Industrial large model performance self-evolution method and system

Abstract

The application belongs to the technical field of neural network optimization, and discloses an industrial large model performance self-evolution method and system. A data perception evaluation module is used to collect multiple types of industrial data in real time, a distribution drift detection operator is used to monitor the difference between input data and a training baseline, and task complexity is quantified from four dimensions of sample information entropy, feature manifold dimension, reasoning error entropy and task output space dimension. The two architecture operation indexes of activation saturation and gradient utilization rate are monitored, and evolution trigger and architecture capacity saturation signals can be autonomously generated. The architecture gradual reconstruction module is combined to calculate the optimal capacity interval of the network, the residual bypass form is matched with the orthogonal initialization strategy to smooth the newly added calculation unit, and the structured pruning is performed based on the gradient contribution decay rate and the activation frequency, so that the dynamic reallocation of computing power quota is realized. In cooperation with the active learning strategy, high-value samples are screened to complete small sample fine tuning, and the capacity demand of the differentiated industrial scene is effectively coped with.

Description

Technical Field

[0001] This invention belongs to the field of neural network optimization technology, specifically a method and system for self-evolution of the performance of large industrial models. Background Technology

[0002] Existing large-scale industrial models typically employ an offline pre-training paradigm combined with scenario-based fine-tuning. Once deployed, the network architecture and parameter weights are essentially fixed, making it difficult to adapt to the dynamically changing operating conditions, product categories, and process requirements of industrial environments. Current technical solutions mainly suffer from the following technical problems: The network structure of existing industrial large models is determined during the training phase. After deployment, it cannot automatically adjust the network depth, width, or neuron connection relationship according to changes in the complexity of industrial tasks and data distribution drift. When facing differentiated scenarios such as small sample anomaly detection and multi-process collaborative reasoning, either there will be redundant computing power leading to resource waste, or insufficient capacity will lead to a decrease in inference accuracy.

[0003] New data from industrial scenarios are constantly being generated, but improving model performance still requires manual data screening, sample labeling, hyperparameter tuning, retraining, and manual evaluation before going online. This makes it impossible to respond in a timely manner to real-time changes such as production line changes, process modifications, and equipment degradation, resulting in a lag between model performance and actual working conditions.

[0004] When the model is incrementally trained for new industrial tasks, the inference performance of the original tasks often suffers irreversible degradation. Existing methods based on replay and parameter isolation have limited effectiveness in industrial multi-task scenarios and are difficult to balance the absorption of new knowledge with the retention of old knowledge.

[0005] Most existing solutions only implement a one-way process from data collection to model training, lacking automated performance evaluation, evolution decision-making, and security verification. They cannot autonomously determine when evolution is needed, what evolution strategy to adopt, or whether the evolution meets industrial reliability requirements, and are highly dependent on manual intervention. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for self-evolution of the performance of large industrial models, so as to solve one or more problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an industrial large model performance self-evolution system, comprising a data perception and evaluation module, an architecture incremental reconstruction module, a gradient incremental optimization module, a hyperparameter evolution optimization module, a knowledge forgetting suppression module, a two-dimensional closed-loop evaluation module, and an evolution security management module. Furthermore, the data perception and evaluation module is used to collect data in real time, including industrial site time-series sensor data, production logs, process parameters, quality inspection results, and equipment alarms. Through the data distribution drift detection operator, it monitors the difference between the input data and the model training baseline distribution by statistical testing and feature distribution distance calculation. When the drift exceeds the set threshold, it outputs an evolution trigger signal. The data perception and evaluation module quantifies and scores the complexity of the current industrial task from four dimensions: sample information entropy, feature manifold dimension, inference error entropy, and task output space dimension. It comprehensively calculates the model capacity demand coefficient and monitors two architecture operation indicators: network activation saturation and gradient utilization. It comprehensively calculates the architecture saturation index and outputs an architecture capacity saturation signal when the index exceeds the threshold. The architecture operation metrics include two underlying monitoring values: activation saturation and gradient utilization. The architecture saturation metric is a quantitative value calculated by combining the two operation metrics. The architecture capacity saturation signal is a trigger signal generated when the architecture saturation metric exceeds a threshold. An active learning strategy combining uncertainty sampling and representative sampling is adopted to select the sample subset with the largest information gain from industrial data, and simultaneously generate sample difficulty distribution labels. The evolution trigger signal, the architecture capacity saturation signal, the task complexity score, the architecture saturation index, the selected sample subset, and the sample difficulty distribution labels are output to the architecture progressive reconfiguration module.

[0008] Furthermore, the incremental architecture reconfiguration module receives the evolution trigger signal, the task complexity score, the architecture saturation index, and the sample subset. Based on the task complexity score and the architecture saturation index, it calculates the optimal capacity range of the current network. When the existing capacity is below the lower threshold, incremental growth is initiated; when it is above the upper threshold, structured pruning is initiated. Incremental growth and structured pruning can be triggered independently based on capacity deviation, or they can be planned and executed in parallel within the same adjustment cycle to ensure dynamic matching between network capacity and the computing power requirements of industrial tasks. In a specified layer of the Transformer encoder, at the granularity of attention head and feedforward network sub-layer, computational units are gradually added in the form of residual bypass. The newly added computational units adopt an orthogonal initialization strategy, first connecting the original computational flow with low-weight bypass, then gradually increasing the bypass weight ratio, and finally smoothly integrating into the main network. During the synchronous cycle of new unit growth, based on the gradient contribution decay rate and activation frequency of the attention head and feedforward network sublayers, low-value redundant units and connections are identified and structured pruning is performed. The computing power quota released by pruning is preferentially allocated to newly added high-value units. After the architecture adjustment is completed, small-sample fine-tuning is automatically performed, and the reconstructed target network structure is output to the gradient incremental optimization module.

[0009] In the small-sample fine-tuning phase after the architecture adjustment, lightweight knowledge anchoring constraints are enabled simultaneously. Based on the updated set of knowledge anchors, the variation range of core parameters is limited to avoid accidental loss of historical task knowledge during the architecture adaptation process. In the fine-tuning phase, the replay feature stream is not enabled, and only the selected high-value new samples are used to complete the architecture fusion.

[0010] Furthermore, the gradient incremental optimization module receives the reconstructed target network structure, calculates the gradient mutual information and gradient propagation path weights between network parameters before incremental training starts, identifies the gradient flow critical path that carries the core knowledge of the historical task, strongly locks the parameters on the gradient flow critical path, and only opens the parameters of non-critical paths for gradient updates. The learning rate of each layer is adjusted according to the gradient norm and gradient fluctuation coefficient of each layer parameter. A decaying small learning rate is used for layers with large gradient fluctuations and located at the edge of the critical path of the gradient flow, while an increasing large learning rate is used for non-critical path layers with stable gradients and significant gains. The combined gradient clipping and weight decay mechanism is used to adjust the gradient clipping threshold and weight decay coefficient based on the noise intensity assessment results of industrial data, thereby suppressing gradient explosion and parameter drift caused by noisy data. During training, the gradient state data of each level is output to the hyperparameter evolution optimization module, and the parameter update state data is synchronized to the knowledge forgetting suppression module.

