Micro bearing production and manufacturing detection data analysis method and system based on big data

Abstract

The application discloses a micro bearing production and manufacturing detection data analysis method and system based on big data, relates to the technical field of data processing, and pre-trains a recurrent time sequence network through collection of vibration signals of bearings in a known state, obtains an evolution feature vector of bearings in the same batch through the network combined with a transient noise suppression operator, extracts a static feature vector combined with a machining time sequence and an end face image, corrects a visual missing sample after splicing, constructs a reconstructed kernel feature matrix, obtains a feature transformation and a mapping matrix based on two-way alternate optimization of the matrix and a real label, establishes a bearing state database, collects multi-modal data of a bearing to be detected to generate an initial to-be-detected feature vector, calculates the similarity of the to-be-detected feature vector with database samples after mapping, and finally outputs a detection result. The application deeply fuses multi-modal features, effectively overcomes interference such as workshop visual pollution, thermal expansion and group tolerance drift, and greatly improves detection accuracy and robustness.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for analyzing micro-bearing production and testing data based on big data. Background Technology

[0002] Miniature bearings are widely used in many high-end manufacturing fields such as aerospace, medical devices, electronic equipment, and precision instruments. The performance and reliability of these devices place extremely high demands on the precision, quality, and stability of miniature bearings. In the production and manufacturing process of miniature bearings, timely and accurate quality inspection is a key link in controlling the yield rate and preventing unqualified products from entering the next process or market. However, traditional miniature bearing inspection methods mostly rely on manual sampling or offline laboratory testing equipment, which is not only inefficient and difficult to meet the pace of large-scale production, but also the inspection results are easily affected by subjective factors, with a large probability of misjudgment and omission. It is impossible to achieve real-time monitoring and feedback of products on the production line. In order to improve the efficiency of online inspection of miniature bearings, some automated inspection solutions based on machine vision have been proposed in the industry. Chinese invention patent document CN120450205B discloses a rapid batch online inspection method and system for miniature bearings. This solution reconstructs and dynamically registers three-dimensional point clouds of multi-dimensional datasets, extracts defect features by traversing key inspection areas, generates defect probability distribution maps and type label groups, and finally guides the online sorting of bearings. In real continuous cutting or grinding environments, the end faces of miniature bearings are prone to adhesion of cutting fluid and large areas of metal dust. This contamination can lead to failure in optical edge feature extraction or severe visual artifacts. Existing technologies that rely solely on visual images lack effective feature compensation mechanisms when faced with local visual defects, making them prone to misjudgment. Current technologies are limited to static optical images and geometric features, ignoring the dynamic vibration acceleration response of miniature bearings during operation. Vibration signals often contain more fundamental evolutionary features such as early fatigue cracks. Existing technologies lack effective cross-dimensional verification of multimodal signals such as vision and vibration. As the continuous machining time of the machine tool increases, the spindle inevitably experiences nonlinear cumulative thermal expansion, and the tool also undergoes microscopic wear. These time-accumulated factors can cause a collective tolerance drift background noise in the entire batch of bearings. Existing detection schemes fail to incorporate the time interval factor of machining operations into the process compensation model, making it easy to misjudge this systematic distribution deviation caused by changes in equipment thermal balance as a batch of product defects. Summary of the Invention

[0003] The technical problem solved by this invention is that existing technologies that rely solely on visual images lack effective feature compensation mechanisms when faced with local visual defects, which can easily lead to misjudgments. Existing technologies are limited to static optical images and geometric structural features, ignoring the dynamic vibration acceleration response of micro-bearings during operation. Existing technologies lack effective cross-dimensional verification of multi-modal signals such as vision and vibration. As the continuous processing time of the machine tool increases, the accumulation of time-series factors can cause the entire batch of bearings to generate collective tolerance drift background noise. Existing detection schemes fail to incorporate the time interval factor of processing operations into the process compensation.

[0004] To address the aforementioned technical problems, this invention provides the following technical solution: a method for analyzing micro-bearing manufacturing and testing data based on big data, comprising the following steps: Step S1: Collect vibration acceleration signals of the first type of bearing samples and pre-train the recursive time series network. Step S2: Collect vibration acceleration signals of the second type of bearing samples, obtain the evolution feature vector of the second type of bearing samples by combining the sliding interception operation of the pre-trained recursive temporal network with the transient noise suppression operator, collect the end face visual image and cumulative processing time of the second type of bearing samples, obtain the static feature vector, concatenate the static feature vector and the evolution feature vector into the initial sample feature vector, perform feature correction on samples with local visual defects to obtain the final sample feature vector, and construct the reconstructed kernel feature matrix; Step S3: Based on the reconstructed kernel feature matrix and the true labels of the second type of bearing samples, construct the objective function, perform bidirectional alternating optimization, solidify the feature transformation weight matrix and the mapping matrix, recalculate the sample feature vectors of all second type of bearing samples using the solidified feature transformation weight matrix, and construct the bearing state database. Step S4: Collect the cumulative processing time, end face visual image and vibration acceleration signal of the micro bearing to be tested, obtain the static feature vector and evolution feature vector of the bearing to be tested, and splice them into the initial feature vector to be tested; Step S5: Based on the solidified feature transformation weight matrix, the initial feature vector to be detected is processed, and the similarity between the processed initial feature vector to be detected and the sample feature vector in the bearing status database is calculated. The final detection result of the bearing to be detected is obtained by combining the mapping matrix.

[0005] Preferably, in step S1, Vibration signals of bearing samples in various known states were collected based on a large database to obtain vibration acceleration sequences; The recursive temporal network is pre-trained using miniature bearing samples with known states to obtain the pre-trained recursive temporal network. The recursive temporal network includes a feature extraction layer, a feature transition layer, and a feature output layer.

[0006] Preferably, in step S2, the second type of bearing sample is a sample from the same batch of bearings to be tested; The process of obtaining the evolutionary feature vector of the second type of bearing sample specifically includes: Vibration acceleration signals of the second type of bearing samples are collected and input into a pre-trained recursive temporal network. After processing through the feature extraction layer to output the first hidden state sequence matrix, the sliding interception operation of the transient noise suppression operator is performed on the first hidden state sequence matrix to obtain the filter sequence matrix. The process of the sliding interception operation of the transient noise suppression operator specifically includes: In the row direction of the first-level hidden state sequence matrix, the first-level hidden state sequence matrix is ​​divided by a preset sliding window according to a preset sliding step size to obtain multiple local state sub-matrices. The impact energy score of the local state submatrix is ​​calculated, and the mathematical expression for the impact energy score is as follows: ; in, For the first The impact energy score of a local state submatrix, where m takes values ​​of 1, 2, ..., M. Representing the In the local state submatrix, the th... Line 1 Column elements; Let be the row number of the local state submatrix. is the number of columns in the local state submatrix; If the impact energy score is less than the preset gap vibration threshold, the local state submatrix will be transformed into a zero mask matrix. If the impact energy score is greater than or equal to the preset clearance vibration threshold, the local state submatrix remains unchanged. The local state submatrix and the all-zero mask matrix are concatenated along the time dimension in their original order. The all-zero row vectors added during the segmentation stage are removed to obtain a filter sequence matrix with the same dimension as the first-layer hidden state sequence matrix. After the filtering matrix is ​​processed through the feature transition layer and feature output layer of the recursive temporal network, the evolution feature vector of the second type of bearing sample is obtained.

