Brushless motor assembly test method, device and equipment

Abstract

The invention relates to the technical field of brushless motor testing, and discloses an assembly testing method, device and equipment of a brushless motor. The method comprises the following steps: performing data stream segmentation processing on multi-source sensor data in the brushless motor assembling process to obtain a normal data block and a drift data block; carrying out dimension reduction processing to obtain a low-dimensional feature vector; training through a deep multilayer sparse regularization time-varying transfer learning network model to obtain a first assembly state prediction model; processing the assembly characteristic data collected in real time to obtain a multi-dimensional index prediction value and an assembly quality score; performing multi-mode anomaly detection and diagnosis to obtain an anomaly report; reinforcement learning and incremental learning are carried out based on the abnormal report, a second assembly state prediction model is obtained, and a corresponding brushless motor assembly strategy is generated, the brushless motor assembly test method is adaptive and intelligent, and the test accuracy, efficiency and self-optimization capability are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of brushless motor testing, and in particular to an assembly testing method, device and equipment for a brushless motor. Background Art

[0002] Brushless motors are widely used in industrial applications due to their high efficiency, long life and low noise. However, traditional assembly test methods mainly rely on manual experience and fixed test processes, which are difficult to adapt to the complex and changing production environment and the ever-increasing quality requirements. At the same time, although a large amount of sensor data contains rich assembly status information, how to effectively process and utilize these high-dimensional, multi-modal data remains a difficult problem.

[0003] In addition, the abnormal conditions in the assembly process of brushless motors are often subtle and diverse, and traditional single thresholds or simple statistical methods are difficult to accurately capture these anomalies. Especially when production conditions change or new models are put into production, existing test models often require a lot of manual intervention to adapt, which greatly limits the flexibility and efficiency of production. Another key issue is how to effectively convert detected anomalies into improvement measures. At present, most anomaly detection systems can only provide simple alarms, but cannot give specific optimization suggestions or automatically adjust test strategies. This leads to a clear gap between the detection results and the actual improvement actions, limiting the closed-loop optimization capabilities of the entire assembly test system. Summary of the invention

[0004] The present invention provides a brushless motor assembly test method, device and equipment, which are used for an adaptive and intelligent brushless motor assembly test method to improve the accuracy, efficiency and self-optimization capability of the test.

[0005] In a first aspect, the present invention provides an assembly and testing method for a brushless motor, the assembly and testing method for a brushless motor comprising: The multi-source sensor data in the brushless motor assembly process is processed by data stream segmentation to obtain normal data blocks and drift data blocks; Based on the normal data block and the drift data block, a dual-metric constrained Laplace eigenmapping algorithm is used to perform dimensionality reduction processing to obtain a low-dimensional feature vector; Inputting the low-dimensional feature vector into a deep multi-layer sparse regularized time-varying transfer learning network model for training to obtain a first assembly state prediction model; Processing the assembly feature data collected in real time based on the first assembly state prediction model to obtain a multi-dimensional index prediction value and an assembly quality score; Performing multi-modal anomaly detection and diagnosis on the assembly process of the brushless motor based on the multi-dimensional index prediction value and the assembly quality score to obtain an anomaly report; Based on the abnormal report, reinforcement learning and incremental learning are performed on the first assembly state prediction model to obtain a second assembly state prediction model, and a corresponding brushless motor assembly strategy is generated.

[0006] In a second aspect, the present invention provides an assembly and testing device for a brushless motor, the assembly and testing device for the brushless motor comprising: A segmentation module is used to perform data stream segmentation processing on multi-source sensor data in the brushless motor assembly process to obtain normal data blocks and drift data blocks; A dimensionality reduction module, used to perform dimensionality reduction processing based on the normal data block and the drift data block by using a Laplace eigenmapping algorithm with dual metric constraints to obtain a low-dimensional feature vector; A training module, used for inputting the low-dimensional feature vector into a deep multi-layer sparse regularized time-varying transfer learning network model for training to obtain a first assembly state prediction model; A processing module, used to process the assembly feature data collected in real time based on the first assembly state prediction model to obtain a multi-dimensional index prediction value and an assembly quality score; A detection module, used for performing multi-modal anomaly detection and diagnosis on the assembly process of the brushless motor based on the multi-dimensional index prediction value and the assembly quality score to obtain an anomaly report; A generation module is used to perform reinforcement learning and incremental learning on the first assembly state prediction model based on the abnormal report to obtain a second assembly state prediction model and generate a corresponding brushless motor assembly strategy.

[0007] The third aspect of the present invention provides an assembly and testing device for a brushless motor, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the assembly and testing device for the brushless motor executes the above-mentioned assembly and testing method for the brushless motor.

[0008] A fourth aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the above-mentioned assembly and testing method for a brushless motor.

[0009] In the technical solution provided by the present invention, high-dimensional multi-source sensor data is effectively processed through data stream segmentation processing and dual-metric constrained Laplace eigenmapping algorithm, thereby improving the accuracy and efficiency of feature extraction. A deep multi-layer sparse regularized time-varying transfer learning network model is adopted to enhance the adaptability of the model to different working conditions and new models of products, thereby reducing the need for manual intervention. A multimodal anomaly detection method based on multidimensional indicator prediction values ​​and assembly quality scores improves the recognition accuracy of abnormal states and reduces false positives and false negatives. By constructing an abnormal pattern library and a diagnostic rule library, the fault diagnosis capability of the system is enhanced, and a more specific and accurate fault cause analysis can be provided. By using reinforcement learning and incremental learning methods, continuous optimization of the model is achieved, so that the system can continuously adapt to new production conditions and requirements. By improving the accuracy and timeliness of anomaly detection, unnecessary downtime and scrap rate are reduced, thereby reducing production costs. The stability and consistency of the assembly quality of brushless motors are ensured, which helps to improve the overall quality level of products. The automated anomaly detection and diagnosis process reduces manual intervention and improves the operating efficiency of the production line. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying any creative work.

[0011] Figure 1 A schematic diagram of a flow chart of an assembly and testing method for a brushless motor provided in an embodiment of the present application; Figure 2 A schematic block diagram of the structure of an assembly and testing device for a brushless motor provided in an embodiment of the present application. DETAILED DESCRIPTION

[0012] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0013] The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may also be decomposed, combined or partially merged, so the actual execution order may change based on actual conditions.

[0014] It should also be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in this application specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0015] It should be further understood that the term “and / or” used in the specification and appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0016] In conjunction with the accompanying drawings, some embodiments of the present application are described in detail below. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0017] See also Figure 1 , Figure 1 A schematic diagram of a flow chart of an assembly and testing method for a brushless motor provided in an embodiment of the present application, such as Figure 1 As shown, the assembly and testing method of the brushless motor provided in the embodiment of the present application includes steps S100 to S600.

[0018] Step S100, performing data stream segmentation processing on multi-source sensor data in the brushless motor assembly process to obtain normal data blocks and drift data blocks; It is understandable that the execution subject of the present invention may be a brushless motor assembly and testing device, or a terminal or a server, which is not specifically limited here. The embodiment of the present invention is described by taking a server as the execution subject as an example.

[0019] Specifically, the multi-source sensor data collected during the assembly process of the brushless motor is segmented by a sliding time window to obtain an initial data block sequence. By selecting an appropriate time window length, it is ensured that each data block can capture sufficient dynamic information while maintaining the real-time nature of data processing. The divergence of adjacent data blocks in the initial data block sequence is calculated to measure the dissimilarity between different data blocks. As an indicator of the difference between data blocks, divergence can effectively capture the possible change trend between data blocks. The initial data block sequence is segmented by an adaptive threshold according to the dissimilarity measure to generate a set of candidate data blocks. The selection of the adaptive threshold is dynamically adjusted according to the specific characteristics of the data, and can accurately capture the changes in the data in different assembly stages. For the generated set of candidate data blocks, the time domain statistical features and wavelet packet energy features of each data block are extracted to form a multidimensional feature vector set. These feature vectors not only reflect the time domain characteristics of the data, but also capture the frequency domain information of the data through the wavelet packet energy features, thereby enriching the expression of the data. Based on the multidimensional feature vector set, probability density estimation is performed to obtain a probability distribution model for each data block. The purpose of probability density estimation is to establish a mathematical model so that the distribution of data blocks can be described. On this basis, by calculating the log-likelihood value and distance of each data block, an anomaly measurement sequence is generated to reveal the potential anomalies between data blocks. The anomaly measurement sequence is detected to identify the possible potential change point set. The potential change point set represents the sudden change point of the data block on the time axis and is an important basis for identifying the differences between data blocks. Based on the potential change point set, the candidate data block set is hierarchically clustered to obtain the partition result of the target data block. Hierarchical clustering gradually merges similar data blocks to make the partition result more consistent with the intrinsic structural characteristics of the data. The inter-block similarity matrix calculation and spectral clustering are performed on the target data block partition result to obtain the final evolutionary segmentation result. Spectral clustering can optimize the data block segmentation result and better reflect the evolutionary relationship between data blocks. By analyzing the evolutionary segmentation result, normal data blocks and drift data blocks are finally generated.