[0011] Furthermore, the hyperparameter evolution optimization module receives the gradient state data at each level and encodes key hyperparameters, including learning rate coefficient, batch size, regularization coefficient, and gradient locking range ratio, into individual chromosomes to form an initial population. Hyperparameter optimization is embedded in the batch update process of incremental training. After each batch of gradient update is completed, the individual fitness is calculated based on two indicators: the training loss of the current batch and the historical task verification accuracy. Based on the individual fitness results, selection, crossover, and mutation operations are performed to generate the next generation of hyperparameter population, which is then applied to the training of the next batch. The optimal hyperparameter configuration is retained during the evolution process and synchronously fed back to the gradient increment optimization module.

[0012] Furthermore, the knowledge forgetting suppression module receives the parameter update status data, establishes a set of knowledge anchors for each connected industrial task, records the activation patterns and weight distribution characteristics of the corresponding task on the key neurons of the network, and updates the knowledge anchor set by mapping according to the sub-layer topology inheritance relationship and the corresponding mapping rules of neurons, based on the sub-layer topology inheritance relationship and the corresponding mapping rules of neurons, and updates the corresponding activation threshold. During incremental training for a new task, the offset of the current parameter update relative to the set of knowledge anchors for the historical task is calculated in real time. When the offset exceeds a safety threshold, a knowledge anchoring regularization term is added to the loss function. The larger the offset, the higher the weight of the knowledge anchoring regularization term, thus constraining the variation of key parameters. The core feature stream of historical tasks is distilled into a fixed-dimensional feature vector sequence using feature distillation, which is stored in the experience replay buffer. During incremental training, the mixing ratio of the replay feature stream is adjusted according to the data drift rate to form a streaming mixed training with the new task data. When a new task is detected to have overlapping features with a historical task, the anchor feature extraction weights of the corresponding task are automatically reused, and the constraint rules are applied in real time to the incremental training process.

[0013] Furthermore, the dual-dimensional closed-loop evaluation module incorporates neural network performance indicators and industrial business performance indicators. The neural network performance indicators include parameter update efficiency, architecture utilization, historical task knowledge retention rate, and gradient stability. The industrial business performance indicators include task accuracy, inference timeliness, robustness, and business value. An evaluation method combining online and offline approaches is adopted. The online approach compares the inference results of the model before and after evolution in parallel through shadow traffic, and statistically analyzes the consistency and improvement of business performance. At the same time, the operating indicators of the neural network are monitored online. The offline approach tests the model based on an industrial benchmark dataset and generates sub-performance data. After the assessment is completed, the rate of change and confidence level of each indicator are calculated, and the assessment results and rollback recommendations are output to the evolution security management module.

[0014] Furthermore, the evolution security management module receives performance data output by the two-dimensional closed-loop evaluation module, and combines it with three types of inputs: the evolution trigger signal, the architecture capacity saturation signal, and available computing resources. Through a deep reinforcement learning algorithm, it outputs the evolution start-up judgment result, evolution combination strategy, computing power allocation quota, and execution time window. Two security checks are set up: a neural network-level robustness check and an industrial business security check. The neural network-level robustness check sequentially undergoes weight perturbation test, feature boundary test, and adversarial gradient stability test to verify the inference stability of the model itself. The industrial business security verification undergoes boundary condition testing, abnormal input testing, and operational constraint compliance testing to verify the business security in industrial scenarios. The evolved target model can only be launched after all verifications are passed. The evolution security management module maintains the evolution version chain, supports the rollback of architecture and parameters of any historical version, generates traceable logs for each evolution including data source, architecture change record, policy parameters, performance changes and operation subject, reserves manual intervention interface, configures manual approval process for key evolution nodes, and synchronously issues management instructions to each execution module.

[0015] This invention also provides a method for the self-evolution of performance in large industrial models, based on the above-described system, comprising the following specific steps: Step S1: Collect data including industrial field time-series sensor data, production logs, process parameters, quality inspection results, and equipment alarms. Calculate the distribution difference between the monitored input data and the training baseline through statistical testing and feature distribution distance calculation. Quantify the complexity of the current industrial task from four dimensions: sample information entropy, feature manifold dimension, inference error entropy, and task output space dimension. Monitor two architecture operation indicators: network activation saturation and gradient utilization. Use an active learning strategy to screen high-value samples and generate difficulty distribution labels. Output evolution trigger signal, architecture capacity saturation signal, task complexity score, architecture saturation index, and the selected high-value sample subset and sample difficulty distribution labels. Step S2: Calculate the optimal capacity range of the network based on the task complexity score and architecture saturation index. When the capacity deviates from the threshold, start architecture adjustment. Add computing units in the specified layer of the Transformer encoder with attention head and feedforward network sub-layer as the granularity, in the form of residual bypass. Identify redundant units based on gradient contribution decay rate and activation frequency and perform structured pruning. After completing small sample fine-tuning, output the reconstructed network structure. Step S3: Initiate incremental training, calculate the gradient mutual information and propagation path weights between parameters of each layer, identify key gradient flow paths and lock the corresponding parameters, adjust the learning rate layer by layer according to the gradient features of each layer, and integrate gradient pruning and weight decay mechanisms; encode the key training hyperparameters into an evolutionary population, calculate the individual fitness online and perform evolutionary operations during the iteration process, and feed back the optimal hyperparameter configuration to the training stage; synchronously establish a set of knowledge anchor points for each task, update the anchor point mapping when the architecture changes, apply regularization constraints according to the anchor point offset during training, and mix in the replay feature flow to participate in training in proportion; Step S4: Set two evaluation indicators: neural network performance and industrial business performance. Use a combination of online shadow traffic parallel testing and offline benchmark testing to complete model testing, calculate the rate of change and confidence of each indicator, and output the evaluation results and rollback suggestions. Step S5: Combining performance evaluation data, evolution trigger signals, architecture capacity saturation signals, and available computing resources, generate an evolution execution strategy using a deep reinforcement learning algorithm. After two security checks, deploy the model online, maintain the evolution version chain, generate traceable logs, and issue control instructions.

[0016] The beneficial effects of this invention are as follows: 1. This invention collects various types of industrial data in real time through a data perception and evaluation module, monitors the difference between the input data and the training baseline using a distribution drift detection operator, and quantifies the task complexity from four dimensions: sample information entropy, feature manifold dimension, inference error entropy, and task output space dimension. It monitors two architecture operation indicators: activation saturation and gradient utilization, and can autonomously generate evolution trigger and architecture capacity saturation signals. Combined with an architecture progressive reconstruction module, it calculates the optimal capacity range of the network, smoothly adds new computing units using residual bypass and orthogonal initialization strategies, and performs structured pruning based on gradient contribution decay rate and activation frequency to achieve dynamic reallocation of computing power quotas. With the help of an active learning strategy, it selects high-value samples to complete small-sample fine-tuning, effectively addressing the capacity requirements of differentiated industrial scenarios, and ensuring that the model always maintains an appropriate network capacity and inference accuracy.

