Meat sorting pretreatment system based on AI vision

Abstract

The application relates to an AI vision-based meat sorting pretreatment system, in particular to the field of intelligent pretreatment of meat processing, which quantitatively monitors feature distribution drift, intelligently triggers active selection of high-value samples and safe and efficient online incremental updating of a model, simultaneously combines digital twin technology to perform forward-looking data enhancement and model reinforcement, and effectively overcomes the performance degradation problem of a static model caused by changes in raw materials and processes, significantly reduces manual intervention and labeling costs, improves the consistency of sorting precision and the generalization ability to new batches and new features, and thus guarantees long-term stable, efficient and intelligent operation of the pretreatment link.

Description

Technical Field

[0001] This invention relates to the field of intelligent pre-processing of meat processing, and more specifically, to a meat sorting and pre-processing system based on AI vision. Background Technology

[0002] In the meat processing industry, the increasing levels of industrialization and automation have placed higher demands on production efficiency and product quality consistency. Among these, sorting and pre-processing, a crucial step in accurately classifying raw meat by cut, grade, and quality, directly impacts subsequent processing value and market benefits. Traditional manual sorting methods suffer from inherent limitations such as low efficiency, inconsistent standards, and rising labor costs. Therefore, automated sorting systems based on artificial intelligence vision have become an important direction for industrial upgrading. These systems collect visual information from raw meat, utilize image processing and deep learning algorithms to automatically identify and classify its external features, and then drive the actuators to complete the sorting. In practical applications, raw meat comes from a wide range of sources, including pigs, cattle, and sheep from different origins, breeds, and rearing cycles, exhibiting significant differences in natural color, texture, intramuscular fat distribution, and size. Furthermore, post-slaughter cooling processes, aging processes, and physicochemical changes in the storage and transportation environment cause dynamic evolution in the surface color and water retention of meat products. Production lines typically operate at high speeds continuously, requiring vision systems to maintain real-time, stable, and high-precision sorting performance under these complex and variable conditions.

[0003] However, current mainstream deep learning-based visual sorting technologies face severe challenges in practical deployment. The core issue lies in the inherent contradiction between the static nature of the model and the dynamic nature of production data. Specifically, existing solutions typically rely on fixed models trained on raw material data from specific periods and batches, such as convolutional neural network models. Once deployed, the parameters and decision boundaries of these models are fixed. In actual production environments, due to the diversity of raw material sources, subtle fluctuations in production processes, and uncontrollable changes in the supply chain, the visual feature distribution of raw meat on the production line will continuously change slowly or drastically—a phenomenon known as concept drift in machine learning. This leads to a deviation between the data distribution learned during training and the actual data distribution encountered during online operation, resulting in a discrepancy between the static and dynamic nature of the model and the actual data distribution encountered during online operation. The recognition accuracy and generalization ability of the dynamic model significantly degrade over time. To address this performance degradation, existing technologies typically employ a traditional model update method involving periodic production stoppages, recollection of new samples, manual annotation, and retraining of the model. This approach is not only cumbersome and time-consuming, but also leads to production line downtime and economic losses. More importantly, its response is severely delayed, making it unable to adapt to the continuous and potentially rapidly changing working conditions on the production line. Therefore, the industry urgently needs to develop an online adaptive visual sorting technology that enables the system to continuously and efficiently optimize and adjust its model using real-time production data without downtime or with minimal human intervention. This will overcome concept drift and ensure the long-term stability and reliability of sorting accuracy throughout the entire production cycle. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing an AI vision-based meat sorting and preprocessing system. The system utilizes a feature monitoring module, a sample selection module, a model update module, and a digital twin module to solve the problems mentioned in the background.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: specifically, it includes: a feature monitoring module, a sample selection module, a model update module, and a digital twin module connected in sequence, wherein; Feature monitoring module: acquires images of meat products flowing on the production line in real time, extracts the depth feature vector of the meat product images, compares the depth feature vector with a dynamically maintained historical feature distribution database, and calculates the feature distribution difference degree; when the feature distribution difference degree exceeds a preset threshold, it generates and sends a drift signal containing the current feature distribution difference degree and feature distribution change trend data. Sample selection module: Receives drift signals, calls a pre-built sorting model to predict the meat images subsequently acquired by the feature monitoring module, and jointly calculates the uncertainty measure of the prediction results and the representative distance of the depth feature vectors corresponding to the subsequent meat images in the feature space. Priority scores are calculated based on the uncertainty measure and the representative distance. Specific meat images are selected according to their priority scores, and the real labels of these images are obtained through an online annotation mechanism, forming incremental sample pairs consisting of the specified meat images and their corresponding real labels. These incremental sample pairs are then stored in an incremental sample pool. Model update module: At a preset time, incremental sample pairs are obtained from the incremental sample pool, and the sorting model is incrementally trained based on the elastic weight consolidation algorithm; wherein, the parameters of the sorting model are constrained based on the weights of historical task importance, and the parameters of different network layers of the sorting model are updated using differentiated learning rates; Digital Twin Module: Includes a parameterized meat appearance simulator, used to receive drift signals, dynamically adjust the parameter distribution of the meat appearance simulator according to the feature distribution differences contained in the drift signals to generate corresponding synthetic training data; and use the synthetic training data to reinforce the sorting model updated by the model update module.

[0006] In a preferred embodiment, the specific process for calculating the feature distribution difference degree in the feature monitoring module is as follows: The depth feature vector of meat images is extracted using the built-in coding network of the feature monitoring module; the historical feature distribution library is modeled using a Gaussian mixture model to model the distribution of historical depth feature vectors. A list of key feature dimensions is constructed by identifying the dimensions with the largest absolute weights from the fully connected layer weights of the pre-built sorting model. When calculating the feature distribution difference, firstly, multiple depth feature vectors corresponding to a predetermined number of meat product images in the current batch are obtained. These multiple depth feature vectors form the depth feature vector set for the current batch, and the following operations are performed sequentially: A1, calculate the Wasserstein distance between the current deep feature vector distribution represented by the deep feature vector set and the historical deep feature vector distribution represented by the historical feature distribution library, as the first difference component; A2. For each dimension in the key feature dimension list, calculate the feature mean of the deep feature vector set in that dimension, and calculate the difference between the feature mean and the historical feature mean corresponding to that dimension in the historical feature distribution library. Then divide the difference by the sum of the historical feature standard deviation corresponding to that dimension in the historical feature distribution library and a very small positive number to obtain the relative offset of that dimension. Then sum the relative offsets of all dimensions in the key feature dimension list to obtain the second difference component. A3. Multiply the first difference component by the first weight coefficient, and multiply the second difference component by the second weight coefficient. The sum of the first and second weight coefficients is always one. Finally, add the two products together. The result is the characteristic distribution difference degree.

[0007] In a preferred embodiment, the specific process of generating the drift signal and dynamically maintaining the historical feature distribution database is as follows: The feature distribution difference is compared with a preset drift judgment threshold. If it exceeds the threshold, concept drift is determined to have occurred, and a structured drift signal is generated. The drift signal includes a trigger flag, the value of the feature distribution difference, the key dimension offset vector, the timestamp, and the batch index. The key dimension offset vector is calculated based on the historical mean and historical standard deviation of each key dimension in the historical feature distribution library, and the feature mean of the current batch depth feature vector set on the corresponding key feature dimension, representing the direction and magnitude of change of each key feature dimension; The historical feature distribution library uses a Gaussian mixture model to model the probability distribution of historical deep feature vectors; the list of key feature dimensions is derived from the weight analysis of the fully connected layers of the sorting model and is recalculated synchronously as the sorting model is updated. After the feature monitoring module generates the drift signal, its historical feature distribution library is independently updated based on feedback from the sample selection module. Specifically, when the sample selection module confirms the true label of the meat image through the online annotation mechanism and forms an incremental sample pair, the depth feature vector corresponding to the incremental sample pair is input into the historical feature distribution library to incrementally update the model parameters of the Gaussian mixture model. The model parameters of the Gaussian mixture model include the mixing weights, mean vector, and covariance matrix of each Gaussian component in the Gaussian mixture model, so that the historical feature distribution library tracks the distribution of the depth feature vector that has been correctly recognized, serving as a new baseline for subsequent calculations.

[0008] In a preferred embodiment, the specific process of calculating the uncertainty measure after receiving the drift signal in the sample selection module is as follows: A fusion weight coefficient is dynamically set based on the numerical value of the difference in the feature distribution in the drift signal. The initial value of the fusion weight coefficient is set to be directly proportional to the numerical value of the difference in the feature distribution, and it decreases according to a preset attenuation function after the selection begins. The attenuation target value is 0.5. For each frame of meat image and its corresponding depth feature vector acquired after the drift signal is emitted, the sorting model is invoked for prediction. The uncertainty measure is calculated using a forward propagation mutation method based on random deactivation: a random deactivation layer is enabled in the sorting model inference, and forward propagation is repeatedly performed on the same depth feature vector to generate multiple prediction probability distributions. The entropy value of the average distribution of these multiple prediction probability distributions is calculated as the uncertainty measure. The larger the entropy value, the more inconsistent the prediction results are and the higher the uncertainty.