[0020] Step S200: Based on the normal data block and the drift data block, a dual-metric constrained Laplace eigenmapping algorithm is used to perform dimensionality reduction processing to obtain a low-dimensional feature vector; Specifically, the normal data block and the drift data block are merged to form a comprehensive data set. The comprehensive data set contains sensor data of the brushless motor in different states, which helps to fully consider the global characteristics of the data in subsequent analysis. The similarity matrix is ​​generated by calculating the cosine similarity of the samples in the comprehensive data set. As an indicator of the similarity between samples, cosine similarity can effectively capture the intrinsic connection between data. A K-nearest neighbor graph is constructed based on the similarity matrix to determine the local connection relationship between samples. The K-nearest neighbor graph connects each sample with the K samples closest to it in similarity to construct a local neighborhood relationship, which better reflects the distribution characteristics of samples in high-dimensional space. Based on the local connection relationship, the geodesic distance and Euclidean distance between samples are calculated to generate a dual-metric distance matrix. The geodesic distance reflects the distance between samples on the nonlinear manifold, while the Euclidean distance is the straight-line distance between samples in the high-dimensional space. These two distances jointly constrain the relationship between samples. The affinity matrix is ​​constructed using the dual-metric distance matrix, and the affinity matrix reveals the global geometric structure between samples. The global geometric structure not only considers the connection relationship of the local neighborhood, but also incorporates the distribution characteristics of the samples in the entire data set. Based on the affinity matrix, the Laplacian matrix is ​​calculated. The Laplacian matrix is ​​an important representation of the graph structure and can reflect the overall structural characteristics of the data. The Laplacian matrix is ​​subjected to eigenvalue decomposition to obtain the eigenvalue sequence and eigenvector matrix. Eigenvalue decomposition provides multiple possible dimensionality reduction methods for data by decomposing the Laplacian matrix into a series of eigenvalues ​​and corresponding eigenvectors. In order to determine the optimal number of dimensionality reduction, the maximum eigenvalue interval criterion is applied according to the eigenvalue sequence. The maximum eigenvalue interval criterion determines the number of dimensions most suitable for dimensionality reduction by finding the maximum interval between eigenvalues, ensuring that the dimension is reduced to the greatest extent while retaining the data structure information as much as possible. According to the determined optimal number of dimensionality reduction, the corresponding eigenvectors are selected from the eigenvector matrix to construct the projection matrix. As the core tool of linear transformation, the projection matrix generates low-dimensional eigenvectors by projecting high-dimensional data into low-dimensional space.

[0021] Step S300, inputting the low-dimensional feature vector into a deep multi-layer sparse regularized time-varying transfer learning network model for training to obtain a first assembly state prediction model; Specifically, the low-dimensional feature vector is segmented in time series to obtain the source domain dataset and the target domain dataset. A multi-layer sparse autoencoder is constructed based on the source domain dataset. The structure of the autoencoder is designed to be five encoding layers and five decoding layers. Each layer uses the ReLU activation function and the sparsity constraint is implemented through L1 regularization. The ReLU activation function can effectively capture nonlinear features, while the L1 regularization helps the model extract important features while suppressing irrelevant or redundant features. Through unsupervised pre-training, the source domain dataset is input into the multi-layer sparse autoencoder to obtain the initial feature representation. The pre-training method helps the model learn the underlying structure of the data without labels. A time-varying transfer learning network is constructed based on the initial feature representation. The design of the network includes three key parts: feature extractor, domain classifier and task classifier. The feature extractor adopts the ResNet structure, which contains four residual blocks. Each residual block includes two convolutional layers, a short-circuit connection and an adaptive instance normalization layer. The ResNet structure solves the gradient vanishing problem in deep networks through residual connections, making feature extraction more profound and effective. The domain classifier adopts a fully connected structure, including three hidden layers. Each layer uses LeakyReLU activation function and Dropout regularization. LeakyReLU activation function can handle negative gradient flow, while Dropout prevents overfitting by randomly discarding some neurons. The task classifier adopts an attention-enhanced LSTM structure, including two LSTM layers and one self-attention layer. The hidden state dimension of the LSTM layer is set to 128. This structure can effectively capture the temporal dependency in sequence data and focus on the most important features for the task through the attention mechanism. The source domain dataset and the target domain dataset are simultaneously input into the time-varying transfer learning network, and domain-invariant features are extracted through adversarial training and multi-task learning strategies. Adversarial training minimizes the difference between domains through the game between domain classifiers and feature extractors, while multi-task learning processes multiple tasks simultaneously by sharing feature extractors, thereby improving the generalization ability of the model. After obtaining the domain-invariant features, a dynamic Kullback-Leibler divergence weight module is constructed. This module contains a fully connected layer and a Softmax layer to dynamically balance the distance metric loss and the adversarial metric loss. The dynamic balance mechanism can adaptively adjust the weights of different losses according to the actual situation during the training process to ensure that the model achieves the best results in inter-domain confrontation and feature learning. The time-varying transfer learning network and the dynamic Kullback-Leibler divergence weight module are combined to obtain the first assembly state prediction model through end-to-end training. This model can not only accurately predict the assembly state of the brushless motor, but also effectively transfer knowledge between different domains to adapt to possible changes in the assembly process.

[0022] Step S400: Processing the assembly feature data collected in real time based on the first assembly state prediction model to obtain a multi-dimensional index prediction value and an assembly quality score; Specifically, the assembly feature data collected in real time is preprocessed to ensure the consistency and comparability of the data. The data is standardized to eliminate the influence caused by the dimension difference between different features and obtain standardized feature data. The standardized feature data is input into the multi-layer sparse autoencoder of the first assembly state prediction model, and the second initial feature representation is obtained through the forward propagation calculation of the five encoding layers. By extracting features layer by layer, the original data is gradually compressed into a low-dimensional initial representation, retaining the most important information in the data, and suppressing the interference of irrelevant features through sparse regularization. The second initial feature representation is input into the feature extractor of the first assembly state prediction model, and the deep feature representation is obtained through the convolution operation of four residual blocks and adaptive instance normalization. The residual block solves the common gradient vanishing problem in the deep network by introducing short-circuit connection, ensuring the depth and accuracy of feature extraction. The adaptive instance normalization improves the consistency of the features, so that the model can still maintain stable performance when processing data with large changes. The deep feature representation is input into the domain classifier, and the domain probability distribution is obtained through the calculation of three fully connected layers. Evaluate the domain attribution of features and determine the similarity between the current input data and the data seen during model training. Based on the deep feature representation, the task classifier of the first assembly state prediction model is used for time series modeling. The task classifier extracts the time series context features of the data through the calculation of two LSTM layers and one self-attention layer. The LSTM layer can effectively capture the dependency of the input data in the time dimension, while the self-attention layer enhances the model's attention to key time points and improves the expressiveness of time series features by giving different weights to features at different time steps. The time series context features are input into the dynamic Kullback-Leibler divergence weight module. The weight distribution of each feature is obtained through the calculation of the fully connected layer and the Softmax layer. The output of the Softmax layer provides a basis for dynamically weighing the importance of features, so that the model can adaptively adjust the influence of each feature according to the characteristics of the current data. Based on the feature weights, the time series context features are weighted and summed to obtain a comprehensive feature representation. Based on the comprehensive feature representation, the multidimensional indicator prediction value is calculated. The multidimensional indicator prediction value reflects the key performance indicators in the assembly process. Based on the multidimensional indicator prediction value, an initial quality score is generated. The initial quality score is dynamically adjusted to obtain the final assembly quality score. The threshold is dynamically modified according to the actual situation during the adjustment process, so that the score can more accurately reflect the quality status of the current assembly process.

