Field

[0001] This invention belongs to the field of machine vision and edible fungi deep processing technology, specifically relating to an intelligent prediction and monitoring system for the mushroom drying process. Background Technology

[0002] As a food and medicinal resource with both nutritional and economic value, the quality of dried mushrooms directly determines their market acceptance and commercial value. The drying process is a core element in ensuring the stable quality of dried mushroom products. During the mushroom drying process, moisture content is a key quality indicator for measuring the drying endpoint and ensuring product storage stability and food safety. Its precise control directly affects the taste, nutrient retention rate, and shelf life of the mushrooms.

[0003] Currently, the core parameters of mushroom drying processes in the industry (such as temperature, humidity, and wind speed) still heavily rely on the subjective experience of operators. Drying progress and moisture content are judged by visually observing the mushrooms' color depth, shrinkage, and other appearance characteristics. This traditional method has significant drawbacks: firstly, the judgment criteria vary from person to person, lacking a unified quantitative basis, leading to inconsistent moisture content across different batches and even within the same batch, resulting in significant quality fluctuations; secondly, manual observation inherently has a lag, making continuous monitoring of the drying process impossible, easily leading to over-drying or under-drying, resulting in resource waste or product spoilage.

[0004] To overcome the limitations of manual operation, some studies have attempted to apply computer vision and machine learning technologies to the field of agricultural product drying monitoring. However, in the case of mushroom drying, several technical bottlenecks remain: First, mushrooms undergo drastic morphological and appearance changes during drying, such as color transitioning from light to dark, surface shrinking from smooth to wrinkled, and significant volume reduction. Furthermore, when densely packed, they are prone to severe mutual occlusion and adhesion. General image segmentation models (such as the standard YOLO series models) struggle to adapt to this dynamic and complex scenario, failing to achieve accurate separation and effective region extraction of individual mushrooms, thus undermining the reliable foundation for subsequent feature analysis. Second, existing technologies lack a systematic approach to feature extraction, failing to address the specific moisture content of the mushrooms. The key features are not accurately selected, but there are problems of feature redundancy or missing core features, which affects the input quality of subsequent prediction models. Third, the selection of moisture content prediction models is accidental, and many use single models such as SVM and BP neural networks, without fully considering the complex nonlinear relationship between the visual features of mushrooms and moisture content, and without targeted optimization of model hyperparameters, resulting in low prediction accuracy, which is difficult to meet the high-precision requirements of industrial production for moisture content monitoring. Fourth, existing solutions mostly stay at the level of single parameter monitoring or prediction, and lack a closed-loop system that combines image information, environmental data, weight change data and drying equipment control, so as not to achieve intelligent adaptive control of the drying process.

[0005] In addition, the few existing technologies that have been attempted to be applied to monitoring agricultural product drying either rely on expensive hyperspectral imaging equipment, which has too high a hardware cost and is difficult to promote in small and medium-sized production scenarios; or they lack targeted optimization in feature processing and model design, and are not robust enough when facing special scenarios of mushroom drying (such as changes in light, dynamic changes in morphology, and high-density shading), thus limiting their practical application effects.

[0006] In conclusion, it is essential and has practical application value to design a dedicated automated system that can adapt to the special scenario of drying deer antler mushrooms and can accurately predict their core quality indicators. Summary of the Invention

[0007] To address the aforementioned problems in the existing technology, this invention provides an intelligent prediction and monitoring system for the mushroom drying process. The objective of this invention can be achieved through the following technical solution: An intelligent prediction and monitoring system for the mushroom drying process, comprising: an image acquisition and region separation module, a visual feature screening module, a drying stage identification module, and a moisture content prediction and closed-loop control module. The image acquisition and region separation module acquires images of the mushrooms in real time during the drying process, and separates the mushroom area from the background area using image segmentation technology to extract the effective area image of the mushrooms. The visual feature filtering module extracts initial visual features based on the effective region image, eliminates redundant features through correlation analysis, and filters out the optimal subset of visual features related to the moisture content of mushrooms by combining a dynamic feature importance ranking mechanism. The intelligent recognition module for the drying stage constructs a drying stage recognition model, performs mushroom segmentation and real-time recognition of the drying stage, and outputs the recognition results of the drying stage. The moisture content prediction and closed-loop control module takes the optimal visual feature subset as the core input, combines the drying environment status data and mushroom weight change data, and performs non-contact prediction of mushroom moisture content through a regression model optimized by hyperparameters. Based on the moisture content prediction results and the preset drying standards, a dynamic closed-loop control mechanism is constructed to generate control commands for adjusting the process parameters of the drying equipment and execute closed-loop control of the drying process.

[0008] Specifically, the image segmentation technique includes: performing segmentation processing on the image based on a pre-trained deep learning model to separate the mushroom region from the background region, extracting the effective mushroom region image, and simultaneously performing edge optimization and interference pixel removal.

[0009] Specifically, the process of simultaneously performing edge optimization and interference pixel removal is as follows: after image segmentation, the edges of the mushroom region are refined and smoothed, and isolated background residual interference points are identified and removed by combining pixel connectivity analysis, while preserving the fine structural features of the mushroom.

[0010] Specifically, the process of extracting initial visual features is as follows: Based on the effective region image, in different color representation spaces, extract the mean and dispersion information of hue-related features, saturation-related features, brightness-related features, and gray-level-related features that reflect the dynamic changes in mushroom color; based on the contour information of the effective region image, extract relevant features that can quantify basic morphological attributes and relevant features that can characterize complex morphological changes; based on a specific texture analysis matrix, calculate contrast-related features, correlation-related features, energy-related features, uniformity-related features, and entropy-related features that can describe the degree of wrinkles on the mushroom surface, and integrate them to generate an initial visual feature set.

[0011] Specifically, the process of eliminating redundant features through correlation analysis is as follows: using the Pearson correlation coefficient analysis method, pairing analysis is performed on the features in the initial visual feature set, the linear correlation strength between features is calculated and the corresponding correlation coefficient is generated, a correlation coefficient judgment threshold is set, the absolute value of the correlation coefficient is compared with the correlation coefficient judgment threshold, and based on the comparison result, redundant features are eliminated to generate a feature set after correlation screening.

[0012] Specifically, the dynamic feature importance ranking mechanism includes: using the random forest algorithm to quantify the contribution of features in the feature set after correlation screening to water content prediction, generating importance scores for the features, sorting all features from high to low importance scores, selecting features with high scores that play a core role in characterizing the water content of mushrooms, and generating the optimal visual feature subset.