[0007] Preferably, the process of obtaining the static feature vector of the second type of bearing sample in step S2 specifically includes: The cumulative processing time of the second type of bearing samples is collected, and the process compensation coefficient is calculated. The mathematical expression of the process compensation coefficient is as follows: ; in, This is the process compensation coefficient. The preset thermal decay constant This is the cumulative processing time; Acquire visual images of the end faces of the second type of bearing samples and convert the visual images of the end faces into grayscale matrices; Based on the global pheromone evaporation factor and the grayscale matrix, the pheromone concentration of each pixel is obtained through the ant colony algorithm, and pixels with pheromone concentrations greater than a preset concentration threshold are designated as coarse edge points. If the number of coarse edge points is less than the preset lower limit threshold, the visual feature extraction is deemed to have failed, and the static feature vector of the bearing to be detected is a vector of all zeros. If the number of coarse edge points is greater than or equal to the preset lower limit threshold, then for each coarse edge point, calculate the Zernike orthogonal moment, and use the Zernike orthogonal moment to solve for the normal distance and edge angle corresponding to the pixel. The sub-pixel offsets of the pixel coordinates of the coarse edge point in the horizontal and vertical directions are calculated using trigonometric functions and then superimposed onto the pixel coordinates of the coarse edge point to obtain the sub-pixel coordinates of the coarse edge point. Construct an objective function that minimizes the sum of squared errors, fit a circular equation to the sub-pixel coordinates of coarse edge points, solve for the parameters that minimize the objective function using the least squares method, and output a static feature vector. The static feature vector and the evolution feature vector of each sample in the second type of bearing sample are concatenated to generate the initial sample feature vector.

[0008] Preferably, step S2, the process of obtaining the final sample feature vector, specifically includes: If the sum of the absolute values ​​of the first three elements of the initial sample feature vector is equal to zero, the current second type of bearing sample is marked as a locally visually missing sample, and the initial sample feature vector of the locally visually missing sample is corrected to obtain the final sample feature vector of the locally visually missing sample. If the sum of the absolute values ​​of the first three elements of the initial sample feature vector is not equal to zero, then the current second type of bearing sample is marked as a visually complete sample. The initial sample feature vector of the visually complete sample is multiplied by the feature transformation weight matrix, and after mapping through the Sigmoid activation function, the final sample feature vector of the visually complete sample is obtained. The process of obtaining the final sample feature vector for any locally visually missing sample specifically includes: The arithmetic mean of the evolution feature vectors of all normal second-class bearing samples is used as the baseline feature vector. The sum of squared Euclidean distances between the baseline feature vector and the evolution feature vectors of locally visually missing samples is calculated. The sum of squared Euclidean distances is input into the exponential decay function to calculate the prior probability of damage. Collect the timestamp of the end of processing of the local visual missing sample. For the local visual missing sample, obtain the second type of bearing sample with adjacent relationship based on the timestamp of the end of processing, and calculate the initial attention correlation score between the local visual missing sample and its adjacent second type of bearing sample. Calculate the prior damage probability of adjacent second-type bearing samples. Adjust the initial attention relevance score using the prior damage probabilities of locally visually missing samples and adjacent second-type bearing samples, obtaining the adjustment weight coefficients. The adjustment rule is as follows: If the prior probability of damage of a local visual missing sample and the prior probability of damage of an adjacent second-class bearing sample are both greater than or equal to a preset probability threshold, or both are less than a preset probability threshold, then the adjustment weight coefficient of the adjacent second-class bearing sample is 1. If the prior probability of damage of a locally visually missing sample and the prior probability of damage of an adjacent second-class bearing sample are both greater than or equal to a preset probability threshold and the other is less than a preset probability threshold, then the adjustment weight coefficient of the adjacent second-class bearing sample is 1 minus the absolute value of the difference between the prior probability of damage of the locally visually missing sample and the adjacent second-class bearing sample. The product of the adjustment weight coefficient of each adjacent second-class bearing sample and the initial attention relevance score is used as the modulation attention relevance score. Softmax normalization is performed on all modulation attention relevance scores to obtain the aggregate weight corresponding to each adjacent second-class bearing sample. Weighted summation is performed on each adjacent second-class bearing sample of the local visual missing sample to obtain the final sample feature vector of the local visual missing sample.

[0009] Preferably, the process of constructing the reconstructed kernel feature matrix in step S3 specifically includes: Extract the final sample feature vectors of all second-class bearing samples, reassemble them by row vectors according to the processing order to form a corrected sample feature matrix, and use the radial basis kernel function to construct the initial estimated kernel matrix; Based on the processing timestamps of the second type of bearing samples, obtain pairs of second type bearing samples with adjacent relationships, calculate the processing time difference between the bearing sample pairs, and calculate the average and variance of all processing time differences; Calculate fractional parameters using the inverse proportional mapping function; Perform orthogonal eigenvalue decomposition on the initial estimated kernel matrix, and perform exponential shrinkage mapping on each main diagonal element of the diagonal matrix obtained after orthogonal eigenvalue decomposition to obtain the reconstructed diagonal matrix; The reconstructed diagonal matrix, orthogonal eigenvector matrix, and transpose of the orthogonal eigenvector matrix are used to restore the original matrix and generate the reconstructed kernel feature matrix.

[0010] Preferably, the bidirectional alternating optimization process in step S3 specifically includes: The second type of bearing samples are divided into a support set and a query set according to a preset ratio; The feature transformation weight matrix and mapping matrix are obtained through random initialization using a normal distribution; The weight matrix of the locked feature transformation does not participate in the differentiation. From the reconstructed kernel feature matrix, all row vectors of the second type of bearing samples whose row index belongs to the support set are extracted and concatenated into the support kernel matrix. Extract the state labels of the second type of bearing samples within the support set, construct a support label matrix, and calculate the first loss function. The mathematical expression of the first loss function is: ; in, For the first loss function, To support the set, To support the first in the kernel matrix row vectors The mapping matrix to be updated, To support the first in the tag matrix row vectors This is the regularization penalty coefficient; To support the total number of second-class bearing samples in the set; The gradient descent algorithm is used to update all elements in the mapping matrix along the negative gradient direction of the first loss function with a preset inner learning rate, so as to obtain the mapping matrix of the current iteration step.

[0011] Preferably, in step S3, The inner updated mapping matrix is ​​locked and not differentiated. From the reconstructed kernel feature matrix, all row vectors whose row indices belong to the query set are extracted to form the query kernel matrix. The query kernel matrix is ​​multiplied by the updated mapping matrix and then subjected to Softmax normalization to obtain the prediction probability matrix of the query set. Calculate the second loss function, whose mathematical expression is: ; in, For the second loss function, For query set, For the query set The probability that a sample is in a normal state. For the query set The probability that a sample state is abnormal. and These are extracted from the prediction probability matrix. The probability that a sample state is predicted to be normal or abnormal; To query the total number of bearing samples in the set; Using the chain rule, backpropagation is performed to the feature transformation weight matrix and the attention projection vector, and the elements of the feature transformation weight matrix and the attention projection vector are updated with a preset outer learning rate; Calculate the absolute value of the difference between the first loss function and the second loss function. If the absolute value of the difference is greater than or equal to the preset convergence threshold, then regenerate the corrected sample feature matrix and reconstruct the kernel feature matrix using the updated feature transformation weight matrix, and solve the mapping matrix again. If the absolute value of the difference is less than the preset convergence threshold or the total number of alternating iterations reaches the set number, bidirectional convergence is determined, training is terminated, and the feature transformation weight matrix and mapping matrix are solidified.