[0023] Step S500, performing multi-modal anomaly detection and diagnosis on the assembly process of the brushless motor based on the multi-dimensional index prediction value and the assembly quality score to obtain an anomaly report; Specifically, the predicted values ​​of multi-dimensional indicators are analyzed in time series to identify the changing trends of each indicator in the time dimension. Through in-depth analysis of these trends, the statistical characteristics of each indicator, such as mean, variance, skewness, etc., are calculated to evaluate the abnormality of each indicator and reveal the potential abnormal trend in the assembly process. At the same time, the threshold analysis of the assembly quality score is performed to determine whether there is a quality abnormality. When the quality score exceeds the set threshold, it is marked as a quality abnormality. Combined with the indicator abnormality evaluation and quality abnormality marking, multimodal data fusion is performed to obtain a comprehensive abnormality index. The comprehensive abnormality index is clustered with historical data to generate an abnormal pattern library. The abnormal pattern library is constructed based on known abnormal situations in historical data and can help identify the type of current abnormality. By matching the current comprehensive abnormality index with the abnormal pattern library, the current abnormality type is preliminarily determined. According to the preliminary abnormality type determination, the features related to the type are extracted from the preset fault feature library to form a target feature set. Based on the target feature set, these multi-source sensor data are secondary analyzed to extract the target abnormality features related to the current abnormality. The target abnormality features are compared with the preset diagnostic rule library to obtain specific fault diagnosis results. The diagnostic rule base contains rules based on a lot of experience and experiments, which can help quickly identify and confirm specific fault types. The specific fault diagnosis results are integrated with the previous comprehensive abnormality index and preliminary abnormality type determination to generate a complete abnormality report. The abnormality report records the nature and cause of the abnormality and also provides possible solutions and suggestions.

[0024] Step S600: Perform reinforcement learning and incremental learning on the first assembly state prediction model based on the abnormal report to obtain a second assembly state prediction model, and generate a corresponding brushless motor assembly strategy.

[0025] Specifically, an abnormal data matrix is ​​generated according to the abnormal report. Based on the abnormal data matrix, a multi-dimensional reward and punishment function is constructed. The design of the reward and punishment function is to reward correct decisions and punish wrong decisions during the model learning process, so as to guide the model to develop in a better direction. The parameter sensitivity analysis of the first assembly state prediction model is performed. By perturbing the parameters of the model one by one and observing the output changes, the sensitivity coefficient of each parameter is calculated, and a parameter importance list sorted by sensitivity is obtained. According to the parameter importance list, the top N most sensitive parameters are selected, and discretized adjustment intervals are designed for these parameters respectively. A high-dimensional discrete action space is formed through Cartesian product combination to generate a set of candidate solutions for model adjustment. The action space contains various possible parameter adjustment solutions, providing multiple possible paths for model optimization. Using the reward and punishment function as an evaluation criterion, a Monte Carlo search tree is constructed based on the candidate solution set. The Monte Carlo search tree gradually explores possible adjustment paths through iterative expansion, simulation and backtracking processes, and finally obtains the optimal parameter adjustment path. Along the optimal parameter adjustment path found, the parameters of the first assembly state prediction model are gradually updated. After each parameter update, the performance of the model is evaluated using the validation set. Through continuous adjustment and verification, until the preset number of iterations is reached or the model performance converges, the second assembly state prediction model is obtained. The optimized second assembly state prediction model is back-analyzed to obtain the annotated optimized sample set. Based on the optimized sample set, a multi-layer decision tree is constructed, in which each layer represents a key decision point in the assembly process, and each node of the tree contains decision conditions, operation instructions, and expected results. Through recursive splitting and pruning, a streamlined and efficient assembly process decision model is obtained. The model can effectively guide each decision point in the assembly process to ensure that each step of the operation is optimal. The assembly process decision model is simulated and tested. By simulating the distribution of assembly results under different decision paths and calculating the expected return and risk coefficient of each path, the impact of different decisions on the final assembly results can be deeply understood. On this basis, a multi-objective optimization algorithm is applied to balance efficiency and stability, and finally the optimal brushless motor assembly strategy is obtained.

[0026] In the embodiment of the present invention, high-dimensional multi-source sensor data is effectively processed through data stream segmentation processing and dual-metric constrained Laplace eigenmapping algorithm, and the accuracy and efficiency of feature extraction are improved. A deep multi-layer sparse regularized time-varying transfer learning network model is adopted to enhance the adaptability of the model to different working conditions and new models of products, and reduce the need for manual intervention. A multimodal anomaly detection method based on multidimensional indicator prediction values ​​and assembly quality scores improves the recognition accuracy of abnormal states and reduces false positives and false negatives. By constructing an abnormal pattern library and a diagnostic rule library, the fault diagnosis capability of the system is enhanced, and a more specific and accurate fault cause analysis can be provided. By using reinforcement learning and incremental learning methods, continuous optimization of the model is achieved, so that the system can continuously adapt to new production conditions and requirements. By improving the accuracy and timeliness of anomaly detection, unnecessary downtime and scrap rate are reduced, thereby reducing production costs. The stability and consistency of the assembly quality of the brushless motor are ensured, which helps to improve the overall quality level of the product. The automated anomaly detection and diagnosis process reduces manual intervention and improves the operating efficiency of the production line.

[0027] In a specific embodiment, the process of executing step S100 may specifically include the following steps: The multi-source sensor data in the brushless motor assembly process is segmented by sliding time window to obtain the initial data block sequence, and the divergence of adjacent data blocks in the initial data block sequence is calculated to obtain the dissimilarity measure between the data blocks. According to the dissimilarity measure, the initial data block sequence is segmented by adaptive threshold to obtain a set of candidate data blocks, and the time domain statistical features and wavelet packet energy features are extracted from each data block in the set of candidate data blocks to obtain a multi-dimensional feature vector set; Based on the multi-dimensional feature vector set, the probability density estimation is performed to obtain the probability distribution model of the data block, and the log-likelihood value and distance of each data block are calculated according to the probability distribution model to obtain the abnormal measurement sequence; Detect the abnormal measurement sequence to obtain a potential change point set, and based on the potential change point set, perform hierarchical clustering on the candidate data block set to obtain the target data block division result; The target data block partition results are subjected to inter-block similarity matrix calculation and spectral clustering to obtain evolutionary segmentation results, and normal data blocks and drift data blocks are generated according to the evolutionary segmentation results.

[0028] Specifically, the multi-source sensor data in the brushless motor assembly process is segmented by sliding time windows. The continuous time series data is divided into several data blocks, each of which contains sensor readings within a certain time range. Assume that the time series data is ,in represents the total length of the time series, and the length of the sliding window is , the sliding step length is , then the initial data block sequence can be expressed as ,in: ; Each data block The corresponding subsequence is This segmentation method ensures the continuity of the time series data and preserves the time correlation within each data block. The divergence of adjacent data blocks in the initial data block sequence is calculated to measure the dissimilarity between different data blocks. Divergence can be calculated in many ways, one of the most commonly used methods is the divergence measure based on probability distribution, such as Kullback-Leiblen divergence. For any two adjacent data blocks and , their respective probability distributions are and , then the Kullback-Leibler divergence between them It can be expressed as: ; in, Indicated in the data block The observed value is The probability of It indicates the distribution difference of the two data blocks. The larger the value, the more significant the distribution difference between the two data blocks, which may indicate that some changes have occurred in the assembly process of the brushless motor. Based on the dissimilarity metric, the initial data block sequence is segmented by adaptive threshold to form a set of candidate data blocks. Adaptive threshold segmentation dynamically adjusts the segmentation threshold according to the size of the dissimilarity metric to adapt to the changes in the differences between data blocks in different time periods. For each pair of adjacent data blocks and , if the divergence between them Exceeds the preset adaptive threshold , it is considered that there are significant change points between the two data blocks, and then they are divided into independent candidate data blocks. Extract features from each data block in the candidate data block set to form a multidimensional feature vector set. Feature extraction mainly includes two categories: time domain statistical features and wavelet packet energy features. Time domain statistical features include mean, standard deviation, skewness, kurtosis, etc. These features can describe the basic statistical characteristics of data blocks in the time domain. The wavelet packet energy feature decomposes the data block into different frequency bands by performing wavelet packet transform on the data block, thereby extracting the energy distribution characteristics of each frequency band. These energy features can reflect the change information of the data in the frequency domain, thereby helping to identify potential anomalies. Probability density estimation is performed based on the multidimensional feature vector set to obtain the probability distribution model of each data block. A statistical model is established for each data block, which can describe the distribution of the features of the data block in the feature space. Commonly used probability density estimation methods include kernel density estimation and Gaussian mixture model. Assume that the data block The eigenvector of , then its probability density estimate It can be expressed as: ; in, is the kernel function, is the bandwidth parameter. By probability density estimation, the log-likelihood value of each data block is calculated and distance , and obtain the abnormal measurement sequence. The abnormal measurement sequence is detected to identify the potential change point set. The purpose of change point detection is to find the moments of significant changes in time series data. These change points often correspond to sudden changes in the system state. Based on the change point set, the candidate data block set is hierarchically clustered to aggregate similar data blocks together. The result of hierarchical clustering is the target data block partitioning, which reflects the different states of the system in different time periods. The target data block partitioning results are subjected to inter-block similarity matrix calculation and spectral clustering to obtain the evolutionary segmentation result. The calculation of the inter-block similarity matrix can be based on Euclidean distance, cosine similarity or other similarity measurement methods. Spectral clustering finds the group structure between data blocks by analyzing the spectral characteristics of the similarity matrix, thereby achieving a more reasonable segmentation. Normal data blocks and drift data blocks are generated according to the evolutionary segmentation results. Normal data blocks represent the stable state in the assembly process, while drift data blocks imply changes or abnormalities in the system state.

