Method, device and storage medium for controlling the browning state of food

Abstract

This application discloses a method, device, and storage medium for controlling the browning state of food, relating to the field of data processing technology. The method includes: in response to a browning control command, acquiring multimodal data of processed food; extracting features from the multimodal data according to a modal differentiation processing strategy to obtain multimodal feature data; calculating the correlation weight between the multimodal feature data and the degree of browning through a cross-modal attention mechanism, and weighted fusing the high-order feature representations of the multimodal data to obtain a browning state parameter set; and calculating parameter adjustment amounts based on the browning state parameter set and the target browning curve to obtain a process parameter adjustment strategy to control the browning state of the processed food. This application solves the problem of unstable quality during food baking by combining modal differentiation processing and cross-modal attention fusion with an optimal decision algorithm to generate a control strategy and achieve closed-loop control, thereby improving the accuracy of browning control and the consistency of food processing quality.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method, apparatus and storage medium for controlling the browning state of food. Background Technology

[0002] In intelligent food baking and processing scenarios, precise control of the baking process and ensuring consistent quality of finished products directly affect the edible quality of food and production efficiency.

[0003] In related technologies, the baking process is controlled by fixed-time-temperature open-loop control, single-point photoelectric sensor detection, or single-visual recognition to determine browning. This type of method relies on preset fixed process parameters and single-dimensional state information for open-loop control, making it susceptible to environmental interference and difficult to adapt to complex baking scenarios with large differences in raw materials and fluctuating temperature and humidity, thus leading to poor food baking quality.

[0004] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main objective of this application is to provide a method, apparatus, and storage medium for controlling the browning state of food, aiming to solve the technical problem of poor food baking quality.

[0006] To achieve the above objectives, this application proposes a method for controlling the browning state of food, the method comprising: In response to browning control instructions, multimodal data of processed foods are obtained by associating the collected modal data according to a unified time reference. The multimodal data of the processed food are processed by feature extraction according to the modal differentiation processing strategy to obtain the multimodal feature data of the processed food. The correlation weights between the multimodal feature data and the degree of browning are calculated by a cross-modal attention mechanism, and the high-order feature representations of the multimodal data are weighted and fused to obtain the browning state parameter set of the processed food. Based on the comparison results between the browning state parameter set and the target browning curve, the parameter adjustment amount is calculated through the optimal decision algorithm to obtain the process parameter adjustment strategy, so as to control the browning state of the processed food through the process parameter adjustment strategy.

[0007] In one embodiment, in response to the browning control command, the surface condition of the processed food, cavity environment parameters, and processing sequence data are collected according to a unified time reference, and the raw data are processed to obtain the raw data. The timestamps of the original data are calibrated according to the time base, and the matching relationship of cross-modal time series is cross-validated to remove abnormal time series data and obtain time series modal data. Based on the time-series modal data, feature association mapping relationships between each modal data are constructed, and the data dimensions of each modal data are unified to obtain the multimodal data of the processed food.

[0008] In one embodiment, based on the modality differentiation processing strategy and the multimodal data type, the modality type processing strategy of the multimodal data is classified and matched. According to the modality type processing strategy, modality-specific preprocessing is performed on the multimodal data, and the data deviation of the multimodal data is corrected to obtain each initial feature data; Each of the initial feature data is subjected to modality-specific feature enhancement operations to extract core feature information strongly correlated with the browning process, thereby obtaining each enhanced feature data; The features of each of the enhanced feature data are aligned, and the dimensions of each of the enhanced feature data are unified to obtain the multimodal feature data of the processed food.

[0009] In one embodiment, based on the modality type processing strategy, the image data in the multimodal data is preprocessed, and image feature data associated with browning is extracted from the preprocessed image data; Based on the modality processing strategy, the temperature data processing strategy performs spatial interpolation and anomaly removal on the temperature data in the multimodal data to obtain temperature feature data of continuous temperature distribution on the food surface. Based on the modality-type processing strategy, the spectral data processing strategy performs baseline correction and principal component dimensionality reduction on the spectral data in the multimodal data to extract the spectral feature data of browning-related chemical markers. The humidity and time data in the multimodal data are synchronously smoothed and filtered to extract the characteristics of environmental humidity changes and processing time sequence. Based on the modal differentiation correction rule, the data deviations of the environmental humidity change, the processing time sequence features and the image feature data, the temperature feature data and the spectral feature data are corrected to obtain the initial feature data of each modality.

[0010] In one embodiment, the multimodal feature data is input into the browning state assessment model and mapped to the high-level feature representation space to obtain multimodal deep feature vectors; Guided by the browning feature discrimination criterion, the correlation weights between the multimodal deep feature vectors and the degree of browning are calculated dimension by dimension through the cross-modal attention mechanism to determine the correlation weight matrix; The multimodal deep feature vectors are finely weighted and superimposed dimension by dimension according to the correlation weight matrix, and the global modal weight coefficients are adjusted according to the current processing stage to obtain the global browning comprehensive features. The global browning comprehensive characteristics are analyzed in multiple dimensions and numerically converted to generate the browning state parameter set.

[0011] In one embodiment, the browning state parameter set is compared with the target browning curve of the corresponding processing stage in a dimension-by-dimensional manner to calculate the deviation value of the browning process and obtain multi-dimensional browning deviation data. Based on the multi-dimensional browning deviation data, combined with equipment operating limits, food safety thresholds, and process stability constraints, the optimal decision-making algorithm is used to solve for the globally optimal process adjustment that satisfies multi-objective optimization. The global optimal process adjustment amount is mapped to each controllable process parameter dimension to generate an initial process parameter adjustment strategy. The execution effect and stability of the initial process parameter adjustment strategy are pre-simulated and verified. The adjustment parameters that cause process fluctuations are corrected to obtain the process parameter adjustment strategy.

[0012] In one embodiment, the multidimensional composition of the globally optimal process adjustment is analyzed, and the individual browning correction parameters are decomposed. Based on the aforementioned sub-item browning correction parameters, and combining the influence weights and interaction relationships of individual process parameters on different browning indices, a coupling correlation matrix between the sub-item adjustment targets and each of the controllable process parameters is constructed. Based on the coupling correlation matrix and the equipment operating parameter boundaries, the optimal adjustment range and direction of each controllable process parameter are calculated, and the independent adjustment command of each controllable process parameter is determined. By integrating the independent instructions corresponding to each controllable process parameter and setting the execution sequence and execution priority, a multi-dimensional controllable initial process parameter adjustment strategy is obtained.

[0013] In one embodiment, after the process parameter adjustment strategy is executed, feedback multimodal data of the processed food is collected according to a unified time base to generate a control effect feedback dataset associated with the timing of the current control command. The real-time browning parameters in the control effect feedback dataset are analyzed, and the corresponding browning features are extracted and correlated to obtain the browning feedback state parameters; If the browning control effect does not reach the preset browning threshold, the browning state association weight, process parameter coupling relationship and optimal decision algorithm decision boundary are updated based on the actual effect data of this control, and the updated data is associated with the system model to obtain the system model parameters. Based on the system model parameters, a new process parameter adjustment strategy is generated through optimization. The process then returns to the multimodal data acquisition step to continue executing closed-loop control until the browning state reaches the target requirements.

[0014] In addition, to achieve the above objectives, this application also proposes a control device for the browning state of food, the control device for the browning state of food comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the control method for the browning state of food as described above.

[0015] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the food browning state control method as described above.

[0016] This application provides a method for controlling the browning state of food, comprising: generating multimodal data of processed food in response to browning control commands and collecting them with a unified time reference; extracting features from the multimodal data using a modal differential processing strategy to obtain multimodal feature data; calculating the correlation weight between the multimodal feature data and the degree of browning through a cross-modal attention mechanism and weighted fusing high-order feature representations to obtain a browning state parameter set; calculating the parameter adjustment amount based on the comparison results between the browning state parameter set and the target browning curve using an optimal decision algorithm to obtain a process parameter adjustment strategy to control the browning state of processed food; and combining the control effect feedback multimodal data to achieve model self-learning and closed-loop control. This method solves the technical problems of data time synchronization, inaccurate browning state assessment, delayed control response, and inability to achieve dynamic and precise adjustment caused by traditional food browning control relying on human experience or single sensor monitoring. It improves the accuracy and real-time performance of food browning state assessment, realizes dynamic closed-loop precise control of browning state throughout the food processing process, improves the consistency and yield of food processing quality, and reduces the cost of manual intervention.

[0017] In summary, this application solves the technical problem of poor food baking quality by combining modal differentiation processing and cross-modal attention fusion with the optimal decision algorithm to generate control strategies and achieve closed-loop control, thereby improving the accuracy of browning control and the consistency of food processing quality. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the first embodiment of the method for controlling browning state of food according to this application. Figure 2 This is a flowchart illustrating the process control for the browning state in this application. Figure 3 This is a schematic flowchart of the seventh embodiment of the method for controlling browning state of food in this application; Figure 4 This is a flowchart illustrating the eighth embodiment of the method for controlling browning state of food according to this application. Figure 5 This is a flowchart illustrating the output of the browning state of food products in this application. Figure 6 This is a schematic diagram of the structure of the control device for the browning state of food in this application.

[0021] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0022] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0023] In related technologies, the baking process is controlled by fixed-time-temperature open-loop control, single-point photoelectric sensor detection, or single-visual recognition to determine browning. This type of method relies on preset fixed process parameters and single-dimensional state information for open-loop control, making it susceptible to environmental interference and difficult to adapt to complex baking scenarios with large differences in raw materials and fluctuating temperature and humidity, thus leading to poor food baking quality.

[0024] This application provides a solution: First, in response to a browning control command, multimodal data of processed food are obtained by associating collected modal data according to a unified time reference. Then, feature extraction processing is performed on the multimodal data of processed food according to a modal-differentiated processing strategy to obtain multimodal feature data of processed food. Next, the correlation weight between the multimodal feature data and the degree of browning is calculated through a cross-modal attention mechanism, and the high-order feature representation of the multimodal data is weighted and fused to obtain a browning state parameter set of processed food. Finally, based on the comparison result between the browning state parameter set and the target browning curve, the parameter adjustment amount is calculated through an optimal decision algorithm to obtain a process parameter adjustment strategy, so as to control the browning state of processed food through the process parameter adjustment strategy.

[0025] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device capable of performing the above functions, a food browning state control device, etc. The following description uses a food browning state control device as an example to illustrate this embodiment and the subsequent embodiments.

[0026] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0027] This application provides a method for controlling the browning state of food, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the method for controlling browning state of food according to this application.

[0028] In this embodiment, the method for controlling the browning state of food includes steps S10 to S40: Step S10: In response to the browning control command, the collected modal data are correlated according to a unified time reference to obtain multimodal data of the processed food.

[0029] Browning control commands are the core control commands that trigger browning monitoring and data acquisition tasks in food. Modal data consists of multi-dimensional monitoring data characterizing the real-time state of processed foods, such as surface visual images, infrared temperature distribution, spectral characteristics, cavity humidity, and cumulative heating time acquired through visual sensors. A unified time reference is a time series calibration standard used to calibrate the timing of various monitoring data and ensure data synchronization. The multi-modal data of processed foods is a unified raw dataset that integrates all monitoring dimensions after time-series correlation and integration.

[0030] In this embodiment, the browning control command can be triggered in six ways. First, it is triggered at processing stage nodes. When the system reaches a preset processing stage node, it automatically generates and issues a browning control command to conduct phased status monitoring. Second, it is triggered by periodic polling. The system periodically generates browning control commands at preset time intervals to achieve normalized data collection throughout the entire processing flow. Third, it is triggered by parameter anomalies. When basic process parameters such as chamber humidity and heating power fluctuate beyond thresholds, the system automatically generates and issues a browning control command to track and collect abnormal status data. Fourth, it is triggered by manual intervention. Operators issue monitoring commands through the equipment's interactive panel, and the system receives the request, generates, and executes the browning control command. Fifth, it is triggered by batch start-up. When a batch of food enters the processing chamber and begins heating, the system synchronously generates a browning control command to initiate initial status data collection. Sixth, it is triggered by early warning. When the system anticipates a localized browning risk in the food, it generates a browning control command in advance and increases the frequency of data collection to control status changes.

[0031] After receiving the browning control command, the food browning control equipment begins to synchronously collect modal data output by different monitoring modules and completes time-series correlation integration based on a unified time reference.

