Garlic slice hot drying image risk early warning system based on deep learning

Abstract

The application discloses a garlic slice hot drying image risk early warning system based on deep learning, and relates to the technical field of agricultural products processing.The system comprises an image acquisition module, a process parameter acquisition module, a data preprocessing module, a feature extraction module, a bidirectional coupling risk prediction module, a process adjustment scheme generation module, an early warning and control module and a man-machine interaction module.The image acquisition module and the process parameter acquisition module are arranged to cooperate with each other, so that the image data and the process parameter data in the garlic slice hot drying process can be synchronously and real-timely collected.The bidirectional coupling risk prediction model based on the graph neural network is constructed, the image features and the drying process parameters are deeply fused, the risk development trend in the drying process can be accurately predicted, the specific process parameter deviation leading to the risk occurrence can be reversely deduced, and the targeted process adjustment scheme can be automatically generated through the process adjustment scheme generation module.

Description

Technical Field

[0001] This invention relates to the field of agricultural product processing technology, specifically to a risk warning system for garlic slices heat-drying images based on deep learning. Background Technology

[0002] Garlic slices are thin-sheet agricultural products made from fresh garlic through washing, peeling, and slicing. They are a core primary product in the garlic deep-processing industry chain, widely used in food seasoning, health food production, and pharmaceutical raw material preparation. Garlic slice heat-drying images refer to a continuous sequence of images acquired by an image acquisition device deployed inside the drying equipment during the garlic slice heat-drying process. These images visually reflect the changes in the physical characteristics of the garlic slices, such as color, shape, and texture, over drying time. Heat drying is an indispensable key step in garlic slice processing. Its main function is to rapidly reduce the moisture content of the garlic slices, inhibit the growth and reproduction of microorganisms, thereby effectively extending the product's shelf life and facilitating long-distance transportation and long-term storage.

[0003] The heat drying process for garlic slices is a complex heat and mass transfer process, influenced by various process parameters such as temperature, humidity, wind speed, and heating power. During drying, defects such as surface browning, micro-cracks, edge curling, and localized scorching are highly likely to occur. These defects not only severely affect the appearance and commercial value of the garlic slices but also lead to significant loss of heat-sensitive nutrients such as allicin and vitamin C, reducing the product's intrinsic nutritional value and, in severe cases, even rendering the entire batch unusable. Therefore, real-time risk monitoring and early warning systems for the garlic slice heat drying process, enabling timely detection and handling of potential quality problems, are of paramount importance for improving the pass rate of garlic slice products, reducing production costs, and enhancing the economic benefits for enterprises.

[0004] In existing technologies, quality control during the hot drying process of garlic slices mainly relies on the operator's experience and observation, as well as sampling inspection of the finished product after drying. Some solutions that incorporate image detection technology can only identify and classify defects in the dried product, failing to detect early risk signs and intervene in a timely manner during the drying process. Furthermore, existing technologies typically use image data as a single basis for quality judgment, without fully integrating it with key process parameters such as temperature, humidity, and wind speed during the drying process for comprehensive analysis. This results in insufficient accuracy and timeliness of risk warnings, and the inability to automatically generate targeted process parameter adjustment plans based on the root causes of risks. Consequently, it is difficult to achieve intelligent control and proactive intervention in the drying process. Therefore, developing a deep learning-based image-based risk warning system for the hot drying of garlic slices is of great significance. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a deep learning-based image risk early warning system for hot-drying garlic slices. This system can solve the problems of existing technologies being unable to detect early risks in a timely manner, having insufficient accuracy in early warning, and being unable to achieve proactive intervention, thereby significantly improving the quality stability of hot-drying garlic slices and the level of intelligence in the production process.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a risk warning system for garlic slices heat drying image based on deep learning, the system comprising: an image acquisition module, a process parameter acquisition module, a data preprocessing module, a feature extraction module, a bidirectional coupled risk prediction module, a process adjustment scheme generation module, a warning and control module, and a human-computer interaction module;

[0007] The image acquisition module is used to acquire surface state images of garlic slices during the heat drying process and send the acquired image data to the data preprocessing module;

[0008] The process parameter acquisition module is used to acquire process parameters inside the drying oven and send the acquired process parameter data to the data preprocessing module;

[0009] The data preprocessing module is used to preprocess the image data and process parameter data and send the preprocessed data to the feature extraction module;

[0010] The feature extraction module is used to extract image features and process features and send the extracted features to the bidirectional coupled risk prediction module;

[0011] The bidirectional coupling risk prediction module is used to fuse image features and process features to perform risk prediction and process parameter deviation derivation, and send the prediction results to the process adjustment scheme generation module.