[0029] In a specific embodiment, the process of executing step S200 may specifically include the following steps: The normal data blocks and the drift data blocks are merged to obtain a comprehensive data set, and the cosine similarity between samples is calculated based on the comprehensive data set to obtain a similarity matrix; A K-nearest neighbor graph is constructed based on the similarity matrix to obtain the local connection relationship between samples, and the geodesic distance and Euclidean distance between samples are calculated based on the local connection relationship to obtain a dual-metric distance matrix; An affinity matrix is ​​constructed based on the bimetric distance matrix to obtain the global geometric structure between samples, and the Laplacian matrix is ​​calculated based on the affinity matrix; Perform eigenvalue decomposition on the Laplace matrix to obtain eigenvalue sequence and eigenvector matrix, and apply the maximum eigenvalue interval criterion according to the eigenvalue sequence to obtain the optimal dimension reduction; The corresponding eigenvectors are selected from the eigenvector matrix based on the optimal dimensionality reduction to obtain a projection matrix, and a linear transformation is performed on the comprehensive data set based on the projection matrix to obtain a low-dimensional eigenvector.

[0030] Specifically, the normal data blocks and drift data blocks are integrated. The normal data blocks representing the stable state during the assembly of the brushless motor and the drift data blocks that may indicate faults or abnormalities are combined to form a more comprehensive integrated data set. The integrated data set can be regarded as a combination of multi-source and multi-state data, including the performance of the system in different states. The cosine similarity between samples is calculated based on the integrated data set. Cosine similarity is a commonly used method to measure the similarity between two vectors, which is suitable for similarity analysis of high-dimensional data. Assume that the integrated data set contains samples, each sample can be represented as a dimensional feature vector ( ), then the two samples and The cosine similarity between can be expressed as: ; in, yes and The inner product of and They are vectors and By calculating the cosine similarity between all sample pairs, we get a The similarity matrix ,in Representation sample and samples Based on the similarity matrix , construct a K nearest neighbor graph. The K nearest neighbor graph is constructed by selecting each sample The most similar neighbors are connected to the samples to form a graph structure. The graph structure reflects the local connection relationship of the samples in the feature space and helps to capture the local geometric characteristics of the data. , retaining its closest The edge weights between neighbors are set to zero, and the weights of other edges are set to zero. The resulting graph can be represented as the adjacency matrix ,in Representation sample and samples The connection relationship between them. Based on the local connection relationship, the geodesic distance and Euclidean distance between samples are calculated to obtain a dual-metric distance matrix. The geodesic distance is a distance defined on the graph structure, which measures the shortest path length of the sample on the graph and reflects the nonlinear relationship between samples. The Euclidean distance is the traditional straight-line distance that reflects the direct distance between samples in the feature space. For any two samples and , a dual-metric distance matrix It can be defined as: ; in, Representation sample and The geodesic distance between represents the Euclidean distance between them, is a weight parameter used to balance the influence of these two distance metrics. Through the dual-metric distance matrix, the local and global structural information of the sample is comprehensively considered. , construct affinity matrix , which describes the global geometric structure between samples. Each element of the affinity matrix It can be calculated by Gaussian kernel function: ; in, is the kernel width parameter, which controls the decay rate of similarity. By converting the distance between samples into similarity, the global geometric structure of the data is encoded into the matrix. The Laplacian matrix is ​​calculated based on the affinity matrix. Laplacian matrix It is a matrix that reflects the structure of the graph and can be defined by the following formula: ; in, is the degree matrix, which is a diagonal matrix with diagonal elements Representation sample The degree of is the sum of the weights of the edges connected to it. It reflects the overall structure of the graph and the global characteristics of data distribution. Perform eigenvalue decomposition on the Laplace matrix to obtain the eigenvalue sequence and the corresponding eigenvector matrix These eigenvalues ​​and eigenvectors represent different structural levels of the data. In order to determine the optimal dimensionality reduction, the maximum eigenvalue interval criterion is applied, that is, to find the pair of eigenvalues ​​with the largest interval in the eigenvalue sequence. and ,Will As the optimal dimension reduction number. Based on the determined optimal dimension reduction number , from the eigenvector matrix Before the selection eigenvectors, construct the projection matrix The projection matrix It is used to project high-dimensional data into low-dimensional space to achieve dimensionality reduction. , whose low-dimensional feature vector It can be calculated by the following formula: ; Low dimensional feature vector It represents the representation of each sample in the comprehensive data set in the space after dimensionality reduction. These low-dimensional feature vectors retain the main structural information of the data and eliminate the redundant information and noise that may exist in the high-dimensional space during the dimensionality reduction process.

[0031] In a specific embodiment, the process of executing step S300 may specifically include the following steps: Perform time series segmentation on the low-dimensional feature vector to obtain the source domain data set and the target domain data set; A multi-layer sparse autoencoder is constructed based on the source domain dataset. The multi-layer sparse autoencoder includes 5 encoding layers and 5 decoding layers. Each layer uses the ReLU activation function and L1 regularization. The source domain dataset is input into a multi-layer sparse autoencoder, and the first initial feature representation is obtained through unsupervised pre-training. A time-varying transfer learning network is constructed based on the first initial feature representation. The time-varying transfer learning network includes a feature extractor, a domain classifier, and a task classifier. The feature extractor adopts a ResNet structure, which contains 4 residual blocks, each of which contains two convolutional layers, a short-circuit connection, and an adaptive instance normalization layer. The domain classifier adopts a fully connected structure, which contains 3 hidden layers, and each layer uses a LeakyReLU activation function and Dropout regularization. The task classifier adopts an attention-enhanced LSTM structure, which contains 2 LSTM layers and 1 self-attention layer. The hidden state dimension of the LSTM layer is 128. The source domain dataset and the target domain dataset are input into the time-varying transfer learning network, and domain-invariant features are obtained through adversarial training and multi-task learning. Based on the domain-invariant features, a dynamic Kullback-Leibler divergence weight module is constructed, which includes a fully connected layer and a Softmax layer to dynamically balance the distance metric loss and the adversarial metric loss. The time-varying transfer learning network and the dynamic Kullback-Leibler divergence weight module are combined, and the first assembly state prediction model is obtained through end-to-end training.