[0009] In a preferred embodiment, the specific process of calculating the representative distance is as follows: For a depth feature vector, its representative distance is calculated. This representative distance is used to measure the degree of deviation of the depth feature vector from the historical feature distribution represented by the historical feature distribution library, and a biased measure is specifically combined with the direction indicated by the key dimension offset vector in the drift signal. The calculation of the representative distance consists of two parts: the first part is to calculate the global Mahalanobis distance from the depth feature vector to the historical feature distribution database. The global Mahalanobis distance is calculated as follows: first, obtain the global mean vector and covariance matrix of the distribution represented by the historical feature distribution database; then, calculate the difference vector between the depth feature vector and the global mean vector; finally, by transposing the difference vector, multiplying it by the inverse of the covariance matrix, and then multiplying it by the difference vector itself, the square root of the product is taken to obtain the global Mahalanobis distance. The second part is to calculate the difference vector between the deep feature vector and the global mean vector of the historical feature distribution library, and the scalar value of the projected distance in the direction of the key dimension offset vector. The scalar value of the projected distance is obtained by calculating the dot product of the difference vector and the key dimension offset vector, and then dividing by the Euclidean norm of the key dimension offset vector. The final representative distance is the product of the overall Mahalanobis distance and a first distance weight, plus the product of the projected distance scalar value and a second distance weight. The sum of the first distance weight and the second distance weight is one. The value of the second distance weight is calculated by normalizing the Euclidean norm of the received key dimension offset vector. The normalization result is used as the value of the second distance weight, so that when the drift in a specific direction is significant, the calculation of the representative distance focuses more on the frontier of the sample in that direction.

[0010] In a preferred embodiment, the specific process of calculating the priority score, forming incremental sample pairs, and storing them in the incremental sample pool is as follows: For each frame of meat product image, its uncertainty measure and representative distance are normalized respectively to obtain the normalized uncertainty value and the normalized representative distance value; Based on the current fusion weight coefficients, the normalized uncertainty value and the normalized representative distance value are nonlinearly fused to obtain the priority score of the meat product image. Nonlinear fusion calculation uses a weighted power average method, which involves a preset power parameter greater than zero; The weighted power average is calculated as follows: First, the normalized uncertainty value is raised to the power parameter, and the result is multiplied by the current fusion weight coefficient to obtain the first term. Second, the normalized representative distance value is raised to the power parameter, and the result is multiplied by the difference between one and the current fusion weight coefficient to obtain the second term. Then, the first and second terms are added to obtain the sum. Finally, the sum is raised to the power parameter by the reciprocal, and the result is the priority score. Internally, a fixed-capacity priority queue is maintained to temporarily store meat images, corresponding depth feature vectors, and calculated priority scores. The queue is updated according to the score order to ensure that the queue stores the sample with the highest score. When the preset stopping condition is met, select the highest-scoring samples from the priority queue as the specified meat product images, obtain their real labels through the online annotation mechanism, form incremental sample pairs composed of the corresponding real labels, and store them in the incremental sample pool.

[0011] In a preferred embodiment, the specific process of incrementally training the sorting model at a preset time in the model update module is as follows: When the number of sample pairs in the incremental sample pool reaches the first quantity threshold or enters a preset fixed time interval, the sorting model is updated. A new training dataset is formed by obtaining the latest incremental sample pairs from the incremental sample pool in a number equal to the first quantity threshold. Simultaneously, the current sorting model parameters and their historical importance weight values ​​are read from the internal memory, and the importance weight values ​​constitute the historical importance weight vector; Calculate a dynamic importance weight for each model parameter. This dynamic importance weight is obtained by weighted summation of three components: the first component is the stored historical importance weight value. The second component is the decay coefficient between zero and one obtained by mapping the absolute value of the gradient in the most recent incremental training of the sorting model parameters through the logistic function; The third component is a binary indicator function. It takes one value when the sorting model parameter is a weight parameter in the fully connected layer of the sorting model that is directly connected to the dimension in the list of key feature dimensions, and zero value otherwise. The sum of the three weighting coefficients is one, which forms the dynamic importance weight vector.

[0012] In a preferred embodiment, the specific process of updating the parameters of the sorting model is as follows: The total loss function for this incremental training is constructed as follows: First, the standard cross-entropy classification loss on the new training dataset is calculated as the first term; Then, for each sorting model parameter, the dynamic importance weight of the sorting model parameter is multiplied by the square of the difference between the initial value of the sorting model parameter before training and the current value during training to obtain the regularization contribution value of the sorting model parameter. The regularization contribution values ​​of all sorting model parameters are summed, and the summation result is multiplied by half of a preset regularization strength coefficient as the second term. Finally, add the first term to the second term to obtain the total loss function; A layered differentiated learning rate strategy is adopted: the network layers of the sorting model are divided into a low-level feature encoding layer, a high-level semantic feature layer, and a classification layer; a lower learning rate is set for the low-level feature encoding layer; and a higher learning rate is set for the high-level semantic feature layer and the classification layer. Using the stochastic gradient descent algorithm, with the total loss function as the objective, the model is trained iteratively on a new training dataset with differentiated learning rates to update the parameters of the sorting model. After training is completed, the Fisher information content of each sorting model parameter is recalculated using the new training dataset to obtain new importance weight values. Then, the stored historical importance weight vector and the new importance weight vector are weighted and fused according to a preset forgetting factor to obtain the updated importance weight vector and store it. At the same time, the updated sorting model parameters are sent back to the feature monitoring module to recalculate the list of key feature dimensions.

[0013] In a preferred embodiment, the specific process of generating corresponding synthetic training data in the digital twin module is as follows: Receive drift signals and parse them to obtain the numerical value of feature distribution difference degree and key dimension offset vector; calculate the distribution inference intensity coefficient based on the numerical value of feature distribution difference degree, ensuring that it is positively correlated with the numerical value of feature distribution difference degree and does not exceed the preset upper limit; The parameterized meat appearance simulator defines multiple simulator parameters that follow a preset prior probability distribution. A transfer mapping is constructed based on the inferred intensity coefficient and the key dimension offset vector. This transfer mapping acts on the original prior distribution of the simulator parameters, transforming them into an adjusted parameter distribution. The process of constructing the transport map includes: estimating the global mean vector and covariance matrix of the original prior distribution based on historical sampling data of simulator parameters, and calculating the square root matrix of the inverse of the covariance matrix for normalization; The key dimension offset vector is linearly transformed into the expected adjustment direction vector through a preset mapping matrix. The transfer mapping is defined as adding the following two terms to the value of each simulator parameter in its original prior distribution: the first term is the calculated distribution inference intensity coefficient, and the second term is the component in the expected adjustment direction vector corresponding to the simulator parameter after normalization by the square root matrix of the inverse of the covariance matrix.

[0014] In a preferred embodiment, the specific process of using synthetic training data to reinforce the updated sorting model is as follows: Sampling is performed from the adjusted simulator parameter distribution, with a bias towards selecting simulator parameter values ​​corresponding to the tail region of the distribution to generate more challenging synthetic meat images; Input the simulator parameter values ​​into the parameterized meat appearance simulator to generate synthetic images and completely known pseudo-labels, forming a synthetic training dataset; use this dataset to enhance the training of the updated sorting model. The total loss function includes the standard classification loss of the synthetic meat image and the consistency regularization loss, which is obtained by calculating the Jensen-Shannon divergence between the predicted distributions of the two different augmented views of the synthetic meat image. The training can employ a course learning strategy to gradually increase the distribution inference intensity coefficient corresponding to the synthetic data used; it also includes an online fine-tuning mechanism: when a newly labeled incremental sample pair is obtained, the ideal simulator parameters are deduced by optimizing the depth feature vector corresponding to the incremental sample pair, and the mapping matrix is ​​fine-tuned by gradient descent in conjunction with the latest drift signal and the covariance information of the simulator parameter distribution, so that the mapping relationship is closer to the real physical relationship.