[0012] The process adjustment scheme generation module is used to generate process parameter adjustment schemes and send the adjustment schemes to the early warning and control module.

[0013] The early warning and control module is used to issue early warning signals, perform process adjustments, and interact with the control system of the drying equipment.

[0014] The human-computer interaction module is used to display system information, receive operation instructions, and interact with various modules.

[0015] Furthermore, the data preprocessing module performs the following operations during data preprocessing:

[0016] Receive raw image data sent by the image acquisition module and raw process parameter data sent by the process parameter acquisition module;

[0017] The original image data is sequentially processed by Gaussian filtering for noise reduction, pixel value normalization, and histogram equalization.

[0018] Outliers were removed from the original process parameter data using the 3σ criterion. Based on the acquisition timestamps of the image data and process parameter data, each set of process parameter data was matched one-to-one with the image data at the corresponding time point.

[0019] Furthermore, the feature extraction module performs the following operations during feature extraction:

[0020] The system receives preprocessed image data and process parameter data sent by the data preprocessing module, and uses a convolutional neural network to extract the color change rate, shrinkage rate, and surface texture roughness features of garlic slices from the preprocessed image data through the image feature extraction unit.

[0021] The process feature extraction unit uses a temporal convolutional network to extract the temporal variation features of process parameters from the preprocessed process parameter data, and converts the extracted image features and process features into feature vectors of the same dimension.

[0022] The image feature vector and the process feature vector are initially weighted according to the adaptive feature weight fusion formula, which is as follows: ,in, Assign weights to the i-th type of feature. The variance of the offline experiment for the i-th feature is calibrated, where n is the total number of feature categories. The determination was made through statistical analysis of multiple sets of offline characteristic data under the standard process of hot drying of garlic slices.

[0023] Furthermore, the bidirectional coupled risk prediction module performs the following operations when performing risk prediction and process parameter deviation derivation:

[0024] The system receives image feature vectors and process feature vectors sent by the feature extraction module, and inputs the image feature vectors and process feature vectors into a pre-trained graph neural network model.

[0025] Image feature vectors and process feature vectors are deeply fused using graph convolution operations. The fusion process follows a bidirectional coupled feature calculation formula: ,in, For the first Layer fusion feature matrix, For graph structure adjacency matrix, Let L be the input feature matrix of the Lth layer. Let B be the feature fusion weight matrix, B be the bias vector, and σ be the activation function. B is determined iteratively through the backpropagation algorithm during model training;

[0026] Based on the fused feature vector, the risk level of the current drying process and the future risk development trend are output. The contribution of each process parameter to the occurrence of risk is analyzed through backpropagation, and the type of process parameter deviation that leads to the occurrence of risk is output.

[0027] Furthermore, the graph neural network model is trained using a dataset of garlic slices undergoing hot drying, labeled with risk levels and process parameter deviations. The dataset contains multiple sets of image data under different drying conditions and corresponding process parameter data. During training, the graph neural network model continuously adjusts its parameters using a backpropagation algorithm until the model's prediction accuracy reaches a preset training threshold.

[0028] Furthermore, the image acquisition module includes multiple sets of high-speed industrial cameras, which are respectively deployed at the top, middle and bottom of the drying chamber. The multiple sets of high-speed industrial cameras face different areas inside the drying chamber. The multiple sets of high-speed industrial cameras continuously acquire surface state images of garlic slices at preset time intervals. The images acquired by the multiple sets of high-speed industrial cameras cover the entire material placement area inside the drying chamber.