[0032] For example, there are two methods for acquiring and associating modal data with time series. The first is global time series binding synchronous acquisition. After the acquisition task starts, the image acquisition module, temperature detection module, spectrum acquisition module, humidity detection module, and timing module are started synchronously with a unified timestamp as the reference. All modules output corresponding monitoring data at the same time node, directly completing the time series binding and summarizing to obtain the original multimodal data. Then, the integrity of the entire set of data is checked, and invalid data with acquisition interruption or missing signal are removed. This method has high time series alignment accuracy and strong module coordination, and is suitable for scenarios with stable processing conditions and strict time series synchronization requirements. The second method is time-division acquisition combined with time series offset completion acquisition. Each monitoring module is started sequentially according to the preset time series order to complete data acquisition, and the original timestamp corresponding to each set of modal data is recorded. Then, the time series offset is calculated based on the inherent delay parameters of each module, and the time series acquisition data is time series compensation and position correction is performed to calibrate all data to the same time reference and summarize to form complete multimodal data. This method can avoid signal interference problems caused by the synchronous operation of multiple modules, reduce hardware load, and is suitable for processing scenarios with densely arranged multiple sensors and complex electromagnetic environments.

[0033] After completing the acquisition and timing calibration of each modality's data, the data can be integrated in the following ways to form multimodal data of processed food.

[0034] In one alternative approach, a fixed time-scale association scheme is adopted. Equal-interval time-scales are pre-set, and the calibrated modal data are matched and archived one by one according to the time-scale. Each time-scale corresponds to a complete set of full-modal data, which is directly concatenated to form structured multimodal data. This scheme has fixed rules, simple computational logic, and well-organized data arrangement, facilitating subsequent batch parsing and processing.

[0035] In another alternative approach, a dynamic temporal interval correlation scheme is employed. This scheme dynamically divides temporal intervals based on the food heating rhythm, aggregating all modal data collected within the same processing period into the corresponding temporal interval. This preserves the original data acquisition density, performing only cross-modal temporal alignment without compressing data points. This approach can fully retain details of short-term state changes, resulting in higher data information retention, and is suitable for fine-grained monitoring scenarios studying the rapid browning stage.

[0036] In an exemplary scheme for determining multimodal data of processed food, the system first loads preset global time reference parameters to clarify the timing calibration standards and allowable delay range. Then, all monitoring modules are simultaneously activated to perform data acquisition, recording the acquisition timestamp and original monitoring values ​​of each modal data stream in real time. Next, an initial timing verification is performed, comparing the timestamps of each modal data stream with the reference time, and marking data exceeding the allowable delay range as data to be corrected. For the data to be corrected, timing offset compensation is performed by considering sensor transmission delay and signal processing time, unifying all data to the standard time reference. Afterwards, data integrity screening is performed, eliminating abnormal single-modal data with signal interruptions or value jumps, and supplementing missing points with adjacent timing data. Finally, all compliant data is integrated according to timing sequence, data storage units are divided, and multimodal data of processed food with a unified format and synchronized timing is generated.

[0037] It should be noted that in some special cases, if the hardware delay of all monitoring modules is consistent and the acquisition timing is completely synchronized, the timing offset correction step can be skipped, and the acquired data can be directly summarized as multimodal data of processed food.

[0038] Step S20: Perform feature extraction processing on the multimodal data of the processed food according to the modal differentiation processing strategy to obtain the multimodal feature data of the processed food.

[0039] The modal differentiation strategy is a set of dedicated preprocessing rules and feature extraction logic configured for different modalities of data, including image, temperature, spectrum, cavity humidity, and cumulative heating time. Feature extraction is a data parsing operation used to filter out interference noise and invalid information from the raw data, and to extract effective information related to the browning state of food. Multimodal feature data is a set of specialized features that characterize the browning properties of food, extracted from each modal of data after differentiation processing.

[0040] In this embodiment, the feature extraction processing task can be initiated in six ways. First, data reception triggering: after the system has fully received the multimodal data output from the previous step, the feature extraction process is initiated, achieving seamless process integration. Second, data volume threshold triggering: when the storage volume of multimodal data to be processed in the buffer reaches a preset threshold, the feature extraction task is initiated in batches to balance the system's computational load. Third, process idle triggering: when food processing enters non-core operation periods such as heating breaks or cavity ventilation, idle computing power is used to centrally execute feature extraction. Fourth, timed batch triggering: the system performs batch feature extraction on multiple sets of cached multimodal data at fixed time intervals. Fifth, status expedited triggering: when the system identifies an abnormal browning trend in food, computing power is prioritized to expedite feature extraction, shortening analysis time. Sixth, batch completion triggering: after the entire batch of food processing is completed, feature extraction is uniformly performed on all multimodal data throughout the process for later process review and analysis.

[0041] After the system acquires multimodal data of processed food, it calls the modality-specific differentiation processing strategy to perform feature extraction processing on different modal data.

[0042] For example, there are two ways to implement feature extraction based on a modal differentiation processing strategy. The first is modal independent pipeline extraction, which builds an independent processing pipeline for each type of modal data. Image data undergoes denoising, color space conversion, and region segmentation operations in sequence; temperature data undergoes spatial interpolation and anomaly removal operations in sequence; spectral data undergoes baseline correction and dimensionality reduction operations in sequence; and humidity and time-series data undergo smoothing filtering operations in sequence. Each pipeline runs independently and does not interfere with the others, outputting the corresponding modal features. This method has independent process logic, a small impact range from failures, and is easy to debug and maintain individually, making it suitable for scenarios with complex modal data types that require individual control. The second is modal collaborative correlation extraction, which uses browning correlation as a link to process modal data with coupled influences in a linked manner. While processing image data, it combines temperature distribution features to correct the region segmentation results, and while parsing spectral data, it refers to heating time series to optimize the dimensionality reduction logic. It relies on the correlation between modalities to improve feature effectiveness and outputs all modal features simultaneously. This method can uncover potential correlations between modalities, improve the matching degree between features and browning states, and is suitable for refined analysis scenarios with high requirements for feature correlation.

[0043] Before performing feature extraction, the applicable modality differentiation strategy can be determined based on the following methods.

[0044] In one alternative approach, a direct matching scheme based on preset rules is employed. The system incorporates multiple differentiated processing strategies for various processing conditions. Based on parameters such as the current food category, heating process, and cavity environment, the preset strategies are directly retrieved and applied to the feature extraction. This approach offers fast retrieval speed, high execution efficiency, mature and stable rules, and is suitable for standard food processing scenarios.

[0045] Another alternative approach employs an adaptive processing strategy. This involves first analyzing the basic attributes of the current multimodal data, such as noise level, data density, and signal quality. Then, combining real-time processing parameters, it calls upon historical optimization models from the rule base to dynamically generate a customized processing strategy adapted to the current data state. The generated strategy is then validated for feasibility before being implemented. This approach allows for flexible adjustments to the processing logic for special scenarios such as data anomalies and fluctuations in processing conditions, offering a wider range of adaptability and higher feature extraction accuracy.

[0046] In an exemplary scheme for determining multimodal feature data, multimodal data of processed food is first read, and the signal quality, data density, and noise type of each modality are analyzed. Then, a modal-specific differentiated processing strategy is matched or dynamically generated, assigning corresponding preprocessing and extraction rules to each type of data. Next, specific processing is performed on each modality of data according to the strategy requirements: for image data, denoising, color space conversion, and region segmentation are performed to extract surface browning-related features. For temperature data, spatial interpolation and anomaly removal are performed to extract continuous temperature distribution features. For spectral data, baseline correction and dimensionality reduction are performed to extract chemical biomarker-related features. For cavity humidity and cumulative heating time data, smoothing filtering is performed to extract environmental and temporal variation features. Afterward, the validity of all extracted single-modal features is verified, distorted and invalid feature values ​​are removed, and missing feature points are supplemented. Finally, all compliant single-modal features are integrated, stored in a standardized manner according to modality classification, and a complete multimodal feature data of processed food is generated.

[0047] It should be noted that in some special cases, if a certain modality data experiences a continuous acquisition failure, the feature extraction of that modality will be skipped, and the corresponding modality missing identifier will be marked in the multimodal feature data for subsequent identification and processing.

[0048] Step S30: Calculate the correlation weight between the multimodal feature data and the degree of browning through a cross-modal attention mechanism, and then weight and fuse the high-order feature representations of the multimodal data to obtain the browning state parameter set of the processed food.

[0049] Cross-modal attention mechanisms are feature analysis rules used to quantify the proportion of different modal features' influence on the browning degree of food and to clarify feature relationships. Association weights are quantitative values ​​representing the contribution of a single modal feature to the overall browning state. High-level feature representations are deep feature information obtained after in-depth mining and dimensionality enhancement of the basic features of each modality. The browning state parameter set is a comprehensive set of parameters that, after integrating all deep modal features, can fully quantify the current browning level, trend, and uniformity of food.

[0050] In this embodiment, the aforementioned correlation weight calculation and feature fusion tasks can be triggered in six ways. First, feature data readiness trigger: after receiving complete multimodal feature data, the system immediately initiates the attention calculation and feature fusion process. Second, segmented fusion trigger: according to the food processing segment nodes, the feature data for the current time period is fused in stages to achieve staged state determination. Third, risk warning trigger: when preliminary screening detects abnormal browning in the feature data, the high-precision feature fusion process is immediately initiated. Fourth, timed aggregation trigger: at fixed time intervals, multiple sets of feature data accumulated within a period are uniformly fused. Fifth, manual judgment trigger: after the operator initiates a state depth analysis command, the system initiates the cross-modal feature fusion process. Sixth, closed-loop feedback trigger: after receiving backend control feedback signals, the system initiates a new round of feature fusion to support dynamic control.

[0051] After the system acquires multimodal feature data of processed food, it uses a cross-modal attention mechanism to calculate association weights and completes weighted fusion of high-order feature representations.

[0052] For example, there are two implementation methods for association weight calculation and feature fusion. The first is local attention modal coupling calculation, which takes a single modal feature as the core, calculates the local association relationship between this modality and all other modal features in turn, solves the association weight between each group of modalities step by step, and then performs weighted fusion of the corresponding high-order feature representations based on each group of weights, gradually integrating them into the overall feature. This method promotes the calculation process step by step, which can verify the rationality of the weights segment by segment, facilitates the location of single-modal association anomalies, and has strong controllability in the operation process. The second is global attention global linkage calculation, which incorporates all modal features into a unified analysis space, calculates the comprehensive impact of all modal features on the degree of browning, generates a complete set of association weights at once, and then performs overall weighted fusion of all high-order feature representations simultaneously based on the global weights. This method can coordinate the complex coupling relationship between modalities, the fusion result is more holistic, the operation time is shorter, and it is suitable for scenarios of rapid processing of large batches of feature data.

[0053] Before calculating the association weights, the cross-modal attention calculation rules to be used can be determined based on the following methods.

[0054] In one alternative approach, a static weighted rule invocation scheme is adopted. The system pre-stores standardized attention calculation rules adapted to different food categories and processing technologies. Based on the current processing scenario, a fixed rule is directly invoked, and the weight calculation logic and influencing factors within the rule remain unchanged. This scheme features unified rules, simple calculation logic, stable operation, and adaptability to standardized production lines with long-term unchanged processes.

[0055] In another alternative approach, a dynamic iterative weight optimization scheme is employed. This scheme retrieves historical browning sample data and corresponding feature weights, combines them with the distribution characteristics of the current multimodal feature data, and iteratively optimizes the influencing factors and constraints in the attention calculation rules. It then dynamically generates specific calculation rules and performs validity verification. This approach can continuously optimize weight accuracy based on data changes and adapt to complex production scenarios such as raw material fluctuations and changes in operating conditions.

[0056] In an exemplary scheme for determining a set of browning state parameters, firstly, cross-modal attention calculation rules adapted to the current scenario are loaded, clarifying the weight calculation constraints and feature fusion standards. Then, multimodal feature data is mapped to a deep feature space, generating high-order feature representations for each modality. Next, according to the selected calculation method, the correlation weight between each modal feature and the degree of browning in the food is calculated, while verifying whether the weight values ​​are within a reasonable range and removing abnormal weights that deviate from the normal range. Then, based on the compliant correlation weights, a weighted superposition is performed on the high-order feature representations of all modalities to complete cross-modal feature fusion, forming a comprehensive feature vector. Next, the fused comprehensive features undergo dimensionality analysis and numerical transformation, decomposing them into sub-parameters such as browning level, browning rate, browning uniformity, and local browning risk. Finally, an overall compliance verification is performed on all sub-parameters, standardizing the parameter format and value range, and summarizing to generate a set of browning state parameters for processed foods.