[0015] The beneficial effects of this invention are as follows: by quantitatively monitoring feature distribution drift, and thereby intelligently triggering the active selection of high-value samples and safe and efficient online incremental updates of the model, and combining digital twin technology for forward-looking data augmentation and model strengthening, this closed-loop technology effectively overcomes the performance degradation problem of static models caused by changes in raw materials and processes, significantly reduces the cost of manual intervention and annotation, improves the consistency of sorting accuracy and the generalization ability to new batches and new features, thereby ensuring the long-term stable, efficient and intelligent operation of the preprocessing stage. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a block diagram of the system structure of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0019] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application. Example

[0020] This embodiment provides, for example Figure 1-2 The system illustrates a meat sorting and preprocessing system based on AI vision, specifically comprising: a feature monitoring module, a sample selection module, a model update module, and a digital twin module connected in sequence; wherein; Feature monitoring module: acquires images of meat products flowing on the production line in real time, extracts the depth feature vector of the meat product images, compares the depth feature vector with a dynamically maintained historical feature distribution database, and calculates the feature distribution difference degree; when the feature distribution difference degree exceeds a preset threshold, it generates and sends a drift signal containing the current feature distribution difference degree and feature distribution change trend data. Sample selection module: Receives drift signals, calls a pre-built sorting model to predict the meat images subsequently acquired by the feature monitoring module, and jointly calculates the uncertainty measure of the prediction results and the representative distance of the depth feature vectors corresponding to the subsequent meat images in the feature space. Priority scores are calculated based on the uncertainty measure and the representative distance. Specific meat images are selected according to their priority scores, and the real labels of these images are obtained through an online annotation mechanism, forming incremental sample pairs consisting of the specified meat images and their corresponding real labels. These incremental sample pairs are then stored in an incremental sample pool. Model update module: At a preset time, incremental sample pairs are obtained from the incremental sample pool, and the sorting model is incrementally trained based on the elastic weight consolidation algorithm; wherein, the parameters of the sorting model are constrained based on the weights of historical task importance, and the parameters of different network layers of the sorting model are updated using differentiated learning rates; Digital Twin Module: Includes a parameterized meat appearance simulator, used to receive drift signals, dynamically adjust the parameter distribution of the meat appearance simulator according to the feature distribution differences contained in the drift signals to generate corresponding synthetic training data; and use the synthetic training data to reinforce the sorting model updated by the model update module.

[0021] In this embodiment, the specific process for calculating the feature distribution difference degree in the feature monitoring module is as follows: The feature detection module uses its built-in encoding network to extract features from each real-time captured meat image to obtain the corresponding deep feature vector. The encoding network is a pre-trained, lightweight convolutional neural network, such as MobileNet or EfficientNet with the original classification head removed. Its function is to map high-dimensional image data into a fixed-length low-dimensional deep feature vector to reduce the computational complexity of subsequent distribution comparison and focus on semantic features related to the sorting task. The dynamically maintained historical feature distribution library uses a Gaussian mixture model to model the deep feature vector distribution of correctly sorted meat samples in history. The Gaussian mixture model is learned through the expectation-maximization algorithm. The number of Gaussian components can be preset according to the clustering characteristics of historical data, for example, between 5 and 20, to fit the multimodal distribution characteristics that may exist in the deep feature space, thereby more accurately representing the feature distribution under the historical normal state. From the weights of the fully connected layers of the pre-built sorting model, identify the dimensions with the largest absolute weight values ​​to form a list of key feature dimensions. The number of "several dimensions" can be determined based on experience or experiments. For example, select the dimensions whose sum of absolute weight values ​​accounts for the top 80% of the total absolute weight values, or select the top 50 dimensions. The purpose of this is to focus on the feature dimensions that have the greatest impact on the current sorting model's decision-making, so that the difference calculation can sensitively capture feature changes that may lead to a decline in the performance of the sorting model. When calculating the feature distribution difference, firstly, multiple depth feature vectors corresponding to a predetermined number of meat images in the current batch are obtained. These multiple depth feature vectors form the depth feature vector set for the current batch. The predetermined number is usually a batch size sufficient for stable statistics, such as 64, 128, or 256. The specific value can be set in a balance between production line speed and real-time calculation requirements. Then, the following operations are performed in sequence: A1 calculates the Wasserstein distance between the current deep feature vector distribution represented by the deep feature vector set and the historical deep feature vector distribution represented by the historical feature distribution library, as the first difference component. The Wasserstein distance, also known as the bulldozer distance, is used to measure the difference between two probability distributions. Unlike KL divergence, the Wasserstein distance can still provide a meaningful and smooth difference measure even if the support sets of the two distributions do not overlap. This makes it more robust to gradual changes and abrupt changes in feature distributions and can effectively capture the overall structural drift of the feature space. A2. For each dimension in the list of key feature dimensions, calculate the mean of the deep feature vector set in that dimension, and calculate the difference between this mean and the historical mean of the corresponding historical feature in the historical feature distribution database. Then, divide the difference by the sum of the historical feature standard deviation of the corresponding dimension in the historical feature distribution database and a very small positive number to obtain the relative offset of that dimension. Then, sum the relative offsets of all dimensions in the list of key feature dimensions to obtain the second difference component. The "very small positive number" is used to prevent division by zero error and can usually be set to 1e-8. By calculating the relative offset (difference divided by historical standard deviation), the influence of the units and fluctuation range of different feature dimensions can be eliminated, making the offset of each dimension comparable. The second difference component obtained after summing can quantify the standardized average offset amplitude of the key feature dimension relative to its historical fluctuation, and is therefore specifically used to detect concentrated offsets occurring in the feature subspace that is crucial to classification decisions. A3. Multiply the first difference component by the first weight coefficient, and multiply the second difference component by the second weight coefficient. The sum of the first and second weight coefficients is always one. Finally, add the two products together. The result is the feature distribution difference. The specific values ​​of the first and second weight coefficients can be configured according to the sensitivity requirements for overall drift and local key drift. For example, in scenarios that emphasize the stability of the model decision boundary, the second weight coefficient can be set to be greater than the first weight coefficient (e.g., the first weight coefficient is set to 0.3 and the second weight coefficient is set to 0.7), so that the difference calculation is more inclined to respond to changes in key feature dimensions. This weighted summation method makes synergistic use of information on overall distribution difference and local key dimension difference, so that the final feature distribution difference index can not only perceive global changes, but also keenly capture targeted changes that may threaten model performance, thus improving the accuracy and relevance of drift detection. The specific process of generating drift signals and dynamically maintaining the historical feature distribution database is as follows: The calculated feature distribution difference is compared with a preset drift judgment threshold. The drift judgment threshold can be determined by the empirical distribution of the feature distribution difference during historical stable operation periods. For example, it can be set as the value obtained by adding three times the standard deviation of the historical difference average. As a specific example, the threshold can be set to 0.15. The setting of the threshold needs to achieve a balance between detection sensitivity and noise resistance. When the difference in feature distribution is greater than the drift determination threshold, a significant concept drift is determined to have occurred, and a structured drift signal is generated. The drift signal includes: a trigger flag indicating the occurrence of drift, the current value of the feature distribution difference, a key dimension offset vector, and the timestamp and data batch index of the generated signal. The key dimension offset vector is calculated based on the historical mean and standard deviation of each key feature dimension in the historical feature distribution database, and the feature mean of the current batch's depth feature vector set on the corresponding key feature dimension. It represents the direction and magnitude of change in each key feature dimension. The key dimension offset vector is a vector with the same length as the list of key feature dimensions, and each element is the relative offset calculated for the corresponding dimension. This vector provides fine-grained guidance on the direction of change for the subsequent sample selection module. For example, the sample selection module can prioritize selecting samples with abnormal feature values ​​on dimensions with the largest positive or negative offsets for labeling. The historical feature distribution library uses a Gaussian mixture model to model the probability distribution of historical deep feature vectors; the list of key feature dimensions is derived from the weight analysis of the fully connected layers of the sorting model and is recalculated synchronously as the sorting model is updated; this means that when the model update module completes an update of the sorting model, the feature monitoring module will immediately regenerate the list of key feature dimensions based on the new sorting model weights, thereby ensuring that the "key dimensions" used for monitoring are always consistent with the latest model decision logic, and realizing the synchronization of monitoring focus and model evolution; After the feature monitoring module generates the drift signal, its internal historical feature distribution database will still be updated independently based on feedback from the sample selection module. Specifically, after the sample selection module confirms the true label of the meat image through an online annotation mechanism, thus forming incremental sample pairs, the depth feature vector corresponding to the incremental sample pair is input into the update process of the historical feature distribution database. This is used to incrementally update the model parameters of the Gaussian mixture model. The model parameters of the Gaussian mixture model include the mixing weights, mean vector, and covariance matrix of each Gaussian component in the Gaussian mixture model. The trigger condition for incremental update is the acquisition of incremental sample pairs. The update goal is to enable the historical feature distribution database to track the correctly recognized depth feature vector distribution, thereby allowing the historical feature distribution database to slowly track the correctly recognized depth feature vector distribution and serve as a new baseline for subsequent calculation of feature distribution differences. Incremental update can be implemented using a variant of the online expectation-maximization algorithm. The key point is that the update of the historical feature distribution database is independent of the generation of drift signals. This design achieves "monitoring baseline". Decoupling from "drift decision": The monitoring baseline (historical feature distribution library) evolves gradually and smoothly as the system learns from newly confirmed correct samples, reflecting the system's accumulated "known knowledge" distribution. The drift decision compares the current data with this gradually evolving baseline. This avoids drift detection failure due to excessively rapid baseline updates, and also avoids errors in reporting all changes due to a stagnant baseline. A drift signal is triggered only when there is a sufficiently large deviation between the current data distribution and this gradually evolving "known knowledge" distribution. This mechanism enables the system to distinguish between normal, adaptive data changes and abnormal, significant drifts that require model readjustment, improving the long-term stability of the system. "Slow tracking" is achieved by setting a small learning rate or using only a small number of new samples to partially update the model parameters, thereby controlling the speed of baseline evolution.