[0012] Preferably, the process of obtaining the test results of the bearing to be tested specifically includes: Collect the cumulative processing time and end face visual images of the bearing to be inspected to obtain the static feature vector of the bearing to be inspected; The vibration acceleration signal of the bearing to be detected is collected and input into a pre-trained recursive temporal network. After processing through the feature extraction layer and outputting the first hidden state sequence matrix, the sliding interception operation of the transient noise suppression operator is performed on the first hidden state sequence matrix to obtain the filter sequence matrix. After the filtering matrix is ​​processed through the feature transition layer and feature output layer of the recursive temporal network, the evolution feature vector of the bearing to be detected is obtained. The static feature vector and the evolved feature vector of the bearing to be tested are concatenated to form the initial feature vector to be tested; The initial feature vector to be detected is processed by the solidified feature transformation weight matrix to obtain the feature vector to be detected. The exponential mapping value of the Euclidean distance between the feature vector to be detected and the sample feature row vector of each second type of bearing sample in the bearing status database is calculated by using the Gaussian radial basis kernel function. The initial similarity vector is obtained by concatenating the two vectors. The mathematical expression for the exponential mapping value is: ; in, For exponential mapping values, For the bearing status database The feature row vector of each sample This is the radial basis kernel width parameter; The feature vector to be detected; All the index mapping values ​​are concatenated sequentially to form the initial similarity vector; Retrieve the orthogonal feature vector matrix, project the initial similarity vector onto the orthogonal feature space, and perform the same exponential shrinkage mapping in combination with the fractional parameter. Then, perform an inverse transformation to generate a reconstructed similarity vector. Multiply the reconstructed similarity vector by the solidified mapping matrix to obtain the judgment result of the bearing to be detected.

[0013] The big data-based micro-bearing manufacturing and testing data analysis system includes a pre-training module, a sample feature extraction module, a database construction module, a feature acquisition module, and a testing module. The pre-training module is used to collect vibration acceleration signals of the first type of bearing samples and pre-train the recursive time series network. The sample feature extraction module is used to collect vibration acceleration signals of the second type of bearing samples, obtain the evolution feature vector of the second type of bearing samples through the sliding interception operation of the pre-trained recursive temporal network combined with the transient noise suppression operator, collect the end face visual image and cumulative processing time of the second type of bearing samples, obtain the static feature vector, concatenate the static feature vector and the evolution feature vector into the initial sample feature vector, perform feature correction on samples with local visual defects to obtain the final sample feature vector, and construct the reconstructed kernel feature matrix; The database construction module is used to construct an objective function based on the reconstructed kernel feature matrix and the true labels of the second type of bearing samples, perform bidirectional alternating optimization, solidify the feature transformation weight matrix and the mapping matrix, recalculate the sample feature vectors of all second type of bearing samples using the solidified feature transformation weight matrix, and construct a bearing state database. The module for acquiring the features to be detected is used to collect the cumulative processing time, end face visual image and vibration acceleration signal of the micro bearing to be detected, obtain the static feature vector and evolution feature vector of the bearing to be detected, and splice them into an initial feature vector to be detected. The detection module is used to process the initial feature vector to be detected based on the solidified feature transformation weight matrix, calculate the similarity between the processed initial feature vector to be detected and the sample feature vectors in the bearing status database, and obtain the final detection result of the bearing to be detected by combining the mapping matrix.

[0014] The beneficial effects of this invention are as follows: By introducing a transient noise suppression operator and a sliding interception operation, this invention can effectively filter out normal assembly clearance noise of miniature bearings based on the impact energy score of the local state submatrix and a preset clearance vibration threshold; this invention fully considers the physical changes of spindle thermal expansion and tool micro-wear nonlinearly accumulating over time during actual continuous cutting, and combines the cumulative machining time of the miniature bearing to calculate the process compensation coefficient. This process compensation coefficient is used to dynamically and adaptively adjust the global pheromone evaporation factor of the ant colony algorithm, extracting the physical machining boundary morphology and static feature vector of the miniature bearing; when a sample is determined to have local visual defects, this invention utilizes damage prior knowledge... By employing a weighted modulation and aggregation of attention mechanisms, this invention cleverly utilizes the temporal continuity of adjacent states and the reliability of vibration signals to repair missing visual features and improve detection robustness under harsh conditions with severe occlusion. Addressing the collective tolerance drift background noise of the entire batch of bearings caused by changes in equipment thermal balance, this invention constructs a reconstructed kernel feature matrix and introduces a fractional parameter based on the variance of the processing time difference. This fractional parameter is used to perform exponential shrinkage mapping on the diagonal matrix after orthogonal eigenvalue decomposition, enabling nonlinear compression of the maximal eigenvalues ​​representing global tolerance drift while retaining the minute eigenvalues ​​representing early minor anomalies in a very small number of bearings. Attached Figure Description

[0015] Figure 1 A flowchart illustrating the steps of a big data-based method for analyzing the production and testing data of miniature bearings, as provided in one embodiment of the present invention. Detailed Implementation

[0016] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0017] Example 1, referring to Figure 1 This paper provides a data analysis method for the production and testing of miniature bearings based on big data, including the following steps: Step S1: Collect vibration acceleration signals of the first type of bearing samples and pre-train the recursive time series network. Step S2: Collect vibration acceleration signals of the second type of bearing samples, obtain the evolution feature vector of the second type of bearing samples by combining the sliding interception operation of the pre-trained recursive temporal network with the transient noise suppression operator, collect the end face visual image and cumulative processing time of the second type of bearing samples, obtain the static feature vector, concatenate the static feature vector and the evolution feature vector into the initial sample feature vector, perform feature correction on samples with local visual defects to obtain the final sample feature vector, and construct the reconstructed kernel feature matrix; Step S3: Construct an objective function based on the reconstructed kernel feature matrix and the true labels of the second type of bearing samples, perform bidirectional alternating optimization, solidify the feature transformation weight matrix and the mapping matrix, recalculate the sample feature vectors of all second type of bearing samples using the solidified feature transformation weight matrix, and construct a bearing state database. Step S4: Collect the cumulative processing time, end face visual image and vibration acceleration signal of the micro bearing to be tested, obtain the static feature vector and evolution feature vector of the bearing to be tested, and splice them into the initial feature vector to be tested; Step S5: Based on the solidified feature transformation weight matrix, the initial feature vector to be detected is processed, and the similarity between the processed initial feature vector to be detected and the sample feature vector in the bearing status database is calculated. The final detection result of the bearing to be detected is obtained by combining the mapping matrix.

[0018] In step S1, vibration signals of bearing samples in various known states are collected based on a large database to obtain vibration acceleration sequences; The known states include normal and abnormal. The normal label for a bearing sample in the known state is 0, and the abnormal label is 1. The recursive temporal network is pre-trained using miniature bearing samples with known states to obtain the pre-trained recursive temporal network. Recursive temporal networks consist of a feature extraction layer, a feature transition layer, and a feature output layer.