[0017] 2. The gradient incremental optimization module of this invention identifies the key gradient flow path carrying core knowledge and locks the corresponding parameters by calculating the gradient mutual information and propagation path weights between parameters. Combined with the hierarchical dynamic learning rate and gradient pruning and weight decay combination mechanism, it can suppress gradient explosion and parameter drift caused by industrial noise, and improve the stability and convergence efficiency of incremental training. With the hyperparameter evolution optimization module, hyperparameter optimization is embedded into the training batch iteration. Based on the training loss and the verification accuracy of historical tasks, online evolution optimization is performed to continuously optimize the training configuration. The knowledge forgetting suppression module establishes a set of knowledge anchor points for each industrial task. When the architecture changes, the anchor point mapping is updated. During training, regular constraints are applied according to the anchor point offset, and the distilled replay feature flow is mixed in proportionally. This can accelerate the convergence of new tasks while constraining the variation of key parameters and reducing the performance degradation of historical tasks.

[0018] 3. The dual-dimensional closed-loop evaluation module of this invention sets multiple sub-indicators from two dimensions: neural network performance and industrial business performance. It adopts a combination of online shadow traffic parallel testing and offline benchmark testing to quantify the change rate and confidence level of each indicator, and outputs objective evaluation results and rollback suggestions. The evolution security management module combines performance data, evolution trigger signals and available computing resources, and generates the optimal evolution execution strategy through deep reinforcement learning algorithms. It sets two verification checkpoints: neural network robustness and industrial business security, to verify the stability of model inference and business compliance. At the same time, it maintains a complete evolution version chain, generates traceable logs, and reserves interfaces for manual intervention and approval, ensuring the automation and timeliness of the evolution process, and ensuring that the evolved model meets the reliability and security requirements of industrial scenarios. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the overall closed-loop process of the performance self-evolution of the industrial large-scale model of this invention. Figure 2 This is a flowchart illustrating the progressive refactoring of the architecture of this invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] like Figures 1 to 2 As shown, this embodiment of the invention provides an industrial large model performance self-evolution system, including a data-aware evaluation module, an architecture incremental reconstruction module, a gradient incremental optimization module, a hyperparameter evolution optimization module, a knowledge forgetting suppression module, a two-dimensional closed-loop evaluation module, and an evolution security management module. The industrial large model baseline on which this invention is based adopts the standard Transformer encoder architecture, with a default configuration of 6 layers of encoder stacked. Each layer contains 8 multi-head attention heads and two layers of feedforward network. The intermediate layer of the feedforward network has a dimension of 2048, the input embedding dimension is 512, and the layer normalization adopts the Pre-LN structure placed at the input end of each sub-layer.

[0022] The model is configured with a data normalization layer and a sinusoidal position encoding layer. The wavelength range of the sinusoidal position encoding is set from 1 to 10000, which corresponds one-to-one with the input embedding dimension. The output layer is adapted to either a classification head or a regression head according to the type of industrial task. The classification head uses the Softmax activation function to output the category probability distribution, while the regression head uses a linear layer to output continuous values.

[0023] For multi-source data from industrial sites, channel-specific preprocessing is performed. Time-series sensor data is truncated into a fixed-length sequence using a sliding window of length 128. After Z-score standardization to remove dimensional differences, it is converted into a 512-dimensional embedding vector through a linear projection layer. Structured data such as production logs and process parameters are first filled with missing values ​​and truncated with 3σ for outliers. Discrete fields are mapped to high-dimensional sparse vectors using one-hot encoding. Continuous fields are processed by min-max normalization and then compressed into feature vectors of the same dimension as the temporal embedding through a feature mapping layer. The two types of features are concatenated along the sequence dimension and then input into the model encoder.

[0024] The model outputs the inference results for the corresponding task. For classification tasks, the output dimension is consistent with the number of task categories, and for regression tasks, the output dimension is consistent with the number of parameters to be predicted. The output results are connected to industrial field control systems or quality inspection and control platforms through standardized interfaces.

[0025] In this embodiment of the invention, the data perception and evaluation module is used to collect real-time data such as industrial site time-series sensor data, production logs, process parameters, quality inspection results, and equipment alarms. Through the data distribution drift detection operator, the difference between the input data and the model training baseline distribution is monitored by statistical testing and feature distribution distance calculation. When the drift exceeds a set threshold, an evolution trigger signal is output. The statistical test uses the Kolmogorov-Smirnov test to verify the distribution consistency of each single-dimensional feature. The test statistic D is calculated as the maximum value of the difference between the cumulative distribution functions of the two samples, and a significance level of 0.05 is used as the threshold for single-feature drift judgment. The deviation of the joint distribution of multiple features is calculated by using the maximum mean difference. The kernel function is a Gaussian kernel, and the kernel bandwidth is set to 0.5 times the median of the feature dimension.

[0026] The drift degree is quantized by weighting the single-feature KS test failure rate and MMD distance value. Both indicators are normalized to the [0,1] interval by min-max, with weights of 0.4 and 0.6 respectively. The drift degree threshold is set to 0.3. When the quantized value exceeds the threshold, the evolution trigger signal is output.

[0027] The data perception and evaluation module quantifies and scores the complexity of the current industrial task from four dimensions: sample information entropy, feature manifold dimension, inference error entropy, and task output space dimension, and comprehensively calculates the model capacity requirement coefficient. Sample information entropy is calculated based on the feature probability distribution of the input samples. The formula for sample information entropy is as follows: ; The entropy of sample information is used to measure the richness of information in an input sample, and its unit is bits. This represents the specific value of the i-th feature; This represents the probability density corresponding to the value of the i-th feature.

[0028] The eigendimensionality of the feature manifold is estimated using the local linear embedding method, with the number of neighborhood samples set to 15. The intrinsic dimension of the low-dimensional manifold is solved by minimizing the reconstruction error. The inference error entropy is calculated based on the error distribution of the current inference result of the model. The error interval is discretized into 10 equal-width bins before the discrete entropy is calculated. The dimension of the task output space is determined by the number of task categories or the number of discrete bins in the range of regression values.

[0029] The task complexity score is obtained by summing the four dimensions after min-max normalization with weights of 0.3, 0.3, 0.25, and 0.15. The model capacity requirement coefficient is positively linearly correlated with the task complexity score, and the coefficient ranges from [0.5 to 2.0].

[0030] Simultaneously monitor two architectural operation indicators of the network: activation saturation and gradient utilization. Calculate the architecture saturation indicator by combining the two indicators. When the indicator exceeds the threshold, output the architecture capacity saturation signal. Activation saturation is the percentage of samples whose activation values ​​of network neurons exceed a preset threshold of 0.7 during the inference phase, and the statistical range covers feedforward network neurons in all encoder layers; gradient utilization is the percentage of parameters whose absolute gradient value is greater than 1e-6 during backpropagation, and the statistical range covers all trainable weight parameters.

[0031] The architecture saturation index is calculated by weighting activation saturation and gradient utilization by 0.6 and 0.4 respectively. The threshold is set to 0.85. When the threshold is exceeded, an architecture capacity saturation signal is output.