[0029] Furthermore, the process parameter acquisition module includes a temperature sensor, a humidity sensor, a wind speed sensor, and a power sensor. The temperature sensor is installed inside the drying chamber at the air inlet, air outlet, and above the material layer. The humidity sensor is installed inside the drying chamber at the air inlet and air outlet. The wind speed sensor is installed inside the drying chamber above the material layer. The power sensor is electrically connected to the heating device of the drying chamber.

[0030] Furthermore, the process adjustment scheme generation module has a built-in process adjustment rule library, which stores process parameter adjustment methods corresponding to different types of process parameter deviations. Based on the risk level and process parameter deviation type output by the bidirectional coupled risk prediction module, the process adjustment scheme generation module retrieves the corresponding process parameter adjustment method from the process adjustment rule library and generates a process adjustment scheme containing specific adjustment parameters according to the process parameter adjustment amount calculation formula. The process parameter adjustment amount calculation formula is as follows: ,in, This represents the adjustment amount for the k-th process parameter. Let k be the deviation value of the process parameter. This is the proportionality coefficient. The integral coefficient is... and The determination was made by fitting and calibrating historical process data of garlic slices during heat drying.

[0031] Furthermore, the early warning and control module has built-in early warning methods and control strategies corresponding to different risk levels. The early warning and control module triggers the corresponding early warning method according to the risk level output by the bidirectional coupled risk prediction module. The early warning and control module sends the process adjustment scheme output by the process adjustment scheme generation module to the control system of the drying equipment. The control system of the drying equipment adjusts the corresponding process parameters according to the process adjustment scheme.

[0032] Furthermore, the human-machine interaction module includes a display unit and an input unit. The display unit displays real-time process parameters of the drying process, real-time surface state images of garlic slices, current risk level, early warning information, and process adjustment schemes. The input unit receives system parameter setting instructions, process adjustment rule modification instructions, and manual control instructions input by the operator.

[0033] Compared with existing technologies, this deep learning-based risk warning system for garlic slices undergoing heat drying images has the following advantages:

[0034] This invention achieves synchronous real-time acquisition of image data and process parameter data during the hot drying of garlic slices by setting up mutually cooperating image acquisition modules and process parameter acquisition modules. By constructing a bidirectional coupled risk prediction model based on graph neural networks, image features and drying process parameters are deeply integrated. This not only accurately predicts the risk development trend during the drying process, but also reversely deduces the specific process parameter deviations that lead to the risk. The process adjustment scheme generation module automatically generates targeted process adjustment schemes, and the early warning and control module sends the adjustment schemes to the drying control system for execution, realizing proactive intervention in the drying process. This effectively improves the accuracy and timeliness of risk early warning, solves the problems of existing technologies that cannot detect early risks in a timely manner, have insufficient early warning accuracy, and cannot achieve proactive intervention, and significantly improves the quality stability of hot-dried garlic slices and the level of intelligence in the production process.

[0035] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0037] Figure 1 This is a schematic diagram of a deep learning-based risk warning system for garlic slices in heat-drying images.

[0038] Figure 2 A flowchart illustrating the workflow of a deep learning-based risk warning system for garlic slices dried by heat.

[0039] Figure 3 This is a flowchart of the feature extraction module during feature extraction. Detailed Implementation

[0040] 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 below.

