Intelligent Control Method, Device and Chip Based on Processing Process Management

Abstract

The present invention relates to the field of intelligent control technology, and discloses an intelligent control method, device and chip based on processing process management. The method includes: setting an anomaly detection instruction through an intelligent control chip and writing it into a control status register to create an anomaly detection infrastructure; constructing a data acquisition protocol, and performing a feature extraction operation on the processing process data to obtain multi-dimensional feature training data; constructing a first anomaly detection sub-model for multiple processing stages, writing pre-trained parameters into the first anomaly detection sub-model to obtain multiple initialized anomaly detection models; deploying the multiple initialized anomaly detection models to the intelligent control chip for model parallel training to obtain multiple second anomaly detection sub-models; calculating the sub-model weight coefficients of the multiple second anomaly detection sub-models, and performing optimized integration and incremental learning to obtain a processing anomaly detection model. The present invention ensures the continuous optimization and performance improvement of the anomaly detection model for processing process management.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent control, and in particular, to an intelligent control method, device, and chip based on processing process management. Background Art

[0002] With the intelligent development of industrial manufacturing, higher requirements are put forward for the accuracy and real-time performance of anomaly detection in processing process management. Traditional anomaly detection methods mainly rely on a single monitoring model, which is difficult to cope with complex and changeable processing environments, and has low utilization efficiency of computing resources, unable to meet the needs of real-time monitoring.

[0003] In the actual production process, due to the complexity of processing procedures and the diversity of process parameters, it is difficult for a single detection model to comprehensively capture the anomaly characteristics of different processing stages. At the same time, the existing detection methods lack an effective parameter sharing mechanism, resulting in low model training efficiency and difficulty in adapting to dynamically changing processing environments. Summary of the Invention

[0004] The present invention provides an intelligent control method, device, and chip based on processing process management, which ensures the continuous optimization and performance improvement of the anomaly detection model for processing process management.

[0005] In a first aspect, the present invention provides an intelligent control method based on processing process management, and the intelligent control method based on processing process management includes:

[0006] Set an anomaly detection instruction through an intelligent control chip and write it into a control status register to create an anomaly detection infrastructure;

[0007] Construct a data acquisition protocol according to the anomaly detection infrastructure, and perform a feature extraction operation on the processing process data to obtain multi-dimensional feature training data;

[0008] Based on the multi-dimensional feature training data, construct a first anomaly detection sub-model for multiple processing stages, and write pre-trained parameters into the first anomaly detection sub-model to obtain multiple initialized anomaly detection models;

[0009] Deploy the multiple initialized anomaly detection models to the intelligent control chip for model parallel training to obtain multiple second anomaly detection sub-models;

[0010] Calculate the sub-model weight coefficients of the multiple second anomaly detection sub-models, and perform optimization integration and incremental learning to obtain a processing anomaly detection model.

[0011] In a second aspect, the present invention provides an intelligent control device based on processing process management, and the intelligent control device based on processing process management includes:

[0012] A creation module is used to set an exception detection instruction through an intelligent control chip and write it into a control status register, thereby creating an exception detection infrastructure;

[0013] A feature extraction module is used to construct a data acquisition protocol according to the exception detection infrastructure, and perform feature extraction operations on the processing process data to obtain multi-dimensional feature training data;

[0014] A pre-training module is used to construct first exception detection sub-models for multiple processing stages based on the multi-dimensional feature training data, and write pre-training parameters into the first exception detection sub-models to obtain multiple initialized exception detection models;

[0015] A parallel training module is used to deploy the multiple initialized exception detection models to the intelligent control chip for model parallel training to obtain multiple second exception detection sub-models;

[0016] An optimization integration module is used to calculate sub-model weight coefficients of the multiple second exception detection sub-models, and perform optimization integration and incremental learning to obtain a processing exception detection model.

[0017] A third aspect of the present invention provides a chip, in which instructions are stored, and when it runs on a computer, it causes the computer to execute the above-mentioned intelligent control method based on processing process management.

[0018] In the technical solution provided by the present invention, by setting an exception detection instruction in an intelligent control chip and configuring an exception detection circuit, an infrastructure that supports two detection modes of data redundancy and time redundancy is constructed, which improves the reliability and flexibility of exception detection. Multi-dimensional feature training data is used to construct multi-stage exception detection sub-models, and a shared parameter layer is set, which reduces the redundancy of model parameters and optimizes the utilization efficiency of computing resources. A training strategy of data parallelism and model parallelism is introduced. Through dynamically adjusting training parameters and a cross-validation mechanism, the exception detection model has stronger generalization ability. An integration strategy based on resource allocation and energy efficiency optimization, combined with a dynamic adjustment mechanism of weight coefficients, realizes the reasonable allocation of computing resources and the optimization of energy efficiency. Through an incremental learning and parameter update mechanism, a complete model optimization and performance evaluation system is established, which ensures the continuous optimization and performance improvement of the exception detection model. A management method of hierarchical storage and parameter indexing is adopted, which improves the access efficiency of model parameters and facilitates the dynamic update and maintenance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the accompanying drawings required for the description of the embodiments. Obviously, the accompanying drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can also be obtained based on these drawings.

[0020] Figure 1 It is a schematic diagram of the steps of the intelligent control method based on processing process management in the embodiments of the present invention;

[0021] Figure 2 It is a schematic diagram of the structure of the intelligent control device based on processing process management in the embodiments of the present invention. Specific embodiments

[0022] The embodiments of the present invention provide an intelligent control method, device and chip based on processing process management. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above accompanying drawings are used to distinguish similar objects, and do not necessarily need to be used to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments described here can be implemented in an order other than that illustrated or described here. In addition, the term "comprising" or "having" and any variation thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device comprising a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0023] For ease of understanding, the following describes the specific process of the embodiments of the present invention. Please refer to Figure 1 , an embodiment of the intelligent control method based on processing process management in the embodiments of the present invention includes:

[0024] Step S1: Set an exception detection instruction through the intelligent control chip and write it into the control status register to create an exception detection infrastructure;

[0025] It can be understood that the execution subject of the present invention can be an intelligent control device based on processing process management, or a terminal or a server. Specifically, it is not limited here. The embodiments of the present invention are described by taking the server as the execution subject as an example.

[0026] Specifically, an exception detection instruction set and a control status register address space are added to the intelligent control chip to expand the instruction system architecture of the existing chip, enabling it to support the functional requirements of exception detection. By expanding the instruction system architecture, new instruction encodings are created for exception detection tasks, the functions of each instruction are clarified, and sufficient address space is allocated for the control status register to accommodate all necessary configuration data. After completing the expansion of the instruction system architecture, data redundancy detection thresholds and time redundancy detection thresholds are set according to the new architecture to form an exception detection parameter matrix. The generation of this matrix needs to comprehensively consider the data characteristics during the processing and the distribution law of potential exceptions to ensure that the set thresholds have practical significance and detection capabilities. The exception detection parameter matrix is written into the control status register to form exception detection configuration data. These configuration data not only cover all the detection thresholds in the parameter matrix but also include the definitions of parameters such as the detection task priority and enabling conditions in the control status register. To adapt to different processing scenarios and requirements, the exception detection configuration data is divided into operating modes, specifically into a data redundancy mode and a time redundancy mode. In the data redundancy mode, the consistency of the processing data among different sampling points is concerned, and by configuring multiple groups of data comparison logics, the real-time collected data is compared and analyzed to detect existing exceptions. In the time redundancy mode, the focus is on the time continuity during the processing, and by configuring the timing comparison logic, whether there is an unreasonable time deviation between the operations and results is detected. After completing the division of the operating modes, operating mode configuration parameters are generated based on these modes, and a data comparison unit and a result output unit are configured. The core function of the data comparison unit is to perform real-time comparison on the data collected during the processing, including comparing the differences of the same data among different sampling points through hardware logic or comparing the redundancy relationship of different data channels. The task of the result output unit is to process and output the results generated by the comparison logic. Whether it is through signal feedback to control the chip or through a data interface to transfer to the upper-level system, its design must meet the requirements of real-time and accuracy. By cascading the data comparison unit and the result output unit, an exception detection circuit is formed. The function of the exception detection circuit is verified. The process of function verification includes simulating various possible exception situations in different processing scenarios to detect whether the exception detection circuit can accurately identify and output the expected results. The results of the function verification are fed back into the intelligent control chip to optimize the configuration parameters of the exception detection instruction set and the exception detection circuit. If the verification results show insufficient detection performance or misjudgment, relevant parameters or logics are adjusted, such as redefining the detection thresholds, optimizing the comparison logic, or expanding the coverage of the operating modes. Through the above iterative optimization process, an exception detection infrastructure is finally formed.