[0013] Specifically, the process of constructing the drying stage identification model is as follows: based on the mainstream segmentation model framework, a multi-scale feature capture mechanism and an efficient attention mechanism are integrated in the feature extraction module; at the same time, a large kernel separable attention mechanism is introduced in the fast spatial pyramid pooling module, and an optimized loss function is adopted. The model is trained and optimized using a mushroom image dataset labeled with information on different drying stages.

[0014] Specifically, during the training of the recognition model in the drying stage, the labeled image dataset is expanded, and diverse training samples are generated by simulating different light intensities, shooting angles and slight occlusion scenarios. At the same time, an early stop mechanism is introduced to monitor the performance of the model validation set.

[0015] Specifically, the process of constructing the dynamic closed-loop control mechanism is as follows: the moisture content prediction result is compared with the preset drying standard in real time, and the control demand priority is analyzed in combination with the drying stage identification result; an initial control scheme is generated based on the comparison result, and the control command is generated; during the execution of the control command, feedback data of system operation is continuously acquired, and the updated moisture content prediction result is regenerated; the change in moisture content before and after control is compared, and based on the comparison result and the feedback data, the initial control scheme is corrected and executed again.

[0016] Specifically, the hyperparameter-optimized regression model is constructed based on the gradient boosting framework, and the hyperparameters of the regression model are iteratively optimized using a genetic algorithm. Through iterative calculation and performance evaluation, the optimal combination of hyperparameters suitable for the mushroom moisture content prediction scenario is selected.

[0017] Specifically, the non-contact prediction process for mushroom moisture content is as follows: using the optimal subset of visual features as the core input, simultaneously acquiring drying environment status data and mushroom weight change data, and simultaneously performing collaborative integration and standardization processing to generate a comprehensive data set; inputting the comprehensive data set into the hyperparameter-optimized regression model, mining the complex nonlinear correlation between the data and mushroom moisture content and completing the computational processing; and outputting the current mushroom moisture content prediction result.

[0018] Specifically, the process of generating the comprehensive data set is as follows: first, the optimal visual feature subset, the drying environment status data, and the mushroom weight change data are checked for consistency and abnormal data are removed. Then, the data is normalized using a unified quantification standard. The processed data is then integrated according to a preset data structure framework to generate the comprehensive data set.

[0019] Specifically, the process of generating control instructions for adjusting the process parameters of the drying equipment is as follows: based on the moisture content prediction results and the current drying stage identification results of the mushrooms, a comparison and analysis is performed with the preset moisture content threshold range and the standard values ​​of the process parameters corresponding to the drying stage; based on the degree of deviation generated by the comparison, the optimal process parameter adjustment amount for the current mushroom drying state is generated; based on the optimal process parameter adjustment amount, the control instructions are generated.

[0020] Compared with the prior art, the present invention has the following significant advantages: Achieving non-contact prediction of mushroom moisture content: Image segmentation is optimized through a pre-trained deep learning model, combined with edge refinement and interference pixel removal techniques to ensure accurate extraction of the effective mushroom region, laying a reliable foundation for feature analysis. Redundant features are eliminated using correlation analysis, and core moisture content-related features are selected using a random forest algorithm. This is then combined with a gradient boosting regression model optimized by a genetic algorithm, integrating visual features, drying environment data, and weight change data to fully explore the complex non-linear relationships between multiple data sources. This improves the accuracy of moisture content prediction, with prediction errors lower than traditional single-model and manual judgment methods. Non-contact detection avoids damage to the mushroom morphology.

[0021] A closed-loop intelligent control system for the drying process is constructed: Based on a drying stage identification model, the drying status of mushrooms is determined in real time. This model integrates multi-scale feature capture and attention mechanisms, exhibiting excellent robustness after training on diverse datasets and adaptability to different drying scenarios. Based on the moisture content prediction results and drying stage identification conclusions, a real-time comparison is performed with preset process standards to automatically generate optimal process parameter adjustment instructions. This enables dynamic adaptive adjustment of parameters such as temperature, humidity, and wind speed, completely changing the traditional open-loop control mode that relies on manual experience. This effectively avoids over-drying or under-drying issues, ensuring consistent moisture content across different batches of products.

[0022] Enhancing the stability and robustness of monitoring in complex scenarios: Addressing the complexities of mushroom drying processes, such as morphological shrinkage, surface wrinkling, and dense occlusion, image segmentation optimization and multi-scale feature fusion techniques are employed to achieve precise separation of mushroom regions and preservation of fine structures. The model training process incorporates diverse scenario simulations including lighting, angle, and occlusion, combined with an early-stop mechanism and data augmentation strategies, significantly improving the system's adaptability to environmental fluctuations in actual production. This addresses the weakness of traditional visual monitoring technologies in lacking robustness in dynamic and complex scenarios.

[0023] Lowering application costs and operational barriers, and adapting to industrial deployment: It eliminates the need for expensive hardware such as hyperspectral imaging; core monitoring functions can be achieved using conventional image acquisition devices, reducing equipment investment costs. Through automated feature extraction, intelligent prediction, and closed-loop control, it reduces reliance on the experience of professional operators, lowers labor intensity and human error, and simplifies system operation. It is suitable for large-scale industrial production lines as well as small- to medium-scale processing scenarios, possessing broad promotional value.

[0024] Ensuring the quality of dried mushroom products and improving production efficiency: Precise moisture content control and dynamic adjustment of process parameters can maximize the retention of nutrients and flavor compounds in mushrooms, reduce spoilage, discoloration, and nutrient loss caused by improper drying, and improve product shelf life and market acceptance. The closed-loop control mode optimizes hot air utilization efficiency, avoids energy waste, and significantly reduces energy consumption compared to traditional drying equipment. It also reduces mushroom loss rate, increases unit yield, and balances ecological and economic benefits, promoting the upgrading of the mushroom processing industry towards standardization, intelligence, and green practices. Attached Figure Description

[0025] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0026] Figure 1 This is a system flowchart of the present invention; Figure 2 This is a system structure diagram of the present invention. Detailed Implementation

[0027] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0028] Please see Figures 1-2 A smart prediction and monitoring system for mushroom drying process includes: an image acquisition and region separation module, a visual feature screening module, a drying stage identification module, and a moisture content prediction and closed-loop control module. The image acquisition and region separation module acquires images of the mushrooms in real time during the drying process, and separates the mushroom area from the background area using image segmentation technology to extract the effective area image of the mushrooms. The visual feature filtering module extracts initial visual features based on the effective region image, eliminates redundant features through correlation analysis, and filters out the optimal subset of visual features related to the moisture content of mushrooms by combining a dynamic feature importance ranking mechanism. The intelligent recognition module for the drying stage constructs a drying stage recognition model, performs mushroom segmentation and real-time recognition of the drying stage, and outputs the recognition results of the drying stage. The moisture content prediction and closed-loop control module takes the optimal visual feature subset as the core input, combines the drying environment status data and mushroom weight change data, and performs non-contact prediction of mushroom moisture content through a regression model optimized by hyperparameters. Based on the moisture content prediction results and the preset drying standards, a dynamic closed-loop control mechanism is constructed to generate control commands for adjusting the process parameters of the drying equipment and execute closed-loop control of the drying process.