[0057] It should be noted that in some special cases, if the integrated features after fusion can only output a single overall browning parameter, then this parameter can be directly used as a simplified browning state parameter set to simplify the subsequent comparison process.

[0058] Step S40: Based on the comparison results between the browning state parameter set and the target browning curve, the parameter adjustment amount is calculated through the optimal decision algorithm to obtain the process parameter adjustment strategy, so as to control the browning state of the processed food through the process parameter adjustment strategy.

[0059] The target browning curve is a pre-set standard browning state change curve corresponding to each stage of food processing, serving as a reference benchmark for evaluating the actual browning state. The comparison result is multi-dimensional deviation information obtained by comparing the current browning state parameter set with the target browning curve. The optimal decision algorithm is an intelligent computational model that combines equipment operating limits and process constraints to solve for the optimal process variation range. Parameter adjustment amounts are the numerical changes in various processing parameters required to compensate for browning deviations and correct the browning state of food. The process parameter adjustment strategy is a complete and implementable control scheme that includes the adjustment object, adjustment range, execution sequence, adjustment priority, and execution duration.

[0060] In this embodiment, the aforementioned state comparison and strategy generation tasks can be triggered in six ways. First, parameter set generation triggering: after the system completes the construction of the browning state parameter set, the curve comparison and strategy calculation process is immediately initiated. Second, deviation exceeding limits triggering: if the previous comparison result shows that the browning deviation is within the allowable range, the calculation is delayed. When the deviation exceeds a preset threshold, the entire process calculation is initiated. Third, processing stage switching triggering: after the food enters a new processing stage, the target browning curve for the corresponding stage is switched, and the comparison and strategy generation are re-executed. Fourth, batch control triggering: for multiple processing devices in a batch, multiple sets of state comparisons and strategy calculations are uniformly initiated. Fifth, emergency control triggering: when a serious browning deviation problem is detected, the calculation process is expedited, and an adjustment strategy is quickly output. Sixth, timed review triggering: at fixed time intervals, the state comparison is repeatedly executed, and the process parameter adjustment strategy is dynamically updated.

[0061] After the system obtains the set of browning state parameters of the processed food, it compares them with the target browning curve of the corresponding processing stage, and then uses the optimal decision algorithm to solve the parameter adjustment amount and generate a process parameter adjustment strategy.

[0062] For example, there are two ways to compare the browning state parameter set with the target browning curve. The first is a piecewise curve point-by-point comparison. The target browning curve is divided into multiple continuous intervals according to the processing sequence. The parameter values ​​of each time-series node in the browning state parameter set are extracted and compared one by one with the standard points of the corresponding interval curve. The individual deviations at each point are calculated, and the results are summarized to form a full-time-series multi-dimensional comparison. This method provides detailed point comparison and can accurately locate browning deviations in local time periods and local points, with strong deviation tracing capabilities. The second method is an overall curve trend fitting comparison. The overall change trend, fluctuation amplitude, and average value of the browning state parameter set are extracted and fitted with the overall trend characteristics of the target browning curve. The overall trend deviation and mean deviation are calculated to form an overall comparison result. This method focuses on macroscopic state assessment, has less computation, and is suitable for scenarios requiring overall processing quality control without the need for single-point tracing.

[0063] Before calculating the parameter adjustment amount, the optimal decision algorithm to be used can be determined based on the following methods.

[0064] In one alternative approach, a fixed algorithm invocation scheme is adopted. Based on the food category and processing equipment type, the system directly retrieves a preset optimal decision-making algorithm. The algorithm's internal operation rules and constraint thresholds remain fixed over a long period and are directly used for parameter adjustment calculations. This scheme offers convenient invocation, fast computation speed, stable and consistent rules, and is suitable for parameter control under normal processing conditions.

[0065] In another alternative approach, a dynamic selection algorithm based on operating conditions is employed. This involves first analyzing the type and magnitude of the deviation obtained from the current comparison, while simultaneously reading real-time equipment load, process safety limits, and other operating conditions. Then, the optimal decision algorithm is matched from the algorithm library to fit the current deviation characteristics and equipment status. If necessary, the algorithm's constraint parameters are fine-tuned before computation. This approach allows for flexible algorithm selection for different deviation scenarios and equipment states, resulting in higher rationality and adaptability of parameter adjustments.

[0066] In an exemplary scheme for determining a process parameter adjustment strategy, the target browning curve of the corresponding version is first retrieved as a comparison benchmark based on the current food processing stage. Then, using a selected comparison method, the set of browning state parameters is compared item by item with the target browning curve, multidimensional browning deviation is calculated, and a complete comparison result is generated. Next, considering equipment operating boundaries, food safety requirements, and process stability constraints, the corresponding optimal decision algorithm is selected and loaded. The multidimensional browning deviation is input into the optimal decision algorithm to solve for the adjustment amounts of various process parameters such as heating power, cavity humidity, and heating time. Then, the influence relationships of the adjustment amounts of each parameter are analyzed, adjustment priorities are assigned, execution sequence and adjustment step size are set, and an initial process parameter adjustment strategy is generated. The initial strategy is then validated for compliance, checking whether the adjustment amounts exceed equipment limits and whether there are process conflicts in the adjustment logic; any unreasonable aspects are corrected and optimized. Finally, all compliant adjustment instructions, execution rules, and timing arrangements are integrated to form the final process parameter adjustment strategy, which is then sent to the equipment for execution, completing the automatic control of the browning state of the processed food.

[0067] It should be noted that in some special cases, if the comparison results show that the current browning state perfectly matches the target browning curve and the deviation is within the allowable error range, the parameter adjustment calculation and strategy generation steps are skipped, and the existing process parameters are maintained to continue processing.

[0068] Further, please refer to Figure 2 , Figure 2This is a flowchart illustrating the browning state process control of this application. A closed-loop control process for the browning state of food begins with the user setting a target browning curve as the control benchmark. Then, a comparator compares the actual browning state from the fusion model with the target browning curve, generating an error signal (e(t)). This error signal is input to a dynamic controller, which can employ algorithms such as fuzzy proportional-integral-derivative control (Fuzzy PID), model predictive control (MPC), or reinforcement learning (RL). Based on the error signal, the dynamic controller calculates and generates process adjustment commands, which are then output to the execution unit. The execution unit adjusts controllable process parameters such as heating power, humidity, airflow, or processing time, acting on the food processing process that follows the dynamics of browning to generate a new actual browning state. This newly generated actual browning state is then fed back to the comparator via a feedback loop for comparison with the target browning curve again, thus forming a continuous closed-loop control that drives the food's browning state to gradually approach and stabilize at the target browning curve.

[0069] Second Embodiment This embodiment provides an exemplary scheme for constructing multimodal data of processed food. In this example, raw data is first obtained by synchronously collecting food surface state, cavity environment parameters, and processing time series data in response to browning control commands. Then, time stamp calibration, cross-modal time series verification, and abnormal data removal are performed based on a unified time reference to obtain time series modal data. Finally, feature association mapping relationships between various modalities are established and data dimensions are unified, thereby finally obtaining time-synchronized and dimensionally regular multimodal data of processed food. Step S10 includes steps A11 to A13: Step A11: In response to the browning control command, collect the food surface condition, cavity environment parameters and processing time sequence data of the processed food according to the unified time reference, and organize them to obtain the raw data.

[0070] Step A12: Calibrate the timestamps of the original data according to the time base, cross-validate the matching relationship of cross-modal time series, remove abnormal time series data, and obtain time series modal data.

[0071] Step A13: Based on the time-series modal data, construct the feature association mapping relationship between each modal data, and unify the data dimension of each modal data to obtain the multimodal data of the processed food.

[0072] Food surface condition data, acquired through visual sensors and optical monitoring modules, directly reflects the appearance, color changes, and surface features of processed food, serving as core data for assessing the degree of browning. Examples include surface image information, surface temperature distribution, surface color parameters, local morphological features, and surface texture changes. Cavity environment parameters are environmental data obtained by monitoring the internal operating environment of the food processing cavity, reflecting external working conditions during processing. Examples include internal humidity, airflow velocity, real-time temperature, air pressure, and ventilation frequency. Processing time-series data records the overall processing progress and duration of the food, marking the processing stage corresponding to each state data point. Examples include cumulative heating time, stage processing time, process switching time, single operation interval, and batch processing time sequence. Raw data refers to the complete set of monitoring data after acquisition, without time-series correction, anomaly screening, or format standardization. This data set includes three basic types of data: food surface condition, cavity environment parameters, and processing time-series data, serving as the original data carrier for subsequent data processing. A timestamp is a temporal identifier marking the specific collection time of a single data point. It is used to determine the collection order and time interval of different modalities and is the core identifier for achieving temporal alignment. Temporal modal data is a standardized monitoring data set where the temporal sequence of each modality remains consistent after timestamp calibration, temporal verification, and outlier removal, ensuring collaborative analysis of multiple data types. Feature association mapping relationships are corresponding association rules established by mining the inherent influence patterns and state linkages between different modalities, used to strengthen the expression of cross-modal data correlations. Data dimensions refer to the data arrangement, feature quantity, and data format specifications of a single modality. Unified data dimensions are a prerequisite for achieving multimodal data fusion and analysis. Multimodal data is an integrated comprehensive data set with standardized format, synchronized temporal sequence, and complete associations after temporal calibration, outlier removal, association construction, and dimension unification, providing a compliant data source for subsequent feature extraction and state analysis.

[0073] In this example, when collecting food surface condition, cavity environment parameters, and processing time sequence data in response to browning control commands to obtain raw data, it can be done either through global synchronous collection or through time-sharing polling collection. Time-sharing polling collection involves activating different monitoring modules in a preset order, with each module sequentially collecting food surface condition, cavity environment parameters, and processing time sequence data. Data is temporarily aggregated during collection, and the raw data is obtained by integrating all data collected from all modules after completion, thus completing the collection of comprehensive raw monitoring data.

[0074] After the raw data collection is completed, a time-series calibration and anomaly screening process is initiated. For each data point in the raw data, the timestamp is rigorously aligned and calibrated according to a unified time benchmark. Simultaneously, cross-module verification is performed to check the time-series matching relationship between food surface conditions, cavity environmental parameters, and processing time-series data. Abnormal time-series data with time-series offsets, signal interruptions, and abrupt distortions are identified and removed, resulting in time-consistent and valid time-series modal data. Subsequently, based on the time-series modal data, the linkage patterns of different monitoring dimensions are deeply mined, feature association mapping relationships between various modal data are constructed, and dimensional transformation and normalization operations are performed on various modal data to unify the data dimensions of all modal data. Finally, processed food multimodal data with a unified format, synchronized time series, and complete associations is generated. This progressive processing logic of layered collection, time-series error correction, association construction, and dimensional normalization comprehensively ensures the validity and standardization of multimodal data, avoiding analytical biases in subsequent feature extraction and browning state assessment caused by time-series misalignment, inconsistent dimensions, and missing modal associations.

[0075] For example, there are two processing methods for sequentially obtaining time-series modal data and multimodal data of processed food from raw data. The first method is sequential temporal calibration and dimensional normalization of each modality. Following a fixed order of food surface state, cavity environment parameters, and processing time-series data, single-class data is extracted and processed one by one. First, for the first class of data, all timestamps of this class are calibrated according to a unified time benchmark. Then, abnormal time-series data within this class are screened and removed. After processing a single modality, the next class of data is retrieved, and the above temporal calibration and abnormal removal operations are repeated until all modal data have been processed, initially forming intermediate data with unified temporal sequence. Subsequently, cross-modal temporal secondary cross-validation is performed group by group to confirm that all modal data are completely matched in time sequence, and then the time-series modal data is summarized. Then, feature association mapping relationships are built between each class of data and the other modal data according to the modal order, and dimensional transformation is performed simultaneously on the single class of data. After all modalities have completed association construction and dimensional unification, the processed food multimodal data is summarized and integrated. This method employs a single-modal serial processing and step-by-step verification execution logic. The processing flow is clearly broken down, which can accurately locate timing errors and dimensional anomalies in a certain modality. It is highly convenient for fault diagnosis and parameter debugging and is suitable for complex processing conditions where monitoring modules are scattered and data failures occur frequently.