[0027] Step S2: Construct a data acquisition protocol according to the exception detection infrastructure, and perform feature extraction operations on the processing process data to obtain multi-dimensional feature training data;

[0028] Specifically, parameters related to data sampling are set based on the anomaly detection infrastructure, including key metrics such as sampling frequency, sampling time window length, and sampling accuracy. By analyzing the dynamic change characteristics in the processing, the sampling frequency is reasonably set to ensure that the collected data has sufficient time resolution, while the length of the sampling time window determines the time range to be observed, and the sampling accuracy is related to the balance between sensor capabilities and data volume. After the parameter setting is completed, the sampling parameters are written into the data acquisition control unit in the intelligent control chip to form the basic configuration of data acquisition. Constraint conditions are set for the basic configuration of data acquisition. By combining the actual requirements and process characteristics of the processing, hierarchical constraints are imposed on the data acquisition range, frequency, and accuracy, and data acquisition levels are divided for different types of processing parameters. Key parameters with high priority, such as temperature, pressure, or speed, require higher sampling frequencies and accuracies, while secondary parameters use lower sampling requirements to reduce the system load. After the acquisition level division is completed, data filtering thresholds are set for each acquisition level to eliminate the outliers or noise data that appear, ensuring that the collected data has high quality and usability. These constraints and level divisions are finally integrated into the data acquisition protocol, which clarifies the acquisition methods, priorities, and preliminary rules for data processing of various parameters. The data acquisition protocol is imported into the data acquisition module of the intelligent control chip, and multi-dimensional sampling of the processing data is completed through the intelligent control chip to form a preliminary raw data matrix. The implementation of multi-dimensional sampling includes simultaneously collecting data signals from multiple sensors, such as temperature sensors, vibration sensors, and flow meters, to ensure comprehensive recording of multi-faceted dynamic information in the processing. Denoising and normalization processing are performed on the raw data matrix to obtain preprocessed data. Denoising processing uses filtering or signal smoothing algorithms to eliminate high-frequency noise or abrupt outliers in the data, while normalization processing unifies the data into a dimension range convenient for comparison, such as by normalization or Z-Score transformation, mapping all parameters to the same order of magnitude. Time-domain features, frequency-domain features, and statistical features of the preprocessed data are extracted respectively. Time-domain feature extraction focuses on the change patterns of the data in the time dimension, such as calculating the mean, maximum, minimum, mean square deviation, etc., to capture the amplitude characteristics and change trends of the signal. Frequency-domain feature extraction uses Fourier transform or wavelet transform to convert the time signal into a frequency signal and extracts features such as spectral density, main frequency, and band energy, which is suitable for analyzing the dynamic characteristics of vibration or periodic change data. Statistical feature extraction focuses on the distribution characteristics of the data, such as skewness, kurtosis, and probability density distribution, to help identify whether the signal has an abnormal distribution or an atypical change pattern. The time-domain features, frequency-domain features, and statistical features are combined to form multi-dimensional feature training data.

[0029] Step S3: Construct first anomaly detection sub-models for multiple processing stages based on multi-dimensional feature training data, and write the pre-trained parameters into the first anomaly detection sub-models to obtain multiple initialized anomaly detection models;

[0030] Specifically, according to the order of processing procedures, the multi-dimensional feature training data is reasonably divided, and the data in different stages is allocated to the corresponding processing stage data sets, forming multiple processing stage feature data sets. Considering the uniqueness of each processing stage, for example, some procedures focus on the changes in pressure and temperature, while others focus on the monitoring of vibration signals and processing time. By specifically allocating multi-dimensional feature data, it is ensured that the subsequent constructed anomaly detection model can be optimized according to the characteristics of each processing stage. During the data division process, ensure the integrity and balance of the data set to avoid a decline in model performance due to uneven data volume distribution or overly skewed feature distribution. Based on the computing resources of the intelligent control chip, determine the structure of the sub-model for each processing stage feature data. Dynamically adjust the number of network layers and neurons of the sub-model according to the complexity and computing requirements of the feature data in each stage. For processing stages with higher feature dimensions or faster dynamic changes, design deeper network structures to capture complex non-linear relationships; while for relatively stable stages, choose shallower network structures to reduce computing overhead. Through the resource-adaptive sub-model design, make the most of the computing power of the intelligent control chip, and at the same time ensure that the structure of each sub-model adapts to the actual needs, generating multiple first anomaly detection sub-models. Conduct parameter space planning for the multiple first anomaly detection sub-models, dividing the parameter space into an independent parameter area and a shared parameter area. The independent parameter area is responsible for processing the specific features of each processing stage, while the shared parameter area is used to capture the correlation between adjacent processing stages. Establish a parameter sharing channel between the first anomaly detection sub-models of adjacent processing stages, enabling these stages to effectively cooperate in the parameter space. Through this design, utilize the continuity characteristics of the processing procedures to partially transfer the knowledge of the previous stage to the next stage, improving the overall efficiency and accuracy of anomaly detection. Based on the division result of the parameter space, generate a parameter space distribution map. Construct a shared parameter layer according to the parameter space distribution map and insert the shared parameter layer into the network structures of multiple first anomaly detection sub-models. The core role of the shared parameter layer is to establish a parameter transfer matrix to achieve efficient parameter sharing between adjacent stages. The parameter transfer matrix defines the specific transfer method and weight ratio of the shared parameters. For example, through a specific linear combination or non-linear mapping mechanism, the shared parameters not only retain the characteristics of the previous stage but also can adapt to the feature distribution of the current stage. Initialize the pre-training parameters of the anomaly detection model with a shared layer. The goal of pre-training is to provide a good initial state for the model according to the feature distribution of each processing stage, accelerating the subsequent training process. By analyzing the distribution characteristics of the feature data in each stage, determine the pre-training parameter values of each shared layer and the independent parameter area, and generate a pre-training parameter matrix based on these values. The setting of the pre-training parameters is based on empirical rules, data statistical analysis, or the transfer results of pre-trained models, ensuring its strong adaptability to the anomaly detection tasks in each stage.Write the pre-trained parameter matrix into the anomaly detection model with shared layers to complete the parameter loading and model compilation processes. When loading the parameters, map the independent part and the shared part of the parameter matrix to the corresponding network layers respectively, and verify the correctness and adaptability of the parameters. During the model compilation process, optimize the computational graph and execution logic of the model for efficient operation on the intelligent control chip. After the optimization and loading operations, multiple initialized anomaly detection models are formed.

[0031] Step S4: Deploy the multiple initialized anomaly detection models to the intelligent control chip for model parallel training to obtain multiple second anomaly detection sub-models;

[0032] Specifically, the training tasks for multiple initialization anomaly detection models are reasonably partitioned. According to the amount of training data required by each model and its computational complexity, the training data is allocated to different computing units in batches. At the same time, training parameters are set for each computing unit, including the initial learning rate, gradient update rule, batch size, etc., to form a data parallel training scheme. Based on the data parallel training scheme, the training data in different batches is processed in parallel. Each computing unit independently processes the allocated data and completes the calculations of forward propagation and error backpropagation within a local scope. After each computing unit completes the training task for the current batch of data, the gradient information of all computing units is aggregated, and the training results of each sub-model are synchronized through global gradient updates to achieve the consistency and global optimization of the sub-model parameters. In this process, the gradient aggregation algorithm performs weighted integration on the gradient information of different computing units to ensure that the direction and amplitude of parameter updates can correctly guide the optimization of the model. Through multiple rounds of parallel processing and gradient aggregation, a data parallel training model is formed. The data parallel training model is split in terms of model structure to make full use of the hardware architecture characteristics of the intelligent control chip. According to the network structure of each sub-model, sub-models with different structures are allocated to independent computing units, enabling each sub-model to complete efficient training tasks on independent hardware resources. At the same time, to ensure the collaborative work between sub-models, a communication link between sub-models is established, and real-time exchange of parameters or intermediate results is achieved through a high-speed communication channel. For example, if adjacent-stage sub-models need to share certain parameter update information or gradient feedback, the communication link is required to support low-latency and high-bandwidth data transmission. After the model structure is split and the communication link is established, a model parallel training scheme is formed. Based on the model parallel training scheme, the specific training tasks of the sub-models are executed. The learning rate and batch size of each sub-model are dynamically adjusted to adapt to the dynamic changes during the training process. For example, when the performance metrics of the model fluctuate at a certain stage, the learning rate is reduced to avoid gradient oscillation, and when the model is approaching convergence, the batch size is gradually increased to accelerate the convergence process. At the same time, the performance metrics during the training process are recorded in real time, such as training loss, accuracy, and gradient norm, etc. By dynamically adjusting the training parameters and monitoring the training process, it is ensured that each sub-model maintains an efficient and stable optimization state during the training process, improving the overall performance of the model. After the sub-model training is completed, cross-validation is performed on the model training data generated during the training process. Cross-validation effectively detects problems of overfitting or underfitting by evaluating the performance of the model on an independent data set. According to the cross-validation results, the training parameters are optimized and adjusted, including resetting the learning rate, updating the regularization parameters, and re-dividing the allocation scheme of the training data. The optimized parameters are written into each sub-model, thereby improving the anomaly detection ability and generalization performance of the model. Through the optimization and validation process, multiple second anomaly detection sub-models are finally generated.