[0019] In one specific embodiment of the present invention, a vibration acceleration sequence is obtained. Each element represents the vibration acceleration at the corresponding time point, and the length of the vibration acceleration sequence is the total time step. It is the product of the sensor's sampling frequency and the duration of a single continuous sampling. The feature extraction layer uses a standard gated recursive unit, with 64 hidden samples. The input to the feature extraction layer is a vibration acceleration sequence. Its hidden state vector (dimension is) The updated mathematical expression for ) is: ; ; ; ; in, For the corresponding time The vibration acceleration value below, To reset the gate vector, To update the gate vector, For candidate state vectors, The hidden state vector; , , This is the weight matrix, with all dimensions being 1. ; , , The weight matrix for the hidden state has dimensions of 1. , , , The bias vector has dimensions of 1. ; It is the Sigmoid activation function. It is the hyperbolic tangent function. This represents the element-wise multiplication of matrix elements; All The hidden state vectors are transposed and concatenated as row vectors to output the first-level hidden state sequence matrix. The dimension of the first-level hidden state sequence matrix is... ; The first hidden state sequence matrix is ​​input into the feature transition layer, the number of hidden samples in the feature transition layer is 64, and the output of the feature transition layer is the second hidden state sequence matrix. The second hidden state sequence matrix output by the feature transition layer is composed of the hidden state vector of the feature transition layer. The hidden state vector of the feature transition layer is obtained by concatenating row vectors. The calculation logic is the same as that of the first layer. The weight matrix of the feature transition layer... , , The dimension is ; The second hidden state sequence matrix is ​​input into the feature output layer. The number of hidden samples in the feature output layer is 128. The hidden state vector of the feature output layer is... The calculation logic is the same as that of the first layer, where the weight matrix... , , The dimension is Hidden state weight matrix , , The dimension is , the last hidden state vector The transposed value is used as the output of the feature output layer, denoted as the evolved feature vector, with the dimension of the evolved feature vector being... ; Using a normal distribution with a variance of 0.01, all weight matrices, hidden state weight matrices, and bias vectors in the recursive temporal network are randomly initialized. At this time, the interlayer assembly gap transient noise suppression operator does not perform masking interception, so as to allow the network to access complete waveform features in the early stage. The bearing samples in known states are input into a recursive temporal network to obtain evolutionary feature vectors. These feature vectors are then multiplied by the classification dimensionality reduction vector to be trained, with the dimension being... The prediction probability of the bearing sample being abnormal is calculated using the Sigmoid activation function. The pre-training loss function is calculated using the binary cross-entropy formula. The gradient of the pre-training loss function with respect to all weight matrices and bias vectors in the recursive temporal network is calculated using the backpropagation algorithm and gradient descent optimizer. The pre-trained recursive temporal network is then iteratively updated with a set learning rate of 0.001 until convergence, at which point training stops and the pre-trained recursive temporal network is obtained.

[0020] This invention utilizes big data and standard gated recursive units to construct a multi-layer feature extraction and transition network. In the early stages of pre-training, masking is not performed to allow the network to fully learn complete waveform features. This design enables the model to establish a fundamental ability to deeply mine the microstructural evolution features from complex nonlinear vibration signals based on massive amounts of real historical samples, laying a solid data-driven foundation for accurately distinguishing between normal and abnormal states.

[0021] In step S2, the second type of bearing sample is a sample from the same batch as the bearing to be tested; The process of obtaining the evolutionary feature vector of the second type of bearing sample specifically includes: Vibration acceleration signals of the second type of bearing samples are collected and input into a pre-trained recursive temporal network. After processing through the feature extraction layer to output the first hidden state sequence matrix, the sliding interception operation of the transient noise suppression operator is performed on the first hidden state sequence matrix to obtain the filter sequence matrix. The process of the sliding interception operation of the transient noise suppression operator specifically includes: In the row direction of the first-level hidden state sequence matrix, the first-level hidden state sequence matrix is ​​divided by a preset sliding window according to a preset sliding step size. In this embodiment, the preset sliding window length is 256 and the preset sliding step size is 256, that is, non-overlapping division to obtain multiple local state sub-matrices. The mathematical expression for the total number of local state sub-matrices is: ; in, Let T be the total number of local state submatrices, and T be the length of the vibration acceleration sequence. The length of the sliding window. This is the sliding step size; During the segmentation process, if the number of rows in the last local state submatrix is ​​less than the length of the sliding window, then a row vector of all zeros is appended to the last row of the first-level hidden state sequence matrix until the number of rows in the last local state submatrix reaches the length of the sliding window. The impact energy score is calculated for the local state submatrix. The mathematical expression for the impact energy score is as follows: ; in, For the first The impact energy score of a local state submatrix, where m takes values ​​of 1, 2, ..., M. Representing the In the local state submatrix, the th... Line 1 Column elements; Let be the row number of the local state submatrix. is the number of columns in the local state submatrix; If the impact energy score is less than the preset clearance vibration threshold, it means that the time period covered by the sliding window only contains normal assembly clearance noise, and the local state submatrix is ​​transformed into an all-zero mask matrix. If the impact energy score is greater than or equal to the preset clearance vibration threshold, the local state submatrix remains unchanged. The local state submatrix and the all-zero mask matrix are concatenated along the time dimension in their original order. The all-zero row vectors added during the segmentation stage are removed to obtain a filter sequence matrix with the same dimension as the first-layer hidden state sequence matrix. After the filtering matrix is ​​processed through the feature transition layer and feature output layer of the recursive temporal network, the evolution feature vector of the second type of bearing sample is obtained.

[0022] In a specific embodiment of the present invention, the preset clearance vibration threshold is determined by collecting the state vibration acceleration signal of historical micro-bearing samples known to be in normal state through a big data platform. For each normal bearing sample, the signal is input into a pre-trained recursive temporal network. After processing through the feature extraction layer and outputting the first layer hidden state sequence matrix, the first layer hidden state sequence matrix is ​​divided into multiple local state sub-matrices through a preset sliding window and according to a preset sliding step size. The impact energy score of the local state sub-matrices is calculated. In this embodiment, the 95th percentile value of the statistical distribution of the impact energy score is preferred as the clearance vibration threshold. This invention compares the impact energy score of the local state submatrix with a preset clearance vibration threshold and performs a zero-masking operation, which can accurately filter out normal assembly clearance noise of micro-bearings at the algorithm level, avoid feature confusion caused by normal mechanical interference impact, and improve the purity of evolution features in representing real early fatigue damage.

[0023] Step S2, the process of obtaining the static feature vector of the second type of bearing sample, specifically includes: The cumulative processing time of the second type of bearing samples was collected, and the process compensation coefficient was calculated. The mathematical expression for the process compensation coefficient is as follows: ; in, This is the process compensation coefficient. The preset thermal decay constant is set to 0.05 in this embodiment. This is the cumulative processing time; Acquire visual images of the end faces of the second type of bearing samples and convert the visual images of the end faces into grayscale matrices; Based on the global pheromone evaporation factor and grayscale matrix, the pheromone concentration of each pixel is obtained through the ant colony algorithm, and pixels with pheromone concentrations greater than a preset concentration threshold (0.6 in this embodiment) are designated as coarse edge points. The mathematical expression for the global pheromone volatile factor is: ; in, The baseline pheromone evaporation factor is set to 0.3 in this embodiment; To compensate for the amplification factor, this embodiment uses a value of 0.2; This is the process compensation coefficient; If the number of coarse edge points is less than the preset lower limit threshold, the visual feature extraction is deemed to have failed, and the static feature vector of the bearing to be detected is a vector of all zeros; the lower limit threshold is set to 20 in this embodiment. If the number of coarse edge points is greater than or equal to the preset lower limit threshold, then for each coarse edge point, calculate the Zernike orthogonal moment, and use the Zernike orthogonal moment to solve for the normal distance and edge angle corresponding to the pixel. The sub-pixel offsets of the pixel coordinates of the coarse edge point in the horizontal and vertical directions are calculated using trigonometric functions and then superimposed onto the pixel coordinates of the coarse edge point to obtain the sub-pixel coordinates of the coarse edge point. Construct an objective function that minimizes the sum of squared errors, and fit a circular equation to the sub-pixel coordinates of coarse edge points. The mathematical expression of the objective function is: ; in, For the first The x and y coordinates of the sub-pixel coordinates of each coarse edge point. Let x be the x-coordinate of the fitted circle center. Let be the ordinate of the fitted circle center. Let be the radius of the fitted circle to be determined. Given the total number of coarse edge points, the least squares method is used to find the parameter solution that minimizes the objective function, and the static feature vector is output. The static feature vector and the evolution feature vector of each sample in the second type of bearing sample are concatenated to generate the initial sample feature vector.