[0032] Specifically, the low-dimensional feature vector is time-series segmented, and the data is divided into two different parts in chronological order: the source domain dataset and the target domain dataset. The source domain dataset usually contains historical data or data in a known state, while the target domain dataset contains current or future unknown state data. This segmentation can help the model learn and predict in different time periods, thereby better adapting to the changing characteristics of time-varying systems. Based on the source domain dataset, a multi-layer sparse autoencoder is constructed. The multi-layer sparse autoencoder is an unsupervised learning method in deep learning for feature extraction and data dimensionality reduction. In this model, the autoencoder consists of five encoding layers and five decoding layers. The role of the encoding layer is to map the high-dimensional features of the input to a low-dimensional hidden representation space, while the decoding layer attempts to reconstruct the original input data from the hidden representation. Each encoding layer and decoding layer uses the ReLU activation function, and the expression of the ReLU activation function is: ; in, is the input value, It means that the activation function retains positive values ​​and suppresses negative values ​​in the output. This property makes the model perform better when dealing with nonlinear problems. In order to ensure that the model can learn sparse features, L1 regularization terms are added to each layer. L1 regularization encourages the model to generate more zero weights, that is, sparsity, by adding the sum of the absolute values ​​of weights to the loss function. The expression of the L1 regularization term is: ; in, is the weight vector of the model, is the regularization strength coefficient. By introducing L1 regularization in each layer, a more concise feature representation is constructed to avoid model overfitting. The source domain dataset is input into a multi-layer sparse autoencoder, and the first initial feature representation is obtained through unsupervised pre-training. In the unsupervised pre-training process, the model learns the intrinsic structure of the data by minimizing the error between the input and the reconstruction, obtains a good initial feature representation, and provides an effective feature space for subsequent supervised learning. A time-varying transfer learning network is constructed based on the first initial feature representation. The network includes three main modules: feature extractor, domain classifier, and task classifier. The feature extractor is used to extract deep features from the input data and adopts the ResNet structure. ResNet (residual network) effectively solves the problem of gradient vanishing in deep neural network training by introducing residual connections. In the feature extractor, the ResNet structure contains four residual blocks, each of which includes two convolutional layers, a short-circuit connection, and an adaptive instance normalization layer. The convolutional layer is used to extract local patterns in the feature map, and the short-circuit connection skips one or more layers and passes the input directly to the output, thereby alleviating the gradient vanishing problem of the deep network. The adaptive instance normalization layer is used to normalize the feature map of each sample, thereby improving the robustness and generalization ability of the model. The domain classifier is a key module in the transfer learning network, which is used to distinguish the data distribution of the source domain and the target domain. The domain classifier adopts a fully connected structure and contains three hidden layers. Each hidden layer uses the LeakyReLU activation function, which is an improved ReLU function, expressed as: LeakyReLU ; in, is a small positive number (usually 0.01) used to retain the negative part of the input, thereby alleviating the "dead neurons" problem of the ReLU function in the negative area. At the same time, Dropout regularization is added to each layer. Dropout prevents the model from overfitting by randomly discarding some neurons. The task classifier uses an attention-enhanced LSTM structure to model time series data. LSTM (Long Short-Term Memory Network) is a neural network structure suitable for sequence data that can capture long-range dependencies in the data. In the task classifier, there are two LSTM layers and one self-attention layer. The hidden state dimension of the LSTM layer is 128, indicating that the state information at each moment is represented by 128 dimensions. The self-attention layer automatically adjusts the weight of each element by calculating the correlation between the elements in the input sequence, thereby highlighting important time points or features and improving the prediction ability of the model. The source domain dataset and the target domain dataset are input into the time-varying transfer learning network, and domain-invariant features are obtained through adversarial training and multi-task learning. Adversarial training introduces a game mechanism between the domain classifier and the feature extractor so that the features learned by the feature extractor remain consistent between the source domain and the target domain, while multi-task learning uses the relationship between tasks to enhance the generalization ability of the model by learning multiple tasks at the same time. After obtaining the domain-invariant features, a dynamic Kullback-Leibler divergence weight module is constructed. This module contains a fully connected layer and a Softmax layer to dynamically balance the distance metric loss and the adversarial metric loss. Kullback-Leibler divergence is an indicator that measures the difference between two probability distributions, and its formula is: ; in, and are two probability distributions. By dynamically adjusting the weight of the Kullback-Leibler divergence, the model can dynamically balance the distribution difference between the source domain and the target domain as needed at different training stages. The time-varying transfer learning network and the dynamic Kullback-Leibler divergence weight module are combined to obtain the first assembly state prediction model through end-to-end training. End-to-end training means that the entire model is trained in the same optimization process, thereby ensuring the coordinated work of each part and ultimately obtaining a prediction model that performs well in both the source domain and the target domain.

[0033] In a specific embodiment, the process of executing step S400 may specifically include the following steps: Preprocess the assembly feature data collected in real time to obtain standardized feature data; The standardized feature data is input into the multi-layer sparse autoencoder of the first assembly state prediction model, and the second initial feature representation is obtained through forward propagation calculation of 5 encoding layers; The second initial feature representation is input into the feature extractor of the first assembly state prediction model, and a deep feature representation is obtained through convolution operation of four residual blocks and adaptive instance normalization; The deep feature representation is input into the domain classifier of the first assembly state prediction model, and the domain probability distribution is obtained through calculations in three fully connected layers; Based on deep feature representation, the task classifier of the first assembly state prediction model is used for time series modeling, and the time series context features are obtained through calculations of two LSTM layers and one self-attention layer; The temporal context features are input into the dynamic Kullback-Leibler divergence weight module of the first assembly state prediction model, and the feature weights are obtained through calculations of the fully connected layer and the Softmax layer; Based on the feature weights, the temporal context features are weighted and summed to obtain a comprehensive feature representation, and the multi-dimensional indicator prediction value is calculated based on the comprehensive feature representation; An initial quality score is generated based on the predicted values ​​of the multi-dimensional indicators, and the initial quality score is dynamically adjusted to obtain the assembly quality score.

[0034] Specifically, the multi-source sensor data collected in real time during the assembly process of the brushless motor is preprocessed to eliminate the dimensional differences of each feature data, so that the data can be compared on the same scale to obtain standardized feature data. The standardized feature data is input into the multi-layer sparse autoencoder of the first assembly state prediction model. The multi-layer sparse autoencoder extracts the core features of the data through forward propagation of multiple encoding layers. The multi-layer sparse autoencoder consists of five encoding layers, and the function of each encoding layer is to gradually reduce the dimension of the input data and learn a more abstract feature representation. In the process of forward propagation, the input data is sequentially transformed by the weight matrix and processed by the nonlinear activation function of each layer to generate a second initial feature representation. The sparse autoencoder introduces L1 regularization during the training process to encourage the model to activate only a small number of neurons, thereby generating a sparse feature representation. This sparsity helps prevent overfitting and can extract the most important features in the data. The second initial feature representation is input into the feature extractor of the first assembly state prediction model. The feature extractor adopts a ResNet structure and includes four residual blocks. Each residual block consists of two convolutional layers, a short-circuit connection, and an adaptive instance normalization layer. The main function of the convolution layer is to extract local patterns in the input features through convolution operations, while the short-circuit connection skips certain layers and directly connects the input and output. The purpose of this design is to solve the common gradient vanishing problem in deep neural networks. The adaptive instance normalization layer is used to normalize the features of each sample to reduce the impact of differences between samples in training on the model. After being processed by the feature extractor, the deep feature representation of the input data is extracted. These deep features can capture more complex patterns and relationships in the data. The deep feature representation is input into the domain classifier for domain classification. The domain classifier consists of three fully connected layers, each layer uses the LeakyReLU activation function and Dropout regularization. The expression of the LeakyReLU activation function is: ; in, is the input value, is a coefficient less than 1, usually 0.01. The activation function allows a small number of negative values ​​to pass through, thereby solving the "dead neuron" problem of the ReLU function. At the same time, Dropout regularization reduces the overfitting of the model to the training data by randomly discarding some neurons. After calculation by the domain classifier, the model obtains the domain probability distribution of each sample, which helps the model distinguish whether the data is from the source domain or the target domain and guides the subsequent transfer learning process. Based on the deep feature representation, the data is input into the task classifier for time series modeling. The task classifier adopts an attention-enhanced LSTM (Long Short-Term Memory Network) structure, which contains two LSTM layers and one self-attention layer. The LSTM network is a special recurrent neural network (RNN) that can effectively capture the time dependency in sequence data. The LSTM layer processes and saves important information in the time series through memory units and gate mechanisms, thereby avoiding the gradient vanishing problem in traditional RNNs. The hidden state dimension of each LSTM layer is set to 128, indicating that the output information of each time step is expressed by 128 neurons. The self-attention layer automatically adjusts the weights by calculating the correlation of different time steps in the input sequence, highlighting the most important time steps or features for the task, and generating temporal context features. The temporal context features are input into the dynamic Kullback-Leibler divergence weight module. The module contains a fully connected layer and a Softmax layer to calculate feature weights. Kullback-Leibler divergence is an indicator that measures the difference between two probability distributions, and the expression is: ; in, and There are two distributions, is a random variable. By dynamically adjusting the weights of these features, the model can focus more on the most valuable features in the prediction process and improve the accuracy and robustness of the prediction. The Softmax layer converts the weights into probability distributions so that the sum of the weights is 1, thereby weighted summing the temporal context features to obtain a comprehensive feature representation. Based on the comprehensive feature representation, the model calculates the multidimensional index prediction values, which reflect the estimated results of key performance indicators in the assembly process. Multidimensional indicators are important reference points in the assembly process of brushless motors, such as current stability and speed consistency, which are directly related to the quality of assembly. Based on the multidimensional index prediction values, the model generates an initial quality score. The initial quality score is an overall evaluation of the current assembly process, but due to the complexity of the assembly process, the initial score may need to be further adjusted. The model corrects the initial quality score through a dynamic threshold adjustment mechanism to obtain the final assembly quality score. The dynamic threshold adjustment mechanism takes into account the volatility of the current assembly state and ensures that the final quality score is more accurate and objective by adaptively adjusting the evaluation criteria.