[0033] Step S5: Calculate the sub-model weight coefficients of multiple second anomaly detection sub-models, and perform optimized integration and incremental learning to obtain a processing anomaly detection model.

[0034] Specifically, formulate a resource allocation strategy based on the operation status data of multiple second anomaly detection sub-models. By monitoring the operation status data of each sub-model, combining the utilization rate of computing resources and the occupancy rate of storage resources, analyze the real-time load conditions of each sub-model, and allocate appropriate resource weights for them. During this process, dynamically adjust the allocation ratio of resources to ensure the efficient utilization of computing resources and the reasonable occupancy of storage resources, and generate a sub-model resource configuration plan. Based on the sub-model resource configuration plan, perform a computational load analysis on multiple second anomaly detection sub-models. Through quantitative analysis of the energy consumption index and computing resource consumption index of each sub-model, evaluate its operation efficiency. Input these indicators into an energy efficiency evaluation function to calculate the energy efficiency optimization parameters of each sub-model. According to the energy efficiency optimization parameters, perform an integrated weight calculation on multiple second anomaly detection sub-models. Considering the detection accuracy and energy efficiency optimization parameters of the sub-models comprehensively, input these factors into a weight calculation matrix, and obtain a weight allocation table for the sub-models through normalization processing. Based on the weight allocation table, establish an optimized integration strategy to combine multiple second anomaly detection sub-models into an overall integrated detection model. Combine the sub-models in stages according to the order of the processing stages, so that the anomaly detection tasks in each processing stage can be undertaken by the most suitable sub-model. At the same time, to improve the dynamic adaptation ability of the integrated model, establish a model switching mechanism and set triggering conditions. When there are phased changes during the processing or the anomaly detection requirements are adjusted, the switching mechanism dynamically selects the optimal sub-model according to the triggering conditions to perform the detection task, realizing the flexibility and robustness of the integrated model. After the dynamic integrated model is constructed, perform incremental learning on it. By analyzing the feature distribution of the newly added data samples, formulate a parameter update strategy to enable the model to adapt to the new changes during the processing. During the incremental learning process, gradually update the model parameters for the newly added samples, and simultaneously evaluate the performance of the updated model in real time to ensure that the model maintains a high detection accuracy and generalization ability when processing new data. The introduction of incremental learning enables the dynamic integrated model to have an adaptive ability, so that it can continuously optimize its detection ability as the processing process changes. Divide the storage unit for the updated anomaly detection model. Write the structural parameters and weight parameters of the model into the fixed storage area and the dynamic storage area respectively to form a clear parameter storage structure. The fixed storage area is used to save the stable model structure, while the dynamic storage area is used to store the weight parameters that are continuously updated with incremental learning. At the same time, by establishing a parameter index table, quickly locate and access the required model parameters to improve the management efficiency of the model. Finally, generate a complete processing anomaly detection model.

[0035] The model parameters of the dynamic integration model are classified into two categories according to their functions: structural parameters and weight parameters. Structural parameters mainly define the topological structure of the model, such as the number of layers of a neural network, the node connection method, etc. These parameters have a decisive impact on the overall framework of the model. Weight parameters, on the other hand, are the specific values learned by the model during the training process and are used for feature transformation and prediction of input data. After the parameters are classified, according to the sub-model weight distribution table, all parameters are sorted according to their importance to generate a parameter update priority table. This priority table is the core basis for model updates. By preferentially updating the parameters that have a greater impact on the model performance, the efficiency of incremental learning is significantly improved, and the interference of irrelevant parameter adjustments on the model performance is reduced. After the parameter sorting is completed, a parameter update strategy is constructed based on the parameter update priority table. The new data samples are preprocessed and classified and labeled, and appropriate labels are assigned to each sample according to the characteristics of the processing stage. The labeled data samples are assigned to the corresponding second anomaly detection sub-model according to the processing stage to form an incremental training data set. The construction of this data set needs to consider the balance and representativeness of the sample distribution to ensure that each sub-model can obtain sufficient training data while avoiding being overly biased towards certain specific processing stages. After the data distribution is completed, the incremental training data is input into the corresponding sub-model for parameter update. Based on the incremental training data set, the model parameters are updated layer by layer in the order of parameter importance. During the update process, the change amount of each parameter and the change index of the model performance are recorded to comprehensively evaluate the impact of parameter adjustment on the model performance. For example, the performance change is quantified by calculating indicators such as the detection accuracy, precision, and recall rate of the updated model, and the change in resource consumption is recorded to evaluate the computational cost of the update. These records constitute the parameter update results. After the parameter update is completed, a performance evaluation of the parameter update results is carried out. The performance evaluation calculates the difference in detection accuracy before and after the update to determine whether the model achieves the expected effect on the new data. And the changes in resource consumption of the updated model are evaluated, such as calculating the resource utilization rate, storage resource occupancy rate, and runtime latency, etc., to ensure that the model update does not cause excessive resource overhead while improving the performance. Based on the model evaluation results, a parameter rollback judgment is performed. By comparing the performance evaluation indicators with the preset thresholds, it is identified which parameter adjustments lead to a performance decline. If it is found that the update of certain parameters makes the detection performance of the model lower than the expected threshold, a rollback operation is performed on these parameters to restore them to the state before the update. The implementation of the parameter rollback combines the parameter change amount and performance change index saved in the historical record to ensure that the rollback operation can accurately locate the problem parameters and effectively solve the performance decline problem. After the parameter rollback is completed, a parameter optimization plan is generated, and the optimized parameters are written into the dynamic integration model to improve the stability and adaptability of the model. Through the above process, the updated anomaly detection model is finally obtained.This model can achieve continuous performance improvement under the influence of newly added data samples, and maintain high reliability and resource utilization rate with the support of performance evaluation and parameter optimization.

[0036] In the embodiments of the present invention, by setting an exception detection instruction and configuring an exception detection circuit in the intelligent control chip, an infrastructure supporting two detection modes of data redundancy and time redundancy is constructed, improving the reliability and flexibility of exception detection. A multi-stage exception detection sub-model is constructed using multi-dimensional feature training data, and a shared parameter layer is set, reducing the redundancy of model parameters and optimizing the utilization efficiency of computing resources. The training strategies of data parallelism and model parallelism are introduced. Through dynamically adjusting training parameters and cross-validation mechanisms, the exception detection model has stronger generalization ability. Based on the integrated strategy of resource allocation and energy efficiency optimization, combined with the dynamic adjustment mechanism of weight coefficients, reasonable allocation of computing resources and optimization of energy efficiency are achieved. Through incremental learning and parameter update mechanisms, a complete model optimization and performance evaluation system is established, ensuring the continuous optimization and performance improvement of the exception detection model. The management method of hierarchical storage and parameter indexing is adopted, improving the access efficiency of model parameters and facilitating the dynamic update and maintenance of the model.

[0037] In a specific embodiment, the process of executing step S1 may specifically include the following steps:

[0038] Add an exception detection instruction set and a control status register address space to the intelligent control chip to obtain an extended instruction system architecture, and set data redundancy detection thresholds and time redundancy detection thresholds according to the extended instruction system architecture to generate an exception detection parameter matrix;

[0039] Write the exception detection parameter matrix into the control status register to obtain exception detection configuration data, and perform operation mode division on the exception detection configuration data to obtain a data redundancy mode and a time redundancy mode;

[0040] Configure multiple groups of data comparison logics for the data redundancy mode and configure a timing comparison logic for the time redundancy mode to obtain operation mode configuration parameters;

[0041] Configure a data comparison unit and a result output unit based on the operation mode configuration parameters, and cascade-connect the data comparison unit and the result output unit to obtain an exception detection circuit;

[0042] Perform functional verification on the exception detection circuit to obtain a circuit verification result, and feedback the circuit verification result to the intelligent control chip. Adjust the exception detection instruction and the configuration parameters of the exception detection circuit according to the circuit verification result to obtain an exception detection infrastructure.