[0041] This invention provides a deep learning-based risk warning system for garlic slices during heat drying. The entire system is based on multi-source synchronous data acquisition, with deep learning feature fusion as its core, bidirectional coupled risk prediction as its key, and intelligent process control as its goal. Through image acquisition and process parameter acquisition modules, it acquires real-time surface state image data and process parameter data such as temperature, humidity, wind speed, and heating power during the garlic slice heat drying process. The data preprocessing module performs image noise reduction, normalization, equalization, and outlier removal and temporal matching of process parameters. Then, the feature extraction module uses a convolutional neural network to extract image features such as garlic slice color change rate, shrinkage rate, and surface texture roughness. A temporal convolutional network is used to extract temporal variation features of process parameters and achieves feature vector dimension unification and adaptive weight allocation. Subsequently, the system... A bidirectional coupled risk prediction module based on graph neural networks achieves deep fusion of image features and process features, accurately determining the drying risk level, predicting risk trends, and inversely deducing process parameter deviations. The process adjustment scheme generation module generates quantified process parameter adjustment schemes based on deviation results and a built-in rule base. The early warning and control module triggers corresponding early warning signals according to the risk level and coordinates the drying equipment to perform adjustment operations. The human-machine interaction module realizes full-process data visualization and human command interaction. Ultimately, a fully closed-loop intelligent early warning and control system is constructed, encompassing data acquisition, feature analysis, risk prediction, and proactive regulation. This system completely solves the industry pain points of lagging risk monitoring, insufficient early warning accuracy, and inability to proactively intervene in process parameters during the traditional garlic slice heat drying process, comprehensively improving the intelligence level and product quality stability of garlic deep processing production. A detailed description is provided below with reference to specific embodiments.

[0042] This embodiment applies to a continuous hot drying production line in a large-scale garlic deep processing enterprise. The production line is equipped with a standardized hot air drying box with an effective internal volume of 10m³. 3It can process 500 kg of garlic slices at a time. After the system hardware and software are customized and deployed according to the production scenario, they are started and run according to the established process. The complete implementation process is as follows.

[0043] After the system starts up, it first completes initialization operations, such as... Figure 1 As shown, this system consists of eight core modules: image acquisition module, process parameter acquisition module, data preprocessing module, feature extraction module, bidirectional coupling risk prediction module, process adjustment scheme generation module, early warning and control module, and human-machine interaction module. Each module establishes a stable data interaction link through industrial Ethernet. During the initialization phase, sensor calibration, camera parameter debugging, model loading, and rule base calling are completed to ensure that all modules are in a ready state, laying the foundation for subsequent data acquisition and analysis.

[0044] After initialization, the system enters the core stage of synchronously acquiring images of the garlic slice surface and drying process parameters, such as... Figure 2 As shown, this step is the foundation of the entire system's data input and directly determines the accuracy of subsequent analysis. The image acquisition module uses six sets of high-speed industrial cameras with a resolution of 20 million pixels and a frame rate of 30 frames per second. These cameras are deployed at three height levels inside the drying chamber: top, middle, and bottom. Each set of cameras faces different material placement areas within the drying chamber. This multi-angle layout ensures that the acquired images completely cover all areas where garlic slices are placed within the drying chamber, eliminating blind spots. The cameras continuously capture images of the garlic slices' surface condition at preset time intervals, capturing subtle changes in color, shape, and texture during the drying process. After capture, the raw image data is transmitted in real-time via Ethernet to the data preprocessing module, ensuring the real-time performance and integrity of the image data.

[0045] The process parameter acquisition module integrates four types of sensors: temperature, humidity, wind speed, and power. Six temperature sensors are installed at different locations within the drying chamber, including the air inlet, outlet, and above the material layer, enabling multi-point temperature monitoring. Two humidity sensors are installed at the air inlet and outlet to accurately monitor humidity changes in the drying medium. The wind speed sensor is installed at the center above the material layer inside the drying chamber to collect real-time hot air velocity data. The power sensor is directly electrically connected to the heating element of the drying chamber to collect real-time power data. All sensors synchronously acquire process parameter data once per second. After acquisition, the raw process parameter data and image data are synchronously transmitted to the data preprocessing module, strictly ensuring that the acquisition time of the image data and process parameter data is perfectly aligned to avoid timing discrepancies affecting subsequent data matching and analysis.

[0046] After the data reception is complete, the system enters the image and process data preprocessing stage, such as... Figure 2 As shown, this step aims to eliminate noise and outliers in the raw data and improve data quality. The data preprocessing module first receives the raw image data transmitted from the image acquisition module and performs three processing operations on the image data in sequence: Gaussian filtering for noise reduction, pixel value normalization, and histogram equalization. Gaussian filtering for noise reduction can effectively remove image noise caused by environmental factors such as dust and moisture in the drying oven, while preserving the true features of the garlic slices' surface; pixel value normalization transforms all pixel values ​​of the image to a uniform range of zero to one, eliminating numerical differences caused by different shooting light and camera parameters; histogram equalization can enhance the contrast of the image, highlight key features such as the color changes and surface texture of the garlic slices, making subsequent feature extraction more accurate.