[0032] The architecture operation metrics include two underlying monitoring values: activation saturation and gradient utilization. The architecture saturation metric is a quantitative value calculated by combining the two operation metrics. The architecture capacity saturation signal is a trigger signal generated when the architecture saturation metric exceeds a threshold. An active learning strategy combining uncertainty sampling and representative sampling is adopted to select the sample subset with the highest information gain from industrial data and generate sample difficulty distribution labels simultaneously. Uncertainty sampling prioritizes samples with model inference confidence in the middle range, covering boundary conditions and abnormal edge samples where model judgment is ambiguous; representative sampling first performs feature clustering on all newly added data, dividing it into a fixed number of clusters according to feature distribution, and then selects the sample closest to the cluster center from each cluster.

[0033] The final sample subset consists of a 7:3 mixture of uncertain samples and representative samples, with the sample size controlled at 15% to 20% of the total new data.

[0034] The sample difficulty distribution labels are divided into three levels: easy, medium, and difficult, based on the magnitude of the model inference error. The error ranges from low to high, each corresponding to a different level. The labels are bound to the samples and output synchronously.

[0035] The evolution trigger signal, architecture capacity saturation signal, task complexity score, architecture saturation index, and the selected sample subset and sample difficulty distribution label are output to the architecture incremental reconstruction module; the evolution trigger signal and architecture capacity saturation signal are simultaneously output to the evolution security management module as input basis for evolution decision.

[0036] In this embodiment of the invention, the incremental architecture reconfiguration module receives the evolution trigger signal, the task complexity score, the architecture saturation index, and the sample subset. Based on the task complexity score and the architecture saturation index, it calculates the optimal capacity range of the current network. When the existing capacity is below the lower threshold, incremental growth is initiated; when it is above the upper threshold, structured pruning is initiated. Incremental growth and structured pruning can be triggered independently according to the capacity deviation type, or they can be planned and executed in parallel within the same adjustment period to achieve precise adaptation of network capacity. The optimal capacity range of the network uses the total number of attention heads and the total number of parameters of the feedforward network as the capacity measurement units. The corresponding baseline capacity value is matched according to the task complexity score. The ratio of the baseline capacity value to the baseline model capacity is equal to the model capacity requirement coefficient. The lower and upper limits of the range are set by 20% above and below the baseline capacity. The range calculation result is rounded to the nearest integer multiple of the attention heads.

[0037] In a specified layer of the Transformer encoder, at the granularity of attention head and feedforward network sub-layer, computational units are gradually added in the form of residual bypass. The newly added computational units adopt an orthogonal initialization strategy, first connecting the original computational flow with low-weight bypass, and then gradually increasing the bypass weight ratio with short-cycle fine-tuning, and finally smoothly integrating into the main network. The residual bypass is deployed between the output of the multi-head attention sublayer and the input of the feedforward network sublayer in the Transformer encoder layer. The bypass contains an independent feedforward network sublayer with the same structure as the original feedforward network. The output of the newly added computing unit is weighted and summed with the input of the original feedforward network sublayer before being connected to the original feedforward network sublayer.

[0038] The new unit adopts an orthogonal initialization strategy based on SVD decomposition to ensure the column orthogonality of the initial weight matrix. In the initial stage, the weight of the bypass output is set to 0.1, and it is linearly increased by 0.05 every 100 steps during fine-tuning training until the weight reaches 0.5 and then the increase stops, thus completing the fusion of the bypass and the main network.

[0039] The designated layer is selected based on the feature abstraction level and gradient contribution ratio of each encoder layer. The intermediate and deep encoder layers of the network are preferentially selected as the layers for architecture adjustment, while the shallow encoder retains its original structure without adjustment.

[0040] For the default six-layer encoder configuration, layers three through six are selected as adjustable layers, while layers one and two maintain a fixed structure.

[0041] During the synchronous cycle of new unit growth, based on the gradient contribution decay rate and activation frequency of the attention head and feedforward network sublayers, low-value redundant units and connections are identified and structured pruning is performed. The gradient contribution decay rate is the percentage decrease in the gradient contribution value of the unit within the last 10 training batches. The gradient contribution is approximated by a first-order Taylor expansion, which is the product of the gradient of the unit output with respect to the loss function and the unit activation value. The activation frequency is the percentage of samples in the validation set whose activation value exceeds the 0.1 threshold. The statistical window covers all validation samples.

[0042] When the gradient contribution decay rate of a cell exceeds 70% and the activation frequency is below 20%, it is identified as a redundant cell and pruning is performed, simultaneously removing all connection weights corresponding to that cell; the computing power quota released by pruning is measured in floating-point operations.

[0043] The computing power quota released by pruning is preferentially allocated to newly added high-value units to achieve dynamic balance of network capacity and efficient utilization of computing resources. After the architecture adjustment is completed, small-sample fine-tuning is automatically performed, and the reconstructed target network structure is output to the gradient incremental optimization module.

[0044] The small-sample fine-tuning uses a high-value sample subset obtained through active learning screening, accounting for 15%-20% of the total new data. The optimizer used is AdamW, with a weight decay coefficient of 1e-4, a training batch size of 16, and an initial learning rate of 1×10⁻⁴. -5 A linear warm-up strategy is adopted, gradually increasing the learning rate to the target rate in the first 10% of training steps; the number of training rounds is preset to 5-10 rounds, the loss function is kept consistent with the original task, and an early stopping strategy is adopted, terminating fine-tuning when the validation set loss does not decrease for 3 consecutive rounds.

[0045] Growth and pruning operations are planned in parallel and executed step by step within the same adjustment cycle. First, redundant units are pruned and corresponding computing resources are released. Then, the scale of new computing units is determined based on the total amount of released computing power, ensuring that the total computing power consumption of the network before and after a single architecture adjustment does not exceed a preset threshold.

[0046] When the computing power demand of newly added units exceeds the amount released by pruning, they are grown in batches according to the priority of gradient contribution, with priority given to the computing units with the highest gradient gain, and the remaining gaps are left to be filled in the next adjustment cycle.

[0047] In this embodiment of the invention, the gradient incremental optimization module receives the reconstructed target network structure, calculates the gradient mutual information and gradient propagation path weights between network parameters before incremental training starts, identifies the gradient flow critical path that carries the core knowledge of the historical task, strongly locks the parameters on the gradient flow critical path, and only opens the parameters of non-critical paths for gradient updates. Gradient mutual information is calculated based on the joint probability distribution and marginal probability distribution of parameter gradients. The kernel density estimation method is used to fit the probability distribution of the gradient. The higher the mutual information value, the stronger the gradient dependency between parameters. The gradient propagation path weight is obtained by accumulating the gradient flow magnitude during the backpropagation process and statistically summing the gradient magnitudes of each parameter layer by layer.

[0048] The network layers and sub-layers containing the top 30% of gradient mutual information and the top 30% of path weights are selected as the key paths of gradient flow. All weight parameters on this path are frozen, their gradients are set to zero, and they do not participate in the parameter updates of incremental training.

[0049] The critical path locking range is not fixed throughout the entire process. After completing a fixed number of training batches, the gradient mutual information and path weights are recalculated, and the coverage range and locked parameter set of the critical path are updated. Only the parameters that continuously carry the core knowledge are frozen, while the parameters that have lost their core role are unlocked and included in the update range. This approach retains sufficient parameter optimization space while suppressing knowledge forgetting.