[0043] Specifically, an exception detection instruction set is added to the intelligent control chip. The exception detection instruction set is the core of the intelligent chip control system. By expanding the instruction system architecture, dedicated instruction support is provided for the exception detection function. For example, a new instruction DETECT_ERR is added, which is used to trigger data comparison and exception judgment operations. And the address space of the control status register is expanded to store the configuration parameters and real-time status information of exception detection. Assume that the address range of the expanded control status register is , where and represent the start address and end address of the register respectively. Then the capacity of the expanded register is:

[0044] ;

[0045] Among them, represents the capacity of the register, which can accommodate all exception detection parameters and operation status data. According to the expanded instruction system architecture, set the data redundancy detection threshold and time redundancy detection threshold. The role of these thresholds is to define the specific judgment conditions for exception detection. For example, the data redundancy detection threshold is used to judge the consistency between multiple groups of redundant data, and the calculation formula is:

[0046] ;

[0047] Among them, represents the th sampling data, represents the reference data, represents the size of the data group. When exceeds the set threshold, it is determined that there is data exception. Similarly, the time redundancy detection threshold is defined as the time interval error between multiple operations:

[0048] ;

[0049] Among them, represents the time interval between two consecutive operations. By combining these thresholds, an exception detection parameter matrix is generated:

[0050] ;

[0051] The matrix contains all possible combinations of detection thresholds. The exception detection parameter matrix After writing to the control status register, abnormal detection configuration data is obtained. To optimize the detection efficiency, the abnormal detection configuration data is divided into operating modes, specifically including a data redundancy mode and a time redundancy mode. In the data redundancy mode, the consistency of multiple groups of data is compared, which is achieved by configuring the comparison logic for multiple groups of data. For example, for groups of redundant data , the comparison logic is defined as:

[0052]

[0053] . In the time redundancy mode, the deviation of time intervals is compared, and the timing comparison logic is defined as:

[0054] ;

[0055] where represents the preset threshold for time redundancy detection. Based on the operating mode configuration parameters, the data comparison logic and the result output logic are configured into the data comparison unit and the result output unit, and the cascade of the two is realized through a hardware connection method to form an abnormal detection circuit. For example, the data comparison unit realizes the real-time comparison of input data through Boolean logic, and the result output unit is responsible for feeding back the comparison result to the control chip in the form of a signal. When the input data meets the detection conditions, the output unit will trigger an alarm signal to indicate a system abnormality. After completing the design of the abnormal detection circuit, its function is verified. The function verification is carried out through simulation tests or actual operation tests. For example, a set of preset test data is input, and some of the data are deliberately set as abnormal values, and the accuracy and reliability of the circuit are verified by observing whether the abnormal detection circuit can accurately output the expected abnormal signal. If the verification result shows insufficient detection accuracy or false alarms, the abnormal detection instruction set and configuration parameters are adjusted according to the feedback result. For example, the threshold range of data comparison is redefined or the sensitivity of the timing comparison logic is optimized to improve the circuit performance. After multiple rounds of verification and adjustment, an abnormal detection infrastructure is finally formed.

[0056] In a specific embodiment, the process of executing step S2 may specifically include the following steps:

[0057] Set data sampling parameters based on the abnormal detection infrastructure, and write the sampling frequency, sampling time window length, and sampling accuracy of the data sampling parameters into the data acquisition control unit to obtain the basic configuration of data acquisition;

[0058] Set the constraint conditions for the basic configuration of data acquisition, divide the data acquisition levels according to the types of processing parameters, and set the data filtering thresholds for each data acquisition level to obtain the data acquisition protocol;

[0059] Import the data acquisition protocol into the data acquisition module, perform multi-dimensional sampling on the processing process data, generate the original data matrix, and perform denoising and standardization processing on the original data matrix to obtain the preprocessed data;

[0060] Extract the time-domain features, frequency-domain features, and statistical features of the preprocessed data respectively, and combine the time-domain features, frequency-domain features, and statistical features into multi-dimensional feature training data.

[0061] Specifically, set the data sampling parameters based on the anomaly detection infrastructure to ensure that the collected processing process data can comprehensively reflect the dynamic behavior of the system. The data sampling parameters include key indicators such as sampling frequency, sampling time window length, and sampling accuracy. The sampling frequency refers to the number of samples collected per second and determines the highest signal frequency that the system can capture. According to the Nyquist theorem, the sampling frequency needs to satisfy is the highest frequency of the signal. The sampling time window length refers to the time range of each sampling and determines the length of the time series that the system can analyze. The calculation formula is:

[0062] ;

[0063] where is the number of samples obtained per sampling. The sampling accuracy represents the resolution of the sampled data. For example, the sensor output is 12-bit or 16-bit precision. The higher the precision, the more subtle signal changes can be captured, but at the same time, the data storage requirements increase. After these sampling parameters are written into the data acquisition control unit, the basic configuration of data acquisition is formed. Set the constraint conditions for the basic configuration of data acquisition to optimize the acquisition efficiency and improve the data quality. Based on the parameter characteristics in the processing process, classify the processing parameters according to importance and dynamics, and divide them into different data acquisition levels. For example, set key parameters such as temperature, pressure, and vibration signals as high priority, with higher sampling frequency and precision; while set some static or slowly changing parameters, such as ambient humidity or power consumption, as low priority, with lower sampling frequency and precision. After completing the acquisition level division, set the data filtering threshold for each level to eliminate noise or outliers. For example, use the Z-score method, and the calculation formula is:

[0064] ;

[0065] where, represents the standardized score of the sample , is the data mean, is the data standard deviation. When If it exceeds the set threshold (e.g., 3), the sample is determined as an outlier and filtered out. Through these operations, a complete data acquisition protocol including sampling levels and filtering rules is formed. After importing the designed data acquisition protocol into the data acquisition module, multi-dimensional sampling of the processing process data is performed to generate an original data matrix. The original data matrix has a shape of , where represents the types of parameters collected, represents the number of samples for each parameter. For example, for a system with three parameters: temperature, vibration, and pressure, and 1000 samples are taken for each parameter, the original data matrix is represented as:

[0066] ;

[0067] Among them, , and represent the values of temperature, vibration, and pressure at the th sampling point respectively. Denoising and normalization processing are performed on the original data to generate high-quality preprocessed data. For denoising, methods such as low-pass filtering or wavelet transform are used. For example, a low-pass filter is used to eliminate high-frequency noise, and the filter output is:

[0068] ;

[0069] Among them, are the filter coefficients, is the input signal, is the filtered output signal. For normalization processing, normalization or Z-score transformation is used to map all parameters to a unified numerical range. For example, for the normalization operation, the normalized data is:

[0070] ;

[0071] Among them, and are the minimum and maximum values of the data respectively. Through this step, a preprocessed data matrix is obtained. Based on the preprocessed data, time-domain features, frequency-domain features, and statistical features are extracted, and these features are combined into multi-dimensional feature training data. Time-domain feature extraction mainly analyzes the time-series changes of the signal, such as calculating the mean, variance, peak value, skewness, and kurtosis. Taking the mean as an example, its calculation formula is:

[0072] ;

[0073] Frequency-domain feature extraction converts the signal from the time domain to the frequency domain through the fast Fourier transform, and extracts spectral features from it, such as the main frequency, band energy, etc. For example, the frequency-domain signal Expressed as:

[0074] ;

[0075] Statistical feature extraction includes the distribution characteristics of data, such as median, quantile, etc. Combine time-domain features, frequency-domain features and statistical features to form a multi-dimensional feature training data matrix :

[0076] ;

[0077] Among them, 、 and respectively represent time-domain, frequency-domain and statistical features.

[0078] In a specific embodiment, the process of executing step S3 may specifically include the following steps:

[0079] According to the processing procedure sequence, allocate the multi-dimensional feature training data to different processing stage data sets to obtain multiple processing stage feature data;

[0080] Based on the computing resources of the intelligent control chip, determine the sub-model structure for each processing stage feature data, set the number of network layers and neurons of the sub-model to obtain multiple first anomaly detection sub-models;

[0081] Perform parameter space planning on multiple first anomaly detection sub-models, divide the parameter space into independent parameter areas and shared parameter areas, and establish parameter sharing channels between the first anomaly detection sub-models in adjacent processing stages to obtain a parameter space distribution map;

[0082] Construct a shared parameter layer according to the parameter space distribution map, insert the shared parameter layer into the network structure of multiple first anomaly detection sub-models, establish a parameter transfer matrix, and obtain an anomaly detection model with a shared layer;

[0083] Perform pre-training parameter initialization on the anomaly detection model with a shared layer, determine the pre-training parameter values according to the feature distribution of each processing stage, generate a pre-training parameter matrix, and write the pre-training parameter matrix into the anomaly detection model with a shared layer for parameter loading and model compilation to obtain multiple initialized anomaly detection models.