[0076] The second approach involves parallel verification of time-series partitions and unification of global dimensions. First, the raw data is divided into multiple independent time-series partitions according to the processing progress. Each partition contains complete data on food surface conditions, cavity environment parameters, and processing time series data, with no computational dependencies between partitions. Parallel processing is initiated simultaneously for all time-series partitions. Within each partition, timestamp calibration, cross-modal time-series matching verification, and outlier removal are performed in parallel. After parallel processing of all partitions, the data from all partitions are aggregated to obtain global time-series modal data. Then, a global correlation analysis is conducted on the global time-series modal data to establish an overall feature association mapping relationship between all modal data. Finally, based on a globally unified standard, dimensional transformation and format normalization are performed simultaneously on all modal data to unify the overall data dimensions, ultimately generating multimodal data for processed food. This method employs partitioned parallel computation and globally unified and regularized execution logic, making full use of parallel computing power to compress data processing time. The global correlation analysis can also uncover deep linkage patterns across time periods and modalities, resulting in higher data processing efficiency and modal correlation integrity. It is suitable for large-scale food production scenarios with large data collection volumes and continuous and uninterrupted processing flows.

[0077] Third Embodiment This embodiment provides an exemplary scheme for constructing multimodal feature data of processed foods. In this example, a modality-specific differentiation processing strategy and a modality type processing strategy corresponding to the classification and matching of multimodal data types are first combined. Then, based on the matched strategy, modality-specific preprocessing is performed and data deviations are corrected to obtain initial feature data. Subsequently, modality-specific feature enhancement operations are performed to extract browning-related core features to obtain enhanced feature data. Finally, feature alignment and dimensionality unification are completed, thereby finally obtaining well-organized and complete multimodal feature data of processed foods. Step S20 includes steps B11 to B14: Step B11: Based on the modal differentiation processing strategy and the multimodal data type, classify and match the modal type processing strategy of the multimodal data.

[0078] Step B12: According to the modality type processing strategy, perform modality-specific preprocessing on the multimodal data and correct the data deviation of the multimodal data to obtain each initial feature data.

[0079] Step B13: Perform modality-specific feature enhancement operations on each of the initial feature data to extract core feature information strongly correlated with the browning process, and obtain each enhanced feature data.

[0080] Step B14: Align the features of each of the enhanced feature data and unify the dimensions of each of the enhanced feature data to obtain the multimodal feature data of the processed food.

[0081] Multimodal data types are data classification identifiers based on data acquisition sources, representation content, and data format. They are used to determine data ownership and match corresponding processing rules, serving as the fundamental classification basis for modal processing. Examples include visual image data, temperature distribution data, spectral detection data, cavity environment data, and processing time-series data. Modal type processing strategies are refined execution rules adapted to single data types, derived from modal differentiated processing strategies. They include single-type data preprocessing methods, deviation correction standards, and feature extraction rules. Examples include image data denoising and segmentation details, temperature data anomaly removal standards, spectral data baseline correction rules, and time-series data filtering parameters. Modal-specific preprocessing involves basic purification operations such as noise reduction, completion, and format normalization tailored to the signal characteristics of a single modality of data. These operations remove invalid interference information from the original data. Examples include image denoising and region segmentation, temperature spatial interpolation, spectral baseline correction, and humidity and time-series data smoothing filtering. Data bias refers to abnormal errors in multimodal data caused by the acquisition environment, hardware interference, and transmission delays, such as numerical shifts, signal distortion, and temporal misalignments. These are the objects that need to be corrected in the preprocessing stage. Examples include image brightness bias, temperature value jump bias, spectral baseline shift bias, and temporal data delay bias. Initial feature data is the basic feature set initially extracted after multimodal data has undergone dedicated preprocessing and bias correction, preserving the basic characterization information of the data. Modality-specific feature enhancement operations are in-depth processing operations that target and amplify information related to food browning and weaken irrelevant interference information for different modalities of initial feature data, thereby improving feature specificity. Examples include enhancing image browning region features, strengthening temperature gradient features, amplifying spectral chemical features, and highlighting temporal change features. Core feature information is key information selected and extracted from the initial feature data that is highly correlated with the degree, trend, and uniformity of food browning, and is the core basis for characterizing the food processing state. Enhanced feature data is a set of deep features with higher browning correlation and stronger recognition after feature enhancement of the initial feature data. Feature alignment is an operation that calibrates the position, order, and correlation between different modal enhancement features based on a unified temporal sequence and feature correspondence, eliminating feature misalignment issues. Data dimension refers to the arrangement structure, number, and representation of feature data; unified dimension is a necessary prerequisite for achieving cross-modal feature fusion and joint analysis. Multimodal feature data is a standardized feature set formed by integrating all enhancement features after alignment and dimension unification, providing a reliable data source for subsequent browning status assessment and weight calculation.

[0082] In this example, when using a modal differentiation processing strategy and a multimodal data type classification matching modality type processing strategy, a global rule traversal matching approach can be used. Alternatively, a modality tag-based fast retrieval method can be employed, binding a unique type tag to each type of multimodal data, retrieving corresponding sub-rules based on the tags, identifying data types while matching modality type processing strategies, and accumulating strategy matching across the entire dataset to achieve rapid allocation of refined processing rules.

[0083] After matching the modal type processing strategies, the multimodal data processing and feature extraction process is initiated. Based on the matched modal type processing strategies, modal-specific preprocessing is performed on each type of multimodal data, simultaneously correcting various data biases to obtain initial feature data for each data type. Then, modal-specific feature enhancement operations are performed on each group of initial feature data to selectively mine and extract core feature information strongly correlated with the browning process, generating corresponding enhanced feature data. Finally, feature alignment is performed on all enhanced feature data to calibrate the correspondence between cross-modal features and unify the data dimensions of all enhanced feature data. After integration and normalization, multimodal feature data of processed foods is obtained. This progressive processing logic—strategic hierarchical matching, data purification, feature enhancement, and dimension normalization—layer by layer filters and amplifies effective features, improving the accuracy and relevance of the overall feature data and avoiding judgment biases in subsequent browning state analysis caused by improper rule matching, residual data bias, or feature misalignment.

[0084] For example, there are two processing methods for progressively generating multimodal feature data from multimodal data. The first is a sequential step-by-step processing method, which performs the entire process on a single modality of data in a fixed modal order of image, temperature, spectrum, environment, and time series. First, the modality type processing strategy corresponding to the current modality is retrieved to complete the modality-specific preprocessing and data bias correction for that set of data, resulting in a single set of initial feature data. Then, modality-specific feature enhancement operations are performed on this set of initial feature data to extract core feature information and generate enhanced feature data. After all single-modality processing is completed, the current enhanced feature data is temporarily stored, and the next modality of data is retrieved and all the above operations are repeated until all modality data is processed. After all enhanced feature data is generated, feature alignment and dimensionality unification are performed on each group in modal order. After all verifications are successful, the multimodal feature data of processed food is obtained by summarizing the data. This method employs a single-modal serial execution and step-by-step temporary storage processing logic. The processing flow for each type of data is independent of each other, allowing for separate verification of the preprocessing effect and feature enhancement quality of a single modality. This facilitates the location of local data anomalies and rule adaptation issues, and is suitable for food processing monitoring scenarios with multiple modal types and high probability of single-type data failures.

[0085] The second approach is parallel integrated processing of modal groups. Based on data association characteristics, multimodal data is divided into several computationally independent data groups, each containing multiple modal data with interconnected relationships. Parallel processing is initiated synchronously for all data groups. Within each group, the modality-specific processing strategies corresponding to each modality are synchronously invoked, completing modality-specific preprocessing and data bias correction for all data within the group in parallel, resulting in all initial feature data for that group. Immediately afterwards, modality-specific feature enhancement is performed on the initial feature data within the group, generating enhanced feature data in batches. After feature enhancement is completed for each group, feature alignment and initial dimensionality unification are immediately performed within the group. After all parallel processing for all groups is completed, global secondary alignment and dimensionality normalization are performed on the output feature data of each group, ultimately integrating them to generate multimodal feature data for processed foods. This method employs a computational logic of grouped parallel computing and integrated processing within groups, making full use of parallel computing power to reduce the overall processing time. At the same time, it relies on grouped linkage processing to strengthen the feature matching degree between related modalities, thereby improving data processing efficiency and cross-modal feature correlation. It is suitable for large-scale food industrial production scenarios with large data collection scale and continuous processing flow.

[0086] Fourth embodiment This embodiment provides an exemplary scheme for multimodal data classification preprocessing and bias correction. In this example, image, temperature, spectrum, humidity, and time data are first processed separately according to corresponding processing strategies, and corresponding features are extracted. Then, the data bias of each type of feature is uniformly corrected according to modal differentiation correction rules, thereby finally obtaining the initial feature data corresponding to each modality. Step B12 includes steps C11 to C15: Step C11: Based on the image data processing strategy of the modality type processing strategy, perform image preprocessing on the image data in the multimodal data, and extract image feature data related to browning from the preprocessed image data.

[0087] Step C12: Based on the temperature data processing strategy of the modal type processing strategy, spatial interpolation and anomaly removal are performed on the temperature data in the multimodal data to obtain temperature feature data of continuous temperature distribution on the food surface.

[0088] Step C13: Based on the modality type processing strategy, the spectral data processing strategy performs baseline correction and principal component dimensionality reduction on the spectral data in the multimodal data to extract the spectral feature data of browning-related chemical markers.

[0089] Step C14: Perform synchronous smoothing filtering on the humidity and time data in the multimodal data to extract the characteristics of environmental humidity changes and processing time sequence.

[0090] Step C15: Based on the modal differentiation correction rule, correct the data deviations of the environmental humidity change, the processing time sequence features and the image feature data, the temperature feature data and the spectral feature data to obtain the initial feature data of each modality.

[0091] Image data processing strategies are standardized processing specifications specifically designed for image-based monitoring data. They include preprocessing execution standards, feature extraction ranges, and feature selection conditions, serving as the basis for image data processing. Image preprocessing involves basic cleansing operations such as denoising, color space conversion, and region segmentation for food surface images, used to remove image interference and define effective detection areas. Image feature data is a set of specialized features extracted from the preprocessed image data, reflecting the proportion and intensity of browning on the food surface. Temperature data processing strategies are dedicated operational rules configured for temperature monitoring data, specifying execution requirements such as spatial interpolation algorithms, outlier thresholds, and data completion methods. Spatial interpolation is a data completion operation that calculates continuous temperature values ​​across the entire area based on discrete temperature sampling points. Outlier removal is a filtering operation that identifies and removes abrupt or distorted temperature values ​​caused by hardware failures or external interference. Temperature feature data, after interpolation and filtering, is feature data that comprehensively characterizes the temperature distribution pattern on the food surface. Spectral data processing strategies are processing specifications for spectral detection data, defining baseline correction methods, principal component extraction dimensions, and dimensionality reduction constraints. Baseline correction eliminates baseline drift and background noise interference in spectral data. Principal component dimensionality reduction is a data compression method that removes redundant dimensions from spectral data while retaining core, effective information. Spectral feature data is a set of chemical markers extracted from spectral data that characterize changes in internal chemical substances in food and are associated with browning reactions. Synchronous smoothing filtering is a joint noise reduction process for humidity and time data with large fluctuations, used to weaken instantaneous data disturbances and restore true trends. Environmental humidity changes reflect the dynamic fluctuations in humidity within the processing chamber, while processing time-series features record the overall processing rhythm and stage duration changes of food. Modal differentiation correction rules are deviation correction standards and operational logics formulated based on the data characteristics and error sources of different modal features, distinguishing correction thresholds, algorithms, and compensation methods for each modality. Data deviation refers to errors such as numerical offsets, temporal misalignments, and amplitude distortions that occur during the acquisition, transmission, and processing of various feature data. Initial feature data is single-modal standard feature data that has been processed specifically and corrected for uniform deviations, and is in a standardized format, with accurate values, and can be used for subsequent in-depth processing.

[0092] In this example, when performing image data preprocessing and extracting image feature data based on image data processing strategies, a full-domain frame-by-frame preprocessing approach can be adopted. The complete preprocessing process is performed frame by frame on the acquired food surface image, and then browning-related features are extracted frame by frame, summarizing to form continuous image feature data. Alternatively, a local region of interest (ROI) approach can be used, locking onto the area containing the food within the image and performing only local preprocessing and feature extraction, discarding invalid background data, thereby efficiently completing the acquisition of image feature data.