[0047] After image data processing is completed, the data preprocessing module processes the raw process parameter data transmitted from the process parameter acquisition module. It uses the 3σ criterion to remove outliers from the process parameters. Specifically, it calculates the mean and standard deviation of a single type of process parameter within a continuous acquisition period. Data exceeding ±3 times the standard deviation of the mean are identified as outliers and removed to avoid errors caused by sensor malfunctions, electromagnetic interference, etc., from interfering with the analysis results. After outlier removal, each set of process parameter data is matched one-to-one with the image data at the corresponding time point based on the acquisition timestamps of the image data and process parameter data, forming a standardized dataset with complete temporal alignment. All preprocessed data is then transmitted to the feature extraction module to provide high-quality data support for subsequent feature extraction.

[0048] After data transmission is complete, the system proceeds to the stage of extracting image features and process timing features, such as... Figure 3 As shown, this step is crucial for transforming raw data into analyzable features. The feature extraction module first receives standardized image data and process parameter data transmitted from the data preprocessing module. Then, through the image feature extraction unit, it calls a lightweight convolutional neural network to extract three core image features of garlic slices from the preprocessed image data: color change rate, shrinkage rate, and surface texture roughness. The color change rate is used to monitor whether the garlic slices are at risk of browning; the shrinkage rate reflects the morphological changes during the drying process; and the surface texture roughness identifies defects such as microcracks and edge curling.

[0049] Simultaneously, the process feature extraction unit invokes a temporal convolutional network to extract the temporal variation features of four types of process parameters—temperature, humidity, wind speed, and heating power—from the preprocessed process parameter data. This accurately captures the dynamic changes of process parameters during the drying process and uncovers the potential correlation between parameter changes and the quality risk of garlic slices. After feature extraction, the image features and process features are converted into feature vectors of the same dimension, eliminating the dimensional differences between different types of features and providing a unified basis for subsequent feature weight allocation and fusion.

[0050] In the specific implementation of this embodiment, after unifying the feature vector dimensions, adaptive feature weight allocation calculation is performed. An adaptive feature weight fusion formula is used to initially allocate weights to the image feature vector and the process feature vector. The formula is as follows: ,in Assign weights to the i-th type of feature. The variance of the offline experiment for the i-th feature is calibrated, where n is the total number of feature categories. The specific determination method is to select fifty sets of offline feature data under the standard process of hot drying of garlic slices, calculate the numerical variance of each type of feature through statistical methods, and take the average of the variance results of multiple experiments to finally obtain the offline experimental calibration variance of each type of feature. This calibration variance can truly reflect the degree of dispersion of various features under the standard drying process. The weights assigned based on this can maximize the analytical value of key features. After the weight calculation is completed, the feature vector before fusion is stably transmitted to the bidirectional coupling risk prediction module.

[0051] After the feature vector is received, the system enters the core analysis stage of bidirectional coupled risk prediction and process parameter deviation derivation, such as... Figure 2 As shown, this step is crucial for the system to achieve accurate early warning. The bidirectional coupled risk prediction module inputs image feature vectors and process feature vectors into a pre-trained graph neural network model. The model is trained using a dedicated dataset for the garlic slice heat drying process, labeled with risk levels and process parameter deviations. This dataset contains 80 sets of image data and corresponding process parameter data under different drying conditions, covering various common risk scenarios such as excessively high temperature, low humidity, insufficient wind speed, and abnormal power. The risk levels are divided into low, medium, and high risk. During model training, a backpropagation algorithm is used, with prediction accuracy as the optimization goal. The internal parameters of the model are iteratively adjusted until the model's risk prediction accuracy reaches a preset training threshold of over 96%, ensuring the model has predictive capabilities for industrial applications.