[0084] Specifically, according to the sequence of processing procedures, allocate the multi-dimensional feature training data to different processing stage data sets in order to construct an anomaly detection model separately for each stage. Assume that the processing process is divided into stages, and the multi-dimensional feature training data of each stage is denoted as ), then the overall training data is decomposed into:

[0085] ;

[0086] Among them, represents the feature data set of the stage, including time-domain features, frequency-domain features, and statistical features of a specific processing stage. Based on the computing resources of the intelligent control chip, the sub-model structure is determined for the feature data of each processing stage to ensure the adaptability of the model complexity to the chip resources. In the design of the neural network model, the number of network layers and the number of neurons in each layer are two key parameters. Assume that the dimension of the feature data of the stage is , then the number of neurons in the input layer is initially set to , the number of neurons in the output layer is 1 (used for binary classification to judge anomalies), and the number of hidden layers and neurons are dynamically adjusted according to the computing resources. For example:

[0087] ;

[0088] Among them, is the maximum number of neurons allowed by the chip resources, represents the number of neurons decreasing layer by layer. Taking the finish machining stage as an example, assume its feature dimension , the maximum number of neurons allowed by the resources , then the hidden layer is set to 3 layers, and the number of neurons is 32, 16, and 8 respectively, forming the first anomaly detection sub-model of the stage. After completing the sub-model design, the parameter space of multiple first anomaly detection sub-models is planned, and its parameter space is divided into an independent parameter area and a shared parameter area, so as to establish a parameter sharing channel between the sub-models of adjacent processing stages. Assume that the parameter set of the sub-model of the stage is , where represents the independent parameter, represents the shared parameter. The shared parameter is shared with the adjacent stage to capture the correlation between stages. For example:

[0089] ;

[0090] Among them, represents the adjustment amount of the shared parameter of the stage. In this way, a parameter space distribution diagram is generated to describe the independence and sharing relationship of the parameters of each stage. According to the parameter space distribution diagram, a shared parameter layer is constructed and inserted into the network structure of multiple first anomaly detection sub-models. The role of the shared parameter layer is to realize the sharing of parameters between adjacent stages through the parameter transfer matrix . Assume that the parameter transfer matrix The dimension of is , where

[0091] ;

[0092] Among them, and respectively represent the outputs of the shared layer in the th and th stages. Through this connection, the knowledge of the previous stage is passed to the next stage, improving the collaborative ability of the overall anomaly detection. Initialize the pre-trained parameters of the anomaly detection model with a shared layer to improve the initial performance of the model. The pre-trained parameter initialization determines the initial parameter values according to the feature distributions of each processing stage. For example, by calculating the mean and standard deviation of each feature, the parameters are initialized by normalization:

[0093] ;

[0094] Among them, represents a normal distribution with as the mean and as the variance. After writing the generated pre-trained parameter matrix into the anomaly detection model with a shared layer, the parameter loading and running environment optimization are completed through model compilation. The compilation process includes computational graph construction, gradient optimization rule setting, etc., to ensure that the model can run efficiently on the intelligent control chip. Through the above steps, multiple initialized anomaly detection models are finally obtained.

[0095] Among them, the first anomaly detection sub-model includes: dividing the multi-dimensional feature training data into an input data stream and a label data stream according to the time series, and constructing a feature extraction layer; the feature extraction layer includes 5 convolutional units, each convolutional unit consists of a two-dimensional convolutional layer, a batch normalization layer, and a ReLU activation function. The number of input channels of the first convolutional unit is the feature dimension of the multi-dimensional feature training data, the number of output channels is 64, the convolutional kernel size is 3×3, and the stride is 1; the number of input channels of the second to fifth convolutional units doubles in sequence, and the number of output channels is 128, 256, 512, and 1024 in sequence, the convolutional kernel size is 3×3 for all, and the stride is 2; constructing a time series modeling layer based on the output of the feature extraction layer, the time series modeling layer includes 3 bidirectional long short-term memory network units, each unit contains 256 hidden layer nodes, the input dimension of the first unit is 1024, the output dimension is 512, the input dimension of the second unit is 512, the output dimension is 256, and the input dimension of the third unit is 256, the output dimension is 128; sending the output of the time series modeling layer into the attention layer, the attention layer adopts a multi-head self-attention mechanism, including 8 attention heads, the query dimension, key dimension, and value dimension of each attention head are all 64, and the attention weight matrix is calculated through scaled dot-product attention to perform weighted aggregation on the time series features; constructing an anomaly detection layer based on the output of the attention layer, the anomaly detection layer includes 3 fully connected layers, the input dimension of the first fully connected layer is 128, the output dimension is 64, and the ReLU activation function is used; the input dimension of the second fully connected layer is 64, the output dimension is 32, and the ReLU activation function is used; the input dimension of the third fully connected layer is 32, the output dimension is 2, and the Softmax activation function is used; sending the output of the anomaly detection layer into the loss calculation layer, the loss calculation layer uses the cross-entropy loss function to calculate the difference between the prediction result and the label data stream, and combines the L2 regularization term to construct the overall loss function for model parameter optimization; constructing a backpropagation layer based on the overall loss function, the backpropagation layer uses the Adam optimizer for parameter update, the initial learning rate is set to 0.001, and the learning rate is reduced to 0.1 of the original every 50 training epochs, the momentum parameter β1 is set to 0.9, and β2 is set to 0.999; adding residual connections between the feature extraction layer and the time series modeling layer, between the time series modeling layer and the attention layer, and between the attention layer and the anomaly detection layer respectively, and adding layer normalization after each residual connection to obtain the complete structure of the first anomaly detection sub-model.

[0096] In a specific embodiment, the process of executing step S4 may specifically include the following steps:

[0097] Dividing the training tasks for multiple initialized anomaly detection models, allocating the training data to different computing units in batches, setting the training parameters of each computing unit, and obtaining a data parallel training scheme;

[0098] Based on the data parallel training scheme, parallel processing is performed on different batches of training data, the training results of each computing unit are aggregated by gradients, the sub-model parameters are updated, and a data parallel training model is obtained;

[0099] The data parallel training model is split in terms of model structure, sub-models with different structures are assigned to independent computing units, and communication links between sub-models are established to obtain a model parallel training scheme;

[0100] Based on the model parallel training scheme, sub-model training is performed, the learning rates and batch sizes of each sub-model are dynamically adjusted, and the performance metrics during the training process are recorded to obtain model training data;

[0101] Cross-validation is performed on the model training data to obtain cross-validation results, and the training parameters are optimized and adjusted according to the cross-validation results to obtain adjusted parameters, and the adjusted parameters are written into each sub-model to obtain multiple second anomaly detection sub-models.

[0102] Specifically, the training tasks are divided for multiple initialized anomaly detection models, and the training data is reasonably allocated to different computing units in batches to maximize the parallel processing ability of hardware resources. Assume that the total amount of training data is , and its dimension is , where represents the number of data samples, represents the feature dimension of each sample. For the convenience of parallel processing, the data is divided into subsets , where represents the dataset of the th batch, satisfying the following conditions:

[0103] ;

[0104] After the data is divided, the corresponding training tasks are assigned to each computing unit, and training parameters are set, such as the learning rate , the batch size and the optimization algorithm. For example, using stochastic gradient descent (SGD) optimization, the parameter update rule for each computing unit is expressed as:

[0105] ;

[0106] where, is the model parameter at the th iteration, is the loss function on the data of the th batch, represents the gradient of the loss function. Based on the above data parallel training scheme, different computing units simultaneously process their assigned batch data and independently calculate the gradient for each batch. Assume that the system has a total of computing units, and the gradient calculation result is . These results are aggregated into the global gradient through the gradient aggregation mechanism:

[0107] ;

[0108] Then, the global gradient is used to update the model parameters:

[0109] ;

[0110] The gradient aggregation method ensures the global consistency of parameter updates while fully utilizing the efficiency of parallel computing. After completing the data parallel training of the model, the model structure is split. Since there are significant differences in the structures and computational complexities of different sub-models, they are assigned to independent computing units to optimize resource utilization. For example, assume that the neural network structure of the th sub-model consists of layers, and the number of parameters in each layer is ( ), then the total number of parameters of the model is:

[0111] ;

[0112] According to the storage and computing capabilities of each computing unit, the sub-model with larger parameters is assigned to a high-performance computing unit, while the smaller sub-model is assigned to a low-power device. And communication links are established between sub-models to enable the exchange of parameters or intermediate results. The bandwidth and latency of the communication link need to meet the following conditions:

[0113] ;

[0114] Among them, represents the communication bandwidth, represents the number of parameters to be exchanged, represents the communication period, is the communication latency, is the latency tolerance threshold. Based on the model parallel training scheme, when training each sub-model, its learning rate and batch size are dynamically adjusted to adapt to the current training state. For example, a larger learning rate is used at the beginning of training to accelerate convergence, and the learning rate is gradually decreased later to improve accuracy:

[0115] ;

[0116] Among them, is the initial learning rate, is the decay coefficient, is the current iteration number. During the training process, performance metrics such as training loss , accuracy and gradient norm are recorded. These data are used for subsequent cross-validation and optimization adjustment. After the training is completed, cross-validation is performed on the model training data to evaluate the generalization ability of the sub-models. By dividing the data into a training set and a validation set, and evaluating the model performance on the validation set, the validation results are obtained. For example, the validation accuracy is expressed as:

[0117] ;

[0118] where is the indicator function, and are the predicted value and the true value respectively, is the number of samples in the validation set. According to the cross-validation results, the training parameters are optimized and adjusted. If it is found that the performance of some sub-models does not meet the standard, improvements are made by adjusting the learning rate, reallocating data, or expanding the model structure. For example, if the validation loss of the th sub-model exceeds the set threshold, the performance is improved by increasing the number of training rounds of this model or optimizing the initial values of its parameters. The optimized parameters are written into each sub-model to obtain multiple second anomaly detection sub-models.

[0119] In a specific embodiment, the process of executing step S5 may specifically include the following steps:

[0120] Formulate a resource allocation strategy based on the operation status data of multiple second anomaly detection sub-models, and perform resource weight allocation on multiple second anomaly detection sub-models by combining the computing resource utilization rate and the storage resource occupancy rate to obtain a sub-model resource configuration plan;

[0121] Based on the sub-model resource configuration plan, perform a computing load analysis on multiple second anomaly detection sub-models, input the energy consumption index and the computing resource consumption index of each second anomaly detection sub-model into the energy efficiency evaluation function to obtain sub-model energy efficiency optimization parameters;

[0122] Perform an integrated weight calculation on multiple second anomaly detection sub-models, input the detection accuracy and the sub-model energy efficiency optimization parameters into the weight calculation matrix for weight coefficient normalization processing to obtain a sub-model weight allocation table;

[0123] Establish an optimized integration strategy according to the sub-model weight allocation table, combine multiple second anomaly detection sub-models into an integrated detection model in the order of the processing stage, establish a model switching mechanism and set the trigger conditions to obtain a dynamic integration model;

[0124] Perform incremental learning on the dynamic integration model, determine the parameter update strategy, and update the model parameters and evaluate the performance for the newly added data samples to obtain an updated anomaly detection model;

[0125] Perform storage unit partitioning on the updated anomaly detection model, write the model structure parameters and weight parameters into the fixed storage area and the dynamic storage area respectively, establish a parameter index table, and obtain a processed anomaly detection model.

[0126] Specifically, formulate a resource allocation strategy according to the operation status data of multiple second anomaly detection sub-models. The operation status data includes the computing resource utilization rate and the storage resource occupancy rate . Suppose the computing resource and storage resource usage of the -th sub-model are and respectively, and the total system resources are and respectively. Then the resource utilization rate is expressed as:

[0127] ;

[0128] Among them, and respectively represent the computing resource requirement and storage resource requirement of the -th sub-model. By comprehensively analyzing the resource usage of all sub-models and combining the importance indicators of each sub-model (such as detection frequency or critical task priority), allocate resource weights for each sub-model, satisfying the following conditions:

[0129] ;

[0130] Among them, represents the importance of the -th sub-model. After completing the formulation of the resource allocation strategy, perform computing load analysis based on the sub-model resource configuration scheme to evaluate the energy efficiency performance of each sub-model. Suppose the energy consumption of the -th sub-model during operation is , and the computing resource consumption is . Then the energy efficiency ratio is defined as:

[0131] ;

[0132] Among them, is the detection accuracy of the -th sub-model, The higher it is, the higher the detection efficiency of the sub-model under unit energy and computing resources. Input the energy consumption and computing resource consumption of all sub-models into the energy efficiency evaluation function to obtain the energy efficiency optimization parameters of each sub-model, which are used to guide the subsequent model integration optimization. According to the energy efficiency optimization parameters, calculate the integration weights of multiple second anomaly detection sub-models to determine their contribution degrees in the integrated model. Input the detection accuracy and the energy efficiency optimization parameters into the weight calculation matrix, and obtain the weight distribution table after normalization. The weight calculation formula is:

[0133] ;

[0134] where, represents the normalized weight of the th sub-model, and are coefficients that adjust the importance of detection accuracy and energy efficiency ratio, and control the strategic tendency of weight distribution by adjusting them. Based on the weight distribution table, establish an optimized integration strategy and combine multiple second anomaly detection sub-models into an integrated detection model according to the order of processing stages. Assume that the processing process is divided into stages, and each stage is responsible for one or more sub-models. Then, the design of the integrated model organizes the sub-models in a phased order. For example:

[0135] ;

[0136] where, represents the set of sub-models in the nd stage. To adapt to the changes in the dynamic environment, establish a model switching mechanism for the integrated model and set the trigger condition . When the trigger condition is met, dynamically select the optimal sub-model to switch to the current task. For example:

[0137] ;

[0138] where, is the adaptability score of the sub-model under the current conditions. After completing the construction of the dynamic integrated model, perform incremental learning to improve its performance and adaptability. The core of incremental learning is to dynamically update the model parameters through the feature distribution of newly added data samples. Assume that the newly added data samples are , and update the model parameters by calculating the gradient :

[0139] ;

[0140] where, is the learning rate, is the loss function. At the same time, evaluate the performance of the updated model. By comparing the detection accuracy and resource consumption changes before and after the update, determine the optimization effect of incremental learning. To facilitate the long-term storage and management of the updated anomaly detection model, divide its storage unit, and write the model structure parameters and weight parameters into the fixed storage area and the dynamic storage area respectively. For example, the network structure and optimization algorithm of the model are fixed and stored in the fixed storage area, while the weight parameters updated with incremental learning are stored in the dynamic storage area. To quickly retrieve the model parameters, establish a parameter index table , where each entry records the parameter name, storage address, and version number. For example:

[0141] ;

[0142] Through the parameter index table, quickly locate and load the latest parameters of the model.

[0143] In a specific embodiment, the process of performing incremental learning on the dynamic ensemble model, determining the parameter update strategy, and updating the model parameters and evaluating the performance for the newly added data samples to obtain the updated anomaly detection model may specifically include the following steps:

[0144] Divide the model parameters of the dynamic ensemble model into two categories: structure parameters and weight parameters, and sort the parameter importance according to the sub-model weight distribution table to obtain the parameter update priority table;

[0145] Based on the parameter update priority table, construct a parameter update strategy, classify and label the newly added data samples according to the processing stage to obtain the labeled data samples, and allocate the labeled data samples to the corresponding second anomaly detection sub-model to obtain the incremental training dataset;

[0146] Based on the incremental training dataset and the parameter importance order, update the model parameters layer by layer, record the parameter change amount and performance change index to obtain the parameter update result, and evaluate the performance of the parameter update result, calculate the difference in detection accuracy before and after the update and the resource consumption change, and generate the model evaluation result;

[0147] Based on the model evaluation result, perform parameter rollback judgment, compare the performance evaluation index with the preset threshold, perform rollback operations on the parameters with performance degradation to obtain the parameter optimization plan, and write the parameter optimization plan into the dynamic ensemble model for optimal update of the model parameters to obtain the updated anomaly detection model.