[0022] In this embodiment, it is necessary to specifically explain the process of calculating the uncertainty measure after receiving the drift signal in the sample selection module: Upon receiving a drift signal, the system enters an active selection state. First, based on the numerical value of the feature distribution difference contained in the received drift signal, a fusion weight coefficient is dynamically set. The fusion weight coefficient is used to balance the contribution ratio of uncertainty measurement and representative distance in subsequent priority scoring calculations. Its initial value can be determined based on the numerical value of feature distribution difference through a linear or nonlinear mapping function. For example, when the feature distribution difference reaches the set drift judgment threshold, the initial value of the fusion weight coefficient is set to 0.9; when the feature distribution difference is significantly higher than the threshold, it can approach 1.0. This design ensures that in the early stage of drift, when the model is most "confused" by the changing data, the sampling strategy is highly inclined to select the sample with the most uncertainty in the model to quickly locate and correct cognitive blind spots. The initial value of the fusion weight coefficient is set to be directly proportional to the value of the feature distribution difference; that is, the larger the value of the feature distribution difference, the closer the initial value of the fusion weight coefficient is to the value of one. After the sample selection process begins, the fusion weight coefficient decreases over time or as the number of processed samples increases according to a preset decay function, with a decay target value of 0.5. The decay function can take the form of exponential decay, linear decay, or inverse proportional decay. For example, the fusion weight coefficient is set to decrease by 0.1 every 100 samples processed or every minute, until it is no less than 0.5. This dynamic adjustment mechanism simulates the human learning process: when encountering a new problem, the focus is first on the least understood part (high uncertainty), and as the number of observed samples increases, the focus gradually shifts to systematically exploring the whole picture of the new field (high representativeness). For each frame of meat image and its corresponding depth feature vector acquired after the drift signal is emitted, the sorting model is invoked for prediction. To calculate the uncertainty measure of the prediction result, a forward propagation mutation method based on random deactivation is adopted: during the inference process of the sorting model, its internal random deactivation layer is enabled, and forward propagation calculation is repeatedly performed multiple times for the same depth feature vector. Each forward propagation produces a potentially different prediction probability distribution due to the presence of the random deactivation layer. The specific number T of "multiple times" is a preset parameter, usually set between 10 and 50 times, for example, T=20. Sufficient repetitions can ensure sufficient sampling of the randomness of the model's prediction, thereby obtaining a stable uncertainty estimate. The core of this method is to measure the model's inherent cognitive uncertainty for the current input by evaluating the prediction divergence caused by the randomness of the model itself. This is more robust and reliable than simply relying on the Softmax confidence of a single prediction, and is especially effective in identifying "uncertain" samples located near the classification boundary. The uncertainty metric is calculated based on the degree of dispersion of multiple predicted probability distributions generated by multiple forward propagations, and is achieved by calculating the entropy value of the average distribution of multiple predicted probability distributions. The specific calculation process is as follows: For each category among all sorting categories defined by the sorting model, firstly, the arithmetic mean of the corresponding probabilities of that category in multiple predicted probability distributions generated by multiple forward propagations is calculated to obtain the probability of that category in the average distribution; then, based on the calculated average probability of all categories, the entropy of the average probability distribution is calculated. The larger the entropy value, the more inconsistent the prediction results are, and the higher the uncertainty measure value; the formula for calculating the entropy is the standard formula for information entropy; the obtained uncertainty measure is a non-negative scalar, and the higher its value, the less confidence the model has in predicting the input sample. The sample may contain new features that the model has not fully learned or is located near the category decision boundary, so the value of labeling and learning it is higher; The specific process for calculating the representative distance is as follows: For a deep feature vector, its representative distance is calculated. This representative distance is used to measure the degree of deviation of the deep feature vector from the historical feature distribution represented by the historical feature distribution library. In particular, it is combined with the direction indicated by the key dimension offset vector in the drift signal for biased measurement. The representative distance aims to screen out the samples that best represent the "newly emerging" or "not yet mastered" data patterns from the feature distribution level. That is, those samples that are far away from the known historical distribution and may be at the forefront of the new distribution. The calculation of the representative distance consists of two parts: the first part is to calculate the overall Mahalanobis distance from the deep feature vector to the historical feature distribution database. The overall Mahalanobis distance is calculated as follows: first, obtain the global mean vector and covariance matrix of the distribution represented by the historical feature distribution database; then, calculate the difference vector between the deep feature vector and the global mean vector; finally, by transposing the difference vector, multiplying it by the inverse of the covariance matrix, and then multiplying it by the difference vector itself, the square root of the product is obtained to obtain the overall Mahalanobis distance. The Mahalanobis distance is different from the Euclidean distance. It considers the correlation between the dimensions of the features (through the inverse of the covariance matrix) and the variance of each dimension itself. This makes the calculated distance more accurately reflect the degree to which the sample deviates from the historical overall distribution in a statistical sense, eliminates the interference caused by different feature scales and correlations, and can more effectively identify "statistically abnormal" samples. The second part calculates the difference vector between the deep feature vector and the global mean vector of the historical feature distribution library, and the projected distance scalar value in the direction of the key dimension offset vector carried by the drift signal. The projected distance scalar value is obtained by calculating the dot product of the difference vector and the key dimension offset vector, and then dividing by the Euclidean norm of the key dimension offset vector. The final representative distance is the product of the overall Mahalanobis distance and a first distance weight, plus the product of the projected distance scalar value and a second distance weight. This projected distance scalar value is directional, and its sign (positive or negative) indicates whether the sample deviates positively or negatively in the key offset direction, and its absolute value indicates the magnitude of the deviation. The introduction of this directional metric allows the system to purposefully filter those samples that "go further" in the observed drift main direction, thereby more efficiently exploring and characterizing the frontier boundary of the new data distribution. In this system, the sum of the first distance weight and the second distance weight is always equal to one. The value of the second distance weight is determined as follows: the Euclidean norm of the received key dimension offset vector is calculated and normalized. The normalized result is used as the value of the second distance weight. This ensures that when the drift in a specific direction is significant, the calculation of the representative distance focuses more on the leading edge of the sample in that direction. The Euclidean norm of the key dimension offset vector reflects the severity of the overall drift in each key dimension. The larger the norm, the stronger and more concentrated the detected drift directionality. By normalizing it and using it as the second distance weight, the representative distance metric is adaptively adjusted: when the drift direction is ambiguous, the general Mahalanobis distance is relied upon more; when there is a clear and strong directional drift, the projected distance of the sample in that specific direction is given more attention, realizing an intelligent and focused sample exploration strategy. The specific process of calculating priority scores, forming incremental sample pairs, and storing them in the incremental sample pool is as follows: For each frame of meat image, after obtaining the uncertainty measure of its prediction result and the representative distance of its corresponding depth feature vector, both are normalized to obtain the normalized uncertainty value and the normalized representative distance value. The normalization process usually uses the uncertainty measure and representative distance value of all candidate samples in the current batch, and performs min-max normalization or Z-score normalization respectively to eliminate the difference in the units and absolute value range of the two, so that they are in the [0,1] interval or have zero mean and unit variance, which facilitates fair weighted fusion. Based on the currently dynamically set fusion weight coefficients, the normalized uncertainty value and the normalized representative distance value are nonlinearly fused to obtain the priority score of the meat product image. Nonlinear fusion computation employs a weighted power average approach. The parameters involved in the calculation include: a normalized uncertainty value, a normalized representativeness distance value, dynamically set fusion weight coefficients, and a preset power parameter greater than zero. This power parameter is a crucial hyperparameter, its value determining the fusion's "bias." When the power parameter equals 1, the weighted power average degenerates into a simple linear weighted sum. When the power parameter is greater than 1 (e.g., equal to 2), the calculation tends to increase the weight of larger components, prioritizing samples where at least one aspect of uncertainty or representativeness is particularly prominent. When the power parameter is between 0 and 1, the calculation tends to balance both, achieving a good overall score even if one value is low while the other is high. Typically, a power parameter of 2 can be set to prioritize capturing extreme samples with extremely uncertain models or highly novel features. The weighted power average calculation process is as follows: First, the normalized uncertainty value is raised to the power parameter, and the result is multiplied by the current fusion weight coefficient to obtain the first term. Second, the normalized representative distance value is raised to the power parameter, and the result is multiplied by the difference between one and the current fusion weight coefficient to obtain the second term. Then, the first and second terms are added to obtain the sum. Finally, the sum is raised to the power parameter, and the result is the priority score. This calculation process ensures that the priority score comprehensively reflects both uncertainty (model cognition level) and representativeness (data distribution level), and introduces non-linear selection preference through the power average, making the scoring mechanism more flexible and powerful. The sample selection module maintains a fixed-capacity priority queue for temporary storage of meat images, corresponding depth feature vectors, and calculated priority scores. The fixed capacity can be set according to the number of incremental samples expected to be obtained, for example, 2-5 times the planned number of labeled samples, to ensure that there are sufficient candidate samples in the queue for selection. The priority score of the current meat product image is compared with the scores of existing samples in the priority queue. The queue content is updated according to the score order to ensure that the priority queue always contains the batch of samples with the highest scores since entering the active selection state. The update strategy is usually as follows: if the queue is not full, it is directly inserted; if the queue is full and the current sample score is higher than the lowest score in the queue, the current sample replaces the sample with the lowest score. This ensures that the queue is dynamically maintained in the "historically best" state. When the preset stopping condition is met, the highest-scoring samples are selected from the priority queue as designated meat images and sent to the online annotation mechanism to obtain real labels, forming incremental sample pairs consisting of designated meat images and their corresponding real labels, and finally stored in the incremental sample pool. The "preset stopping condition" can be reaching a preset time window (e.g., active selection lasting 5 minutes), or having accumulated a preset number of total samples (e.g., processing 500 frames of images), or the lowest score of the samples in the priority queue exceeding a certain stable threshold. After the condition is met, the top N samples with the highest scores (N is the number of samples to be incrementally updated in this plan, e.g., 50 samples) are selected from the queue for subsequent annotation and storage.