[0050] The learning rate of each layer is adjusted according to the gradient norm and gradient fluctuation coefficient of each layer parameter. A decaying small learning rate is used for layers with large gradient fluctuations and located at the edge of the critical path of the gradient flow, while an increasing large learning rate is used for non-critical path layers with stable gradients and significant gains. The hierarchical base learning rate range is set to 1×10. -6 Up to 5×10 -5 The baseline learning rate is uniformly set to 1×10. -5 The gradient variability coefficient is the coefficient of variation of the gradient norm of the layers within the last 5 batches, i.e., the ratio of the standard deviation to the mean. For critical path edge layers with a gradient variability coefficient exceeding 0.8, a decaying small learning rate of 0.5 multiplied by the base learning rate is adopted, and a cosine annealing scheduling strategy is used to decay the learning rate batch by batch. For layers with gradient variability coefficients in the range of 0.3 to 0.8, a baseline learning rate is used, along with a fixed step size scheduling strategy. For non-critical path layers with gradient variability coefficients below 0.3, an incremental large learning rate of 1.2 multiplied by the base learning rate is adopted, and a linear incremental scheduling strategy is used to gradually increase the learning rate.

[0051] The combined gradient clipping and weight decay mechanism is used to adjust the gradient clipping threshold and weight decay coefficient based on the noise intensity assessment results of industrial data, thereby suppressing gradient explosion and parameter drift caused by noisy data. During training, the gradient state data of each level is output to the hyperparameter evolution optimization module, and the parameter update state data is synchronized to the knowledge forgetting suppression module.

[0052] Industrial data noise intensity is quantified by the standard deviation of the sample label residuals, where the label residuals are the difference between the sample label values ​​and the model inference values. Higher noise intensity corresponds to a smaller gradient clipping threshold and a larger weight decay coefficient. The default gradient clipping threshold is 1.0, and a global gradient norm clipping method is used to uniformly constrain the gradient norm of all trainable parameters. The default weight decay factor is 1×10 -4 This only applies to the weight parameters, not the bias term or the layer normalization parameters; for every 0.1 increase in the normalization value of the noise intensity, the gradient clipping threshold is lowered by 0.1, and the weight decay coefficient is increased by 2 × 10. -5 .

[0053] During incremental training, sample difficulty distribution labels are introduced for batch sampling scheduling. Samples of different difficulties are mixed in a fixed proportion within each training batch. In the early stage of training, the sampling proportion of high-difficulty samples is appropriately increased to accelerate the model's learning of boundary and abnormal scenarios. In the later stage of training, the proportion of high-difficulty samples is gradually reduced to stabilize the model's convergence effect.

[0054] In this embodiment of the invention, the hyperparameter evolution optimization module receives the gradient state data of each level and encodes key hyperparameters such as learning rate coefficient, batch size, regularization coefficient, and gradient locking range ratio in the incremental training process into individual chromosomes to form an initial population. The initial population size was set to 20. Each hyperparameter was encoded using an 8-bit binary encoding method. The learning rate coefficient was encoded in the range of [1e-6, 1e-4], the batch size was encoded in the range of [8, 64] and taken as an integer power of 2, the regularization coefficient was encoded in the range of [1e-5, 1e-3], and the gradient locking range ratio was encoded in the range of [10%, 50%].

[0055] The selection operation uses a roulette wheel method, where the probability of an individual being selected is proportional to its fitness value; the crossover operation uses single-point crossover with a crossover probability of 0.7; and the mutation operation uses bit-flip mutation with a mutation probability of 0.1. Every 5 training batches complete one generation of evolutionary iteration. The top 20% of elite individuals by fitness are retained and directly enter the next generation population, while the remaining individuals are generated through selection, crossover, and mutation.

[0056] During the evolution process, a hyperparameter safety boundary is set. All generated hyperparameter individuals must fall within the preset value range. Individuals that exceed the boundary are directly eliminated and replaced by the best individual of the previous generation. If the optimal fitness of the population decreases for two consecutive generations, the mutation operation is paused, and only the selection and crossover operations are retained. At the same time, the global best individual of the previous generation is introduced into the current population.

[0057] Hyperparameter optimization is embedded into the batch update process of incremental training. After each batch of gradient update is completed, the individual fitness is calculated based on two indicators: the training loss of the current batch and the historical task verification accuracy. The formula for calculating individual fitness is:

[0058] Fitness refers to the overall performance score of an individual with hyperparameters, with a value range of [0, 1]. The higher the value, the better the overall performance of the corresponding hyperparameter configuration. This represents the normalized training loss, which is the ratio of the current batch training loss to the initial batch training loss, normalized to the interval [0, 1]. This represents the normalized historical task verification accuracy, which is the ratio of the historical task verification set accuracy to the baseline accuracy under the current hyperparameters, and has been normalized to the interval [0, 1]. 0.4 and 0.6 represent the weighting coefficients of the training loss and historical task accuracy indicators.

[0059] Based on the individual fitness results, selection, crossover, and mutation operations are performed to generate the next generation of hyperparameter population, which is then applied to the training of the next batch. During the evolution process, the optimal hyperparameter configuration is retained and synchronously fed back to the gradient increment optimization module.

[0060] In this embodiment of the invention, the knowledge forgetting suppression module receives the parameter update status data, establishes a set of knowledge anchor points for each connected industrial task, and records the activation patterns and weight distribution characteristics of the corresponding task on the key neurons of the network. The key neurons are selected independently by encoder layer. The top 10% of neurons in each layer are selected based on their activation values ​​in response to the task, forming a multi-layer anchor set corresponding to the number of encoder layers. The activation mode is recorded as the average activation vector of neurons in the benchmark dataset for the task, with the dimension consistent with the number of neurons in a single layer. The weight distribution features are recorded as the mean and variance of the input connection weights of the corresponding neurons. The knowledge anchor set is generated and solidified after the initial model training and before the first deployment, with each industrial task corresponding to an independent anchor set.

[0061] When the network architecture changes, the knowledge anchor set is mapped and updated. Based on the sub-layer topology inheritance relationship and the mapping rules corresponding to neurons, the old anchors are mapped to the new network topology positions, and the corresponding activation thresholds are updated. For networks after structured pruning, the anchor entries corresponding to the pruned neurons are directly removed, while the remaining unpruned anchors retain their original positions and all feature parameters. For networks after adding computational units, the positions of neurons corresponding to the original anchors remain unchanged, and the newly added neurons are not included in the anchor set of the historical task. After the new task is trained and evaluated, the anchor parameters corresponding to the newly added units will be generated. If the number of network layers increases or decreases, cross-layer mapping is performed according to the correspondence of feature abstraction levels.

[0062] During incremental training for a new task, the offset of the current parameter update relative to the set of knowledge anchors for the historical task is calculated in real time. When the offset exceeds a safety threshold, a knowledge anchoring regularization term is added to the loss function. The larger the offset, the higher the weight of the knowledge anchoring regularization term, thus constraining the variation of key parameters. The anchor offset is the cosine distance between the activation vector of the current key neuron and the activation vector of the reference anchor point, calculated using the following formula: ;

[0063] This represents the anchor point offset, which is essentially the cosine distance. Its value ranges from [0, 2]. The larger the value, the greater the difference in the activation mode between the current network and the reference anchor point. This represents the activation vector of a key neuron in the current network; This represents the baseline activation vector recorded in the task knowledge anchor set; Indicates the L2 norm of the current activation vector; This represents the L2 norm of the baseline anchor activation vector.