[0029] Specifically, the image segmentation technique includes: performing segmentation processing on the image based on a pre-trained deep learning model to separate the mushroom region from the background region, extracting the effective mushroom region image, and simultaneously performing edge optimization and interference pixel removal.

[0030] Specifically, the process of simultaneously performing edge optimization and interference pixel removal is as follows: after image segmentation, the edges of the mushroom region are refined and smoothed, and isolated background residual interference points are identified and removed by combining pixel connectivity analysis, while preserving the fine structural features of the mushroom.

[0031] This embodiment takes the image acquisition and region separation during the drying process of deer antler mushrooms as an example. The specific implementation process is as follows: Image acquisition device: An industrial RGB camera is used, installed in the middle layer inside the drying equipment. The lens is vertically facing the plane where the mushrooms are placed, and the vertical distance between the center of the lens and the area where the mushrooms are placed is L. This ensures that the camera's field of view completely covers the entire material tray area and can clearly capture the fine structures of the mushrooms, such as branches and spines. The camera parameters are set as follows: resolution W×H pixels, frame rate F frames / second, exposure time S milliseconds, and ISO sensitivity G. The equipment's built-in anti-fog and dustproof lens cover reduces the impact of moisture in the drying environment on the image and avoids image blurring caused by condensation on the surface of the mushrooms.

[0032] Auxiliary lighting device: N sets of diffuse LED light sources are symmetrically installed on both sides and above the camera. The wavelength range of the light source is λ1-λ2 nanometers, and the light intensity can be adjusted to Q lux. The side-above oblique lighting method is adopted to avoid the excessive shadow of the branches of the deer antler mushroom or the interference of surface reflection caused by strong direct light. At the same time, it ensures that the brightness contrast between the deer antler mushroom area and the background area is ≥R, highlighting the three-dimensional shape and surface texture characteristics of the deer antler mushroom.

[0033] Data transmission module: Using an industrial Ethernet interface, the image data captured by the camera is transmitted to the back-end processing unit in real time, with a transmission rate of ≥V megabits / second, ensuring that there is no delay or loss of image data during the dynamic drying process of deer antler mushrooms, thus providing a guarantee for subsequent real-time segmentation processing.

[0034] Real-time image acquisition: After the system starts, the image acquisition device continuously captures images of the dried antler mushrooms according to a preset acquisition cycle (one frame per T1 second). During the acquisition process, a hardware synchronization mechanism ensures that the camera and lighting device work together to maintain stable lighting intensity for each frame, avoiding distortion of the antler mushrooms' color and texture due to light fluctuations. The acquired images are stored in a lossless format in the processing unit's buffer, awaiting further processing.

[0035] Image segmentation based on pre-trained deep learning models: The pre-trained U2Net deep learning model is invoked to perform segmentation processing on the acquired raw images: The model has been trained on a labeled image dataset containing P-type edible and medicinal fungi (focusing on different growth states and drying stages of deer antler mushrooms). The training iterations were K times. The optimized model weight file is stored locally in the processing unit and has the ability to adapt to the scene of deer antler mushrooms growing in clusters and branching. After the original image is input into the model, the model performs multi-scale feature extraction on the image through an encoder-decoder structure. Based on the pixel difference between the antler mushroom region (pixel features are represented by specific color ranges and texture density) and the blue background region, a binarized segmentation mask is generated. In the mask, the region with a pixel value of A corresponds to the antler mushroom target region, and the region with a pixel value of B corresponds to the background region. Based on the segmentation mask, an initial effective image containing only the antler mushroom region is cropped from the original image, completing the initial separation of the antler mushroom region from the background region. At this point, the initial effective image has initially eliminated background interference such as the tray and the inner wall of the drying oven.

[0036] Edge optimization processing: The initial effective images of *Deer Antler Mushroom* obtained from preliminary separation were processed to refine and smooth the edges, focusing on preserving the fine structures such as branches and spines of *Deer Antler Mushroom*. An M×M Gaussian convolution kernel is constructed, and the kernel parameters are generated using a normal distribution function. The convolution smoothing intensity is controlled to avoid excessive blurring, taking into account the thin edges of the antler mushroom. Gaussian convolution kernels are applied to the edge regions of the initial effective image of *Mammillaria fuciformis*, and weighted average calculation is performed on the edge pixels to eliminate jagged edge defects generated during the segmentation process, especially optimizing the edge regularity at the branch connection and the root of the thorn of *Mammillaria fuciformis*. A morphological thinning algorithm is used to moderately shrink the optimized edges, accurately preserving the fine structural features of the mushroom, such as the arrangement of the spines and the angle of the branches, avoiding the loss of key morphological information due to edge smoothing, and finally obtaining an effective image of the middle of the mushroom with regular edges and complete details.

[0037] Interference pixel removal: By combining pixel connectivity analysis, isolated background residual interference points in the intermediate effective images are removed, while preserving the interstitial structure of the antler mushroom itself. The pixel connectivity criterion is set as 8-neighbor connectivity. That is, if the number of pixels among the 8 neighboring pixels of a certain pixel that satisfy the difference between the pixel value and the pixel value (A) of the antler mushroom region is ≤D and ≥E, then the pixel is determined to be a valid pixel of the antler mushroom region. This criterion can effectively distinguish between isolated interference points and normal background pixels in the gaps between the branches of the antler mushroom. Traverse all pixels in the intermediate valid image and identify isolated pixels that do not meet the connectivity criteria (these pixels are mostly background noise left over from segmentation, light spots caused by water vapor condensation, and other interference). The isolated interfering pixels identified were replaced, and their pixel values ​​were corrected to the average pixel values ​​of the antler mushroom region. At the same time, the pixel features corresponding to the structure of the antler mushroom, such as the gaps between branches and the gaps between the spines, were strictly preserved. Finally, an effective area image of the antler mushroom with no interference, regular edges, and complete fine structure was obtained, which provides a high-quality data foundation for subsequent visual feature extraction (such as texture features and morphological features).