[0023] In this embodiment, it is necessary to specifically explain the process of incrementally training the sorting model at a preset time in the model update module: The system continuously monitors the number of incremental sample pairs accumulated in the incremental sample pool. When the number of incremental sample pairs reaches a preset first threshold, or when the system enters a preset fixed time interval, the sorting model is triggered to update. The first threshold can be set according to computing resources and model convergence speed, for example, to 50, 100, or 200. The principle for setting it is to accumulate enough samples for an effective gradient descent iteration, while avoiding excessively long training times that could affect system response. The preset fixed time interval can be a natural break in the production line, such as during hourly shift changes or daily planned maintenance periods, thus utilizing non-production time for computation and avoiding competition for computing resources with online sorting tasks. This dual-trigger mechanism ensures the timeliness of model updates and adaptability to production rhythm. Once triggered, the latest incremental sample pairs, whose number is equal to the first threshold, are retrieved from the incremental sample pool to form the new training dataset used for this incremental training. Simultaneously, all sorting model parameters of the current sorting model are read from the internal memory, along with the importance weight value of each sorting model parameter to the historical task. This importance weight value constitutes the historical importance weight vector. The historical importance weight vector is usually obtained and initialized by calculating the diagonal elements of the Fisher information matrix based on the initial training dataset after the initial training of the sorting model. It quantifies the importance of each model parameter to completing the initial sorting task. The higher the importance weight value of a parameter, the less constrained it should be to change in subsequent incremental training to prevent catastrophic forgetting. When calculating the dynamic importance weights used to constrain the changes in the parameters of each sorting model during this incremental training, three components are calculated for each sorting model parameter: the first component is the importance weight value of the stored historical tasks; this component is the core of the elastic weight consolidation algorithm, used to retain the memory of historical knowledge; The second component is the moving average of the absolute gradient values ​​of the sorted model parameters during the most recent incremental training. This moving average is mapped to a value between zero and one using a sigmoid function (logistic function), whose input is the negative of the product of the moving average of the absolute gradient values ​​and a preset positive scaling factor. The introduction of this component is innovative. The moving average of the absolute gradient values ​​reflects the activity level of the parameter in recent learning; a large value indicates that the parameter changes drastically when adapting to new knowledge. Mapping it to a decay factor between 0 and 1 (the larger the gradient, the smaller the decay factor) dynamically reduces the importance constraint of such "active" parameters in the current training, effectively "unsealing" some key parameter space for the model to quickly learn new knowledge, thus achieving a more refined balance between preventing forgetting and promoting adaptation. The preset positive scaling factor controls the sensitivity of the gradient value to the decay factor. The third component is a binary indicator function. When the sorting model parameter is a weight parameter in the fully connected layer of the sorting model that is directly connected to a dimension in the list of key feature dimensions provided by the feature monitoring module, the indicator function takes a value of one; otherwise, it takes a value of zero. This component enables deep collaboration with the upstream feature monitoring module. The list of key feature dimensions identifies features crucial to the current model's decision. The weight parameters of the fully connected layer directly connected to these dimensions directly determine how these key features affect the classification results. In incremental training, imposing stronger importance constraints on such parameters (by adding this component) can particularly protect the model's ability to utilize core discriminative features, ensuring that the model's fundamental and validated decision-making logic remains intact when adapting to new changes. This improves the stability and relevance of incremental learning. The dynamic importance weights of the sorting model parameters are obtained by multiplying each of the three components by three preset weighting coefficients and then summing them. Each of the three weighting coefficients is a real number between zero and one, and their sum is one. The setting of the three weighting coefficients (e.g., 0.5, 0.3, 0.2) determines the strategy preference for calculating the dynamic importance weights. By adjusting them, the system can be controlled to rely more on traditional historical importance (higher first coefficient), to focus more on loosening the constraints on recently active parameters (higher second coefficient), or to emphasize protecting the association weights of key features (higher third coefficient). This configurability allows the system to adapt to different incremental learning scenarios. The dynamic importance weights of each sorting model parameter are calculated in this way, forming a dynamic importance weight vector. The final dynamic importance weight vector integrates information from three dimensions: static historical importance, dynamic learning activity, and correlation of key task features, providing a more refined and adaptive parameter constraint blueprint for this incremental training than traditional methods. The specific process of updating the parameters of the sorting model is as follows: The total loss function for this incremental training is constructed as follows: First, the standard cross-entropy classification loss on the new training dataset is calculated as the first term; this term drives the sorting model to learn the knowledge in the new samples. Then, for each sorting model parameter, the dynamic importance weight of that sorting model parameter is multiplied by the square of the difference between the initial value of the sorting model parameter before training and the current value during training to obtain the regularization contribution value of that sorting model parameter. The regularization contribution values ​​of all sorting model parameters are summed, and the sum is multiplied by half of a preset regularization strength coefficient as the second term; this term is the elastic weight consolidation regularization term; its core function is to penalize excessive modification of important parameters (high dynamic importance weights); the difference square term measures the magnitude of parameter change; the regularization strength coefficient is used to control the overall balance between "memory retention" and "new knowledge learning", and is a key hyperparameter, which usually needs to be tuned through the validation set, for example, set between 1 and 1000; Finally, the first and second terms are added together to obtain the total loss function; the total loss function unifies the goal of learning new tasks and the constraint of protecting old knowledge within an optimization framework. When optimizing the total loss function, a hierarchical differentiated learning rate strategy is adopted: the network layers of the sorting model are divided into a low-level feature encoding layer, a high-level semantic feature layer, and a classification layer. The low-level feature encoding layer includes the first few layers of the convolutional neural network, which is used to extract general visual features. The high-level semantic feature layer includes the later layers of the convolutional neural network, which is used to extract task-related features. The classification layer is a fully connected layer. This division is based on the knowledge representation characteristics of deep learning models: the low-level network usually learns general and transferable visual features (such as edges and textures), which are relatively stable across different tasks; while the high-level network and classification layer learn feature combinations and decision boundaries that are closely related to specific tasks, requiring greater flexibility to adapt to changes. Set a low learning rate for the bottom feature encoding layer. This low learning rate is equal to the product of a preset base learning rate and a preset multiplier coefficient less than one. For example, if the base learning rate is 0.001 and the bottom multiplier coefficient is 0.1, then the bottom learning rate is 0.0001. A small learning rate allows the bottom general features to be slowly and finely adjusted, avoiding the loss of their universality due to adaptation to new data. Set a high learning rate for the high-level semantic feature layer and classification layer. This high learning rate is equal to the result of multiplying the base learning rate by a preset multiplier coefficient greater than or equal to one. For example, if the high-level multiplier coefficient is 1.0 or 2.0, the corresponding learning rate is 0.001 or 0.002. A larger learning rate allows the model to quickly adjust its high-level representation and decision logic to efficiently absorb new patterns and knowledge contained in new samples. Using stochastic gradient descent or adaptive moment estimation, the model is trained iteratively on a new training dataset with a differentiated learning rate, targeting the total loss function, and continuously updating the parameters of the sorting model. The training process stops after a preset number of iterations (epochs) or stops early when the loss on the validation set (which can be a small portion of the new training dataset) no longer decreases, in order to prevent overfitting. After training, the Fisher information content of each sorting model parameter is recalculated using the new training dataset to obtain new importance weight values. Then, the stored historical importance weight vector and the calculated new importance weight vector are weighted and fused according to a preset forgetting factor. The specific calculation process of weighted fusion is as follows: for each sorting model parameter, its corresponding old importance weight value stored in the historical importance weight vector is multiplied by the forgetting factor to obtain the first product. Multiply the corresponding, calculated new importance weight value by the difference obtained by subtracting the forgetting factor to get the second product; Adding the first and second products yields the updated importance weight values ​​for the sorting model parameters, resulting in the updated importance weight vector, which is then stored. This step enables online updating of the importance weight vector. The forgetting factor (e.g., set to 0.8) controls the degree to which historical importance memories are retained. The closer the factor is to 1, the more historical memories are retained, and the more conservative the model becomes. The smaller the factor, the faster the importance information of new tasks is incorporated, and the more flexible the model becomes. Through this smooth update, the importance weight vector can dynamically evolve along with the tasks learned by the model, providing more accurate prior knowledge for the next incremental training. This is the key mechanism for achieving long-term continuous learning without forgetting. Simultaneously, the updated sorting model parameters are sent back to the feature monitoring module, enabling the feature monitoring module to recalculate the list of key feature dimensions based on the updated parameters, thus keeping the monitoring focus synchronized with the updated sorting model state. This feedback loop is crucial; after the sorting model is updated, its decision-making basis (reflected in the weights of the fully connected layers) has changed, therefore the criteria for judging which feature dimensions are "critical" also change accordingly. Feeding the new model parameters back to the feature monitoring module allows it to immediately reanalyze and generate a new list of key feature dimensions based on the latest sorting model, ensuring the consistency of the entire system's "monitoring-learning" cycle; otherwise, the monitoring module may still be using outdated standards to detect drift, leading to detection failure or misleading subsequent sample selection.