[0035] In a specific embodiment, the process of executing step S500 may specifically include the following steps: Perform time series analysis on the predicted values ​​of multi-dimensional indicators to obtain the changing trends of each indicator, and calculate the statistical characteristics of each indicator based on the changing trends to obtain the abnormal degree evaluation of the indicators; Threshold analysis is performed on the assembly quality score to obtain quality anomaly marks, and multimodal data fusion is performed to obtain a comprehensive anomaly index by combining indicator anomaly degree evaluation and quality anomaly marks; Perform cluster analysis on historical data based on the comprehensive anomaly index to obtain an anomaly pattern library, and match the current comprehensive anomaly index with the anomaly pattern library to obtain a preliminary anomaly type determination; According to the preliminary abnormality type determination, relevant features are extracted from the preset fault feature library to obtain the target feature set, and the multi-source sensor data is secondary analyzed based on the target feature set to obtain the target abnormality features; The target abnormality characteristics are compared with the preset diagnostic rule base to obtain specific fault diagnosis results, and the comprehensive abnormality index, preliminary abnormality type determination and specific fault diagnosis results are integrated to generate an abnormality report.

[0036] Specifically, a systematic time series analysis is performed on the predicted values ​​of the multidimensional indicators generated during the assembly process of the brushless motor to identify the changing rules and trends of each indicator in the time series and reveal potential anomalies in the assembly process. By analyzing the indicator values ​​at each time point, the changing trends of each indicator, such as rising, falling or fluctuating, are obtained. Trend analysis can be achieved through a simple linear regression model, or more complex time series analysis methods such as the ARIMA model (autoregressive integrated moving average model) can be used to capture trends and seasonal components. The statistical characteristics of each indicator are calculated based on the changing trend of each indicator. These statistical characteristics include mean, variance, skewness, kurtosis, etc., which can fully describe the distribution of indicators in the time series. For example, if the variance of an indicator increases significantly, it may mean that the indicator fluctuates abnormally in the current assembly process, indicating potential problems. Threshold analysis is performed on the assembly quality score to obtain quality anomaly marking. The core of threshold analysis is to set a reasonable threshold. When the assembly quality score exceeds this threshold, it can be marked as abnormal. The threshold can be set based on historical data or dynamically adjusted to adapt to different production conditions. Multimodal data fusion is performed by combining indicator abnormality evaluation and quality anomaly marking. By integrating information from multiple data sources, the independent abnormal evaluation results are unified to obtain a comprehensive abnormal index. Based on the comprehensive abnormal index, cluster analysis of historical data is performed to build an abnormal pattern library. The construction of the abnormal pattern library depends on the identification and classification of different abnormal patterns in historical data. Through cluster analysis, data with similar abnormal characteristics are clustered together to form different abnormal categories. By matching the current comprehensive abnormal index with the abnormal pattern library, the type of abnormality that may occur in the current assembly process is preliminarily determined. According to the determination results, relevant features are extracted from the preset fault feature library to obtain a target feature set. The fault feature library contains a large amount of historical fault data and feature descriptions. By matching with the current abnormal type, the most likely related features are selected as the target feature set. Based on the target feature set, secondary analysis is performed on multi-source sensor data to deeply explore possible abnormal root causes. Secondary analysis may include more sophisticated timing analysis, frequency domain analysis or other signal processing technologies to extract more specific abnormal features from the data. The target abnormal features obtained through analysis are compared with the preset diagnostic rule library. The diagnostic rule base contains fault diagnosis rules summarized based on engineering experience and historical data. These rules can quickly identify abnormal patterns in the data and provide corresponding fault diagnosis results. The comparison process of diagnostic rules can be implemented through decision trees, rule engines, or machine learning classification models. When the target abnormal feature matches the diagnostic rule, a specific fault diagnosis result is generated, which clearly indicates the nature and possible causes of the current abnormality. The comprehensive abnormality index, preliminary abnormality type determination, and specific fault diagnosis results are integrated to generate an abnormality report.

[0037] In a specific embodiment, the process of executing step S600 may specifically include the following steps: Generate an abnormal data matrix based on the abnormal report, and build a multi-dimensional reward and punishment function based on the abnormal data matrix; Perform parameter sensitivity analysis on the first assembly state prediction model. By perturbing each parameter and observing the change in model output, the sensitivity coefficient of each parameter is calculated to obtain a parameter importance list sorted by sensitivity. According to the parameter importance list, select the top N most sensitive parameters, design a discrete adjustment interval for each parameter, form a high-dimensional discrete action space through Cartesian product combination, and obtain a set of candidate solutions for model adjustment; Using the reward and punishment function as the evaluation criteria, a Monte Carlo search tree is constructed based on the candidate solution set, and the optimal parameter adjustment path is obtained through iterative expansion, simulation and backtracking processes; Along the optimal parameter adjustment path, the parameters of the first assembly state prediction model are gradually updated, and the performance is evaluated using the validation set after each update until a preset number of iterations is reached or the performance converges, thereby obtaining a second assembly state prediction model; The second assembly state prediction model is used for backtracking analysis to obtain an optimized sample set after annotation. Based on the optimized sample set after annotation, a multi-layer decision tree is constructed. Each layer represents a key decision point in the assembly process. Each node of the tree contains decision conditions, operation instructions and expected results. Through recursive splitting and pruning, an assembly process decision model is obtained. The assembly process decision model is simulated and tested to simulate the distribution of assembly results under different decision paths. The expected benefit and risk coefficient of each path are calculated. The efficiency and stability are balanced through a multi-objective optimization algorithm to obtain the brushless motor assembly strategy.