[0064] The knowledge anchoring regularization term adopts the L2 regularization form, and the calculation formula is as follows: ;

[0065] The loss value representing the knowledge anchoring regularization term needs to be added to the total loss function to participate in gradient updates; This represents the weight coefficient of the regularization term. The larger the anchor offset, the higher the coefficient. The initial value is 0.01. For every 0.05 that the offset exceeds the threshold, the value of λ increases by 0.005, with an upper limit of 0.1. This represents the weight vector of the current key neuron; This represents the baseline weight vector recorded in the knowledge anchor set.

[0066] The core feature stream of historical tasks is distilled into a fixed-dimensional feature vector sequence by feature distillation and stored in the experience replay buffer. During incremental training, the mixing ratio of the replay feature stream is adjusted according to the data drift rate to form a streaming mixed training with the new task data, thereby reducing the storage and computing power overhead of the replay stage. The data drift rate is the increment of the drift degree per unit time, with a statistical window of 24 hours; the higher the drift rate, the higher the proportion of replay feature stream mixed in. The replay feature stream is generated by feature distillation, using the current model as the teacher model, extracting the output features of historical task samples at the last layer of the encoder, distilling them into a fixed 512-dimensional feature vector sequence, and storing it in the experience replay buffer; The buffer uses a FIFO update strategy with a maximum capacity of 10,000 feature sequences. The default mixing ratio is 10%, and the mixing ratio increases by 5% for every 0.1 increase in drift, up to a maximum of 30%. Each training batch randomly selects replay features and mixes them with new sample features before inputting them into the model.

[0067] The knowledge anchoring regularization term and the replay feature flow mechanism work together. In the early stage of training, the replay feature flow is mainly mixed in, and the anchoring regularization term is kept with a low weight. The activation distribution of the historical task is maintained first through feature replay. In the middle and late stages of training, the weight of the anchoring regularization term is gradually increased, while the proportion of replay feature mixing is reduced. The weight distribution of historical knowledge is solidified through parameter constraints. The proportion of the two mechanisms is dynamically adjusted according to the training progress.

[0068] Feature distillation uses the current stable version of the model as the teacher network, selects the output of the last layer of the encoder as the distillation feature source, performs full inference on the historical task benchmark dataset, extracts the corresponding features, processes them into a fixed-length feature sequence through mean pooling, and then stores them in the buffer.

[0069] When the network architecture changes and the input dimension changes, a linear adaptation layer is used to map the dimension of the original playback features to the input dimension of the new network. The adaptation layer is fine-tuned synchronously with incremental training.

[0070] When a new task is detected to have overlapping features with a historical task, the anchor feature extraction weights of the corresponding task are automatically reused to accelerate the convergence of the new task and reduce the loss of historical knowledge. The constraint rules are applied in real time to the incremental training process.

[0071] Feature semantic overlap detection is achieved by calculating the similarity of the output features of new task samples and historical task samples in the penultimate layer of the encoder. The cosine value of the feature center of the two types of samples is taken as the overlap index. When the overlap index exceeds the set threshold, it is determined that there is feature semantic overlap. The higher the overlap, the higher the reuse ratio of the anchor feature extraction weight.

[0072] In this embodiment of the invention, the dual-dimensional closed-loop evaluation module incorporates neural network performance indicators and industrial business performance indicators. The neural network performance indicators include parameter update efficiency, architecture utilization, historical task knowledge retention rate, and gradient stability. The industrial business performance indicators include task accuracy, inference timeliness, robustness, and business value. The evaluation method combines online and offline approaches. The online approach compares the inference results of the model before and after evolution in parallel through shadow traffic, and statistically analyzes the consistency and improvement of business performance. At the same time, it monitors the inference state gradient stability, activation distribution and other operating indicators of the evolved model online for comprehensive performance evaluation. The offline approach tests the model based on industrial benchmark datasets and generates sub-performance data. Shadow traffic is mirrored from the actual online production traffic at a fixed ratio. The default split ratio is set to 10%. The splitting process does not affect the normal inference and business execution of the main production traffic. After the traffic is split, it is simultaneously input into the pre-evolution baseline model and the post-evolution target model. The two sets of inference requests use the exact same input data and timestamps. After the inference is completed, the results are aligned sample by sample and various difference indicators are calculated. The output of the post-evolution model does not enter the actual production control link, but is only used for performance data statistics and comparison.

[0073] After the assessment is completed, the rate of change and confidence level of each indicator are calculated, and the assessment results and rollback recommendations are output to the evolution security management module.

[0074] The rate of change of the indicator is the relative difference between the indicator value after evolution and the baseline value before evolution. The calculation formula is: rate of change = (value after evolution - value before evolution) / value before evolution × 100%; the confidence level is based on the test sample size and the variance of the indicator. The ratio of the half-width of the 95% confidence interval to the mean of the indicator is calculated by the t-distribution. The larger the sample size and the smaller the variance, the higher the confidence level.

[0075] When the overall performance change rate is below -5% and the confidence level is above 90%, a Level 1 rollback recommendation is triggered, suggesting a rollback to the previous stable version; when the accuracy of core business metrics decreases by more than 3% and the confidence level is above 95%, an emergency rollback recommendation is triggered, suggesting an immediate termination of the evolution version launch.

[0076] In addition to providing rollback recommendations, the evaluation results also feed back the changes in each sub-indicator to the architecture refactoring, gradient optimization, and hyperparameter optimization modules as baseline reference data for the next evolution. If the assessment shows that the architecture utilization is not as expected, the baseline capacity value will be lowered in the next capacity interval calculation. If the assessment shows that the historical knowledge retention rate is low, the initial weight of the knowledge anchoring regularization term will be increased in the next incremental training.

[0077] In this embodiment of the invention, the evolution security management module receives performance data output by the two-dimensional closed-loop evaluation module, and combines the three types of inputs—the evolution trigger signal, the architecture capacity saturation signal, and available computing resources—to output the evolution start-up judgment result, evolution combination strategy, computing power allocation quota, and execution time window through a deep reinforcement learning algorithm. The deep reinforcement learning method uses the DQN algorithm. Both the evaluation network and the target network adopt a 3-layer fully connected network structure with hidden layer dimensions of 128 and 64, respectively, and the activation function is ReLU.

[0078] The state space includes four continuous state variables: drift degree, architecture saturation, available computing power ratio, and current model comprehensive performance score. The input is normalized to the [0, 1] interval. The action space is discretized into 32 combinations, including four types of actions: whether to start evolution (2 types), evolution strategy type, computing power allocation ratio, and execution period. The evolution strategy type includes four types: incremental training only, architecture adjustment only, architecture adjustment + incremental training, and full-process evolution. The computing power allocation ratio includes two types: 50% and 100%. The execution period includes two types: valley period and flat period.

[0079] The reward function uses the improvement in overall business performance after evolution as a positive reward with a weight of 1.0; and uses the proportion of computing power consumption and the amount of performance degradation of historical tasks as negative rewards with weights of 0.3 and 0.7, respectively.