[0148] Specifically, divide the model parameters of the dynamic ensemble model into structure parameters and weight parameters. The structure parameters define the basic architecture of the model, such as the number of layers of the network , the number of neurons in each layer , the activation function etc., which have a global impact on the overall performance of the model. The weight parameters are the specific values learned by the model during the training process, such as the weight matrix of each layer and the bias vector . Suppose the total parameter set of the model is , then it is expressed as:

[0149] ;

[0150] where represents the set of structural parameters, represents the set of weight parameters. After completing the parameter classification, all parameters are sorted according to the sub-model weight distribution table to generate a parameter update priority table. The weight of each sub-model in the sub-model weight distribution table represents its importance in the dynamic ensemble model, and the update priority is assigned to the parameters of each sub-model according to the size of its weight. For example, for the weight parameter of the th sub-model, its importance is defined as:

[0151] ;

[0152] where is the loss function, represents the sensitivity of the parameter to the loss. The parameters with higher importance are updated first to maximize the effect of incremental learning. Based on the parameter update priority table, a parameter update strategy is constructed, and the newly added data samples are classified and labeled according to the processing stage to form an incremental training data set. The labeling of the newly added data samples is based on the feature distribution of the processing stage. For example, the samples in the rough machining stage mainly contain vibration and pressure features, while the samples in the finish machining stage contain temperature and cutting force features. Suppose the newly added data sample set is Through feature extraction and classification labeling, it is divided into data subsets of each stage:

[0153] ;

[0154] where represents the labeled data subset of the th stage. The labeled data samples are assigned to the corresponding second anomaly detection sub-model for incremental training. Based on the incremental training data set and the parameter importance order, the model parameters are updated layer by layer. For each layer of weight parameter and bias , the gradient descent method is used for updating, and the update rule is:

[0155] ;

[0156] where is the learning rate, and are the gradients of the weights and biases respectively. During the update process, the parameter change and the performance change metrics, such as the detection accuracy change and the change in resource consumption are recorded. After the update is completed, the performance of the parameter update result is evaluated. By comparing the detection accuracy and the change in resource consumption before and after the update, the model evaluation result is generated. For example, the performance evaluation is measured by calculating the accuracy improvement ratio and the resource efficiency improvement ratio :

[0157] ;

[0158] where and are the accuracy after the update and the accuracy before the update respectively, and are the resource consumption after the update and the resource consumption before the update respectively. According to the model evaluation result, the parameter rollback judgment is performed. When the update of some parameters leads to a performance degradation and is lower than the preset threshold, the rollback operation is performed to restore these parameters to the state before the update. For example, when or is detected, the corresponding parameter is restored:

[0159] ;

[0160] where and are the thresholds of the accuracy and the resource efficiency respectively. The rollback operation can effectively avoid performance degradation caused by improper parameter updates. After the parameter rollback is completed, a parameter optimization scheme is generated, and the optimized parameters are written into the dynamic integration model to complete the optimal update of the model parameters. The updated anomaly detection model is divided into storage units, the fixed structure parameters are written into the fixed storage area, while the dynamically changing weight parameters are written into the dynamic storage area, and a parameter index table is established for efficient retrieval. For example, the index table records the storage address and the update timestamp of each parameter:

[0161] ;

[0162] Through the index table, the latest parameter version can be quickly located and loaded to obtain the updated anomaly detection model.

[0163] In this embodiment, after obtaining the processing anomaly detection model, it further includes: designing a distributed architecture for the processing anomaly detection model, dividing the model structure into a training module and an execution module, deploying the training module on a cloud server, deploying the execution module on an edge computing node, establishing a parameter synchronization mechanism between the modules to obtain a cloud-edge collaborative anomaly detection architecture; constructing a feature importance evaluation unit based on the cloud-edge collaborative anomaly detection architecture, calculating the mutual information and contribution analysis of the feature data in each processing stage, generating a feature importance ranking table, screening key features according to an importance threshold to obtain a reduced-dimensional feature subspace; inputting the reduced-dimensional feature subspace into a multi-agent system, configuring corresponding detection agents for each processing stage, where the detection agent includes a state perception module, an action selection module, and a reward calculation module, establishing a collaboration mechanism between the agents to obtain a distributed detection network; performing domain randomization training on the distributed detection network, generating diverse training scenarios by perturbing processing parameters, environmental parameters, and model parameters, constructing a domain randomization data set, inputting the domain randomization data set into the cloud server to obtain robust training data; constructing a hybrid multi-agent reinforcement learning network based on the robust training data, where the reinforcement learning network includes a policy network and a value network, setting independent network parameter spaces and a shared experience pool for each detection agent, optimizing the policy parameters through the soft Actor-Critic algorithm to obtain a reinforcement learning model; deploying the reinforcement learning model to the cloud-edge collaborative anomaly detection architecture, performing policy optimization and parameter update on the cloud side, performing real-time detection and action execution on the edge side, establishing a cloud-edge bidirectional communication channel to achieve dynamic adjustment of the detection policy and obtain a dynamic optimization system; monitoring the running state of the dynamic optimization system in real time, collecting performance metrics such as detection accuracy, response latency, and resource utilization, constructing a performance evaluation matrix, dynamically adjusting the cloud training policy and the edge execution policy based on the evaluation results to obtain a performance optimization plan; writing the performance optimization plan into the cloud-edge collaborative anomaly detection architecture, optimizing the allocation of cloud-edge resources, communication bandwidth, and computing load, performing a robustness verification experiment, and verifying the detection performance of the system in an uncertain environment to obtain a processing anomaly detection system with adaptive capabilities.

[0164] The intelligent control method based on processing process management in the embodiments of the present invention is described above. Next, the intelligent control device based on processing process management in the embodiments of the present invention will be described. Please refer to Figure 2 In an embodiment, the intelligent control device based on processing process management in the embodiments of the present invention includes:

[0165] A creation module, configured to set an anomaly detection instruction through an intelligent control chip and write it into a control status register to create an anomaly detection infrastructure;

[0166] A feature extraction module, configured to construct a data acquisition protocol according to the anomaly detection infrastructure and perform feature extraction operations on the processing process data to obtain multi-dimensional feature training data;

[0167] A pre-training module, configured to construct first anomaly detection sub-models for multiple processing stages based on the multi-dimensional feature training data, and write pre-training parameters into the first anomaly detection sub-models to obtain multiple initialized anomaly detection models;

[0168] A parallel training module, configured to deploy multiple initialized anomaly detection models to an intelligent control chip for model parallel training to obtain multiple second anomaly detection sub-models;

[0169] An optimization integration module, configured to calculate sub-model weight coefficients of multiple second anomaly detection sub-models, and perform optimization integration and incremental learning to obtain a processing anomaly detection model.

[0170] Through the collaborative cooperation of the above-mentioned various components, by setting anomaly detection instructions and configuring an anomaly detection circuit in the intelligent control chip, an infrastructure supporting two detection modes of data redundancy and time redundancy is constructed, improving the reliability and flexibility of anomaly detection. Using multi-dimensional feature training data to construct multi-stage anomaly detection sub-models and setting a shared parameter layer reduces the redundancy of model parameters and optimizes the utilization efficiency of computing resources. Introducing training strategies of data parallelism and model parallelism, through dynamically adjusting training parameters and a cross-validation mechanism, the anomaly detection model has stronger generalization ability. Based on an integration strategy of resource allocation and energy efficiency optimization, combined with a dynamic adjustment mechanism of weight coefficients, the reasonable allocation of computing resources and the optimization of energy efficiency are realized. Through an incremental learning and parameter update mechanism, a complete model optimization and performance evaluation system is established to ensure the continuous optimization and performance improvement of the anomaly detection model. Adopting a management method of hierarchical storage and parameter indexing improves the access efficiency of model parameters and facilitates the dynamic update and maintenance of the model.

[0171] An embodiment of the present invention further provides a chip, on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented.

[0172] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the above-described system, system and unit can refer to the corresponding processes in the foregoing method embodiments and will not be described herein again.

[0173] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0174] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of various embodiments of the present invention.