[0038] Specifically, key information is extracted from the abnormal report, which usually includes the occurrence time of various abnormal events, the characteristics involved, the degree of impact, and the system response. These data can be organized into an abnormal data matrix , whose rows represent different abnormal events and columns represent different characteristics or dimensions, such as temperature, current, vibration, etc. Each element in the matrix Representative The abnormal event is Outliers or indicators on feature dimensions, such as the degree of exceeding the normal range. Based on the abnormal data matrix , construct a multi-dimensional reward and punishment function ,in Represents the model parameter vector. The reward and penalty function is used to guide the optimization direction during the model parameter adjustment process. The specific form can be: ; in, is the number of abnormal events, is the number of feature dimensions, It is The weight of a feature reflects its importance. is a loss function, which means The abnormal event is The degree of penalty on each feature dimension can usually be designed as some form of error or deviation function. By constructing a reward and punishment function, a comprehensive guide for model parameters is provided by taking into account multiple abnormal events and feature dimensions. Parameter sensitivity analysis is performed on the first assembly state prediction model to evaluate the degree of influence of each model parameter on the final output. By perturbing each parameter one by one and observing the changes in the model output, the sensitivity coefficient of each parameter is calculated. ,Right now: ; in, is the model output, It is By calculating the sensitivity coefficient, we can get a parameter importance list sorted by sensitivity, which reflects which parameters have the greatest impact on the model prediction accuracy. The most sensitive parameters are designed with a discrete adjustment interval for each parameter. The adjustment interval represents the possible value range of the parameter, and the high-dimensional discrete action space is formed through the Cartesian product combination. , where each point Represents a parameter combination scheme. These combination schemes constitute a candidate set of model adjustment, providing rich choices for further optimization search. Using the reward and punishment function as the evaluation criterion, based on the candidate set Construct a Monte Carlo search tree. The Monte Carlo search tree obtains the optimal parameter adjustment path by randomly sampling and simulating possible parameter adjustment paths in the parameter space, and gradually optimizing through iterative expansion, simulation and backtracking processes. The Monte Carlo search tree approaches the optimal solution through the statistical properties of a large number of random samples. In each simulation, the benefit of the current path is evaluated by the reward and punishment function, and the subsequent search direction is guided according to this benefit, gradually approaching the optimal solution. Along the optimal parameter adjustment path, the parameters of the first assembly state prediction model are gradually updated. After each parameter update, the model performance is evaluated using an independent validation set. This process usually iterates multiple times until the preset number of iterations is reached or the performance converges. The final model is the second assembly state prediction model, which can more accurately predict the assembly state under the optimized parameters. After obtaining the second assembly state prediction model, the model is used for backtracking analysis to obtain the labeled optimized sample set. Backtracking analysis identifies which samples are judged according to the optimized model by applying the model to historical data, and adds optimized labels to these samples. The labeled sample set reflects the assembly state under the new model. Based on the annotated optimized sample set, a multi-layer decision tree is constructed. Each layer of the decision tree represents a key decision point in the assembly process, and each node of the tree contains decision conditions, operation instructions, and expected results. Through recursive splitting and pruning, a concise and efficient assembly process decision model is obtained. Recursive splitting refers to dividing the data set into different subsets according to the characteristics of the sample, and pruning is to simplify the decision tree by removing unnecessary branches, thereby improving the generalization ability of the model. The assembly process decision model is simulated and tested to simulate the distribution of assembly results under different decision paths. In the simulation test, the expected benefits and risk coefficients of each path are calculated to evaluate the advantages and disadvantages of different paths. The expected benefits usually indicate the benefits that the path can bring in most cases, while the risk coefficient reflects the worst possible result of the path. Through the multi-objective optimization algorithm, a trade-off is made between efficiency and stability, and finally an optimal brushless motor assembly strategy is obtained.

[0039] See also Figure 2 , Figure 2 The structure schematic block diagram of the assembly test device 200 of the brushless motor provided in the embodiment of the present application is as follows: Figure 2 As shown, the assembly test device 200 of the brushless motor includes: The segmentation module 210 is used to perform data stream segmentation processing on the multi-source sensor data in the brushless motor assembly process to obtain normal data blocks and drift data blocks; A dimension reduction module 220 is used to perform dimension reduction processing based on the normal data block and the drift data block using a Laplace eigenmapping algorithm with dual metric constraints to obtain a low-dimensional feature vector; A training module 230 is used to input the low-dimensional feature vector into a deep multi-layer sparse regularized time-varying transfer learning network model for training to obtain a first assembly state prediction model; A processing module 240 is used to process the assembly feature data collected in real time based on the first assembly state prediction model to obtain a multi-dimensional index prediction value and an assembly quality score; A detection module 250 is used to perform multi-modal anomaly detection and diagnosis on the assembly process of the brushless motor based on the multi-dimensional index prediction value and the assembly quality score to obtain an anomaly report; The generation module 260 is used to perform reinforcement learning and incremental learning on the first assembly state prediction model based on the abnormal report to obtain the second assembly state prediction model and generate a corresponding brushless motor assembly strategy.

[0040] Through the collaboration of the above components, the high-dimensional multi-source sensor data is effectively processed through data stream segmentation processing and dual-metric constrained Laplace eigenmapping algorithm, which improves the accuracy and efficiency of feature extraction. The deep multi-layer sparse regularized time-varying transfer learning network model is adopted to enhance the model's adaptability to different working conditions and new models of products, reducing the need for manual intervention. The multimodal anomaly detection method based on multi-dimensional indicator prediction values ​​and assembly quality scores improves the recognition accuracy of abnormal states and reduces false positives and false negatives. By building an abnormal pattern library and a diagnostic rule library, the fault diagnosis capability of the system is enhanced, which can provide more specific and accurate fault cause analysis. Using reinforcement learning and incremental learning methods, the model is continuously optimized, enabling the system to continuously adapt to new production conditions and requirements. By improving the accuracy and timeliness of anomaly detection, unnecessary downtime and scrap rate are reduced, thereby reducing production costs. The stability and consistency of the brushless motor assembly quality are ensured, which helps to improve the overall quality level of the product. The automated anomaly detection and diagnosis process reduces manual intervention and improves the operating efficiency of the production line.

[0041] The present application also provides an assembly and testing device for a brushless motor, the assembly and testing device for the brushless motor comprising a memory and a processor, the memory storing computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the assembly and testing method for the brushless motor in the above-mentioned embodiments.

[0042] The present application also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the assembly and testing method of the brushless motor.

[0043] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0044] 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 application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc., various media that can store program codes.

[0045] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A brushless motor assembly test method, It is characterized in that The method comprises: The multi-source sensor data in the brushless motor assembly process is processed by data stream segmentation to obtain normal data blocks and drift data blocks; Based on the normal data block and the drift data block, a dual-metric constrained Laplace eigenmapping algorithm is used to perform dimensionality reduction processing to obtain a low-dimensional feature vector; Inputting the low-dimensional feature vector into a deep multi-layer sparse regularized time-varying transfer learning network model for training to obtain a first assembly state prediction model; Processing the assembly feature data collected in real time based on the first assembly state prediction model to obtain a multi-dimensional index prediction value and an assembly quality score; Performing multi-modal anomaly detection and diagnosis on the assembly process of the brushless motor based on the multi-dimensional index prediction value and the assembly quality score to obtain an anomaly report; Based on the abnormal report, reinforcement learning and incremental learning are performed on the first assembly state prediction model to obtain a second assembly state prediction model, and a corresponding brushless motor assembly strategy is generated.

2. The assembly and testing method of the brushless motor according to claim 1, It is characterized in that The method of performing data stream segmentation processing on multi-source sensor data in the brushless motor assembly process to obtain normal data blocks and drift data blocks includes: Performing sliding time window segmentation on multi-source sensor data in the brushless motor assembly process to obtain an initial data block sequence, and performing divergence calculation on adjacent data blocks in the initial data block sequence to obtain a dissimilarity measure between the data blocks; According to the dissimilarity measure, the initial data block sequence is segmented by an adaptive threshold to obtain a candidate data block set, and time domain statistical features and wavelet packet energy features are extracted from each data block in the candidate data block set to obtain a multi-dimensional feature vector set; Probability density estimation is performed based on the multidimensional feature vector set to obtain a probability distribution model of the data block, and the log-likelihood value and distance of each data block are calculated according to the probability distribution model to obtain an abnormal measurement sequence; Detecting the abnormal measurement sequence to obtain a potential change point set, and based on the potential change point set, hierarchically clustering the candidate data block set to obtain a target data block division result; The target data block division result is subjected to inter-block similarity matrix calculation and spectral clustering to obtain an evolutionary segmentation result, and a normal data block and a drift data block are generated according to the evolutionary segmentation result.

3. The assembly and testing method of the brushless motor according to claim 1, It is characterized in that The step of performing dimensionality reduction processing based on the normal data block and the drift data block using a dual-metric constrained Laplace eigenmapping algorithm to obtain a low-dimensional feature vector includes: Merging the normal data block and the drift data block to obtain a comprehensive data set, and performing cosine similarity calculation between samples based on the comprehensive data set to obtain a similarity matrix; Constructing a K-nearest neighbor graph according to the similarity matrix to obtain a local connection relationship between samples, and calculating the geodesic distance and the Euclidean distance between samples based on the local connection relationship to obtain a dual-metric distance matrix; constructing an affinity matrix according to the bimetric distance matrix to obtain a global geometric structure between samples, and calculating a Laplace matrix based on the affinity matrix; Performing eigenvalue decomposition on the Laplace matrix to obtain an eigenvalue sequence and an eigenvector matrix, and applying a maximum eigeninterval criterion according to the eigenvalue sequence to obtain an optimal dimension reduction; Based on the optimal dimension reduction number, corresponding eigenvectors are selected from the eigenvector matrix to obtain a projection matrix, and a linear transformation is performed on the comprehensive data set based on the projection matrix to obtain a low-dimensional eigenvector.