[0080] The evolution execution time window prioritizes matching low-load periods of industrial production and planned downtime intervals, avoiding peak production capacity and critical production process stages; all training and verification operations during the evolution process are performed on offline backup computing resources, without occupying the computing power quota for production inference.

[0081] When only the generation evolution trigger signal is triggered and the architecture capacity saturation signal is not triggered, a lightweight evolution process is executed, which only starts incremental training and hyperparameter optimization without adjusting the network architecture. When both the generation evolution trigger signal and the architecture capacity saturation signal are triggered, a full evolution process is executed, which first completes the incremental reconstruction of the architecture to adjust the network capacity, and then starts incremental training and the corresponding knowledge forgetting suppression and hyperparameter optimization. When only the architecture capacity saturation signal is triggered and no obvious data distribution drift is detected, only the architecture capacity adaptation adjustment is executed, and no new data is introduced for incremental training.

[0082] Two security checks are set up: a neural network-level robustness check and an industrial business security check. The neural network-level robustness check sequentially undergoes weight perturbation test, feature boundary test, and adversarial gradient stability test to verify the inference stability of the model itself. The industrial business security verification undergoes boundary condition testing, abnormal input testing, and operational constraint compliance testing to verify the business security in industrial scenarios. The evolved target model can only be launched after all verifications are passed. The criteria for judging the robustness of neural networks are as follows: in the weight perturbation test, after applying a Gaussian perturbation with a standard deviation of 0.05 to the weights, the inference accuracy decreases by no more than 2%; in the feature boundary test, the inference accuracy of boundary samples generated by the extreme value construction method is no less than 85%; and in the adversarial gradient stability test, the accuracy decreases by no more than 5% under adversarial sample attacks generated by the FGSM algorithm.

[0083] The criteria for industrial business safety verification are as follows: under boundary condition testing, the inference results of extreme conditions covering the upper and lower limits meet the process constraint thresholds; under abnormal input testing, the model can identify invalid inputs that are missing, exceed limits, or have incorrect formats and output abnormality flags; under operational constraint compliance testing, the output results do not violate the equipment operation safety thresholds and process specification requirements.

[0084] If the neural network-level robustness check fails, it will revert to the incremental training phase, increase the weight decay coefficient and gradient clipping strength, and then retrain. If the industrial business safety check fails, it will revert to the data and sample screening phase, supplement boundary condition samples and constraint rules, and then re-execute fine-tuning. If the same version fails two consecutive checks, the current evolution process will be terminated, and the current online stable version will remain unchanged.

[0085] The evolution security management module maintains the evolution version chain, supports the rollback of architecture and parameters of any historical version, generates traceable logs for each evolution including data source, architecture change record, policy parameters, performance changes and operation subject, reserves manual intervention interface, configures manual approval process for key evolution nodes, and synchronously issues management instructions to each execution module.

[0086] The evolution version chain links all officially launched model versions in chronological order. Each version corresponds to a unique version number and stores complete network architecture description files, parameter weight files, hyperparameter configuration files, and a complete set of evaluation reports for the corresponding version.

[0087] Version rollback is divided into two categories: automatic rollback and manual rollback. Automatic rollback is triggered by the rollback suggestion output by the evaluation module and is executed automatically after passing the security verification. Manual rollback is initiated through the manual intervention interface and can only be executed after going through the approval process.

[0088] This invention also provides a method for the self-evolution of performance in large industrial models, based on the above-described system, including the following specific steps: Step S1: Collect data including industrial field time-series sensor data, production logs, process parameters, quality inspection results, and equipment alarms. Calculate the distribution difference between the monitored input data and the training baseline through statistical testing and feature distribution distance calculation. Quantify the complexity of the current industrial task from four dimensions: sample information entropy, feature manifold dimension, inference error entropy, and task output space dimension. Monitor two architecture operation indicators: network activation saturation and gradient utilization. Use an active learning strategy to screen high-value samples and generate difficulty distribution labels. Output evolution trigger signal, architecture capacity saturation signal, task complexity score, architecture saturation index, and the selected high-value sample subset and sample difficulty distribution labels. Step S2: Calculate the optimal capacity range of the network based on the task complexity score and architecture saturation index. When the capacity deviates from the threshold, start architecture adjustment. Add computing units in the specified layer of the Transformer encoder with attention head and feedforward network sub-layer as the granularity, in the form of residual bypass. Identify redundant units based on gradient contribution decay rate and activation frequency and perform structured pruning. After completing small sample fine-tuning, output the reconstructed network structure. Step S3: Initiate incremental training, calculate the gradient mutual information and propagation path weights between parameters of each layer, identify key gradient flow paths and lock the corresponding parameters, adjust the learning rate layer by layer according to the gradient features of each layer, and integrate gradient pruning and weight decay mechanisms; encode the key training hyperparameters into an evolutionary population, calculate the individual fitness online and perform evolutionary operations during the iteration process, and feed back the optimal hyperparameter configuration to the training stage; synchronously establish a set of knowledge anchor points for each task, update the anchor point mapping when the architecture changes, apply regularization constraints according to the anchor point offset during training, and mix in the replay feature flow to participate in training in proportion; Step S4: Set two evaluation indicators: neural network performance and industrial business performance. Use a combination of online shadow traffic parallel testing and offline benchmark testing to complete model testing, calculate the rate of change and confidence of each indicator, and output the evaluation results and rollback suggestions. Step S5: Combining performance evaluation data, evolution trigger signals, architecture capacity saturation signals, and available computing resources, generate an evolution execution strategy using a deep reinforcement learning algorithm. After two security checks, deploy the model online, maintain the evolution version chain, generate traceable logs, and issue control instructions.

[0089] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0090] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A self-evolving performance system for large industrial models, characterized in that, Includes the following modules: The data perception and evaluation module is used to collect industrial field data, monitor the distribution difference between input data and training baseline, quantify the complexity of industrial tasks, monitor architecture operation indicators, screen high-value samples, and generate sample difficulty distribution labels. The incremental architecture refactoring module is used to calculate the optimal capacity range of the network. It adds computing units through residual bypass and performs structured pruning. After fine-tuning with a small number of samples, it outputs the refactored network structure. The gradient increment optimization module is used to identify the critical path of the gradient flow and lock the corresponding parameters, adjust the learning rate in layers, integrate gradient clipping and weight decay mechanisms, and output the gradient and parameter update status. The hyperparameter evolution optimization module is used to perform evolution optimization on the hyperparameters of incremental training and feed back the optimal hyperparameter configuration to the gradient incremental optimization module. The knowledge forgetting suppression module is used to establish and update the set of task knowledge anchors, and suppress knowledge forgetting by mixing knowledge anchoring regularization terms with playback feature streams; The dual-dimensional closed-loop evaluation module is used to perform online and offline combined evaluation from both neural network and industrial business dimensions, and output evaluation results and rollback recommendations. The evolution security management module is used to generate evolution execution policies, perform security checks, maintain the evolution version chain, and issue management instructions.