[0038] Specifically, the process of extracting initial visual features is as follows: Based on the effective region image, in different color representation spaces, extract the mean and dispersion information of hue-related features, saturation-related features, brightness-related features, and gray-level-related features that reflect the dynamic changes in mushroom color; based on the contour information of the effective region image, extract relevant features that can quantify basic morphological attributes and relevant features that can characterize complex morphological changes; based on a specific texture analysis matrix, calculate contrast-related features, correlation-related features, energy-related features, uniformity-related features, and entropy-related features that can describe the degree of wrinkles on the mushroom surface, and integrate them to generate an initial visual feature set.

[0039] Specifically, the process of eliminating redundant features through correlation analysis is as follows: using the Pearson correlation coefficient analysis method, pairing analysis is performed on the features in the initial visual feature set, the linear correlation strength between features is calculated and the corresponding correlation coefficient is generated, a correlation coefficient judgment threshold is set, the absolute value of the correlation coefficient is compared with the correlation coefficient judgment threshold, and based on the comparison result, redundant features are eliminated to generate a feature set after correlation screening.

[0040] Specifically, the process of selecting the optimal visual feature subset related to the moisture content of mushrooms is as follows: using the random forest algorithm, the contribution of features in the feature set after correlation screening to the moisture content prediction is quantitatively analyzed, and the importance score corresponding to the feature is generated. All features are sorted from high to low according to their importance scores, and the features with the highest scores and the core role in the characterization of mushroom moisture content are selected to generate the optimal visual feature subset.

[0041] This embodiment continues the scenario of drying deer antler mushrooms, and details the specific implementation process of the visual feature screening module as follows: Data input: The core input data is an image of the effective area of ​​the antler mushroom with no interference, regular edges, and complete fine structure. The drying environment status data (temperature and humidity) and the weight change data of the antler mushroom at the time of image acquisition are synchronously associated as auxiliary references for subsequent feature analysis.

[0042] Processing hardware: It adopts an embedded processing unit, equipped with a multi-core processor and a dedicated graphics processing module, supports parallel computing, and ensures the real-time performance of feature extraction and analysis. The feature processing time for a single frame image is ≤T milliseconds.

[0043] Software support: The processing framework is built on the Python language, integrating the OpenCV library for image feature extraction and the Scikit-learn library for correlation analysis and random forest algorithm implementation, ensuring the efficient implementation of module functions.

[0044] Initial visual feature extraction: Based on the effective area image of *Deer Antler Mushroom*, initial visual features were extracted from three dimensions: color, contour, and texture, and integrated to form an initial visual feature set, as follows: Color feature extraction (reflecting the dynamic color changes of deer antler mushroom): Color representation space selection: covering HSV space, Lab space and grayscale space, to fully capture the color change patterns during the drying process of deer antler mushrooms.

[0045] Feature extraction content: In each color space, extract the mean and dispersion information (standard deviation): HSV space: Extract the mean of hue (H), the standard deviation of hue (H), the mean of saturation (S), the standard deviation of saturation (S), the mean of lightness (V), and the standard deviation of lightness (V); Lab space: Extract the mean value of luminance (L), the standard deviation of luminance (L), the mean value of color channel a, the standard deviation of color channel a, the mean value of color channel b, and the standard deviation of color channel b; Grayscale space: Extract the mean and standard deviation of grayscale values; Color feature set: A total of C1 color-related features were extracted to fully characterize the color gradient process of antler mushroom from wet to dry state.

[0046] Contour feature extraction (quantifying morphological changes in *Mammillaria truncata*): Based on the contour information of the effective region image, morphological features are extracted in two categories: Morphological basic attribute features: extract perimeter (P), area (A), major axis length (X), and minor axis length (Y) to directly quantify the basic morphological size of *Mammillaria fuciformis*. Complex morphological variation characteristics: Based on basic attributes, roundness (RO), thinness (SL), compactness (C), rectangularity (RE), and diagonal length (DL) are calculated. Among them, roundness reflects the overall morphological regularity of deer antler mushroom, thinness characterizes the branching and extension characteristics, and compactness reflects the degree of shrinkage during the drying process. Contour feature set: A total of C2 contour-related features were extracted to accurately capture the volume shrinkage and branching deformation of velvet mushrooms caused by dehydration.

[0047] Texture feature extraction (describing the degree of wrinkling on the surface of the deer antler mushroom): Texture analysis matrix selection: Gray-level co-occurrence matrix (GLCM) is adopted, with the distance parameter set to D and the angle parameter set to θ (including four directions: 0°, 45°, 90°, and 135°) to comprehensively capture the spatial distribution information of the surface texture of the deer antler mushroom; Feature calculation content: Five core texture features are calculated based on GLCM. The calculation formula is as follows: Contrast-related features (Con): Reflect the degree of difference in grayscale values ​​of pixels on the surface of the deer antler mushroom; the more obvious the wrinkles, the higher the contrast. Correlation feature (Corr): Describes the degree of linear correlation between texture pixels. The smoother the surface of the antler mushroom, the higher the correlation. The formula is: i and j represent the gray levels of each pixel in the deer antler mushroom image. This indicates that the gray level in GLCM is The pixels and gray levels are The probability of pixels appearing adjacently in an image. This represents the standard deviation of all pixels in the gray-level co-occurrence matrix of an image of a deer antler mushroom ROI region. This represents the mean value of all pixels in the gray-level co-occurrence matrix of an image of a deer antler mushroom ROI region; Energy-related characteristics (En): Characterize the uniformity of texture. Wet *Deer Antler Mushroom* has a smooth surface and a higher energy value, while in dry conditions, increased wrinkles lead to a lower energy value. The formula is: In the formula, i and j represent the gray levels of each pixel in the deer antler mushroom image. This indicates that the gray level in GLCM is The pixels and gray levels are The probability of adjacent pixels appearing in an image; Homogeneity correlation characteristic (Hom): Positively correlated with energy characteristics but with a different emphasis, the formula is: ; Entropy-related feature (Entr): Reflects the complexity of the texture. The more wrinkles and the rougher the surface, the greater the entropy value. The formula is: ; Texture feature set: Five types of features are extracted in each direction, and a total of C3 texture-related features are extracted to fully describe the dynamic changes of the surface wrinkles of the antler mushroom.