[0024] In one specific embodiment of the present invention, when the production line produces a new batch of bearings and replaces new cutting tools, 1,000 micro-bearings in known condition that were processed for the first time are collected as a second type of bearing sample. The bearing sample data used for pre-training in the recursive temporal network comes from big data collection, and the collected samples are single-mode data. This data is used to enable the network to extract evolution feature vectors from complex vibration signals through massive historical samples. It is also used to statistically score the impact energy of these historical normal bearing samples to establish and solidify the clearance vibration threshold. This threshold serves as the prior physical boundary of the transient noise suppression operator, which can effectively filter out normal assembly clearance noise of bearings at the algorithm level in subsequent production inspection. The bearing samples collected in this embodiment are from the first batch of bearings that are known to be produced when the current workshop production line is producing a new batch of bearings or replacing new tools. This collection is multimodal data, which is mainly used to eliminate systematic deviations caused by the specific physical environment of the current workshop. Extract the cumulative processing time of the second type of bearing samples from the production line control system in the workshop. The cumulative processing time represents the total time from the start of processing the first bearing in the current batch to the completion of the bearing processing. During the continuous cutting process of miniature bearings, the thermal expansion of the spindle and the micro-wear of the tool accumulate nonlinearly over time and eventually tend to a state of thermal equilibrium. This physical change will cause the cutting surface of the bearing end face chamfer to no longer be as sharp as the first piece. In order to quantify this physical change, this embodiment uses an exponential saturation function to calculate the process compensation coefficient. An improved ant colony algorithm was used to find the physical boundaries of the inner and outer diameter contours of the micro-bearing end face in the visual image. An industrial camera is used to capture the end face of a miniature bearing, generating a visual image of the end face, which is then converted into a two-dimensional grayscale matrix, where each element represents the grayscale value of a pixel. As processing time increases, the process compensation coefficient increases, and the global pheromone evaporation factor dynamically increases. The increase in the global pheromone evaporation factor means that the improved algorithm will forget the previous optimization path more quickly, forcing subsequent ants to expand their exploration range in the gray-scale gradient band, thereby filtering out visual artifacts caused by machining errors at the algorithm level. If the number of coarse edge points is greater than or equal to a preset lower threshold, for each coarse edge point, the local coordinate system origin is used as the pixel coordinates of the coarse edge point to truncate it. The gray-level submatrix of the neighborhood is used as input for calculating the Zernike orthogonal moments. It is then subjected to a two-dimensional discrete convolution operation with a preset Zernike integral template to output the orthogonal moment components of the Zernike orthogonal moments of the coarse edge point. , and ; The normal distance and edge angle of the real physical edge line relative to the current pixel center are solved using orthogonal moment components. The mathematical expressions for the normal distance and edge angle are: ; ; in, and Represent The real and imaginary parts, Normal distance, The included angle of the edges, and All are orthogonal moment components; The pixel coordinates of the coarse edge points are obtained by solving trigonometric functions. Subpixel offset of direction and in Subpixel offset of direction They will be respectively direction and in The sub-pixel offset of the direction is arithmetically superimposed onto the pixel coordinates of the coarse edge point to obtain the sub-pixel coordinates of the coarse edge point; The data format of static feature vectors is ,in, The x-coordinate represents the center of the fitted circle. The ordinate represents the center of the fitted circle. Represents the radius of the fitted circle; The static feature vectors, stripped of optical interference, represent the physical machining boundary morphology of the micro-bearing.

[0025] This invention incorporates the cumulative machining time of micro-bearings into the visual feature extraction process. By constructing an exponential process compensation coefficient to dynamically adjust the global pheromone evaporation factor of the ant colony algorithm, this dynamic compensation mechanism conforms to the physical laws of spindle thermal expansion and nonlinear micro-wear of the tool in actual cutting. This forces the optimization path to cross the gray-scale gradient zone. Combined with Zernike orthogonal moments for sub-pixel-level circumferential fitting, the visual artifacts caused by machining errors are successfully filtered out, and the physical machining boundary shape is obtained.

[0026] Step S2, the process of obtaining the final sample feature vector, specifically includes: If the sum of the absolute values ​​of the first three elements of the initial sample feature vector is equal to zero, the current second type of bearing sample is marked as a locally visually missing sample, and the initial sample feature vector of the locally visually missing sample is corrected to obtain the final sample feature vector of the locally visually missing sample. If the sum of the absolute values ​​of the first three elements of the initial sample feature vector is not equal to zero, then the current second type of bearing sample is marked as a visually complete sample. The initial sample feature vector of the visually complete sample is multiplied by the feature transformation weight matrix, and after mapping through the Sigmoid activation function, the final sample feature vector of the visually complete sample is obtained. The process of obtaining the final feature vector of any locally visually missing sample specifically includes: The arithmetic mean of the evolutionary feature vectors of all normal Class II bearing samples is used as the baseline feature vector. The sum of squared Euclidean distances between the baseline feature vector and the evolutionary feature vectors of locally visually missing samples is calculated. This sum of squared Euclidean distances is then input into an exponential decay function to calculate the prior probability of damage. The mathematical expression for the prior probability of damage is: ; in, For the prior probability of damage, The scaling factor is set to 0.05 in this embodiment. The sum of squared Euclidean distances between the baseline eigenvectors and the evolving eigenvectors; The processing completion timestamps of locally visually missing samples are collected. For each locally visually missing sample, second-class bearing samples with adjacent relationships are obtained based on the processing completion timestamps. The initial attention correlation score between the locally visually missing sample and its adjacent second-class bearing samples is calculated. The mathematical expression for the initial attention correlation score is as follows: ; in, An initial attentional relevance score is assigned between locally visually missing samples and adjacent second-class bearing samples. For a leaky linear rectified activation function, This is a learnable attention projection vector. This is the initial sample feature vector for samples with local visual defects. The feature transformation weight matrix, The initial sample feature vectors of adjacent samples. This indicates that the transformed vectors will be horizontally concatenated and merged. Calculate the prior damage probability of adjacent second-type bearing samples. Adjust the initial attention relevance score using the prior damage probabilities of locally visually missing samples and adjacent second-type bearing samples, obtaining the adjustment weight coefficients. The adjustment rule is as follows: If the prior probability of damage of a local visual missing sample and the prior probability of damage of an adjacent second-class bearing sample are both greater than or equal to a preset probability threshold, or both are less than a preset probability threshold, then the adjustment weight coefficient of the adjacent second-class bearing sample is 1. If the prior probability of damage of a locally visually missing sample and the prior probability of damage of an adjacent second-class bearing sample are both greater than or equal to a preset probability threshold and the other is less than a preset probability threshold, then the adjustment weight coefficient of the adjacent second-class bearing sample is 1 minus the absolute value of the difference between the prior probability of damage of the locally visually missing sample and the adjacent second-class bearing sample. In this embodiment, the preset probability threshold is 0.6; The product of the modulated weight coefficient of each adjacent second-class bearing sample and the initial attention relevance score is used as the modulated attention relevance score. Softmax normalization is performed on all modulated attention relevance scores to obtain the aggregate weight corresponding to each adjacent second-class bearing sample. A weighted summation operation is then performed on each adjacent second-class bearing sample of the locally visually missing sample to obtain the final sample feature vector of the locally visually missing sample. The mathematical expression for the final sample feature vector of the locally visually missing sample is: ; in, This is the sample feature vector of a sample with local visual defects. The first of the local visual missing samples The initial sample feature vector of each adjacent second-class bearing sample. The first of the local visual missing samples The aggregate weights corresponding to the adjacent second-class bearing samples. This is the set of indices of all neighboring samples of a locally visually missing sample. The feature transformation weight matrix; The final sample feature vectors of all second-class bearing samples are used to construct a bearing state database.