[0024] In this embodiment, the specific process of generating the corresponding synthetic training data in the digital twin module is as follows: After receiving the drift signal from the feature monitoring module, the drift signal is parsed to obtain the value of the feature distribution difference degree and the key dimension offset vector contained therein. The value of the feature distribution difference degree directly reflects the degree to which the current data distribution deviates from the historical baseline, while the key dimension offset vector indicates the specific direction and magnitude of change in the feature dimensions that are most important in the model decision. A distribution inference intensity coefficient is calculated based on the numerical value of the feature distribution difference. The specific calculation method of the distribution inference intensity coefficient is as follows: the numerical value of the feature distribution difference is divided by a preset drift judgment threshold, and the quotient is used as the input of the hyperbolic tangent function. Then, the output value of the hyperbolic tangent function is multiplied by a preset upper limit coefficient. The final result is the distribution inference intensity coefficient. This calculation ensures that the distribution inference intensity coefficient is positively correlated with the numerical value of the feature distribution difference, and its maximum value does not exceed the upper limit coefficient. The drift judgment threshold is the same as the threshold used in the feature monitoring module, for example, 0.15. The upper limit coefficient is used to limit the maximum inference intensity to prevent the generation of overly bizarre and unrealistic images. For example, it can be set to 0.5. The hyperbolic tangent function (tanh) can map the input to the (0,1) interval to achieve smooth nonlinear saturation, so that when the feature distribution difference is much greater than the threshold, the inference intensity tends to stabilize at the upper limit, avoiding unlimited amplification. The parameterized meat appearance simulator defines multiple simulator parameters that control the generation of meat appearance. Each simulator parameter follows a preset prior probability distribution. The simulator parameters can include a series of low-dimensional controllable variables such as color (e.g., mean and variance of HSV space), texture (e.g., density, direction, and contrast of marble pattern), shape and contour (e.g., aspect ratio and curvature), and surface defects (e.g., type, size, and density). Its prior probability distribution is usually obtained based on the analysis of a large number of normal meat images, for example, it is set as a multivariate Gaussian distribution. Adjust the parameter distribution of the simulator and construct a transfer mapping for all simulator parameters based on the inferred intensity coefficient and the key dimension offset vector. The mathematical essence of this transfer mapping is to perform a translation transformation in the probability distribution space of the simulator parameters. The direction and magnitude of the translation are determined by the drift signal. This transport mapping acts on the original prior distribution of the simulator parameters, transforming it into a new, adjusted distribution of the meat appearance simulator parameters; The specific construction process of the transport mapping is as follows: First, based on the historical sampling data of the simulator parameters, the global mean vector and covariance matrix of its original prior distribution are estimated, and the square root matrix of the inverse of the covariance matrix is ​​calculated. This is used to normalize the simulator parameters to eliminate the influence of dimensions. Calculating the square root matrix of the inverse of the covariance matrix (i.e., the square root of the precision matrix) is a key step in the Mahalanobis distance calculation. Its function is equivalent to "whitening" the parameter space, eliminating the correlation between different parameters and normalizing the variance of each dimension, thereby ensuring that the subsequent superposition of direction vectors is carried out in an isotropic standard space, making the inference more reasonable and balanced. Secondly, by using a pre-defined or learnable mapping matrix, the key dimension offset vector is linearly transformed to obtain the expected adjustment direction vector for the simulator parameter space. The mapping matrix establishes a bridge between the high-dimensional deep feature space (the space where the key dimension offset vector is located) and the low-dimensional interpretable simulator parameter space. It can be pre-defined based on domain knowledge (for example, it is known that the positive offset of the "intramuscular fat reflectivity" feature dimension should mainly correspond to the increase of the "texture contrast" parameter), or it can be initialized as a random matrix and learned from the data through subsequent online fine-tuning mechanisms. The calculation process for the transfer mapping applied to a specific simulator parameter value is defined as follows: For each simulator parameter, its value in its original prior distribution is added to the product of the following two terms: the first term is the calculated distribution projection strength coefficient, and the second term is the component in the expected adjustment direction vector corresponding to the simulator parameter after normalization by the square root matrix of the inverse of the covariance matrix. This calculation clearly defines how to transform a parameter value sampled from the original prior distribution into a parameter value that conforms to the expectations of the new distribution by superimposing a controlled offset; the offset is jointly determined by the projection strength (control amplitude) and the normalized adjustment direction (control direction). By applying this transfer mapping to the original prior distribution of simulator parameters, an adjusted simulator parameter distribution is obtained. This adjusted distribution tends to generate meat appearance images that conform to the changing trends indicated by the drift signal in the feature space. The final adjusted distribution has a systematic shift in its mean vector relative to the original distribution, and the shift direction is associated with the drift direction detected by the feature monitoring module through a mapping matrix. Sampling and generating images from this distribution can systematically produce a batch of synthetic samples that "look" like the drift trend that is occurring, providing the model with forward-looking and targeted reinforcement training materials. The specific process of using synthetic training data to enhance the updated sorting model is as follows: Sample from the adjusted simulator parameter distribution to obtain multiple sets of specific simulator parameter values; During sampling, there is an intentional bias towards selecting simulator parameter values ​​corresponding to the tail regions of the adjusted distribution, that is, selecting parameter combinations with lower probabilities and more extreme occurrences to generate more challenging synthetic meat images that cover potential edge cases. Specific sampling strategies can be importance sampling, or sampling from the distribution in descending order of probability and retaining only samples with probabilities below a certain quantile (e.g., the 20th percentile). This ensures that the generated synthetic data are not "common" samples in the new distribution, but rather "rare" or "difficult" samples, thereby more effectively widening the model's decision boundary and improving its robustness to extreme cases in the new distribution. The parametric meat appearance simulator is an image generation device that takes simulator parameter values ​​as conditional input. Its implementation methods include physically based rendering engines or deep generative models that take simulator parameters as conditional vectors. Physically based rendering engines can provide extremely high fidelity and clear physical interpretation, but have high computational cost. On the other hand, deep generative models, represented by conditional generative adversarial networks or conditional variational autoencoders, can achieve fast and diverse image synthesis after learning a sufficient number of samples. The sampled simulator parameter values ​​are input into the parameterized meat appearance simulator, which generates corresponding synthetic meat images and their fully known pseudo-labels. The pseudo-labels include one or more of the following: meat part, grade, and defect type. Together, they constitute the synthetic training dataset. Using the synthetic training dataset, the sorting model, which has already been updated by the model update module, is subjected to reinforcement training. This step occurs after the model update module has completed the main adaptation using a small number of real incremental samples. It is a kind of "icing on the cake" reinforcement training. Its purpose is to consolidate and generalize the new knowledge learned on the small real samples. The construction process of the total loss function used in reinforcement training is as follows: First, the standard classification loss of the synthetic meat image is calculated. This standard classification loss is the first loss calculated by the cross-entropy loss function based on the pseudo-label of the synthetic meat image and the prediction result of the sorting model. This ensures that the sorting model can also make correct classifications on synthetic data. Secondly, the consistency regularization loss is calculated as follows: two different data augmentation transformations are independently applied to the same synthetic meat image to obtain a first augmented view and a second augmented view. The first and second augmented views are then input into the sorting model to obtain two predicted probability distributions. The Jensen-Shannon divergence between these two predicted probability distributions is then calculated, and the divergence value is used as the second loss term, i.e., the consistency regularization loss. Consistency regularization is an effective self-supervised learning technique. The Jensen-Shannon divergence is a symmetrical measure of the difference between two probability distributions. Minimizing this loss means encouraging the sorting model to produce consistent and stable predictions for different augmented views of the same synthetic sample. This forces the model to learn features that are not robust to subtle changes in color and texture, and to focus on more essential semantic features, significantly improving the model's generalization ability and robustness on unseen real samples with similar distribution shifts. Finally, the results of multiplying the first loss and the second loss by a preset regularization weight coefficient are added together, and the sum is the total loss function used for reinforcement training. The regularization weight coefficient is used to balance classification accuracy and consistency constraints, and is usually set to a positive number less than 1, such as 0.5 or 1.0, and is determined through cross-validation. During training, a curriculum learning strategy can be adopted. Initially, synthetic data based on the adjusted simulator parameter distribution, generated with smaller distribution inference intensity coefficients, is used. Later, the distribution inference intensity coefficients are gradually increased, and synthetic data with larger distribution inference intensity coefficients, deviating more from the original prior distribution, is used for training. The curriculum learning strategy simulates a learning process from easy to difficult. Initially, synthetic data with low inference intensity (closer to the current adapted state) is used to help the model stabilize. Later, data with high inference intensity (closer to the predicted future drift front) is introduced to challenge and expand the model's capability boundaries. This can make the training process smoother and avoid model instability due to overly difficult initial data. Furthermore, the digital twin module includes an online fine-tuning mechanism: when the sample selection module subsequently provides new incremental sample pairs labeled with high confidence, the depth feature vectors of the meat images corresponding to these incremental sample pairs are used to deduce the ideal simulator parameters that can generate similar feature images through optimization methods. This deduction result, along with the latest drift signal sent by the feature monitoring module, is then used to fine-tune and update the mapping matrix. This mechanism enables continuous optimization of the digital twin model. Real incremental samples are valuable "ground realities" used to correct the mapping relationship between simulator parameters and feature space drift. The process of back-deriving the ideal simulator parameters through optimization is as follows: The simulator parameters are iteratively optimized to minimize the mean square error between the depth feature vector of the meat image generated by the parameterized meat appearance simulator and the depth feature vector of the meat image corresponding to the incremental sample pair. The optimized simulator parameter values, when input to the parameterized meat appearance simulator, result in depth features of the generated image that are closest to the depth features of the meat image in the incremental sample pair. The final parameter values ​​obtained are the ideal simulator parameters. This optimization process typically uses gradient descent, leveraging the differentiability of the generator to update the simulator parameters through backpropagation of feature reconstruction loss. This is equivalent to finding simulator parameters that allow the generator to "best mimic" the performance of a given real sample in the feature space. The process of fine-tuning the mapping matrix is ​​as follows: The key dimension offset vectors in the ideal simulator parameters and drift signals, along with the square root of the inverse of the covariance matrix of the estimated prior distribution of the simulator parameters, are used as inputs. Gradient descent is used to adjust the element values ​​in the mapping matrix. The goal of this adjustment is to ensure that the expected adjustment direction vector calculated by the adjusted mapping matrix, when applied to the simulator parameters sampled from the prior distribution of the simulator parameters, more accurately approximates the difference vector between the ideal simulator parameters and the global mean vector of the original prior distribution. This makes the mapping relationship closer to the real physical correlation, thereby improving the quality and relevance of subsequent synthetic data generation. This fine-tuning process defines a supervised learning task: learning a mapping function (i.e., the mapping matrix) such that the input key dimension offset vector can predict the standardized offset of the simulator parameters required to match real samples. By continuously fine-tuning with new real sample-drift signal pairs, the mapping matrix M can increasingly accurately capture the complex nonlinear relationship from feature drift to changes in appearance parameters, making the "deduction" capability of the digital twin module stronger and the simulation of real distribution changes in the generated synthetic data increasingly realistic.