[0093] After extracting various single-modal features, the overall deviation correction process is initiated. First, temperature, spectral, humidity, and time data are processed sequentially according to corresponding strategies, yielding temperature feature data, spectral feature data, environmental humidity changes, and processing time sequence features, respectively. Then, preset modal differentiation correction rules are invoked to compensate for and correct deviations for each type of error in different modal features, eliminating numerical and temporal deviations between various features. Finally, the initial feature data for each modality is output. This end-to-end design, encompassing categorized processing, multi-dimensional feature extraction, and differential deviation correction, balances the inherent characteristics of different modal data with cross-modal data consistency, improving the overall accuracy of all initial feature data and preventing single-type data anomalies and excessive inter-modal deviations from affecting subsequent browning state analysis results.

[0094] For example, there are two ways to obtain the initial feature data of each modality from multimodal raw data processing. The first is single-modal sequential processing and correction, which executes the complete processing flow for each modality in a fixed order of image, temperature, spectrum, humidity, and time. First, the processing strategy for the corresponding modality is retrieved, and the specific calculations and feature extraction for that type of data are completed, and the generated feature data is temporarily stored. After processing one type of modality, the data processing and feature extraction for the next type of modality are started, until all modal features are extracted. After all feature data is summarized, the image feature data, temperature feature data, spectral feature data, environmental humidity changes, and processing time sequence features are sequentially checked and corrected according to the modal differentiation correction rules, and the errors in the data are corrected one by one. Finally, the initial feature data of each modality is output sequentially. This method employs a single-modal serial execution, step-by-step temporary storage, and class-by-class correction operation logic. Each processing step is independently controllable, which can accurately locate data anomalies and correction failures in a single mode. It facilitates on-site debugging and troubleshooting and is suitable for food processing monitoring scenarios where sensors are distributed and single-channel data is easily interfered with.

[0095] The second approach involves multimodal grouping and parallel processing with centralized correction. Based on data correlation, image and temperature data are divided into an appearance temperature group, and spectral, humidity, and time data are divided into a compositional environment group. These two groups operate independently but are processed in parallel. Within each group, preprocessing, interpolation, dimensionality reduction, and filtering are performed in parallel according to corresponding processing strategies, simultaneously extracting feature information corresponding to all modalities within the group. After feature extraction is complete for both groups, all modal feature data are aggregated, and modal differentiation correction rules are invoked to perform full-domain parallel deviation detection on all feature data. Based on the error type of each modality, the corresponding correction algorithm is matched to simultaneously correct the deviation, generating initial feature data for all modalities in batches. This method employs grouped parallel computation and full-domain centralized correction, fully utilizing the parallel computing power of the equipment to compress the overall processing time. Simultaneously, it centrally manages cross-modal deviations, effectively ensuring the matching degree between multiple groups of feature data. This results in stronger data processing efficiency and modal synergy, making it suitable for continuous food industrial processing scenarios with high sampling frequencies and large data volumes.

[0096] Fifth Embodiment This embodiment provides an exemplary scheme for food browning feature fusion and parameter generation based on cross-modal attention. In this example, multimodal feature data is first mapped to a high-level feature representation space to obtain multimodal deep feature vectors. Then, based on the browning feature discrimination criterion and the cross-modal attention mechanism, the modal correlation weight matrix is ​​calculated. Combined with the dynamic weight coefficients of the processing stage, a fine-grained weighted fusion is performed to obtain the global browning comprehensive features. Finally, through multi-dimensional state analysis and numerical conversion, a standardized browning state parameter set is accurately generated. Step S30 includes steps D11~D14: Step D11: Input the multimodal feature data into the browning state assessment model and map it to the high-level feature representation space to obtain multimodal deep feature vectors.

[0097] Step D12: Based on the browning feature discrimination criterion, the correlation weights between the multimodal deep feature vectors and the degree of browning are calculated dimension by dimension through the cross-modal attention mechanism to determine the correlation weight matrix.

[0098] Step D13: Finely weight and superimpose the multimodal deep feature vectors dimension by dimension according to the correlation weight matrix, and adjust the global modal weight coefficients according to the current processing stage to obtain the global browning comprehensive features.

[0099] Step D14: Perform multi-dimensional state analysis and numerical conversion on the global browning comprehensive characteristics to generate the browning state parameter set.

[0100] The browning state assessment model is a pre-trained intelligent feature analysis model adapted to the browning evolution laws of food processing. It is used to map and transform shallow features into deep, high-dimensional features, serving as the core carrier for quantitative assessment of browning state. The high-dimensional feature representation space is a high-dimensional feature computation space that is detached from the original data dimension and can carry multimodal deep correlation information, maximizing the discovery of implicit correlations between features and browning state. The multimodal deep feature vector is a standardized high-dimensional vector obtained after the multimodal feature data is mapped and upgraded by the model, containing deep browning correlation information from images, temperature, spectra, and environmental time series. The browning feature discrimination criterion is based on the dynamic characteristics of food browning and pre-set feature judgment standards from the sample training data. It is used to constrain the calculation direction and value range of attention weights, ensuring that the weight results conform to the actual browning laws. The cross-modal attention mechanism is an adaptive computational rule used to quantify the contribution of different modalities and feature dimensions to the degree of food browning, enabling differentiated weighting and fusion of multimodal features. The correlation weight matrix is ​​a two-dimensional numerical matrix encompassing the correlation weights of all modalities, all feature dimensions, and browning degree, serving as the core basis for refined feature weighted fusion. The global modal weight coefficient is a dynamically adjusted parameter adapted to different food processing stages, used to globally correct the overall contribution ratio of each modality, and adapted to the dynamic changes in browning characteristics throughout the processing. The global browning comprehensive feature is an integrated comprehensive feature obtained after dimension-wise weighted fusion and global weight correction of all deep feature vectors, fully representing the current overall browning state of the food. The browning state parameter set is a multi-dimensional quantitative parameter set obtained after analyzing and converting the comprehensive features, including core quantitative indicators such as browning grade, browning rate, browning uniformity, and local browning risk.

[0101] In this example, when mapping multimodal feature data to a high-dimensional feature representation space to obtain multimodal deep feature vectors, a global batch mapping approach can be used. This involves inputting the entire batch of multimodal feature data into the browning state assessment model at once, uniformly completing the full-dimensional feature upscaling and spatial mapping, and outputting well-structured multimodal deep feature vectors in batches. Alternatively, a modal incremental mapping approach can be used. Feature data is input sequentially according to modality classification, completing the single-modal feature upscaling mapping for each category, and outputting single-modal deep vectors in real time. After all modal mappings are completed, the complete multimodal deep feature vectors are obtained by summing them up, thus accurately completing the refined acquisition of high-dimensional features.

[0102] After generating multimodal deep feature vectors, the cross-modal weight calculation and feature fusion process is initiated. Using a pre-defined browning feature discrimination criterion as a constraint, and relying on a cross-modal attention mechanism, the correlation weights between each deep feature vector and the degree of browning in the food are calculated dimension by dimension. All dimensional weight values ​​are then integrated to construct a complete correlation weight matrix. Subsequently, based on the correlation weight matrix, all deep feature vectors are subjected to precise, dimension-by-dimensional weighted superposition. Simultaneously, considering the current processing stage of the food (heating, isothermal holding, cooling and shaping), the global modal weight coefficients corresponding to each modality are dynamically adjusted to eliminate the shortcomings of fixed weights in adapting to changes in processing conditions, resulting in a more representative global browning comprehensive feature. Finally, multi-dimensional state decomposition, feature analysis, and standardized numerical conversion are performed on the global browning comprehensive feature to generate a quantifiable and comparable set of browning state parameters. Through a hierarchical logic of dimensional mining, adaptive weighting, dynamic parameter tuning, and precise conversion, the comprehensiveness and accuracy of browning state quantification are improved, avoiding browning assessment biases caused by fixed-weight fusion and shallow feature analysis.

[0103] For example, there are two methods for generating global browning comprehensive features from multimodal deep feature vectors. The first is a sequential weighted fusion method, starting from the first feature dimension of the associated weight matrix, sequentially retrieving the single-dimensional weight values, and performing weighted operations and superposition fusion on the corresponding multimodal deep feature vectors dimension by dimension. After completing the fusion of a single dimension, the weighted superposition operation is repeated on the next dimension, and the fusion operation of all feature dimensions is completed sequentially. After the fusion of all dimensions is completed, the current processing stage identifier is read, the corresponding standard global modality weight coefficient is matched, and the preliminary fused features are globally corrected and optimized, finally outputting the global browning comprehensive features. This method adopts a computational logic of sequential operation and step-by-step fusion correction, which can verify the weight matching accuracy and feature fusion effect dimension by dimension, accurately locate abnormal weight dimensions, effectively avoid the overall fusion deviation caused by the failure of local feature weights, and adapt to the fine processing scenarios with complex processing conditions and frequent local browning mutations.

[0104] The second method is a global parallel weighted fusion, which performs global analysis on the correlation weight matrix, extracting the correlation weight values ​​of all modalities and all feature dimensions at once. Simultaneously, it performs global parallel dimension-by-dimensional weighted calculations on all multimodal deep feature vectors, batch-completing the weighted superposition of all feature dimensions to quickly generate initial comprehensive features. Simultaneously, it identifies the current processing stage, real-time browning rate, and environmental parameters, dynamically iteratively optimizing the global modal weight coefficients to adapt to the weight requirements of each modal feature in real time, and globally adaptively correcting the initial comprehensive features, ultimately generating high-precision global browning comprehensive features. This method employs a global parallel weighted and dynamically adaptive parameter tuning calculation logic, greatly reducing the computation time for multi-dimensional feature fusion. It can also dynamically adapt the modal weight ratio according to the processing progress, aligning with the dynamic evolution of food browning during processing, and is suitable for continuous, high-volume, and time-sensitive industrial food processing scenarios.

[0105] Furthermore, a training method for a browning state assessment model used for multimodal feature dimensionality enhancement and deep information mining involves first collecting batches of five categories of raw multimodal data—images, temperature, spectra, cavity humidity, and processing time sequences—under various food categories, processing stages, and multiple on-site conditions. Simultaneously, combining physicochemical test results with manual comprehensive judgment, each data set is labeled with multidimensional standard labels such as browning level, browning rate, browning uniformity, and local browning risk. At the same time, following the data processing logic of this scheme, the time sequence alignment, cross-modal calibration, and modal-specific preprocessing of the full sample are completed. Training, validation, and test sets are divided according to scenarios such as normal production conditions, sensor data deviations, environmental humidity disturbances, and abnormal temperature fluctuations, constructing a mixed sample dataset covering all working conditions and all processing stages. Subsequently, an overall model structure integrating a high-level feature representation module and a cross-modal attention association module was constructed. Dynamic modal weight branches adapted to different processing stages were added to match the application characteristics of the contribution of each modal feature to the dynamic changes in the processing process in actual production. During the training phase, a hierarchical iterative mode of single-modal pre-training plus multi-modal joint training was adopted. First, the parameters of each modal feature mapping branch were optimized independently to ensure the accuracy of deep feature extraction of single-class data. Then, the cross-modal attention module was connected to carry out joint training. Actual production constraints such as equipment operating limits, food safety thresholds, and process stability were incorporated into the loss function. At the same time, small numerical deviations and temporal offsets were randomly superimposed on the training samples to simulate real data errors in the on-site collection and transmission process to enhance the model's anti-interference ability. After the initial training of the model, the matching accuracy between the deep feature vector output and the standard browning parameters is verified using the test set. For samples with large judgment deviations, reverse parameter fine-tuning is performed. Finally, relying on the closed-loop control process of this scheme, the newly added multimodal data collected in real time from the production line and the browning state labels after manual verification are continuously fed back to the sample library to realize online incremental iterative updates of the model. This allows the model to continuously adapt to on-site scenarios such as fluctuations in raw material characteristics, changes in equipment operating conditions, and minor adjustments to the process. Ultimately, a browning state assessment model that can accurately complete multimodal feature dimensionality enhancement, in-depth mining of cross-modal correlations, and adaptation to the entire processing flow and complex on-site conditions is trained, providing stable and reliable model support for the subsequent core links of this scheme, such as correlation weight calculation, feature fusion, and browning parameter generation.