[0027] In a specific embodiment of the present invention, each row of the sample feature matrix is ​​traversed, if the... The sum of the absolute values ​​of the first three elements of the row equals This indicates that the bearing end face was largely obscured by cutting fluid and metal dust, making it impossible to extract an effective visual contour. The row is marked as a sample with local visual defects; However, since the vibration sensor is not affected by oil, the evolutionary feature vector contained in the locally visually missing samples is still complete. This invention uses the evolutionary feature vector to calculate the damage prior probability, which is used to characterize the possibility that the bearing has already suffered fatigue damage. In the mathematical expression for the initial attention relevance score, the feature transformation weight matrix... Used to map initial sample feature vectors to the latent graph feature space, the learnable attention projection vector dimension is... In this embodiment The value is 64. Before optimization begins, and All elements are randomly initialized from a normal distribution with a mean of 0 and a variance of 0.01. The logic for determining the adjacency relationship is that the time window interval of the processing end timestamp is less than or equal to 5 minutes. If the number of adjacent samples retrieved within the above 5-minute time window is zero or less than the preset minimum number of samples (2), the time window restriction is automatically abandoned, and the three closest Class II bearing samples to the missing sample are directly extracted forward and backward according to the processing order as adjacent samples to ensure that the subsequent feature weighting aggregation process based on the attention mechanism can be executed stably. To address the challenge of localized visual loss on end faces caused by extensive obstruction of cutting fluid and metal dust adhesion in workshops, this invention proposes a feature correction strategy based on cross-modal and temporal adjacency fusion. This invention utilizes vibration evolution features unaffected by oil contamination to calculate the prior probability of damage, and uses this as a weight to adjust the attention correlation score of adjacent processing samples. By weighted aggregation of features from adjacent samples, the currently missing visual dimension is repaired, greatly enhancing the continuity and anti-interference capability of the detection method under extreme working conditions.

[0028] Step S3, the process of constructing the reconstructed kernel feature matrix, specifically includes: Extract the final sample feature vectors of all second-class bearing samples, reassemble them by row vectors according to the processing order to form a corrected sample feature matrix, and use the radial basis kernel function to construct the initial estimated kernel matrix; The mathematical expression for the initial estimation kernel matrix is: ; in, For the initial estimation of the kernel matrix, the first... Line 1 The elements of the column represent the first element. The second type of bearing sample and the first Inter-sample similarity among the second-class bearing samples To correct the first feature matrix of the sample row vectors To correct the first feature matrix of the sample Row vectors; The kernel width parameter of the radial basis function is 2.0 in this embodiment. This value is chosen because the feature vectors have been normalized before concatenation, and the element values ​​are distributed in... between, This ensures that the calculated similarity between samples is guaranteed. Uniformly mapped on arrive Within the interval, to prevent extreme polarization of the inner product result; Based on the processing timestamps of the second type of bearing samples, obtain pairs of second type bearing samples with adjacent relationships, calculate the processing time difference between the bearing sample pairs, and calculate the average and variance of all processing time differences; The fractional parameter is calculated using the inverse proportional mapping function. The mathematical expression for the fractional parameter is: ; in, It is a fractional parameter. The variance penalty coefficient is a preset value, which is set to 0.5 in this embodiment. The variance of all processing time differences; Perform orthogonal eigenvalue decomposition on the initial estimated kernel matrix, and perform exponential shrinkage mapping on each main diagonal element of the diagonal matrix obtained after orthogonal eigenvalue decomposition to obtain the reconstructed diagonal matrix; The mathematical expression for reconstructing the elements on the diagonal of the diagonal matrix is: ; in, To reconstruct the first line on the diagonal of the diagonal matrix One element, The first element on the diagonal of a diagonal matrix One element, It is a fractional parameter; The reconstructed kernel feature matrix is ​​generated by restoring the reconstructed diagonal matrix, the orthogonal eigenvector matrix, and the transpose of the orthogonal eigenvector matrix.

[0029] In a specific embodiment of the present invention, an initial estimation kernel matrix is ​​constructed to measure the absolute similarity of the same batch of micro bearings in the feature space. The initial estimation kernel matrix is ​​a symmetric square matrix, and the number of rows and columns are the same as the number of the same batch of micro bearings. Extract the actual timestamp of the miniature bearing off the production line from the workshop production system, calculate the processing time difference between two adjacent bearings, and calculate the average and variance of all processing time differences; The larger the variance, the more unstable the workshop processing cycle, the more drastic the fluctuation of the rate of tool thermal expansion and mechanical wear, and the more severe the collective tolerance drift background noise generated by the entire batch of bearings. Fractional parameters are used to transform the instability of the machining process into a correction force. In the initial estimation kernel matrix, the large eigenvalues ​​physically represent the global tolerance drift background components common to the entire batch of bearings, caused by the slow wear of the tool. The small eigenvalues ​​represent the early fatigue crack abnormalities of a very small number of bearings. Traditional deep learning models are easily biased by these huge background eigenvalues, resulting in batch false alarms. This embodiment uses fractional mathematical operations to filter out huge background noise and highlight tiny abnormal signals. Perform orthogonal eigenvalue decomposition on the initial estimated kernel matrix to obtain the orthogonal eigenvector matrix, the transpose of the orthogonal eigenvector matrix, and the diagonal matrix. The elements on the diagonal of the diagonal matrix are the eigenvalues ​​of the initial estimated kernel matrix arranged in descending order from largest to smallest, and all off-diagonal elements are 0. because It is a floating-point number less than 1. According to the mathematical properties of power functions, exponential contraction mapping will produce nonlinear compression for maximal eigenvalues, while for eigenvalues ​​close to maximal eigenvalues... The small eigenvalues ​​have little impact, thus achieving automatic decay of population tolerance drift and preserving small individual anomalies; Using the reconstructed diagonal matrix, combined with the orthogonal eigenvector matrix and the transpose of the orthogonal eigenvector matrix, a reconstructed kernel feature matrix is ​​generated. The reconstructed kernel feature matrix eliminates the systematic distribution deviation caused by the thermal expansion of workshop equipment.

[0030] This invention introduces a fractional parameter based on the variance of processing time difference and performs orthogonal eigenvalue decomposition and fractional exponential shrinkage mapping on the initial estimated kernel matrix. This enables nonlinear compression of the maximal eigenvalues ​​representing the background noise of the entire batch's group tolerance drift, while retaining the tiny eigenvalues ​​representing a very small number of early fatigue anomalies. This effectively eliminates the systematic distribution deviation caused by changes in the thermal balance of workshop equipment and prevents batch false alarms from the source.

[0031] Step S3, the bidirectional alternating optimization process, specifically includes: The second type of bearing samples are divided into a support set and a query set according to a preset ratio; The feature transformation weight matrix and mapping matrix are obtained through random initialization using a normal distribution. The weight matrix of the locked feature transformation does not participate in the differentiation. From the reconstructed kernel feature matrix, all row vectors of the second type of bearing samples whose row index belongs to the support set are extracted and concatenated into the support kernel matrix. Extract the state labels of the second type of bearing samples within the support set, construct a support label matrix, and calculate the first loss function. The mathematical expression of the first loss function is: ; in, For the first loss function, To support the set, To support the first in the kernel matrix row vectors The mapping matrix to be updated, To support the first in the tag matrix row vectors The regularization penalty coefficient is set to 0.01 in this embodiment to prevent overfitting caused by excessively large elements in the mapping matrix. To support the total number of second-class bearing samples in the set; The gradient descent algorithm is used to update all elements in the mapping matrix along the negative gradient direction of the first loss function with a preset inner learning rate of 0.01, thereby obtaining the mapping matrix of the current iteration step.