[0052] In the specific implementation of this embodiment, the graph neural network model performs deep fusion of image feature vectors and process feature vectors through graph convolution operations. The fusion process strictly follows the bidirectional coupled feature calculation formula, which is as follows: ,in For the first Layer fusion feature matrix, For graph structure adjacency matrix, Let L be the input feature matrix of the Lth layer. Let B be the feature fusion weight matrix, B be the bias vector, and σ be the activation function. The specific method for determining B is as follows: during the model training phase, with the goal of minimizing the prediction error, the values ​​of the weight matrix and bias vector are iteratively updated using the gradient descent method until the model loss function converges to a stable minimum value. At this point, the weight and bias values ​​are the optimal parameters.

[0053] After the feature fusion is completed, the model outputs the risk level of the current drying process and the risk development trend in the next 30 seconds based on the fused feature vector. At the same time, it analyzes the contribution of four types of process parameters—temperature, humidity, wind speed, and heating power—to the occurrence of risk through the backpropagation algorithm, accurately locating the specific process parameter deviation types that cause the risk, such as excessive temperature, insufficient humidity, too low wind speed, and power fluctuation. The risk prediction results and process parameter deviation results are then sent to the process adjustment scheme generation module simultaneously.

[0054] After receiving the risk and deviation data, the system enters the risk assessment and process adjustment plan generation stage, such as... Figure 2 As shown, the system first determines whether there is a quality risk in the current drying process. If no risk is detected, the system returns to the data acquisition stage and continues to cycle through the monitoring process. If a risk is detected, the system immediately initiates the process parameter adjustment scheme generation process. The process adjustment scheme generation module has a built-in process adjustment rule library that has been verified in industrial practice. The rule library stores in detail the standardized adjustment methods corresponding to different types of process parameter deviations, covering the adjustment direction and range of various deviations such as temperature, humidity, wind speed, and power.

[0055] Based on the risk level and process parameter deviation type output by the bidirectional coupled risk prediction module, the module quickly retrieves matching adjustment strategies from the process adjustment rule base. Then, it calls the process parameter adjustment amount calculation formula to generate a process parameter adjustment scheme containing specific values. The formula is: ,in This represents the adjustment amount for the k-th process parameter. Let k be the deviation value of the process parameter. This is the proportionality coefficient. is the integral coefficient. and The specific determination method involves extracting historical data from 1,500 successfully dried garlic slices from the past three years. The correlation between the deviation values ​​of process parameters and the corresponding adjustment amounts is used as the fitting basis. The least squares method is used for linear fitting calculation to obtain the proportional coefficients and integral coefficients corresponding to four types of parameters: temperature, humidity, wind speed, and heating power. These coefficients are in line with the actual production characteristics of the production line, ensuring the accuracy and practicality of the adjustment amount calculation. Once the scheme is generated, it is immediately transmitted to the early warning and control module.

[0056] After the adjustment plan is received, the system enters the stage of triggering an early warning and executing process adjustments, such as... Figure 2 As shown, the early warning and control module incorporates a tiered early warning mechanism and automated control strategy. Low-risk levels trigger flashing lights on the equipment panel; medium-risk levels activate simultaneous audible and visual warnings; and high-risk levels trigger both audible and visual warnings and an emergency load reduction command for the heating device, providing comprehensive alerts to operators regarding potential risks. Simultaneously, the early warning and control module transmits the generated process adjustment plan to the drying equipment's dedicated control system via an industrial bus. The control system adjusts the drying chamber's heating temperature, hot air humidity, airflow velocity, and heating power in real time according to the specific parameter values ​​in the plan, achieving proactive intervention and dynamic control of the drying process and eliminating quality risks at their source. During early warning and control, the module synchronizes all early warning information, control commands, and adjustment results to the human-machine interface module in real time, ensuring full-process data traceability.