2. The industrial large-scale model performance self-evolution system according to claim 1, characterized in that, The industrial field data collected by the data perception and evaluation module includes time-series sensor data, production logs, process parameters, quality inspection results, and equipment alarms. The data perception and evaluation module monitors the distribution difference between the input data and the training baseline by using a data distribution drift detection operator, statistical testing, and feature distribution distance calculation. When the drift exceeds a set threshold, an evolution trigger signal is output. The data perception and evaluation module quantifies the complexity of the industrial task from four dimensions: sample information entropy, feature manifold dimension, inference error entropy, and task output space dimension. It calculates the model capacity requirement coefficient and monitors network activation saturation and gradient utilization to obtain the architecture saturation index. When the architecture saturation index exceeds a set threshold, it outputs an architecture capacity saturation signal. The data perception and evaluation module uses an active learning strategy to filter a subset of samples and outputs the evolution trigger signal, the architecture capacity saturation signal, the industrial task complexity score, the architecture saturation index, the subset of samples, and the sample difficulty distribution label to the architecture progressive reconfiguration module.

3. The industrial large-scale model performance self-evolution system according to claim 2, characterized in that, The incremental architecture refactoring module calculates the optimal network capacity range based on the industrial task complexity score and the architecture saturation index. When the existing network capacity is lower than the lower threshold of the optimal network capacity range, incremental growth is initiated; when it is higher than the upper threshold of the optimal network capacity range, structured pruning is initiated. The progressive growth is carried out at a designated layer of the Transformer encoder, with attention heads and feedforward network sub-layers as the granularity, and new computation units are added through the residual bypass. The new computation units adopt an orthogonal initialization strategy, and after being connected to the original computation flow in the form of low-weight bypass, the bypass weight ratio is gradually increased, and they are smoothly integrated into the main network. The structured pruning identifies and removes redundant units based on gradient contribution decay rate and activation frequency, and the released computing power quota is preferentially allocated to the newly added computing units. After the architecture is adjusted, the small sample fine-tuning is performed, and the reconstructed network structure is output.

4. The industrial large-scale model performance self-evolution system according to claim 3, characterized in that, Before incremental training, the gradient incremental optimization module calculates the gradient mutual information and gradient propagation path weights between network parameters, identifies the key path of the gradient flow and performs strong locking on the corresponding parameters on the path, and only opens the parameters on non-critical paths for gradient updates. The gradient increment optimization module adjusts the learning rate of each layer according to the gradient norm and gradient fluctuation coefficient. For layers located at the edge of the critical path of the gradient flow and with large gradient fluctuations, a decaying small learning rate is used, while for layers on non-critical paths and with stable gradients, an increasing large learning rate is used. The gradient increment optimization module integrates the gradient clipping and the weight decay mechanism, and adjusts the gradient clipping threshold and weight decay coefficient according to the noise intensity of industrial data to suppress gradient explosion and parameter drift. During training, the gradient states of each level are synchronously output to the hyperparameter evolution optimization module, and the output parameter update states are output to the knowledge forgetting suppression module.

5. The industrial large-scale model performance self-evolution system according to claim 4, characterized in that, The hyperparameter evolution optimization module encodes the learning rate coefficient, batch size, regularization coefficient, and gradient locking range ratio into individual chromosomes to form an initial population. The hyperparameter evolution optimization module embeds the evolution optimization into the batch update process. After each batch of gradient update is completed, the individual fitness is calculated based on the current batch training loss and historical task verification accuracy. Selection, crossover, and mutation operations are performed to generate the next generation of hyperparameter population, which is then applied to the training of the next batch. The optimal hyperparameter configuration is retained during the evolution process and fed back to the gradient increment optimization module.

6. The industrial large-scale model performance self-evolution system according to claim 5, characterized in that, The knowledge forgetting suppression module establishes a set of task knowledge anchors for each industrial task and records the activation patterns and weight distribution characteristics of the corresponding industrial task on key neurons of the network. When the architecture changes, the mapping update of the set of task knowledge anchors is completed according to the topology inheritance relationship and the neuron mapping rules. During incremental training, the knowledge forgetting suppression module calculates the offset of the parameter update relative to the set of knowledge anchor points of the historical task in real time. When the offset exceeds the set safety threshold, a knowledge anchoring regularization term is added to the loss function. The larger the offset, the higher the weight of the knowledge anchoring regularization term. The knowledge forgetting suppression module extracts the core feature stream of historical tasks through feature distillation and stores it in the experience replay buffer. During training, the mixing ratio of the replay feature stream is adjusted according to the data drift rate to form a streaming mixed training with new data.

7. The industrial large-scale model performance self-evolution system according to claim 6, characterized in that, The neural network performance metrics of the dual-dimensional closed-loop evaluation module include parameter update efficiency, architecture utilization, historical task knowledge retention rate, and gradient stability; the industrial business performance metrics include task accuracy, inference timeliness, robustness, and business value. Online evaluation compares the inference results of the model before and after evolution in parallel using shadow traffic, statistically analyzes performance changes, and monitors network operation indicators; offline evaluation tests are conducted based on industrial benchmark datasets to generate performance data for each component. After the assessment is completed, the rate of change and confidence level of each indicator are calculated, and the assessment results and the rollback recommendations are output to the evolution security management module.

8. The industrial large-scale model performance self-evolution system according to claim 7, characterized in that, The evolution security management module combines performance data, the evolution trigger signal, the architecture capacity saturation signal, and available computing resources, and outputs the evolution start determination result, evolution combination strategy, computing power allocation quota, and execution time window through deep reinforcement learning. The evolution security management module sets up two checkpoints: neural network robustness verification and industrial business security verification. The neural network robustness verification sequentially undergoes weight perturbation testing, feature boundary testing, and adversarial gradient stability testing. The industrial business security verification sequentially undergoes boundary condition testing, abnormal input testing, and operational constraint compliance testing. The evolved model can only be put into online operation after passing both checkpoints. The evolution security management module maintains the evolution version chain, supports historical version rollback, generates traceable logs, and issues management instructions.

9. A method for self-evolution of performance in a large industrial model, based on the system described in any one of claims 1-8, characterized in that, The specific steps include the following: Collect industrial field data, monitor the distribution difference between input data and training baseline, quantify the complexity of industrial tasks, monitor architecture operation indicators, screen high-value samples and generate sample difficulty distribution labels, and output evolution trigger signals and architecture capacity saturation signals. The optimal capacity range of the network is calculated, new computing units are added through residual bypass and structured pruning is performed, and the network structure is reconstructed by fine-tuning the output with a small number of samples. Initiate incremental training, identify the critical path of gradient flow and lock the parameters, adjust the learning rate in layers, and simultaneously perform hyperparameter evolution optimization and knowledge forgetting suppression. The assessment is conducted both online and offline from the perspectives of neural networks and industrial business, and the assessment results and rollback recommendations are output. Generate an evolution execution strategy, go online after security verification, maintain the evolution version chain, and issue control instructions.

10. The method for self-evolution of performance in a large industrial model according to claim 9, characterized in that, In the incremental training process, the gradient mutual information and gradient propagation path weights between parameters are first calculated to lock the parameters corresponding to the key path of the gradient flow. Then, the learning rate is adjusted according to the hierarchical gradient feature differentiation. The training hyperparameters are encoded into an evolutionary population, and the evolutionary optimization is performed iteratively in batches. At the same time, the knowledge anchoring regularization term constraint and the mixing of the replay feature flow are used to suppress the forgetting of historical knowledge.

Images