Claims

1. An intelligent control method based on machining process management, characterized in that: The method comprises: An abnormality detection instruction is set through an intelligent control chip and written into a control status register to create an abnormality detection infrastructure; specifically comprising: adding an abnormality detection instruction set and a control status register address space to the intelligent control chip to obtain an extended instruction system architecture, and setting a data redundancy detection threshold and a time redundancy detection threshold according to the extended instruction system architecture to generate an abnormality detection parameter matrix; writing the abnormality detection parameter matrix into a control status register to obtain abnormality detection configuration data, and dividing the abnormality detection configuration data into operation modes to obtain a data redundancy mode and a time redundancy mode; configuring multiple groups of data comparison logic for the data redundancy mode, and configuring a timing comparison logic for the time redundancy mode to obtain an operation mode configuration parameter; configuring a data comparison unit and a result output unit based on the operation mode configuration parameter, and cascading the data comparison unit and the result output unit to obtain an abnormality detection circuit; performing functional verification on the abnormality detection circuit to obtain a circuit verification result, and feeding back the circuit verification result to the intelligent control chip, adjusting the abnormality detection instruction and the configuration parameters of the abnormality detection circuit according to the circuit verification result to obtain an abnormality detection infrastructure; Constructing a data collection protocol according to the anomaly detection infrastructure, and performing feature extraction operations on the processing data to obtain multi-dimensional feature training data; Building a first anomaly detection sub-model of multiple processing stages based on the multi-dimensional feature training data, writing pre-trained parameters into the first anomaly detection sub-model, and obtaining multiple initialized anomaly detection models; Deploying the multiple initialized anomaly detection models to the intelligent control chip for model parallel training to obtain multiple second anomaly detection sub-models; The sub-model weight coefficients of the plurality of second anomaly detection sub-models are calculated, and optimization integration and incremental learning are performed to obtain a processing anomaly detection model.

2. The intelligent control method based on machining process management according to claim 1 is characterized in that: The data collection protocol is constructed according to the abnormality detection infrastructure, and feature extraction operations are performed on the processing data to obtain multi-dimensional feature training data, including: Setting data sampling parameters based on the anomaly detection infrastructure, and writing the sampling frequency, sampling time window length and sampling accuracy of the data sampling parameters into a data acquisition control unit to obtain a data acquisition basic configuration; Setting constraints on the data acquisition basic configuration, dividing the data acquisition levels according to the processing parameter types, setting the data filtering threshold of each data acquisition level, and obtaining the data acquisition protocol; Importing the data acquisition protocol into a data acquisition module, performing multi-dimensional sampling on the processing data to generate an original data matrix, and performing denoising and standardization processing on the original data matrix to obtain pre-processed data; The time domain features, frequency domain features and statistical features of the preprocessed data are extracted respectively, and the time domain features, the frequency domain features and the statistical features are combined into multi-dimensional feature training data.

3. The intelligent control method based on machining process management according to claim 1 is characterized in that: The first abnormality detection sub-models of multiple processing stages are constructed based on the multi-dimensional feature training data, and the pre-trained parameters are written into the first abnormality detection sub-model to obtain multiple initialized abnormality detection models, including: According to the processing sequence, the multi-dimensional feature training data is distributed to different processing stage data sets to obtain a plurality of processing stage feature data; Determine the sub-model structure for the characteristic data of each processing stage based on the computing resources of the intelligent control chip, set the number of network layers and the number of neurons of the sub-model, and obtain multiple first abnormality detection sub-models; Performing parameter space planning on the multiple first anomaly detection sub-models, dividing the parameter space into an independent parameter area and a shared parameter area, establishing a parameter sharing channel between the first anomaly detection sub-models in adjacent processing stages, and obtaining a parameter space distribution map; Constructing a shared parameter layer according to the parameter space distribution graph, inserting the shared parameter layer into the network structure of the plurality of first anomaly detection sub-models, establishing a parameter transfer matrix, and obtaining an anomaly detection model with a shared layer; The pre-trained parameters of the anomaly detection model with a shared layer are initialized, the pre-trained parameter values ​​are determined according to the feature distribution of each processing stage, a pre-trained parameter matrix is ​​generated, and the pre-trained parameter matrix is ​​written into the anomaly detection model with a shared layer for parameter loading and model compilation to obtain multiple initialized anomaly detection models.

4. The intelligent control method based on machining process management according to claim 1 is characterized in that: The step of deploying the multiple initialized anomaly detection models to the intelligent control chip for model parallel training to obtain multiple second anomaly detection sub-models includes: Dividing the training tasks of the multiple initialized anomaly detection models, allocating the training data to different computing units in batches, setting the training parameters of each computing unit, and obtaining a data parallel training scheme; Based on the data parallel training scheme, different batches of training data are processed in parallel, the training results of each computing unit are gradient aggregated, and sub-model parameters are updated to obtain a data parallel training model; The data parallel training model is split into model structures, sub-models with different structures are allocated to independent computing units, communication links between sub-models are established, and a model parallel training scheme is obtained; Execute sub-model training based on the model parallel training scheme, dynamically adjust the learning rate and batch size of each sub-model, record the performance indicators during the training process, and obtain model training data; The model training data is cross-validated to obtain a cross-validation result, and the training parameters are optimized and adjusted according to the cross-validation result to obtain adjusted parameters, and the adjusted parameters are written into each sub-model to obtain multiple second anomaly detection sub-models.

5. The intelligent control method based on machining process management according to claim 1 is characterized in that: The calculating of the sub-model weight coefficients of the plurality of second anomaly detection sub-models, and performing optimization integration and incremental learning to obtain a processing anomaly detection model includes: Formulate a resource allocation strategy according to the operation status data of the plurality of second anomaly detection sub-models, and perform resource weight allocation on the plurality of second anomaly detection sub-models in combination with the computing resource utilization rate and the storage resource occupancy rate to obtain a sub-model resource configuration plan; Performing computational load analysis on the plurality of second anomaly detection sub-models based on the sub-model resource configuration scheme, inputting the energy consumption index and the computational resource consumption index of each second anomaly detection sub-model into an energy efficiency evaluation function, and obtaining sub-model energy efficiency optimization parameters; Performing integrated weight calculation on the multiple second anomaly detection sub-models, inputting the detection accuracy and the sub-model energy efficiency optimization parameters into a weight calculation matrix to perform weight coefficient normalization processing, and obtaining a sub-model weight allocation table; Establishing an optimization integration strategy according to the sub-model weight distribution table, combining the multiple second anomaly detection sub-models into an integrated detection model in the order of processing stages, establishing a model switching mechanism and setting trigger conditions to obtain a dynamic integrated model; Performing incremental learning on the dynamic integrated model, determining a parameter update strategy, and updating model parameters and evaluating performance on newly added data samples to obtain an updated anomaly detection model; The updated anomaly detection model is divided into storage units, the model structure parameters and weight parameters are written into the fixed storage area and the dynamic storage area respectively, a parameter index table is established, and the processing anomaly detection model is obtained.

6. The intelligent control method based on machining process management according to claim 5 is characterized in that: The step of performing incremental learning on the dynamic integrated model, determining a parameter update strategy, and performing model parameter update and performance evaluation on newly added data samples to obtain an updated anomaly detection model includes: The model parameters of the dynamic integrated model are divided into two categories: structural parameters and weight parameters, and the parameter importance is sorted according to the sub-model weight allocation table to obtain a parameter update priority table; Constructing a parameter update strategy based on the parameter update priority table, classifying and labeling the newly added data samples according to the processing stage to obtain labeled data samples, and assigning the labeled data samples to the corresponding second anomaly detection sub-model to obtain an incremental training data set; Based on the incremental training data set and the order of parameter importance, the model parameters are updated layer by layer, the parameter changes and performance change indicators are recorded, the parameter update results are obtained, and the performance evaluation is performed on the parameter update results, the detection accuracy difference and resource consumption change before and after the update are calculated, and the model evaluation results are generated; Based on the model evaluation result, parameter rollback judgment is performed, the performance evaluation index is compared with the preset threshold, and a rollback operation is performed on the parameters with degraded performance to obtain a parameter optimization plan, and the parameter optimization plan is written into the dynamic integration model to perform optimal update of the model parameters to obtain an updated anomaly detection model.

7. An intelligent control device based on processing management, characterized in that: Used to execute the intelligent control method based on machining process management as described in any one of claims 1 to 6, the intelligent control device based on machining process management comprises: Create a module to set anomaly detection instructions through the intelligent control chip and write to the control status register to create anomaly detection infrastructure; A feature extraction module is used to construct a data acquisition protocol according to the anomaly detection infrastructure and perform feature extraction operations on the processing data to obtain multi-dimensional feature training data; A pre-training module, used for constructing a first anomaly detection sub-model of multiple processing stages based on the multi-dimensional feature training data, writing pre-training parameters into the first anomaly detection sub-model, and obtaining a plurality of initialized anomaly detection models; A parallel training module, used for deploying the multiple initialized anomaly detection models to the intelligent control chip for model parallel training to obtain multiple second anomaly detection sub-models; The optimization integration module is used to calculate the sub-model weight coefficients of the multiple second abnormality detection sub-models, and perform optimization integration and incremental learning to obtain a processing abnormality detection model.

8. A chip, characterized in that: A computer program is stored thereon, and when the computer program is executed by a processor, the processor is enabled to execute the intelligent control method based on machining process management as claimed in any one of claims 1 to 6.

Images