[0106] Sixth Embodiment This embodiment provides an exemplary scheme for generating a food processing process parameter adjustment strategy. In this example, the browning state parameter set is first compared dimension-by-dimensionally with the target browning curve of the corresponding processing stage, and the deviation value is calculated to obtain multi-dimensional browning deviation data. Then, combined with equipment, safety, and process-related constraints, the globally optimal process adjustment amount is solved using an optimal decision-making algorithm. Subsequently, the adjustment amount is mapped to each controllable process parameter dimension to generate an initial process parameter adjustment strategy. Finally, abnormal parameters are corrected through pre-simulation of the execution effect and stability verification, thereby ultimately obtaining an executable process parameter adjustment strategy. Step S40 includes steps E11~E14: Step E11: Compare the set of browning state parameters with the target browning curve of the corresponding processing stage dimension by dimension, calculate the deviation value of the browning process, and obtain multi-dimensional browning deviation data.

[0107] Step E12: Based on the multi-dimensional browning deviation data, combined with equipment operating limits, food safety thresholds, and process stability constraints, the optimal decision algorithm is used to solve for the globally optimal process adjustment that satisfies multi-objective optimization.

[0108] Step E13: Map the global optimal process adjustment amount to each controllable process parameter dimension to generate an initial process parameter adjustment strategy.

[0109] Step E14: Perform pre-simulation and stability verification of the execution effect of the initial process parameter adjustment strategy, correct the adjustment parameters that cause process fluctuations, and obtain the process parameter adjustment strategy.

[0110] Deviation values ​​are the differences between various indicators within the browning state parameter set and the corresponding standard values ​​of the target browning curve, used to quantify the degree of deviation between the actual and ideal states. Examples include browning grade deviation, browning rate deviation, browning uniformity deviation, and localized browning risk deviation. Multi-dimensional browning deviation data is a complete deviation set formed by integrating deviation values ​​from all dimensions, comprehensively reflecting the deviation of the overall browning state of the food. Equipment operating limits are hard constraints such as the upper and lower limits of parameter adjustment and the upper limit of operating load allowed by the processing equipment hardware itself, used to avoid problems such as parameter adjustments exceeding the equipment's capacity. Examples include the maximum adjustment amplitude of heating power, the limit of cavity humidity adjustment, the airflow speed operating threshold, and the limit of the frequency of continuous equipment adjustment. Food safety thresholds are parameter red lines set based on food processing safety standards and quality requirements, preventing process adjustments from causing food quality deterioration and safety hazards. Examples include the maximum processing temperature limit, the safe humidity range, and the maximum step size for a single parameter adjustment. Process stability constraints are rules set to ensure the continuous and stable operation of the processing flow, avoiding drastic parameter changes that cause operational fluctuations. For example, there are gradual parameter adjustment rules, multi-parameter linkage constraints, and stage parameter maintenance requirements. Controllable process parameter dimensions are the classification dimensions of various process parameters that can be actively adjusted in processing equipment; they are the carriers for the implementation of adjustment quantities. Examples include heating power, cavity humidity, processing time, and internal airflow. The initial process parameter adjustment strategy is a preliminary set of adjustment schemes formed after matching the globally optimal process adjustment quantity to each parameter dimension. Execution effect pre-simulation is a process of predicting the browning state of food and the operating conditions of equipment after the strategy is implemented through simulation calculations before the strategy is officially issued and executed. Stability verification is a check operation to detect whether the initial strategy will cause parameter jumps, operating condition fluctuations, equipment overload, or other problems. The process parameter adjustment strategy is the final executable control scheme that, after simulation verification and parameter correction, is logically sound, operates stably, and meets all constraints.

[0111] In this example, when comparing the browning state parameter set with the target browning curve dimension by dimension and calculating the deviation values ​​to obtain multi-dimensional browning deviation data, a simultaneous comparison across all dimensions can be performed. This involves simultaneously retrieving all indicators within the browning state parameter set and comparing them with the corresponding node values ​​of the target browning curve in one go, batch calculating all deviation values, and integrating them into multi-dimensional browning deviation data. Alternatively, a dimension-by-dimensional comparison can be used. Individual comparisons and deviation calculations are performed sequentially according to the browning evaluation indicators. After processing each individual data item, all results are summarized to complete the collection of multi-dimensional browning deviation data.

[0112] After acquiring multi-dimensional browning deviation data, the process adjustment calculation and strategy generation process is initiated. Using the multi-dimensional browning deviation data as input, three types of constraints are introduced: equipment operating limits, food safety thresholds, and process stability constraints. An optimal decision-making algorithm is run to perform multi-objective optimization, calculating the globally optimal process adjustment. This adjustment is then mapped to each controllable process parameter dimension according to parameter correspondences, generating an initial process parameter adjustment strategy. Subsequently, a pre-simulation of the initial strategy's execution effect is conducted to predict the actual operating conditions and browning trends after strategy implementation. Simultaneously, stability verification is performed, identifying and correcting unreasonable adjustment parameters that easily cause process fluctuations. Finally, a compliant and stable process parameter adjustment strategy is output. Through this layered processing logic of deviation comparison, optimization, parameter mapping, and simulation verification, the rationality, safety, and stability of the strategy are controlled at each level, preventing food browning control failure due to inaccurate deviation calculations, excessive parameter adjustments, or insufficient strategy stability.

[0113] For example, there are two ways to gradually generate the final process parameter adjustment strategy from multi-dimensional browning deviation data. The first is a step-by-step serial calculation and verification throughout the entire process, executed in the order of deviation processing, adjustment amount solution, strategy generation, and simulation verification. First, the multi-dimensional browning deviation data is analyzed item by item, each deviation value is read and its validity is screened, and then the three types of constraints—equipment, safety, and process—are substituted one by one. The optimal decision algorithm is run step by step to solve for the globally optimal process adjustment amount, and the solution results are initially judged for compliance. Then, according to the order of controllable process parameter dimensions, the mapping and matching of adjustment amount and parameter dimension are completed one by one, and the initial process parameter adjustment strategy is built step by step. Afterward, the adjustment instructions in the initial strategy are extracted one by one, and the pre-simulation of the execution effect of single instruction and stability verification are carried out sequentially. Once an adjustment parameter that will cause process fluctuation is detected, it is immediately corrected. After all instructions have been verified and corrected, the final process parameter adjustment strategy is obtained. This method employs a single-stage serial advancement and item-by-item verification and correction logic. Each step of the calculation and verification process is independently controllable, enabling precise location of abnormal problems in each stage, such as deviation calculation, algorithm solution, parameter mapping, and strategy verification. Debugging and troubleshooting are convenient, making it suitable for sophisticated food processing scenarios with complex operating conditions and high requirements for single-parameter control accuracy.

[0114] The second approach involves partitioned parallel computation and global unified verification. Based on parameter functional attributes, multi-dimensional browning deviation data is divided into appearance deviation group, environmental condition deviation group, and time-series change deviation group. Each group operates independently and synchronously. Each group runs its corresponding optimal decision algorithm in parallel, simultaneously solving for the corresponding adjustment components. The resulting integrated algorithm yields the globally optimal process adjustment. This globally optimal process adjustment is then mapped to each controllable process parameter dimension according to parameter category, generating multiple sub-adjustment strategies in parallel. These are then combined to form a complete initial process parameter adjustment strategy. A global parallel execution effect pre-simulation is initiated for the complete initial strategy, simultaneously simulating the operational effects of all adjustment commands. Global stability joint verification is conducted, batch identifying all adjustment parameters with fluctuation risks, and centrally completing parameter correction and strategy optimization. This method employs partitioned parallel computation and global unified verification correction logic, fully utilizing parallel computing power to compress overall computation time. The global simulation and joint verification also take into account the overall operating condition changes caused by multi-parameter linkages, resulting in stronger overall strategy coordination. It is suitable for large-scale continuous food processing scenarios with multiple parameter dimensions, large production batches, and high response time requirements.

[0115] Seventh Embodiment This embodiment provides an exemplary scheme for generating an initial process parameter adjustment strategy based on parameter coupling correlation. In this example, the multidimensional composition of the globally optimal process adjustment is first analyzed and decomposed into individual browning correction parameters. Then, a coupling correlation matrix is ​​constructed by combining the influence of process parameters on browning indicators and the linkage relationship between parameters. Based on the coupling correlation matrix and the boundaries of equipment operating parameters, the adjustment amplitude and direction of each controllable process parameter are calculated, and independent adjustment commands are generated. Finally, all independent adjustment commands are integrated, and the execution sequence and execution priority are configured, thereby ultimately generating a multi-dimensional controllable initial process parameter adjustment strategy. Please refer to... Figure 3 , Figure 3 This is a flowchart illustrating the seventh embodiment of the method for controlling browning state of food according to this application. Step E13 includes steps F11-F14: Step F11: Analyze the multidimensional composition of the global optimal process adjustment and decompose it to obtain the individual browning correction parameters.

[0116] Step F12: Based on the sub-item browning correction parameters, and combining the influence weights and interaction relationships of single process parameters on different browning indices, construct a coupling correlation matrix between the sub-item adjustment targets and each of the controllable process parameters.

[0117] Step F13: Based on the coupling correlation matrix and the equipment operating parameter boundary, calculate the optimal adjustment range and adjustment direction of each controllable process parameter, and determine the independent adjustment command for each controllable process parameter.

[0118] Step F14: Integrate the independent instructions corresponding to each of the controllable process parameters and set the execution sequence and execution priority to obtain the initial process parameter adjustment strategy with multi-dimensional control.

[0119] Sub-parameters for browning correction are detailed control components obtained by breaking down the globally optimal process adjustment amount according to different browning correction targets. Each parameter corresponds to the correction requirement of a type of browning index. Examples include browning grade correction, browning rate correction, browning uniformity correction, and local browning risk suppression. Influence weights are the proportion of the effect of a single controllable process parameter on each browning index, used to quantify the degree of influence of parameter control on the browning state. Examples include the proportion of the influence of heating power on browning rate and the proportion of the influence of cavity humidity on browning uniformity. Interaction relationships are the linkage laws of mutual constraints and synergistic effects between different controllable process parameters, reflecting the mutual interference and synergistic effects when multiple parameters are adjusted synchronously. Examples include the suppression relationship of heating power increase on cavity humidity adjustment and the synergistic relationship between airflow velocity and cavity humidity. The coupling correlation matrix is ​​a two-dimensional data matrix formed by digitally integrating sub-parameter adjustment targets, controllable process parameters, influence weights, and parameter linkage relationships. It is the core computational carrier for achieving accurate parameter conversion. Equipment operating parameter boundaries are hard constraints on the adjustment limits of process parameters allowed by the processing equipment hardware, such as the upper and lower limits, and the single adjustment step size. These constraints are used to avoid problems where adjustment commands exceed the equipment's operating capabilities. Examples include the maximum adjustment amplitude of heating power, the adjustable range of cavity humidity, the adjustment limit of airflow speed, and the threshold for a single parameter change. The optimal adjustment amplitude is the optimal change value of the process parameter calculated by combining correction requirements and equipment boundaries. The adjustment direction is the control direction in which the parameter needs to be adjusted upwards, downwards, or remain unchanged. An independent adjustment command is a dedicated execution command generated for a single controllable process parameter, containing both the adjustment amplitude and the adjustment direction. The execution sequence is the time arrangement rule for the sequential execution of multiple adjustment commands, and the execution priority is the authority ranking standard when multiple commands are executed in parallel, used to ensure the orderly implementation of multi-parameter control. The initial process parameter adjustment strategy is a complete preliminary control scheme formed by integrating all independent adjustment commands, sequence rules, and priority rules.

[0120] In this example, when analyzing the multidimensional composition of the globally optimal process adjustment to obtain the individual browning correction parameters, a global decomposition approach can be used. This involves reading all dimensions of the globally optimal process adjustment, uniformly splitting it according to a preset browning index classification standard, and batch outputting all individual browning correction parameters. Alternatively, a hierarchical decomposition approach can be used. The globally optimal process adjustment is decomposed according to the hierarchy of browning impact, with validity verification performed simultaneously for each type of individual browning correction parameter obtained. This process is repeated layer by layer to complete the entire decomposition, thereby accurately extracting the individual browning correction parameters.