4. The assembly and testing method of the brushless motor according to claim 1, It is characterized in that The step of inputting the low-dimensional feature vector into a deep multi-layer sparse regularized time-varying transfer learning network model for training to obtain a first assembly state prediction model comprises: Performing time series segmentation on the low-dimensional feature vector to obtain a source domain data set and a target domain data set; Constructing a multi-layer sparse autoencoder based on the source domain dataset, wherein the multi-layer sparse autoencoder includes 5 encoding layers and 5 decoding layers, and each layer uses a ReLU activation function and L1 regularization; Input the source domain data set into the multi-layer sparse autoencoder, obtain a first initial feature representation through unsupervised pre-training, and construct a time-varying transfer learning network based on the first initial feature representation, wherein the time-varying transfer learning network includes a feature extractor, a domain classifier, and a task classifier; the feature extractor adopts a ResNet structure, including 4 residual blocks, each residual block includes two convolutional layers, a short-circuit connection, and an adaptive instance normalization layer; the domain classifier adopts a fully connected structure, including 3 hidden layers, each layer uses a LeakyReLU activation function and Dropout regularization; the task classifier adopts an attention-enhanced LSTM structure, including 2 LSTM layers and 1 self-attention layer, and the hidden state dimension of the LSTM layer is 128; Inputting the source domain dataset and the target domain dataset into the time-varying transfer learning network, and obtaining domain-invariant features through adversarial training and multi-task learning; Based on the domain invariant features, a dynamic Kullback-Leibler divergence weight module is constructed, which includes a fully connected layer and a Softmax layer for dynamically balancing the distance metric loss and the adversarial metric loss; The time-varying transfer learning network and the dynamic Kullback-Leibler divergence weight module are combined to obtain a first assembly state prediction model through end-to-end training.

5. The assembly and testing method of the brushless motor according to claim 4, It is characterized in that The assembly feature data collected in real time is processed based on the first assembly state prediction model to obtain a multi-dimensional index prediction value and an assembly quality score, including: Preprocess the assembly feature data collected in real time to obtain standardized feature data; Inputting the standardized feature data into the multi-layer sparse autoencoder of the first assembly state prediction model, and performing forward propagation calculations of 5 encoding layers to obtain a second initial feature representation; Inputting the second initial feature representation into the feature extractor of the first assembly state prediction model, and obtaining a deep feature representation through convolution operation of four residual blocks and adaptive instance normalization; Inputting the deep feature representation into the domain classifier of the first assembly state prediction model, and obtaining domain probability distribution through calculation of three fully connected layers; Based on the deep feature representation, the task classifier of the first assembly state prediction model is used to perform time series modeling, and the time series context features are obtained through calculations of two LSTM layers and one self-attention layer; Inputting the temporal context features into the dynamic Kullback-Leibler divergence weight module of the first assembly state prediction model, and obtaining feature weights through calculations of a fully connected layer and a Softmax layer; Performing weighted summation on the temporal context features based on the feature weights to obtain a comprehensive feature representation, and calculating a multi-dimensional indicator prediction value based on the comprehensive feature representation; An initial quality score is generated based on the multi-dimensional indicator prediction value, and a dynamic threshold adjustment is performed on the initial quality score to obtain an assembly quality score.

6. The assembly and testing method of the brushless motor according to claim 1, It is characterized in that The method of performing multi-modal anomaly detection and diagnosis on the assembly process of the brushless motor based on the multi-dimensional index prediction value and the assembly quality score to obtain an anomaly report includes: Performing time series analysis on the predicted values ​​of the multi-dimensional indicators to obtain the change trend of each indicator, and calculating the statistical characteristics of each indicator based on the change trend to obtain an evaluation of the abnormality degree of the indicator; Performing a threshold analysis on the assembly quality score to obtain a quality anomaly mark, and combining the indicator anomaly degree evaluation and the quality anomaly mark to perform multimodal data fusion to obtain a comprehensive anomaly index; Performing cluster analysis on historical data based on the comprehensive anomaly index to obtain an anomaly pattern library, and matching the current comprehensive anomaly index with the anomaly pattern library to obtain a preliminary anomaly type determination; According to the preliminary abnormality type determination, relevant features are extracted from a preset fault feature library to obtain a target feature set, and a secondary analysis is performed on the multi-source sensor data based on the target feature set to obtain a target abnormality feature; The target abnormality feature is compared with a preset diagnostic rule library to obtain a specific fault diagnosis result, and the comprehensive abnormality index, the preliminary abnormality type determination and the specific fault diagnosis result are integrated to generate an abnormality report.

7. The assembly and testing method of the brushless motor according to claim 1, It is characterized in that The method of performing reinforcement learning and incremental learning on the first assembly state prediction model based on the abnormal report to obtain a second assembly state prediction model and generate a corresponding brushless motor assembly strategy includes: Generate an abnormal data matrix according to the abnormal report, and construct a multi-dimensional reward and punishment function based on the abnormal data matrix; Performing parameter sensitivity analysis on the first assembly state prediction model, calculating the sensitivity coefficient of each parameter by perturbing each parameter and observing the change in the model output, and obtaining a parameter importance list sorted by sensitivity; According to the parameter importance list, the top N most sensitive parameters are selected, a discretized adjustment interval is designed for each parameter, a high-dimensional discrete action space is formed through Cartesian product combination, and a set of candidate solutions for model adjustment is obtained; Using the reward and punishment function as an evaluation criterion, a Monte Carlo search tree is constructed based on the candidate solution set, and an optimal parameter adjustment path is obtained through iterative expansion, simulation and backtracking processes; Along the optimal parameter adjustment path, the parameters of the first assembly state prediction model are gradually updated, and the performance is evaluated using a validation set after each update until a preset number of iterations is reached or the performance converges, thereby obtaining a second assembly state prediction model; Using the second assembly state prediction model to perform backtracking analysis to obtain a labeled optimized sample set, and based on the labeled optimized sample set, construct a multi-layer decision tree, where each layer represents a key decision point in the assembly process, and each node of the tree contains a decision condition, an operation instruction, and an expected result, and obtains an assembly process decision model through recursive splitting and pruning; The assembly process decision model is simulated and tested to simulate the distribution of assembly results under different decision paths. The expected return and risk coefficient of each path are calculated. The efficiency and stability are balanced through a multi-objective optimization algorithm to obtain the brushless motor assembly strategy.

8. An assembly and testing device for a brushless motor, It is characterized in that For performing the assembly test method of the brushless motor according to any one of claims 1 to 7, the device comprises: A segmentation module is used to perform data stream segmentation processing on multi-source sensor data in the brushless motor assembly process to obtain normal data blocks and drift data blocks; A dimensionality reduction module, used to perform dimensionality reduction processing based on the normal data block and the drift data block by using a Laplace eigenmapping algorithm with dual metric constraints to obtain a low-dimensional feature vector; A training module, used for inputting the low-dimensional feature vector into a deep multi-layer sparse regularized time-varying transfer learning network model for training to obtain a first assembly state prediction model; A processing module, used to process the assembly feature data collected in real time based on the first assembly state prediction model to obtain a multi-dimensional index prediction value and an assembly quality score; A detection module, used for performing multi-modal anomaly detection and diagnosis on the assembly process of the brushless motor based on the multi-dimensional index prediction value and the assembly quality score to obtain an anomaly report; A generation module is used to perform reinforcement learning and incremental learning on the first assembly state prediction model based on the abnormal report to obtain a second assembly state prediction model and generate a corresponding brushless motor assembly strategy.

9. A brushless motor assembly and testing device, It is characterized in that The assembly and testing equipment of the brushless motor comprises: a memory and at least one processor, wherein instructions are stored in the memory; The at least one processor calls the instructions in the memory to enable the assembly and testing equipment for the brushless motor to execute the assembly and testing method for the brushless motor according to any one of claims 1 to 7.

10. A computer-readable storage medium having instructions stored thereon, It is characterized in that When the instructions are executed by the processor, the assembly test method of the brushless motor according to any one of claims 1 to 7 is implemented.

Images