[0048] Initial visual feature set integration: The color features (C1), contour features (C2), and texture features (C3) mentioned above are integrated to form an initial visual feature set containing C = C1 + C2 + C3 features, providing a comprehensive data foundation for subsequent screening.

[0049] Redundancy feature elimination by correlation analysis: Adopt the Pearson correlation coefficient analysis method to eliminate redundant features in the initial visual feature set. The specific process is as follows: Feature pairing and correlation coefficient calculation: Pair up the C features in the initial visual feature set pairwise, forming a total of P pairs of paired combinations. For each pair of paired features (F_i, F_j), based on the collection of N groups of Pleurotus eryngii sample data (each group of samples contains complete feature values and corresponding actual moisture contents), calculate the linear correlation strength between the two through the Pearson correlation coefficient formula. The formula is: ;, and are both one-dimensional vectors, representing the th element in each physical sign of Pleurotus eryngii; represents the mean value of each physical sign of Pleurotus eryngii; Calculate the correlation coefficient r ij (the value range is [-1, 1]), where the closer the absolute value of r ij is to 1, the stronger the linear correlation between the two features.

[0050] Set the correlation coefficient judgment threshold: Combining the actual distribution law of Pleurotus eryngii features, set the correlation coefficient judgment threshold to T (0 < T < 1) to judge whether the features are redundant.

[0051] Redundant feature elimination: Compare the absolute value of the correlation coefficient |rij| of each pair of paired features with the threshold T: If |rij| > T, it indicates that the features F_i and F_j are highly linearly correlated and there is serious redundancy. Retain the feature with clearer physical meaning and higher measurement stability between the two (for example, in color features, preferentially retain the features related to the a channel in the Lab space because it is more sensitive to the change of the moisture content of Pleurotus eryngii), and eliminate the other feature; If |rij| ≤ T, it indicates that the features F_i and F_j have weak correlation and no obvious redundancy, and both are retained in the feature set after correlation screening.

[0052] Generate the feature set after correlation screening: After the above screening, eliminate R redundant features, and obtain a feature set after correlation screening containing C_筛 = C - R features. While retaining the core information, this set reduces the computational complexity of the subsequent model.

[0053] Combine the dynamic feature importance ranking mechanism to screen the optimal visual feature subset: Adopt the random forest algorithm to screen out the optimal visual feature subset strongly related to the moisture content of Pleurotus eryngii from the feature set after correlation screening. The specific process is as follows: Construct a random forest model: Build a random forest model with K decision trees, set the maximum depth of each decision tree to D and the minimum number of samples for node splitting to S, and set the remaining hyperparameters to default values; Use the C_screen features in the feature set after correlation screening as the model input, and the actual moisture content of Hericium coralloides as the output label to construct a training data set and a validation data set (the sample division ratio is 7:3).

[0054] Quantitative analysis of feature importance: Input the training data set into the random forest model for training. During the training process, evaluate the contribution degree of each feature to the moisture content prediction result through the out-of-bag sample error: First, calculate the original coefficient of determination R 2 of the model to measure the model fitting effect. The formula is: , where y^_oob is the out-of-bag sample; y^_i is the prediction of the model for the Hericium coralloides sample i, y_i is the target value of the actual Hericium coralloides sample, y_mean is the average of the actual target values of the Hericium coralloides samples; For each feature F_k, randomly shuffle its values in the out-of-bag samples to obtain a new data set , and recalculate the coefficient of determination after shuffling. The formula is: , where y^_oob(j) is the data when the Hericium coralloides feature j is shuffled; Based on the error change amount before and after shuffling, quantitatively obtain the importance score mda(t,j) of the feature F_k in a single decision tree. The formula is: ; Traverse all decision trees and calculate the total importance score P MDA of the feature F_k. The formula is: , where T is the total number of decision trees in the random forest; The larger the value of P MDA , the higher the contribution degree of the feature to the Hericium coralloides moisture content prediction and the stronger the correlation with the moisture content.

[0055] Feature ranking and optimal subset screening: Sort the C_screen features in the feature set after correlation screening according to the importance score P MDA from high to low to obtain the feature ranking sequence [F_1', F_2',..., F_C_screen'] (where F_1' has the highest importance score). Combining with the model prediction performance requirements, select the top M features (M < C_screen) with the highest scores. These features play a core role in characterizing the Hericium coralloides moisture content and have no redundancy and complementary information.

[0056] Generate the optimal visual feature subset: The final M core features selected constitute the optimal visual feature subset. This subset retains key information that is strongly correlated with moisture content and has the characteristics of simplified dimensions and high computational efficiency.

[0057] Specifically, the process of constructing the drying stage identification model is as follows: based on the mainstream segmentation model framework, a multi-scale feature capture mechanism and an efficient attention mechanism are integrated in the feature extraction module; at the same time, a large kernel separable attention mechanism is introduced in the fast spatial pyramid pooling module, and an optimized loss function is adopted. The model is trained and optimized using a mushroom image dataset labeled with information on different drying stages.

[0058] Specifically, during the training of the recognition model in the drying stage, the labeled image dataset is expanded, and diverse training samples are generated by simulating different light intensities, shooting angles and slight occlusion scenarios. At the same time, an early stop mechanism is introduced to monitor the performance of the model validation set.

[0059] This embodiment continues the scenario of drying deer antler mushrooms, detailing the construction, training, and inference process of the identification model during the drying stage. The specific implementation process is as follows: Data input: The original RGB images acquired by the image acquisition module are used as input, covering different stages of the entire drying cycle of deer antler mushrooms (wet stage, semi-dry stage, dry stage). The image resolution is uniformly adjusted to W×H pixels to ensure the consistency of the input data format.

[0060] Drying stage identification model construction process: The mainstream YOLOv8-seg segmentation model is selected as the basic framework. This framework has the advantages of being lightweight, having fast inference speed, and high segmentation accuracy. It is suitable for the real-time monitoring needs of the deer antler mushroom drying scenario. Its core structure includes a backbone feature extraction network, a neck feature fusion network, and a head detection and segmentation network, providing a stable foundation for subsequent module improvements.