[0057] The human-machine interface module participates in the entire system operation process. Its display unit provides real-time high-definition images of the surface condition of garlic slices inside the drying chamber, real-time process parameters such as temperature, humidity, wind speed, and power, current risk level, early warning information, and process adjustment plans, allowing operators to intuitively grasp the status of the entire drying process. The input unit supports operators in real-time input of system parameter setting commands, process adjustment rule modification commands, and manual control commands. In special production scenarios, operators can switch between automatic and manual control modes through the input unit to fine-tune system parameters or modify adjustment rules, achieving refined human-machine collaborative management and further improving the system's adaptability and flexibility.

[0058] In summary, in this embodiment, the system continuously cycles according to the above process, forming a complete closed loop from data acquisition, preprocessing, feature extraction, risk prediction to scheme generation and early warning control. The entire process requires no frequent manual intervention, automatically completing the intelligent monitoring and risk control of the entire garlic slice heat-drying process. This embodiment, through the complete deployment and operation of a deep learning-based garlic slice heat-drying image risk early warning system, achieves synchronous real-time acquisition and deep fusion of image data and process parameter data during the garlic slice heat-drying process. Relying on the deep learning model, it achieves early and accurate prediction of drying quality risks and reverse location of process parameter deviations, effectively avoiding quality defects such as browning, microcracks, edge curling, and localized scorching. The automatically generated quantitative process adjustment scheme can directly link the drying equipment for control, eliminating the need for manual parameter adjustments and significantly improving the timeliness and accuracy of process control.

[0059] 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 deep learning-based risk warning system for garlic slices during heat drying images, characterized in that, The system includes: an image acquisition module, a process parameter acquisition module, a data preprocessing module, a feature extraction module, a two-way coupled risk prediction module, a process adjustment scheme generation module, an early warning and control module, and a human-computer interaction module; The image acquisition module is used to acquire surface state images of garlic slices during the heat drying process and send the acquired image data to the data preprocessing module; The process parameter acquisition module is used to acquire process parameters inside the drying oven and send the acquired process parameter data to the data preprocessing module; The data preprocessing module is used to preprocess the image data and process parameter data and send the preprocessed data to the feature extraction module; The feature extraction module is used to extract image features and process features and send the extracted features to the bidirectional coupled risk prediction module; The bidirectional coupling risk prediction module is used to fuse image features and process features to perform risk prediction and process parameter deviation derivation, and send the prediction results to the process adjustment scheme generation module. The process adjustment scheme generation module is used to generate process parameter adjustment schemes and send the adjustment schemes to the early warning and control module. The early warning and control module is used to issue early warning signals, perform process adjustments, and interact with the control system of the drying equipment. The human-computer interaction module is used to display system information, receive operation instructions, and interact with various modules.

2. The deep learning-based risk warning system for garlic slices during heat drying images according to claim 1, characterized in that, The data preprocessing module performs the following operations during data preprocessing: Receive raw image data sent by the image acquisition module and raw process parameter data sent by the process parameter acquisition module; The original image data is sequentially processed by Gaussian filtering for noise reduction, pixel value normalization, and histogram equalization. Outliers were removed from the original process parameter data using the 3σ criterion. Based on the acquisition timestamps of the image data and process parameter data, each set of process parameter data was matched one-to-one with the image data at the corresponding time point.

3. The deep learning-based risk warning system for garlic slices during heat drying images according to claim 1, characterized in that, The feature extraction module performs the following operations during feature extraction: The system receives preprocessed image data and process parameter data sent by the data preprocessing module, and uses a convolutional neural network to extract the color change rate, shrinkage rate, and surface texture roughness features of garlic slices from the preprocessed image data through the image feature extraction unit. The process feature extraction unit uses a temporal convolutional network to extract the temporal variation features of process parameters from the preprocessed process parameter data, and converts the extracted image features and process features into feature vectors of the same dimension. The image feature vector and the process feature vector are initially weighted according to the adaptive feature weight fusion formula, which is as follows: ,in, Assign weights to the i-th type of feature. The variance of the offline experiment for the i-th feature is calibrated, where n is the total number of feature categories. The determination was made through statistical analysis of multiple sets of offline characteristic data under the standard process of hot drying of garlic slices.