[0032] In a specific embodiment of the present invention, 1000 miniature bearings are divided into two non-overlapping subsets according to a preset ratio, including an initial support set and a query set. In this embodiment, the ratio of the number of bearings in the support set and the query set is 2:8. Each row of the support label matrix corresponds to a bearing sample in the support set. If the bearing is in a normal state, its corresponding behavior is... If the bearing is in an abnormal state, the corresponding behavior is... ; Initialize the mapping matrix, the dimension of which is... The mapping matrix is ​​a linear projection operator that needs to be learned through a neural network. It is used to compress the high-dimensional kernel feature space into a 2-dimensional decision plane. The initial values ​​of the elements in the mapping matrix are based on the mean of... variance is The Gaussian distribution is randomly generated.

[0033] This invention locks the feature transformation weight matrix and uses a specific first loss function containing a regularization penalty term to optimize the mapping matrix. This inner space update strategy can effectively prevent overfitting caused by excessive parameters when compressing the high-dimensional kernel feature space to the two-dimensional decision plane, ensuring the model's generalization ability and classification boundary accuracy during the calibration process for the same batch of samples.

[0034] In step S3, the inner layer updated mapping matrix is ​​locked and does not participate in the differentiation. From the reconstructed kernel feature matrix, all row vectors whose row indices belong to the query set are extracted to form the query kernel matrix. The query kernel matrix is ​​multiplied by the updated mapping matrix and Softmax normalization is performed to obtain the prediction probability matrix of the query set. The second loss function is calculated, and its mathematical expression is as follows: ; in, For the second loss function, For query set, For the query set The probability that a sample is in a normal state. For the query set The probability that a sample state is abnormal. and These are extracted from the prediction probability matrix. The probability that a sample state is predicted to be normal or abnormal; To query the total number of bearing samples in the set; Using the chain rule, backpropagation is performed to the feature transformation weight matrix and the attention projection vector. The elements of the feature transformation weight matrix and the attention projection vector are updated with a preset outer learning rate of 0.005. The outer learning rate is set smaller than that of the inner layer to ensure that the evolution of the underlying feature space is stable and does not cause violent oscillations. Calculate the absolute value of the difference between the first loss function and the second loss function. If the absolute value of the difference is greater than or equal to the preset convergence threshold of 0.001, then use the updated feature transformation weight matrix to regenerate the corrected sample feature matrix and reconstruct the kernel feature matrix, and solve the mapping matrix again. If the absolute value of the difference is less than the preset convergence threshold or the total number of alternating iterations reaches the set number of 50, bidirectional convergence is determined, training is terminated, and the feature transformation weight matrix and mapping matrix are solidified. Using the solidified feature transformation weight matrix, and following the same feature processing rules as in step S2, that is, judging the sum of the absolute values ​​of the first three elements of the initial sample feature vector to determine whether to perform weighted feature correction for local visual defects or to directly perform Sigmoid activation function mapping, the final sample feature vectors of all second-class bearing samples are recalculated, and all recalculated sample feature vectors are used to construct the final bearing state database. In a specific embodiment of the present invention, if the mapping matrix of the current iteration step is used to test the query set, and the classification effect is poor and the error is large, it indicates that the matrix feature transformation weight matrix for repairing local visual missing samples in step S3 is not good. Therefore, the error must be propagated backward from the outer layer to the lowest layer matrix feature transformation weight matrix and modified.

[0035] This invention constructs an outer feedback adjustment mechanism that uses the query set prediction probability to calculate the second loss function. By using the chain rule to update the feature transformation weight matrix and attention projection vector of the lower layer with a small outer learning rate, it achieves precise fine-tuning of the feature space of the lower hidden graph by the high-level classification error. The small learning rate ensures the smooth evolution of the feature graph space without drastic oscillations, and ultimately makes the model perfectly adapt to the physical environment characteristics unique to the current batch.

[0036] The process of obtaining the test results of the bearing under test specifically includes: Collect the cumulative processing time and end face visual images of the bearing to be inspected to obtain the static feature vector of the bearing to be inspected; The vibration acceleration signal of the bearing to be tested is collected and input into a pre-trained recursive temporal network. After processing through the feature extraction layer and outputting the first hidden state sequence matrix, the sliding interception operation of the transient noise suppression operator is performed on the first hidden state sequence matrix to obtain the filter sequence matrix. After the filtering matrix is ​​processed through the feature transition layer and feature output layer of the recursive temporal network, the evolution feature vector of the bearing to be detected is obtained. The static feature vector and the evolved feature vector of the bearing to be tested are concatenated to form the initial feature vector to be tested; The initial feature vector to be detected is processed using the solidified feature transformation weight matrix to obtain the feature vector to be detected; The processing rules for the initial feature vector to be detected are consistent with the process of obtaining the final sample feature vector for the second type of bearing samples. That is, it is determined whether the sum of the absolute values ​​of the first three elements of the initial feature vector to be detected is equal to zero. If it is equal to zero, it is marked as a sample with local visual defects, and samples adjacent to its processing timestamp are extracted and weighted feature correction is performed according to the aforementioned rules based on the prior probability of damage and the initial attention correlation score. If it is not equal to zero, it is multiplied with the solidified feature transformation weight matrix and mapped through the Sigmoid activation function. Using the Gaussian radial basis kernel function, the exponential mapping value of the Euclidean distance between the feature vector to be detected and the sample feature row vector of each second type of bearing sample in the bearing status database is calculated, and the initial similarity vector is obtained by concatenating them. The mathematical expression for the exponential mapping value is: ; in, For exponential mapping values, For the bearing status database The feature row vector of each sample This is the radial basis kernel width parameter; The feature vector to be detected; All the index mapping values ​​are concatenated sequentially to form the initial similarity vector; Retrieve the orthogonal feature vector matrix, project the initial similarity vector onto the orthogonal feature space, and perform the same exponential shrinkage mapping in combination with the fractional parameter. Then, perform an inverse transformation to generate a reconstructed similarity vector. Multiply the reconstructed similarity vector by the solidified mapping matrix to obtain the judgment result of the bearing to be detected. The judgment result includes the probability of abnormality and the probability of normality. If the probability of abnormality is greater than or equal to the probability of normality, the test result of the bearing to be tested will be judged as abnormal. If the probability of an anomaly is less than the probability of a normal anomaly, the test result of the bearing to be tested will be judged as normal.

[0037] This invention maps the multimodal splicing features of the bearing to be tested through an optimized weight matrix, calculates the similarity between the features and samples in the bearing condition database using a Gaussian radial basis kernel function, and finally outputs the exact probabilities of anomalies and normalities by combining the mapping matrix. This can quickly transform the complex results of previous training and optimization into high-precision quantitative detection results, meeting the real-time detection requirements of production lines.