[0121] After acquiring the individual browning correction parameters, the process of constructing coupling relationships, calculating instructions, and integrating strategies is initiated. Based on the obtained individual browning correction parameters, and considering the influence weight of each controllable process parameter on different browning indicators, as well as the interaction relationships between parameters, a coupling correlation matrix is ​​constructed that corresponds one-to-one with each controllable process parameter. Subsequently, using the coupling correlation matrix as the basis for calculation, compliance calculations are performed in conjunction with the equipment operating parameter boundaries to calculate the optimal adjustment range and direction for each controllable process parameter, thereby determining the independent adjustment instructions for each controllable process parameter. Finally, all independent adjustment instructions are summarized, and the execution sequence and priority of the instructions are uniformly set according to the food processing requirements, integrating them to form an initial process parameter adjustment strategy with multi-dimensional control capabilities. This end-to-end design, through parameter hierarchical decomposition, coupling relationship modeling, boundary compliance calculations, and unified instruction orchestration, fully considers the multi-parameter coupling and linkage characteristics and equipment operating limitations, improving the rationality and adaptability of adjustment instructions and avoiding problems such as poor browning control effects and abnormal equipment conditions caused by ignoring parameter linkage relationships or exceeding adjustment limits.

[0122] For example, there are two ways to generate an initial process parameter adjustment strategy from sub-item browning correction parameters. The first is parameter-by-parameter serial modeling and instruction generation, which is carried out sequentially according to the classification order of controllable process parameters. First, the first sub-item browning correction parameter is selected, and the influence weight and interaction relationship of the corresponding process parameters are matched to build a locally coupled correlation unit. Then, the optimal adjustment range and adjustment direction of this group of process parameters are calculated in combination with the boundary of equipment operating parameters, and the corresponding independent adjustment instructions are generated and temporarily stored. After completing the processing of a group of parameters, the next sub-item browning correction parameter is processed in sequence, and the entire process of coupled modeling, parameter calculation, and instruction generation is repeated until all independent adjustment instructions for all controllable process parameters are generated. After all instructions are summarized, the execution sequence and execution priority of each instruction are configured one by one according to the preset order, and the integration work is completed step by step to finally obtain the initial process parameter adjustment strategy. This method employs a single-parameter serial advancement and step-by-step modeling and verification operation logic. The processing flow of each set of parameters is independent of each other, which can quickly locate problems such as incorrect coupling relationship construction, parameter calculation exceeding limits, and abnormal instruction logic. It is convenient for on-site debugging and troubleshooting and is suitable for fine food processing scenarios with a large number of process parameters and high single-parameter control accuracy requirements.

[0123] The second approach involves parallel modeling of parameter groups and integration of global instructions. Based on the functional correlation characteristics of process parameters, controllable process parameters with strong interactions are grouped into the same parameter group. These groups operate independently without computational dependencies and are processed in parallel simultaneously. For each parameter group, the influence weights, interactions, and corresponding browning correction parameters within the group are considered simultaneously to construct a dedicated coupling correlation matrix for each group. Each group, relying on its own coupling correlation matrix and the boundaries of the equipment operating parameters, simultaneously calculates the adjustment range and direction of all process parameters within the group, generating all independent adjustment instructions in batches. After all parameter groups have generated their instructions, a global coordination is implemented for all independent adjustment instructions. The execution sequence of cross-group instructions is uniformly planned, and execution priorities are assigned based on the urgency of browning correction. The overall integration generates the initial process parameter adjustment strategy. This method employs a computational logic of grouped parallel computation and globally unified orchestration, significantly reducing overall computation time. Furthermore, it considers parameter linkage relationships during the grouping stage, resulting in higher integrity of the coupling modeling and better multi-parameter collaborative control effects. It is suitable for large-scale food industrial production scenarios with multiple parameters, strong production continuity, and high requirements for control response time.

[0124] Furthermore, a hardware architecture for a dynamic control system for the browning state of food based on multimodal fusion is presented. This hardware system comprises a multimodal sensing unit, a data acquisition and synchronization module, an edge computing or cloud-based fusion processing unit, an execution unit, and a human-machine interface, all working in a layered and collaborative manner. The multimodal sensing unit, relying on a visual imaging module, an infrared thermal imager, a spectral sensor, a humidity sensor, and a timing module, collects real-time data on the food surface image state, food surface temperature distribution, browning-related spectral information, cavity environmental humidity, and the entire processing timeline. This provides a comprehensive raw data source for the multimodal data construction, feature extraction, and browning state assessment of this technology. The data acquisition and synchronization module is responsible for unifying the acquisition time reference of various sensors in the multimodal sensing unit, completing the time alignment, timing matching, and error removal of multi-source heterogeneous data, ensuring the timing consistency of subsequent cross-modal data association, feature fusion, and state comparison. The edge computing or cloud-integrated processing unit serves as the core computing platform, deploying and running the multimodal preprocessing strategy, cross-modal attention browning state assessment model, optimal decision algorithm, and parameter coupling solution logic of this technology. It independently completes the entire intelligent computing process, including browning feature extraction, browning state parameter generation, deviation calculation, optimal process adjustment solution, and control strategy optimization and verification. The execution unit, through adjustable heaters, humidity control devices, fans, and conveyor belt speed controllers, accurately responds to the control commands issued by the upper-level computing unit, dynamically adjusting the heating power, cavity humidity, airflow state inside the cavity, and food conveying speed to achieve closed-loop correction control of the food browning evolution process. The human-machine interface allows staff to set target browning curves for each processing stage as needed. It also provides real-time visualization of processing conditions, multimodal monitoring data, browning status parameters, and control records. Furthermore, it proactively triggers alarms when browning exceeds tolerances, equipment exceeds limits, or abnormal process fluctuations occur. This enables full-process parameter configuration, status monitoring, and safety warnings. The entire hardware module is interconnected and collaborates in layers, fully adapting to the closed-loop control logic of this technology, which features multimodal perception, intelligent evaluation, and dynamic control. This ensures accurate, stable, and adaptive control of the browning status of food.

[0125] Eighth embodiment This embodiment provides an exemplary scheme for closed-loop feedback and adaptive model updating in food browning control. In this example, a process parameter adjustment strategy is first executed, followed by the collection of multimodal feedback data at a unified time base to construct a time-series correlated control effect feedback dataset. Then, the dataset is parsed to extract real-time browning parameters and browning characteristics to obtain browning feedback state parameters. Next, the control effect is determined based on a preset browning threshold. If the threshold is not met, the relevant model parameters are updated based on measured data to obtain system model parameters. Finally, the control strategy is optimized based on the updated parameters, and the system reverts to the data acquisition stage to continue closed-loop control until the browning state meets the target requirements. Please refer to... Figure 4 , Figure 4This is a flowchart illustrating the eighth embodiment of the method for controlling browning in food according to this application. Following step S40, steps G11-G14 are also included: Step G11: After executing the process parameter adjustment strategy, collect feedback multimodal data of the processed food according to a unified time base, and generate a control effect feedback dataset associated with the timing of the current control command.

[0126] Step G12: Analyze the real-time browning parameters in the control effect feedback dataset, extract the corresponding browning features, and correlate them to obtain the browning feedback state parameters.

[0127] Step G13: If the browning control effect does not reach the preset browning threshold, then based on the actual effect data of this control, update the browning state association weight, process parameter coupling relationship and the decision boundary of the optimal decision algorithm, and associate the updated data with the system model to obtain the system model parameters.

[0128] Step G14: Based on the system model parameters, a new process parameter adjustment strategy is generated through optimization. The process then returns to the multimodal data acquisition step to continue executing closed-loop control until the browning state meets the target requirements.

[0129] Feedback multimodal data refers to various monitoring data, such as images, temperature, spectrum, cavity humidity, and processing sequence, re-collected by the multimodal sensing unit after the execution of process control operations. It is used to intuitively reflect the true state of the food after process adjustment and serves as the original data source for evaluating control effectiveness. Examples include post-control food surface image data, real-time temperature distribution data, spectral detection data, cavity dynamic humidity data, and subsequent processing sequence records. The control effect feedback dataset is a standardized data set formed by integrating feedback multimodal data after time-series calibration according to a unified time benchmark and establishing a time-series binding relationship with the currently issued control command. It can accurately correspond to the state changes caused by each control action. Examples include the full-time multimodal data set corresponding to a single control command and the segmented feedback data set corresponding to staged control. Real-time browning parameters are basic quantitative indicators directly parsed from the feedback multimodal data, which can preliminarily reflect the current browning level of the food. Examples include real-time browning level, instantaneous browning rate, and localized browning degree values.

[0130] Browning features are specialized feature information strongly correlated with the browning evolution law of food, deeply mined from feedback multimodal data. They include single-modal unique features and cross-modal correlated features. Examples include texture features of browned areas in images, temperature gradient change features, spectrochemical marker features, and humidity-time-linked change features. Browning feedback state parameters are a comprehensive set of quantitative parameters formed by fusing real-time browning parameters and browning features, used to comprehensively evaluate the actual effect of a single process control. Preset browning thresholds are the pre-set conditions for judging the acceptable browning state and the allowable deviation range based on food processing quality standards; they are the critical basis for distinguishing whether the control effect meets the standards. Examples include the maximum allowable deviation value for browning grade, the browning rate fluctuation threshold, and the local browning risk warning value. Browning state correlation weights are the contribution ratio parameters of each modal feature to the degree of browning within the cross-modal attention mechanism, directly determining the accuracy of feature fusion and state assessment. Process parameter coupling relationships are the linkage rules of mutual constraints and synergistic effects among various controllable process parameters; they are the core foundation for constructing the coupling correlation matrix and calculating independent adjustment commands. The decision boundary of the optimal decision algorithm refers to the constraint interval, solution range, and judgment rules followed during the operation of the optimal decision algorithm, which limits the algorithm's operational logic and output boundaries. System model parameters refer to the complete set of model operating parameters formed after integrating and updating the browning state correlation weights, process parameter coupling relationships, and the decision boundary of the optimal decision algorithm, providing a new basis for subsequent full-process calculations.

[0131] In this example, when collecting feedback multimodal data and generating a control effect feedback dataset according to a unified time base, a full-domain synchronous acquisition method can be used. All sensor modules are simultaneously activated to complete the acquisition of all-dimensional feedback data at once. Then, the data is bound with the corresponding time-series labels of control commands, and complete control effect feedback datasets are generated in batches. Alternatively, a time-division polling incremental acquisition method can be used, collecting various modal data in turn according to a preset order. While completing data acquisition, time-series calibration and command association matching are performed, gradually summarizing and integrating to obtain the control effect feedback dataset, thereby completing the standardized construction of the feedback data.

[0132] After constructing the control effect feedback dataset, the feedback parsing and closed-loop iteration process is initiated. All content in the dataset is read, and the corresponding real-time browning parameters are extracted one by one. Simultaneously, various browning features are deeply mined and extracted from the data. The real-time browning parameters and browning features are correlated and fused to obtain complete browning feedback state parameters. These parameters are compared item by item with preset browning thresholds to determine if the current browning control effect meets the acceptable standard. If the browning control effect fails to meet the preset threshold, the actual effect data generated by this control is retrieved. The browning state correlation weights, process parameter coupling relationships, and the decision boundary of the optimal decision algorithm are adaptively corrected and iteratively updated sequentially. All updated data are synchronously entered and correlated with the system's built-in model to generate new system model parameters. Based on the updated system model parameters, a new process parameter adjustment strategy is recalculated and optimized. The process then returns to the initial multimodal data acquisition step, and the entire control process is executed cyclically for continuous closed-loop control until the food browning state stably reaches the preset target requirements. By employing a closed-loop operating logic encompassing feedback collection, status analysis, effect judgment, model self-updating, and strategy iteration, the system can autonomously optimize internal parameters based on the actual on-site control effects, continuously correcting evaluation deviations and control errors, and effectively avoiding the problem of decreased control accuracy caused by model solidification and long-term parameter aging.