[0061] Model structure optimization design: To address the complexities of the drying process of deer antler mushrooms, including shrinkage, dense occlusion, and changes in surface texture, the basic framework was improved, and a dedicated drying stage recognition model (CM-YOLO) was constructed. Specific optimizations are as follows: Feature extraction module improvement: In the C2f module of the backbone network, FasterNetBlock is integrated with an efficient multi-scale attention mechanism. FasterNetBlock employs a structure that combines grouped pointwise convolution with depthwise convolution, which reduces computational cost (FLOPs are reduced by F%) while improving feature extraction efficiency and adapting to the hardware resources of edge computing units. The efficient multi-scale attention mechanism enhances the model's ability to identify the differences in the size of branches and the morphological differences at different drying stages of antler mushrooms by capturing feature information at different scales in parallel (receptive field ranges are S1×S1, S2×S2, and S3×S3, respectively), thus avoiding segmentation errors caused by changes in target scale.

[0062] Improvements to the Fast Spatial Pyramid Pooling (SPPF) module: Introduction of Large Kernel Separable Attention (LSKA) mechanism into the SPPF module: This mechanism expands the model's receptive field by using a large-size separating convolutional kernel (K×K kernel size), enhances the ability to capture global contextual information, effectively solves the problem of mutual occlusion when deer antler mushrooms are densely arranged, and improves the individual segmentation accuracy in occluded scenes. Meanwhile, a separable convolutional structure is used to reduce computational complexity, ensuring that the model can still maintain efficient inference speed after introducing large kernel attention, balancing accuracy and real-time requirements.

[0063] Loss function optimization: The EIoU (Enhanced Intersection over Union) loss function is used instead of the original loss function. This function, based on the IoU loss, further considers the distance between the target's center point, aspect ratio differences, and scale differences. Construction and expansion of labeled image datasets: Original dataset construction: Images of the entire drying process of *Deer Antler Mushroom* from different batches and growth stages were collected, totaling N original images. These images were divided into three categories according to the drying stage: wet stage (characterized by light color, smooth surface, and full volume), semi-dry stage (characterized by darker color, slight surface wrinkling, and volume shrinkage), and dry stage (characterized by dark color, severe surface wrinkling, and smallest volume). Professional annotators used the LabelMe tool to annotate the images, including the segmentation mask of the target region of *Deer Antler Mushroom* and the corresponding drying stage label, forming the original annotated dataset.

[0064] Dataset augmentation: To improve the model's generalization ability, the original labeled dataset is augmented in various ways to generate M augmented samples (M=N×K, where K is the augmentation factor). Specific augmentation methods are as follows: Simulate different light intensities: By adjusting the image brightness (brightness adjustment range is L1-L2 times the original brightness) and contrast (contrast adjustment range is C1-C2 times the original contrast), simulate the imaging effect under different lighting conditions inside the drying equipment; Simulate different shooting angles: Rotate the image (rotation angle range is -θ°~+θ°) to simulate changes in shooting angle caused by camera installation deviation or slight offset of material tray; Simulate slight occlusion scenarios: By adding occlusion areas of random shapes (occlusion area ≤ 0% of the total image area), simulate slight occlusion caused by branching intersections or a small amount of impurities in the deer antler mushroom. Other enhancement methods include horizontally flipping the image and slightly scaling it (by a scaling factor of S1 to S2 times the original size) to further enrich the sample diversity.

[0065] Dataset partitioning: The expanded labeled dataset is divided into training set, validation set and test set in a ratio of 7:2:1. The training set is used for model parameter learning, the validation set is used for performance monitoring during training, and the test set is used for final model performance evaluation.

[0066] Model training optimization: Training parameter settings: The total number of model training iterations is set to E, the batch size is B, the initial learning rate is LR, and a cosine annealing learning rate scheduling strategy is adopted to dynamically adjust the learning rate during training to avoid the model getting trapped in local optima; the optimizer is AdamW with a weight decay coefficient of WD to suppress model overfitting.

[0067] Early stopping mechanism introduced: To avoid the decline in generalization ability caused by overtraining, an early stopping mechanism is introduced to monitor validation set performance. The monitoring metrics are set as the mean accuracy (mAP) and loss value of the validation set, and the validation set performance is calculated once every T1 iterations. Set the early stopping patience value to P. If the validation set mAP does not improve (improvement rate ≤ Δ%) and the loss value does not decrease (decrease rate ≤ Δ%) in P consecutive iterations, then automatically stop training and save the current optimal model weights (i.e., the model parameters when the validation set performance is optimal).

[0068] Model training process: The preprocessed training set is input into the improved CM-YOLO model. The prediction results are calculated through forward propagation. The error between the prediction results and the true labels is calculated using the EIoU loss function. The model parameters are then updated through backpropagation. During the training process, the training set loss value, the validation set loss value, and the validation set mAP are recorded in real time to dynamically monitor the model training status and ensure model convergence.

[0069] Real-time identification and reasoning process during the drying stage: Image preprocessing: After inputting the real-time acquired raw image of deer antler mushroom into the model, the first preprocessing operation is performed: the image size is scaled to the model input size (W×H pixels), normalization is performed (pixel value is divided by 255 and mapped to the [0,1] interval), and channel order conversion is performed (to adapt to the model input format) to generate the preprocessed image.

[0070] Model inference: The preprocessed image is input into the trained CM-YOLO model. The model extracts and fuses features through the optimized feature extraction module and SPPF module, and then outputs the segmentation mask of the deer antler mushroom and the predicted probability of the drying stage through the head network.

[0071] Post-processing of results: Post-processing of the model output results: Segmentation result optimization: Morphological opening operation is used to remove tiny noise in the segmentation mask, and pixel connectivity analysis is used to fill in the small holes in the segmentation region to obtain accurate antler mushroom segmentation results; Stage identification and determination: The predicted probabilities of the three drying stages output by the model are normalized by Softmax. The stage with the highest probability value (≥ threshold P_th) is selected as the final identification result. If the probability of all stages is lower than the threshold, the "to be confirmed" prompt is output, and further determination is required based on subsequent data.

[0072] Output results: Real-time output of segmented mask images of deer antler mushrooms and corresponding drying stage identification results (wet stage / semi-dry stage / dry stage), and transmission of the results to the moisture content prediction and closed-loop control module to provide a basis for subsequent process parameter adjustments.