[0025] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0026] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0027] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0028] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0029] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0030] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0031] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A meat sorting and preprocessing system based on AI vision, characterized in that, Specifically, it includes: The module is sequentially connected to a feature monitoring module, a sample selection module, a model update module, and a digital twin module, among which; Feature monitoring module: acquires images of meat products flowing on the production line in real time, extracts the depth feature vector of the meat product images, compares the depth feature vector with a dynamically maintained historical feature distribution database, and calculates the feature distribution difference degree; When the feature distribution difference exceeds a preset threshold, a drift signal containing the current feature distribution difference and feature distribution change trend data is generated and sent. Sample selection module: Receives drift signals, calls a pre-built sorting model to predict the meat images subsequently acquired by the feature monitoring module, and jointly calculates the uncertainty measure of the prediction results and the representative distance of the depth feature vectors corresponding to the subsequent meat images in the feature space. Priority scores are calculated based on the uncertainty measure and the representative distance. Specific meat images are selected according to their priority scores, and the real labels of these images are obtained through an online annotation mechanism, forming incremental sample pairs consisting of the specified meat images and their corresponding real labels. These incremental sample pairs are then stored in an incremental sample pool. Model update module: At a preset time, incremental sample pairs are obtained from the incremental sample pool, and the sorting model is incrementally trained based on the elastic weight consolidation algorithm; wherein, the parameters of the sorting model are constrained based on the weights of historical task importance, and the parameters of different network layers of the sorting model are updated using differentiated learning rates; Digital Twin Module: Includes a parameterized meat appearance simulator, used to receive drift signals, dynamically adjust the parameter distribution of the meat appearance simulator according to the feature distribution differences contained in the drift signals to generate corresponding synthetic training data; and use the synthetic training data to reinforce the sorting model updated by the model update module.

2. The meat sorting and preprocessing system based on AI vision according to claim 1, characterized in that: The specific process for calculating the feature distribution difference in the feature monitoring module is as follows: The depth feature vector of meat images is extracted using the built-in coding network of the feature monitoring module; the historical feature distribution library is modeled using a Gaussian mixture model to model the distribution of historical depth feature vectors. A list of key feature dimensions is constructed by identifying the dimensions with the largest absolute weights from the weights of the fully connected layers of the pre-built sorting model. When calculating the feature distribution difference, firstly, multiple depth feature vectors corresponding to a predetermined number of meat product images in the current batch are obtained. These multiple depth feature vectors form the depth feature vector set for the current batch, and the following operations are performed sequentially: A1, calculate the Wasserstein distance between the current deep feature vector distribution represented by the deep feature vector set and the historical deep feature vector distribution represented by the historical feature distribution library, as the first difference component; A2. For each dimension in the key feature dimension list, calculate the feature mean of the deep feature vector set in that dimension, and calculate the difference between the feature mean and the historical feature mean corresponding to that dimension in the historical feature distribution library. Then divide the difference by the sum of the historical feature standard deviation corresponding to that dimension in the historical feature distribution library and a very small positive number to obtain the relative offset of that dimension. Then sum the relative offsets of all dimensions in the key feature dimension list to obtain the second difference component. A3. Multiply the first difference component by the first weight coefficient, and multiply the second difference component by the second weight coefficient. The sum of the first and second weight coefficients is always one. Finally, add the two products together. The result is the characteristic distribution difference degree.

3. The meat sorting and preprocessing system based on AI vision according to claim 2, characterized in that: The specific process of generating drift signals and dynamically maintaining the historical feature distribution database is as follows: The feature distribution difference is compared with a preset drift judgment threshold. If it exceeds the threshold, concept drift is determined to have occurred, and a structured drift signal is generated. The drift signal includes a trigger flag, the value of the feature distribution difference, the key dimension offset vector, the timestamp, and the batch index. The key dimension offset vector is calculated based on the historical mean and historical standard deviation of each key dimension in the historical feature distribution library, and the feature mean of the current batch depth feature vector set on the corresponding key feature dimension, representing the direction and magnitude of change of each key feature dimension; The historical feature distribution library uses a Gaussian mixture model to model the probability distribution of historical deep feature vectors; The list of key feature dimensions is derived from the weight analysis of the fully connected layers of the sorting model and is recalculated synchronously as the sorting model is updated. After the feature monitoring module generates the drift signal, its historical feature distribution library is independently updated based on feedback from the sample selection module. Specifically, when the sample selection module confirms the true label of the meat image through the online annotation mechanism and forms an incremental sample pair, the depth feature vector corresponding to the incremental sample pair is input into the historical feature distribution library to incrementally update the model parameters of the Gaussian mixture model. The model parameters of the Gaussian mixture model include the mixing weights, mean vector, and covariance matrix of each Gaussian component in the Gaussian mixture model, so that the historical feature distribution library tracks the distribution of the depth feature vector that has been correctly recognized, serving as a new baseline for subsequent calculations.

4. The meat sorting and preprocessing system based on AI vision according to claim 3, characterized in that: In the sample selection module, the specific process of calculating the uncertainty measure after receiving the drift signal is as follows: A fusion weight coefficient is dynamically set based on the numerical value of the difference in the feature distribution in the drift signal. The initial value of the fusion weight coefficient is set to be directly proportional to the numerical value of the difference in the feature distribution, and it decreases according to a preset attenuation function after the selection begins. The attenuation target value is 0.