[0133] For example, there are two ways to achieve the overall implementation of closed-loop feedback, model update, and strategy iteration. The first is to execute the entire process sequentially, following a fixed order of data acquisition, dataset construction, parameter parsing, feature extraction, state fusion, threshold determination, model update, and strategy optimization. First, multimodal feedback data is collected categorically, and time-series calibration and command association are completed step-by-step to gradually build the control effect feedback dataset. Then, real-time browning parameters are parsed item by item, browning features are extracted categorically, and browning feedback state parameters are obtained step-by-step through fusion. Next, each browning feedback state parameter is compared with a preset browning threshold. If the control effect is not met, the parameters are updated sequentially according to the browning state association weight, process parameter coupling relationship, and the decision boundary of the optimal decision algorithm. The validity of each update is verified before integration to obtain the system model parameters. Finally, based on the new system model parameters, the process parameter adjustment strategy is optimized step-by-step, and then the process jumps back to the multimodal data acquisition stage to start the next round of closed-loop circulation. This method employs a single-stage serial advancement and step-by-step verification operation logic, which can locate and respond to problems such as parsing anomalies, model update failures, and unreasonable strategy optimizations segment by segment. It is highly convenient for fault diagnosis and on-site debugging, and is suitable for food processing scenarios with strict requirements for the stability of single-stage operation, such as small-batch trial production.

[0134] The second approach involves multi-module partitioned parallel closed-loop iteration. Based on data function and modal attributes, feedback data is divided into visual temperature partitions and spectral environment time-series partitions. These two partitions are independent and computationally independent, simultaneously performing data acquisition, parameter analysis, and feature extraction. Local browning feedback state parameters within each partition are generated in parallel and then aggregated to obtain global browning feedback state parameters. A parallel comparison between all parameters and preset browning thresholds is performed. If the control effect is confirmed to be substandard, the browning state correlation weights, process parameter coupling relationships, and optimal decision algorithm decision boundaries are updated in parallel across modules, generating system model parameters in batches. Based on the newly generated system model parameters, the optimization of process parameter adjustment strategies is completed globally in parallel, and the process jumps to the multi-modal data acquisition step, entering the next round of closed-loop control. This method employs partitioned parallel computation and global synchronous iteration, significantly shortening the overall time of a single closed-loop cycle. Simultaneously, multi-module parallel updates can take into account the coordinated optimization of parameters across different dimensions, resulting in stronger model adaptive learning capabilities. This method is suitable for industrial food processing production lines with continuous production, large batch sizes, and high requirements for timely control response.

[0135] Further, please refer to Figure 5 , Figure 5 This is a flowchart illustrating the output of food browning status in this application. A food browning status assessment process based on multimodal data first collects three types of basic monitoring data: visual data (Red-Green-Blue / Hyperspectral), thermal data (Infrared thermography), and spectral data (Near-Infrared / Raman). The visual data undergoes image segmentation, denoising, and color space conversion to extract the proportion of browned areas and the browning intensity index. The thermal data is processed through spatial interpolation and outlier removal to obtain the temperature mean and temperature gradient features. The spectral data undergoes baseline correction and normalization, followed by extraction of chemical marker features after dimensionality reduction using Principal Component Analysis (PCA). Subsequently, all features output from the three data channels are input into a multimodal fusion layer for feature-level or decision-level fusion, ultimately outputting a food browning status assessment result including browning grade, browning rate, and browning uniformity.

[0136] This application provides a control device for the browning state of food. The control device for the browning state of food includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the control method for the browning state of food in the first embodiment described above.

[0137] The following is for reference. Figure 6The diagram illustrates a structural schematic of a control device suitable for achieving the browning state of food in the embodiments of this application. The control device for the browning state of food in the embodiments of this application may include, but is not limited to, mobile terminals such as oxidative browning control devices and non-enzymatic browning control devices, as well as fixed terminals such as vacuum degassers, vacuum low-temperature fryers, and vacuum dryers. Figure 6 The illustrated device for controlling the browning state of food is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0138] like Figure 6 As shown, the control device for the browning state of food may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage device 1003 into a random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the control device for the browning state of food. The processing unit 1001, the read-only memory 1002, and the RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the food browning control device to communicate wirelessly or wiredly with other devices to exchange data. Although the figure shows a food browning control device with various systems, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems may be implemented alternatively.

[0139] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0140] The food browning state control device provided in this application, employing the food browning state control method in the above embodiments, can solve the technical problem of poor food baking quality. Compared with the prior art, the beneficial effects of the food browning state control device provided in this application are the same as those of the food browning state control method provided in the above embodiments, and other technical features of the food browning state control device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0141] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0142] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0143] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the food browning state control method in the above embodiments.

[0144] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, radio frequency (RF), etc., or any suitable combination thereof.

[0145] The aforementioned computer-readable storage medium may be included in a control device for controlling the browning state of food; or it may exist independently and not be assembled into a control device for controlling the browning state of food.

[0146] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by a food browning state control device, the control device causes the following: in response to a browning control command, it correlates the collected modal data according to a unified time reference to obtain multimodal data of the processed food; it performs feature extraction processing on the multimodal data of the processed food according to a modal differentiation processing strategy to obtain multimodal feature data of the processed food; it calculates the correlation weight between the multimodal feature data and the degree of browning through a cross-modal attention mechanism, and weightedly fuses the high-order feature representations of the multimodal data to obtain a browning state parameter set of the processed food; based on the comparison result between the browning state parameter set and the target browning curve, it calculates the parameter adjustment amount through an optimal decision algorithm to obtain a process parameter adjustment strategy, so as to control the browning state of the processed food through the process parameter adjustment strategy.

[0147] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0148] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation that may be implemented in systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0149] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0150] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described method for controlling the browning state of food, thereby solving the technical problem of poor food baking quality. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the food browning state control method provided in the above embodiments, and will not be repeated here.

[0151] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for controlling the browning state of food, characterized in that, The method includes: In response to browning control instructions, multimodal data of processed foods are obtained by associating the collected modal data according to a unified time reference. According to the modal differentiation processing strategy, feature extraction processing is performed on the multimodal data of the processed food to obtain the multimodal feature data of the processed food. The correlation weights between the multimodal feature data and the degree of browning are calculated by a cross-modal attention mechanism, and the high-order feature representations of the multimodal data are weighted and fused to obtain the browning state parameter set of the processed food. Based on the comparison results between the browning state parameter set and the target browning curve, the parameter adjustment amount is calculated through the optimal decision algorithm to obtain the process parameter adjustment strategy, so as to control the browning state of the processed food through the process parameter adjustment strategy.

2. The method for controlling the browning state of food as described in claim 1, characterized in that, The step of obtaining multimodal data of processed food by associating the collected modal data according to a unified time base in response to browning control commands includes: In response to the browning control command, the surface condition of the processed food, cavity environment parameters, and processing time sequence data are collected according to the unified time reference, and the raw data are processed. The timestamps of the original data are calibrated according to the time base, and the matching relationship of cross-modal time series is cross-validated to remove abnormal time series data and obtain time series modal data. Based on the time-series modal data, feature association mapping relationships between each modal data are constructed, and the data dimensions of each modal data are unified to obtain the multimodal data of the processed food.

3. The method for controlling the browning state of food as described in claim 1, characterized in that, The step of performing feature extraction processing on the multimodal data of the processed food according to the modal differentiation processing strategy to obtain the multimodal feature data of the processed food includes: Based on the aforementioned modal differentiation processing strategy and multimodal data types, classify and match the modality type processing strategy of the multimodal data; According to the modality type processing strategy, modality-specific preprocessing is performed on the multimodal data, and the data deviation of the multimodal data is corrected to obtain each initial feature data; Each of the initial feature data is subjected to modality-specific feature enhancement operations to extract core feature information strongly correlated with the browning process, thereby obtaining each enhanced feature data; The features of each of the enhanced feature data are aligned, and the dimensions of each of the enhanced feature data are unified to obtain the multimodal feature data of the processed food.

4. The method for controlling the browning state of food as described in claim 3, characterized in that, The steps of performing mode-specific preprocessing on the multimodal data according to the mode type processing strategy, and correcting the data bias of the multimodal data to obtain each initial feature data include: Based on the modality processing strategy, the image data processing strategy preprocesses the image data in the multimodal data and extracts image feature data associated with browning from the preprocessed image data; Based on the modality processing strategy, the temperature data processing strategy performs spatial interpolation and anomaly removal on the temperature data in the multimodal data to obtain temperature feature data of continuous temperature distribution on the food surface. Based on the modality-type processing strategy, the spectral data processing strategy performs baseline correction and principal component dimensionality reduction on the spectral data in the multimodal data to extract the spectral feature data of browning-related chemical markers. The humidity and time data in the multimodal data are synchronously smoothed and filtered to extract the characteristics of environmental humidity changes and processing time sequence. Based on the modal differentiation correction rule, the data deviations of the environmental humidity change, the processing time sequence features and the image feature data, the temperature feature data and the spectral feature data are corrected to obtain the initial feature data of each modality.

5. The method for controlling the browning state of food as described in claim 1, characterized in that, The step of calculating the correlation weights between the multimodal feature data and the degree of browning through a cross-modal attention mechanism, and then weighted and fusing the high-order feature representations of the multimodal data to obtain the browning state parameter set of the processed food includes: The multimodal feature data is input into the browning state assessment model and mapped to the high-level feature representation space to obtain multimodal deep feature vectors. Guided by the browning feature discrimination criterion, the correlation weights between the multimodal deep feature vectors and the degree of browning are calculated dimension by dimension through the cross-modal attention mechanism to determine the correlation weight matrix; The multimodal deep feature vectors are finely weighted and superimposed dimension by dimension according to the correlation weight matrix, and the global modal weight coefficients are adjusted according to the current processing stage to obtain the global browning comprehensive features. The global browning comprehensive characteristics are analyzed in multiple dimensions and numerically converted to generate the browning state parameter set.

6. The method for controlling the browning state of food as described in claim 1, characterized in that, The step of calculating parameter adjustment amounts using an optimal decision algorithm based on the comparison results between the browning state parameter set and the target browning curve, and obtaining the process parameter adjustment strategy to control the browning state of the processed food through the process parameter adjustment strategy includes: The browning state parameter set is compared with the target browning curve of the corresponding processing stage in a dimension-by-dimensional manner to calculate the deviation value of the browning process and obtain multi-dimensional browning deviation data. Based on the multi-dimensional browning deviation data, combined with equipment operating limits, food safety thresholds, and process stability constraints, the optimal decision-making algorithm is used to solve for the globally optimal process adjustment that satisfies multi-objective optimization. The global optimal process adjustment amount is mapped to each controllable process parameter dimension to generate an initial process parameter adjustment strategy. The execution effect and stability of the initial process parameter adjustment strategy are pre-simulated and verified. The adjustment parameters that cause process fluctuations are corrected to obtain the process parameter adjustment strategy.

7. The method for controlling the browning state of food as described in claim 6, characterized in that, The step of mapping the globally optimal process adjustment amount to each controllable process parameter dimension to generate an initial process parameter adjustment strategy includes: The multidimensional composition of the global optimal process adjustment is analyzed, and the individual browning correction parameters are obtained by decomposition. Based on the aforementioned sub-item browning correction parameters, and combining the influence weights and interaction relationships of individual process parameters on different browning indices, a coupling correlation matrix between the sub-item adjustment targets and each of the controllable process parameters is constructed. Based on the coupling correlation matrix and the equipment operating parameter boundaries, the optimal adjustment range and direction of each controllable process parameter are calculated, and the independent adjustment command of each controllable process parameter is determined. By integrating the independent instructions corresponding to each controllable process parameter and setting the execution sequence and execution priority, a multi-dimensional controllable initial process parameter adjustment strategy is obtained.

8. The method for controlling the browning state of food as described in claim 1, characterized in that, Following the step of calculating parameter adjustment amounts using an optimal decision algorithm based on the comparison results between the browning state parameter set and the target browning curve, and obtaining the process parameter adjustment strategy to control the browning state of the processed food through the process parameter adjustment strategy, the method for controlling the browning state of the food further includes: After executing the process parameter adjustment strategy, feedback multimodal data of the processed food are collected according to a unified time base to generate a control effect feedback dataset associated with the timing of the current control command. The real-time browning parameters in the control effect feedback dataset are analyzed, and the corresponding browning features are extracted and correlated to obtain the browning feedback state parameters; If the browning control effect does not reach the preset browning threshold, the browning state association weight, process parameter coupling relationship and optimal decision algorithm decision boundary are updated based on the actual effect data of this control, and the updated data is associated with the system model to obtain the system model parameters. Based on the system model parameters, a new process parameter adjustment strategy is generated through optimization. The process then returns to the multimodal data acquisition step to continue executing closed-loop control until the browning state reaches the target requirements.

9. A device for controlling the browning state of food, characterized in that, The control device for the browning state of food includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the control method for the browning state of food as described in any one of claims 1 to 8.

10. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the method for controlling the browning state of food as described in any one of claims 1 to 8.

Images