[0073] Specifically, the process of constructing the dynamic closed-loop control mechanism is as follows: the moisture content prediction result is compared with the preset drying standard in real time, and the control demand priority is analyzed in combination with the drying stage identification result; an initial control scheme is generated based on the comparison result, and the control command is generated; during the execution of the control command, feedback data of system operation is continuously acquired, and the updated moisture content prediction result is regenerated; the change in moisture content before and after control is compared, and based on the comparison result and the feedback data, the initial control scheme is corrected and executed again.

[0074] Specifically, the hyperparameter-optimized regression model is constructed based on the gradient boosting framework, and the hyperparameters of the regression model are iteratively optimized using a genetic algorithm. Through iterative calculation and performance evaluation, the optimal combination of hyperparameters suitable for the mushroom moisture content prediction scenario is selected.

[0075] Specifically, the non-contact prediction process for mushroom moisture content is as follows: using the optimal subset of visual features as the core input, simultaneously acquiring drying environment status data and mushroom weight change data, and simultaneously performing collaborative integration and standardization processing to generate a comprehensive data set; inputting the comprehensive data set into the hyperparameter-optimized regression model, mining the complex nonlinear correlation between the data and mushroom moisture content and completing the computational processing; and outputting the current mushroom moisture content prediction result.

[0076] Specifically, the process of generating the comprehensive data set is as follows: first, the optimal visual feature subset, the drying environment status data, and the mushroom weight change data are checked for consistency and abnormal data are removed. Then, the data is normalized using a unified quantification standard. The processed data is then integrated according to a preset data structure framework to generate the comprehensive data set.

[0077] Specifically, the process of generating control instructions for adjusting the process parameters of the drying equipment is as follows: based on the moisture content prediction results and the current drying stage identification results of the mushrooms, a comparison and analysis is performed with the preset moisture content threshold range and the standard values ​​of the process parameters corresponding to the drying stage; based on the degree of deviation generated by the comparison, the optimal process parameter adjustment amount for the current mushroom drying state is generated; based on the optimal process parameter adjustment amount, the control instructions are generated.

[0078] This embodiment continues the scenario of drying deer antler mushrooms, integrating the core mechanism of the GA+XGB model and key steps of closed-loop control. The specific process is as follows: GA+XGB prediction model construction: Basic model selection: The XGBoost regression model is used as the core. Its objective function includes a loss term (mean squared error MSE) and a regularization term, which can accurately adapt to the complex nonlinear relationship between visual features, environmental data and moisture content.

[0079] Hyperparameter optimization: Optimization targets: Six key hyperparameters were selected, including learning rate (η), maximum depth of decision tree (max_depth), number of decision trees (n_estimators), minimum weight of leaf nodes (min_child_weight), L2 regularization coefficient (λ), and subsample ratio (subsample).

[0080] Genetic algorithm process: Randomly generate S hyperparameter combinations as the initial population, using the coefficient of determination (R²) as the basis. 2 The fitness function is constructed using the mean absolute error (MAE) and iterated through roulette wheel selection, single-point crossover (crossover probability P_c), and random mutation (mutation probability P_m). The iteration terminates when the fitness improvement over P consecutive generations is ≤Δ%, and the optimal hyperparameter combination is output.

[0081] Model performance: Model determination coefficient R on the test set 2 =0.89, mean absolute error (MAE)=6.599677, meeting the requirements for high-precision prediction in industrial applications.

[0082] Generate comprehensive dataset: Consistency check: Integrate data with the same timestamp, according to the formula Calculate the reference moisture content (MC represents the moisture content of the deer antler mushroom). This indicates the wet weight of the deer antler mushroom. (Indicates the dry weight of deer antler mushrooms), outliers are removed using the 3σ principle, and timestamp consistency is verified.

[0083] Normalization: The Min-Max method is used to map the data to the [0,1] interval to eliminate dimensional differences.

[0084] Data integration: A comprehensive set of data for model input is formed according to the framework of "visual features + environmental data + reference moisture content".

[0085] Non-contact moisture content prediction: Model inference: Input the comprehensive dataset into the optimized XGBoost model, and mine data correlations through gradient boosting operations using multiple decision trees.

[0086] Results output and processing: Output the predicted moisture content (MC_pred) and confidence level. If the confidence level is ≥ C_th (threshold), output it directly; if it is < C_th, trigger an alarm. Combine the reference moisture content (MC_ref) to assist in the judgment. At the same time, store the results and transmit them to the control module.

[0087] Closed-loop control execution: Preset standard configuration: Set corresponding moisture content threshold ranges and standard values ​​for process parameters for each of the three drying stages: Humid phase: Moisture content threshold range [MC_max1,MC_min1], standard temperature (T1), standard relative humidity (RH1), standard wind speed (V1); Semi-dry stage: Moisture content threshold range [MC_max2,MC_min2], standard temperature (T2), standard relative humidity (RH2), standard wind speed (V2); Dry stage: moisture content threshold range [MC_max3, MC_min3], standard temperature (T3), standard relative humidity (RH3), standard wind speed (V3); where MC_maxn and MC_minn are the upper and lower limits of allowable moisture content for each stage, and Tn, RHn, and Vn are the standard values ​​of the optimal process parameters for each stage.

[0088] The specific implementation process of the dynamic closed-loop control mechanism: Comparison and Priority Analysis: The moisture content prediction result (MC_pred) is compared with the preset drying standard in real time, and the priority of control needs is analyzed in combination with the drying stage identification results; Moisture content deviation: ∆MC=|MC_pred-MC_target|, where MC_target is the median of the current stage's moisture content threshold range; Process parameter deviation: Obtain the current actual process parameters of the drying equipment (T_act, RH_act, V_act), ∆T=T_act-T_std, ∆RH=RH_act-RH_std, ∆V=V_act-V_std.

[0089] Initial control scheme and control command generation: Based on the comparison results and deviation analysis, an initial control scheme is generated in combination with the priority of control needs; the specific adjustment rules are as follows: the adjustment is graded according to the size of the moisture content deviation -- when ∆MC≤∆MC1 (small deviation), the process parameters are finely adjusted (adjustment amount ∆T1, ∆RH1, ∆V1); when ∆MC1<∆MC≤∆MC2 (medium deviation), the adjustment is moderate (adjustment amount ∆T2, ∆RH2, ∆V2); when ∆MC>∆MC2 (large deviation), the adjustment is large (adjustment amount ∆T3, ∆RH3, ∆V3), and the adjustment amount does not exceed the equipment operating limits (such as the temperature adjustment range [T_min, T_max]).