5. For each frame of meat image and its corresponding depth feature vector acquired after the drift signal is emitted, the sorting model is invoked for prediction. The uncertainty measure is calculated using a forward propagation mutation method based on random deactivation: a random deactivation layer is enabled in the sorting model inference, and forward propagation is repeatedly performed on the same depth feature vector to generate multiple prediction probability distributions. The entropy value of the average distribution of these multiple prediction probability distributions is calculated as the uncertainty measure. The larger the entropy value, the more inconsistent the prediction results are and the higher the uncertainty.

5. The meat sorting and preprocessing system based on AI vision according to claim 4, characterized in that: The specific process for calculating the representative distance is as follows: For a depth feature vector, its representative distance is calculated. This representative distance is used to measure the degree of deviation of the depth feature vector from the historical feature distribution represented by the historical feature distribution library, and a biased measure is specifically combined with the direction indicated by the key dimension offset vector in the drift signal. The calculation of the representative distance consists of two parts: the first part is to calculate the global Mahalanobis distance from the depth feature vector to the historical feature distribution database. The global Mahalanobis distance is calculated as follows: first, obtain the global mean vector and covariance matrix of the distribution represented by the historical feature distribution database; then, calculate the difference vector between the depth feature vector and the global mean vector; finally, by transposing the difference vector, multiplying it by the inverse of the covariance matrix, and then multiplying it by the difference vector itself, the square root of the product is taken to obtain the global Mahalanobis distance. The second part is to calculate the difference vector between the deep feature vector and the global mean vector of the historical feature distribution library, and the scalar value of the projected distance in the direction of the key dimension offset vector. The scalar value of the projected distance is obtained by calculating the dot product of the difference vector and the key dimension offset vector, and then dividing by the Euclidean norm of the key dimension offset vector. The final representative distance is the product of the overall Mahalanobis distance and a first distance weight, plus the product of the projected distance scalar value and a second distance weight. The sum of the first distance weight and the second distance weight is one. The value of the second distance weight is calculated by normalizing the Euclidean norm of the received key dimension offset vector. The normalization result is used as the value of the second distance weight, so that when the drift in a specific direction is significant, the calculation of the representative distance focuses more on the frontier of the sample in that direction.

6. The meat sorting and preprocessing system based on AI vision according to claim 5, characterized in that: The specific process of calculating priority scores, forming incremental sample pairs, and storing them in the incremental sample pool is as follows: For each frame of meat product image, its uncertainty measure and representative distance are normalized respectively to obtain the normalized uncertainty value and the normalized representative distance value; Based on the current fusion weight coefficients, the normalized uncertainty value and the normalized representative distance value are nonlinearly fused to obtain the priority score of the meat product image. Nonlinear fusion calculation uses a weighted power average method, which involves a preset power parameter greater than zero; The calculation process of the weighted power average is as follows: First, the normalized uncertainty value is raised to the power parameter, and the result is multiplied by the current fusion weight coefficient to obtain the first term; Second, the normalized representative distance value is raised to the power parameter, and the result is multiplied by the difference obtained by subtracting the current fusion weight coefficient to obtain the second term. Then, add the first and second terms to get the sum; finally, raise the sum to the power of the reciprocal of the power parameter, and the result is the priority score. Internally, a fixed-capacity priority queue is maintained to temporarily store meat images, corresponding depth feature vectors, and calculated priority scores. The queue is updated according to the score order to ensure that the queue stores the sample with the highest score. When the preset stopping condition is met, select the highest-scoring samples from the priority queue as the specified meat product images, obtain their real labels through the online annotation mechanism, form incremental sample pairs composed of the corresponding real labels, and store them in the incremental sample pool.

7. The meat sorting and preprocessing system based on AI vision according to claim 6, characterized in that: In the model update module, the specific process of incrementally training the sorting model at preset times is as follows: When the number of sample pairs in the incremental sample pool reaches the first quantity threshold or enters a preset fixed time interval, the sorting model is updated. A new training dataset is formed by obtaining the latest incremental sample pairs from the incremental sample pool in a number equal to the first quantity threshold. Simultaneously, the current sorting model parameters and their historical importance weight values ​​are read from the internal memory, and the importance weight values ​​constitute the historical importance weight vector; Calculate a dynamic importance weight for each model parameter. This dynamic importance weight is obtained by weighted summation of three components: the first component is the stored historical importance weight value. The second component is the decay coefficient between zero and one obtained by mapping the absolute value of the gradient in the most recent incremental training of the sorting model parameters through the logistic function; The third component is a binary indicator function. It takes one value when the sorting model parameter is a weight parameter in the fully connected layer of the sorting model that is directly connected to the dimension in the list of key feature dimensions, and zero value otherwise. The sum of the three weighting coefficients is one, which forms the dynamic importance weight vector.

8. The meat sorting and preprocessing system based on AI vision according to claim 7, characterized in that: The specific process for updating the parameters of the sorting model is as follows: The total loss function for this incremental training is constructed as follows: First, the standard cross-entropy classification loss on the new training dataset is calculated as the first term; Then, for each sorting model parameter, the dynamic importance weight of the sorting model parameter is multiplied by the square of the difference between the initial value of the sorting model parameter before training and the current value during training to obtain the regularization contribution value of the sorting model parameter. The regularization contribution values ​​of all sorting model parameters are summed, and the summation result is multiplied by half of a preset regularization strength coefficient as the second term. Finally, add the first term to the second term to obtain the total loss function; A layered, differentiated learning rate strategy is adopted: the network layers of the sorting model are divided into a low-level feature encoding layer, a high-level semantic feature layer, and a classification layer; a lower learning rate is set for the low-level feature encoding layer. Set a high learning rate for the high-level semantic feature layer and the classification layer; Using the stochastic gradient descent algorithm, with the total loss function as the objective, the model is trained iteratively on a new training dataset with differentiated learning rates to update the parameters of the sorting model. After training is completed, the Fisher information content of each sorting model parameter is recalculated using the new training dataset to obtain new importance weight values. Then, the stored historical importance weight vector and the new importance weight vector are weighted and fused according to a preset forgetting factor to obtain the updated importance weight vector and store it. At the same time, the updated sorting model parameters are sent back to the feature monitoring module to recalculate the list of key feature dimensions.

9. The meat sorting and preprocessing system based on AI vision according to claim 8, characterized in that: The specific process for generating the corresponding synthetic training data in the digital twin module is as follows: Receive drift signals and parse them to obtain the numerical value of feature distribution difference degree and key dimension offset vector; calculate the distribution inference intensity coefficient based on the numerical value of feature distribution difference degree, ensuring that it is positively correlated with the numerical value of feature distribution difference degree and does not exceed the preset upper limit; The parameterized meat appearance simulator defines multiple simulator parameters that follow a preset prior probability distribution. A transfer mapping is constructed based on the inferred intensity coefficient and the key dimension offset vector. This transfer mapping acts on the original prior distribution of the simulator parameters, transforming them into an adjusted parameter distribution. The process of constructing the transport map includes: estimating the global mean vector and covariance matrix of the original prior distribution based on historical sampling data of simulator parameters, and calculating the square root matrix of the inverse of the covariance matrix for normalization; The key dimension offset vector is linearly transformed into the expected adjustment direction vector through a preset mapping matrix. The transfer mapping is defined as adding the following two terms to the value of each simulator parameter in its original prior distribution: the first term is the calculated distribution inference intensity coefficient, and the second term is the component in the expected adjustment direction vector corresponding to the simulator parameter after normalization by the square root matrix of the inverse of the covariance matrix.

10. The meat sorting and preprocessing system based on AI vision according to claim 9, characterized in that: The specific process of using synthetic training data to enhance the training of the updated sorting model is as follows: Sampling is performed from the adjusted simulator parameter distribution, with a bias towards selecting simulator parameter values ​​corresponding to the tail region of the distribution to generate more challenging synthetic meat images; Input the simulator parameter values ​​into the parameterized meat appearance simulator to generate synthetic images and completely known pseudo-labels, forming a synthetic training dataset; use this dataset to enhance the training of the updated sorting model. The total loss function includes the standard classification loss of the synthetic meat image and the consistency regularization loss, which is obtained by calculating the Jensen-Shannon divergence between the predicted distributions of the two different augmented views of the synthetic meat image. The training can employ a course learning strategy to gradually increase the distribution inference intensity coefficient corresponding to the synthetic data used; it also includes an online fine-tuning mechanism: when a newly labeled incremental sample pair is obtained, the ideal simulator parameters are deduced by optimizing the depth feature vector corresponding to the incremental sample pair, and the mapping matrix is ​​fine-tuned by gradient descent method in conjunction with the latest drift signal and the covariance information of the simulator parameter distribution, so that the mapping relationship is closer to the real physical relationship.

Images