4. The deep learning-based risk warning system for garlic slices during heat drying images according to claim 1, characterized in that, The bidirectional coupled risk prediction module performs the following operations when performing risk prediction and process parameter deviation derivation: The system receives image feature vectors and process feature vectors sent by the feature extraction module, and inputs the image feature vectors and process feature vectors into a pre-trained graph neural network model. Image feature vectors and process feature vectors are deeply fused using graph convolution operations. The fusion process follows a bidirectional coupled feature calculation formula: ,in, For the first Layer fusion feature matrix, For graph structure adjacency matrix, Let L be the input feature matrix of the Lth layer. Let B be the feature fusion weight matrix, B be the bias vector, and σ be the activation function. B is determined iteratively through the backpropagation algorithm during model training; Based on the fused feature vector, the risk level of the current drying process and the future risk development trend are output. The contribution of each process parameter to the occurrence of risk is analyzed through backpropagation, and the type of process parameter deviation that leads to the occurrence of risk is output.

5. The deep learning-based risk warning system for garlic slices during heat drying images according to claim 4, characterized in that, The graph neural network model is trained using a dataset of garlic slices undergoing heat drying, labeled with risk levels and process parameter deviations. The dataset contains multiple sets of image data under different drying conditions and corresponding process parameter data. During training, the graph neural network model continuously adjusts its parameters using a backpropagation algorithm until the model's prediction accuracy reaches a preset training threshold.

6. The deep learning-based risk warning system for garlic slices during heat drying images according to claim 1, characterized in that, The image acquisition module includes multiple sets of high-speed industrial cameras, which are respectively deployed at the top, middle and bottom of the drying chamber. The multiple sets of high-speed industrial cameras face different areas inside the drying chamber. The multiple sets of high-speed industrial cameras continuously acquire surface state images of garlic slices at preset time intervals. The images acquired by the multiple sets of high-speed industrial cameras cover the entire material placement area inside the drying chamber.

7. The deep learning-based risk warning system for garlic slices during heat drying images according to claim 1, characterized in that, The process parameter acquisition module includes a temperature sensor, a humidity sensor, a wind speed sensor, and a power sensor. The temperature sensor is installed inside the drying chamber at the air inlet, air outlet, and above the material layer. The humidity sensor is installed inside the drying chamber at the air inlet and air outlet. The wind speed sensor is installed inside the drying chamber above the material layer. The power sensor is electrically connected to the heating device of the drying chamber.

8. The deep learning-based risk warning system for garlic slices during heat drying images according to claim 1, characterized in that, The process adjustment scheme generation module has a built-in process adjustment rule library, which stores process parameter adjustment methods corresponding to different types of process parameter deviations. Based on the risk level and process parameter deviation type output by the bidirectional coupled risk prediction module, the process adjustment scheme generation module retrieves the corresponding process parameter adjustment method from the process adjustment rule library and generates a process adjustment scheme containing specific adjustment parameters according to the process parameter adjustment amount calculation formula. The process parameter adjustment amount calculation formula is as follows: ,in, This represents the adjustment amount for the k-th process parameter. Let k be the deviation value of the process parameter. This is the proportionality coefficient. The integral coefficient is... and The determination was made by fitting and calibrating historical process data of garlic slices during heat drying.

9. The deep learning-based risk warning system for garlic slices during heat drying images according to claim 1, characterized in that, The early warning and control module has built-in early warning methods and control strategies corresponding to different risk levels. The early warning and control module triggers the corresponding early warning method according to the risk level output by the bidirectional coupled risk prediction module. The early warning and control module sends the process adjustment scheme output by the process adjustment scheme generation module to the control system of the drying equipment. The control system of the drying equipment adjusts the corresponding process parameters according to the process adjustment scheme.

10. The deep learning-based risk warning system for garlic slices during heat drying as described in claim 1, characterized in that, The human-machine interaction module includes a display unit and an input unit. The display unit displays real-time process parameters of the drying process, real-time surface state images of garlic slices, current risk level, early warning information, and process adjustment schemes. The input unit receives system parameter setting instructions, process adjustment rule modification instructions, and manual control instructions input by the operator.

Images