[0090] Closed-loop execution: The optimal adjustment is converted into Modbus-RTU protocol control commands and sent to the actuators via the industrial bus; after the equipment is adjusted, new sensor data and image data are collected in real time, and the "prediction-deviation calculation-adjustment" process is repeated. When ∆MC≤∆MC0 (allowable deviation), the current process parameters are maintained, forming closed-loop control.

[0091] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A smart prediction and monitoring system for the mushroom drying process, characterized in that, include: Image acquisition and region separation module, visual feature screening module, drying stage identification module, moisture content prediction and closed-loop control module; The image acquisition and region separation module acquires images of the mushrooms in real time during the drying process, and separates the mushroom area from the background area using image segmentation technology to extract the effective area image of the mushrooms. The visual feature filtering module extracts initial visual features based on the effective region image, eliminates redundant features through correlation analysis, and filters out the optimal subset of visual features related to the moisture content of mushrooms by combining a dynamic feature importance ranking mechanism. The intelligent recognition module for the drying stage constructs a drying stage recognition model, performs mushroom segmentation and real-time recognition of the drying stage, and outputs the recognition results of the drying stage. The moisture content prediction and closed-loop control module takes the optimal visual feature subset as the core input, combines the drying environment status data and mushroom weight change data, and performs non-contact prediction of mushroom moisture content through a regression model optimized by hyperparameters. Based on the moisture content prediction results and the preset drying standards, a dynamic closed-loop control mechanism is constructed to generate control commands for adjusting the process parameters of the drying equipment and execute closed-loop control of the drying process.

2. The system according to claim 1, characterized in that, The image segmentation technique specifically includes: performing segmentation processing on the image based on a pre-trained deep learning model to separate the mushroom region from the background region, extracting the effective mushroom region image, and simultaneously performing edge optimization and interference pixel removal.

3. The system according to claim 2, characterized in that, The specific process of simultaneously performing edge optimization and interference pixel removal is as follows: after the image segmentation process, the edges of the mushroom area are refined and smoothed, and isolated background residual interference points are identified and removed by combining pixel connectivity analysis, while preserving the fine structural features of the mushroom.

4. The system according to claim 1, characterized in that, The specific process of extracting the initial visual features is as follows: Based on the effective area image, in different color representation spaces, extract the mean and dispersion information of hue-related features, saturation-related features, brightness-related features, and gray-level-related features that reflect the dynamic changes in mushroom color; based on the contour information of the effective area image, extract relevant features that can quantify basic morphological attributes and relevant features that can characterize complex morphological changes. Based on a specific texture analysis matrix, contrast-related features, correlation-related features, energy-related features, uniformity-related features, and entropy-related features that can describe the degree of wrinkling on the surface of mushrooms are calculated and integrated to generate an initial set of visual features.

5. The system according to claim 1, characterized in that, The specific process of eliminating redundant features through correlation analysis is as follows: using the Pearson correlation coefficient analysis method, pairing analysis is performed on the features in the initial visual feature set, the linear correlation strength between features is calculated and the corresponding correlation coefficient is generated, a correlation coefficient judgment threshold is set, the absolute value of the correlation coefficient is compared with the correlation coefficient judgment threshold, and based on the comparison result, redundant features are eliminated to generate a feature set after correlation screening.

6. The system according to claim 1, characterized in that, The dynamic feature importance ranking mechanism specifically includes: using the random forest algorithm to quantify the contribution of features in the feature set after correlation screening to water content prediction, generating importance scores for the features, sorting all features from high to low importance scores, selecting features with high scores that play a core role in characterizing the water content of mushrooms, and generating the optimal visual feature subset.

7. The system according to claim 1, characterized in that, The specific process of constructing the drying stage identification model is as follows: based on the mainstream segmentation model framework, a multi-scale feature capture mechanism and an efficient attention mechanism are integrated in the feature extraction module; at the same time, a large kernel separable attention mechanism is introduced in the fast spatial pyramid pooling module, and an optimized loss function is adopted. The model is trained and optimized using a mushroom image dataset labeled with information on different drying stages.

8. The system according to claim 1, characterized in that, The specific process of constructing the dynamic closed-loop control mechanism is as follows: the moisture content prediction result is compared with the preset drying standard in real time, and the control demand priority is analyzed in combination with the drying stage identification result; an initial control scheme is generated based on the comparison result, and the control command is generated. During the execution of the control command, feedback data of system operation is continuously acquired, and updated moisture content prediction results are regenerated; the change in moisture content before and after regulation is compared, and based on the comparison results and feedback data, the initial regulation scheme is corrected and executed again.

9. The system according to claim 1, characterized in that, The hyperparameter-optimized regression model is specifically constructed as follows: a regression model is built based on the gradient boosting framework, and a genetic algorithm is used to iteratively optimize the hyperparameters of the regression model. Through iterative calculation and performance evaluation, the optimal combination of hyperparameters suitable for the mushroom moisture content prediction scenario is selected.

10. The system according to claim 1, characterized in that, The specific process of performing non-contact prediction of mushroom moisture content is as follows: using the optimal visual feature subset as the core input, simultaneously acquiring drying environment status data and mushroom weight change data, and simultaneously performing collaborative integration and standardization processing to generate a comprehensive data set; inputting the comprehensive data set into the hyperparameter-optimized regression model, mining the complex nonlinear correlation between the data and mushroom moisture content and completing the calculation processing; outputting the current mushroom moisture content prediction result.

11. The system according to claim 10, characterized in that, The specific process for generating the comprehensive data set is as follows: First, the optimal visual feature subset, the drying environment status data, and the mushroom weight change data are checked for consistency and abnormal data are removed. Then, the data is normalized using a unified quantification standard. The processed data is then integrated according to a preset data structure framework to generate the comprehensive data set.

12. The system according to claim 1, characterized in that, The specific process of generating control instructions for adjusting the process parameters of the drying equipment is as follows: based on the moisture content prediction results and the current drying stage identification results of the mushrooms, a comparison and analysis is performed with the preset moisture content threshold range and the standard values ​​of the process parameters corresponding to the drying stage. Based on the degree of deviation generated by the comparison, the optimal process parameter adjustment amount for the current mushroom drying state is generated; based on the optimal process parameter adjustment amount, the control instructions are generated.

Images