[0038] The big data-based micro-bearing manufacturing and testing data analysis system includes a pre-training module, a sample feature extraction module, a database construction module, a feature acquisition module, and a testing module. The pre-training module is used to collect vibration acceleration signals of the first type of bearing samples and pre-train the recursive time series network. The sample feature extraction module is used to collect vibration acceleration signals of the second type of bearing samples. It obtains the evolution feature vector of the second type of bearing samples by combining the sliding interception operation of the transient noise suppression operator with the pre-trained recursive temporal network. It also collects the end face visual image and cumulative processing time of the second type of bearing samples to obtain the static feature vector. The static feature vector and the evolution feature vector are concatenated to form the initial sample feature vector. The features of samples with local visual defects are corrected to obtain the final sample feature vector, and a reconstructed kernel feature matrix is ​​constructed. The database construction module is used to construct the objective function based on the reconstructed kernel feature matrix and the true labels of the second type of bearing samples, perform bidirectional alternating optimization, solidify the feature transformation weight matrix and the mapping matrix, recalculate the sample feature vectors of all second type of bearing samples using the solidified feature transformation weight matrix, and construct the bearing state database. The module for acquiring the features to be detected is used to collect the cumulative processing time, end face visual image and vibration acceleration signal of the micro bearing to be detected, obtain the static feature vector and evolution feature vector of the bearing to be detected, and splice them into the initial feature vector to be detected. The detection module is used to process the initial feature vector to be detected based on the solidified feature transformation weight matrix, calculate the similarity between the processed initial feature vector to be detected and the sample feature vectors in the bearing condition database, and obtain the final detection result of the bearing to be detected by combining the mapping matrix.

[0039] This invention introduces a transient noise suppression operator and a sliding interception operation, which can effectively filter out normal assembly clearance noise of micro-bearings based on the impact energy score of the local state submatrix and the preset clearance vibration threshold. This invention fully considers the physical changes of spindle thermal expansion and tool micro-wear nonlinearly accumulated over time during actual continuous cutting. It combines the cumulative machining time of the micro-bearing to calculate the process compensation coefficient. The process compensation coefficient is used to dynamically and adaptively adjust the global pheromone evaporation factor of the ant colony algorithm. By dynamically adjusting the pheromone evaporation factor, the algorithm's edge search adaptability in gray-scale fuzzy areas caused by thermal expansion is enhanced, thereby extracting the physical machining boundary morphology and static feature vector of the micro-bearing. To address the issue of localized visual loss on end faces caused by large-area obstruction of cutting fluid and adhesion of metal dust under complex workshop conditions, this invention utilizes vibration evolution characteristics unaffected by oil contamination to calculate the prior probability of fatigue damage when a sample is determined to have localized visual loss. It also combines the features of adjacent samples extracted by the processing timestamp with weighted modulation and aggregation of attention mechanisms. This ingeniously utilizes the temporal continuity of adjacent states and the reliability of vibration signals to repair the missing visual features and improve the detection robustness under harsh conditions with severe obstruction. To address the background noise caused by the collective tolerance drift of bearings in a batch due to changes in equipment thermal balance, this invention constructs a reconstructed kernel feature matrix and introduces a fractional parameter based on the variance of the processing time difference. This fractional parameter is used to perform exponential shrinkage mapping on the diagonal matrix after orthogonal eigenvalue decomposition. This nonlinearly compresses the maximal eigenvalues ​​representing global tolerance drift while retaining the minute eigenvalues ​​representing early minor anomalies in a very small number of bearings. This operation automatically attenuates the collective tolerance drift, effectively avoiding the batch false alarm phenomenon that is easily generated by traditional deep learning models. Combined with a bidirectional alternating optimization strategy with different learning rates for inner and outer layers, it ensures the smooth evolution of the underlying feature space without drastic oscillations, achieving high-precision, low-false-alarm online detection and quality control of micro-bearings in complex real-world physical environments.

[0040] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0041] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. A method for analyzing manufacturing and testing data of miniature bearings based on big data, characterized in that, Includes the following steps: Step S1: Collect vibration acceleration signals of the first type of bearing samples and pre-train the recursive time series network. Step S2: Collect vibration acceleration signals of the second type of bearing samples, obtain the evolution feature vector of the second type of bearing samples through the sliding interception operation of the pre-trained recursive temporal network combined with the transient noise suppression operator, collect the end face visual image and cumulative processing time of the second type of bearing samples, obtain the static feature vector, concatenate the static feature vector and the evolution feature vector into the initial sample feature vector, perform feature correction on samples with local visual defects to obtain the final sample feature vector, and construct the reconstructed kernel feature matrix; Step S3: Based on the reconstructed kernel feature matrix and the true labels of the second type of bearing samples, construct the objective function, perform bidirectional alternating optimization, solidify the feature transformation weight matrix and the mapping matrix, recalculate the sample feature vectors of all second type of bearing samples using the solidified feature transformation weight matrix, and construct the bearing state database. Step S4: Collect the cumulative processing time, end face visual image and vibration acceleration signal of the micro bearing to be tested, obtain the static feature vector and evolution feature vector of the bearing to be tested, and splice them into the initial feature vector to be tested; Step S5: Based on the solidified feature transformation weight matrix, the initial feature vector to be detected is processed, and the similarity between the processed initial feature vector to be detected and the sample feature vector in the bearing status database is calculated. The final detection result of the bearing to be detected is obtained by combining the mapping matrix. In step S2, the second type of bearing sample is a sample from the same batch as the bearing to be tested. The process of obtaining the evolutionary feature vector of the second type of bearing sample specifically includes: Vibration acceleration signals of the second type of bearing samples are collected and input into a pre-trained recursive temporal network. After processing through the feature extraction layer and outputting the first hidden state sequence matrix, the sliding interception operation of the transient noise suppression operator is performed on the first hidden state sequence matrix to obtain the filter sequence matrix. The process of the sliding interception operation of the transient noise suppression operator specifically includes: In the row direction of the first-level hidden state sequence matrix, the first-level hidden state sequence matrix is ​​divided by a preset sliding window according to a preset sliding step size to obtain multiple local state sub-matrices. The impact energy score of the local state submatrix is ​​calculated, and the mathematical expression for the impact energy score is as follows: ； in, For the first The impact energy score of a local state submatrix, where m takes values ​​of 1, 2, ..., M. Representing the In the local state submatrix, the th... Line number Column elements; Let be the row number of the local state submatrix. is the number of columns in the local state submatrix; If the impact energy score is less than the preset gap vibration threshold, the local state submatrix will be transformed into a zero mask matrix. If the impact energy score is greater than or equal to the preset clearance vibration threshold, the local state submatrix remains unchanged. The local state submatrix and the all-zero mask matrix are concatenated along the time dimension in their original order. The all-zero row vectors added during the segmentation stage are removed to obtain a filter sequence matrix with the same dimension as the first-layer hidden state sequence matrix. After the filtering matrix is ​​processed through the feature transition layer and feature output layer of the recursive temporal network, the evolution feature vector of the second type of bearing sample is obtained.

2. The method for analyzing micro-bearing manufacturing and testing data based on big data as described in claim 1, characterized in that, In step S1 Vibration signals of bearing samples in various known states were collected based on a large database to obtain vibration acceleration sequences; The recursive temporal network is pre-trained using miniature bearing samples with known states to obtain the pre-trained recursive temporal network. The recursive temporal network includes a feature extraction layer, a feature transition layer, and a feature output layer.

3. The method for analyzing micro-bearing manufacturing and testing data based on big data as described in claim 1, characterized in that, The process of obtaining the static feature vector of the second type of bearing sample in step S2 specifically includes: The cumulative processing time of the second type of bearing samples is collected, and the process compensation coefficient is calculated. The mathematical expression of the process compensation coefficient is as follows: ； in, This is the process compensation coefficient. The preset thermal decay constant This is the cumulative processing time; Acquire visual images of the end faces of the second type of bearing samples and convert the visual images of the end faces into grayscale matrices; Based on the global pheromone evaporation factor and the grayscale matrix, the pheromone concentration of each pixel is obtained through the ant colony algorithm, and pixels with pheromone concentrations greater than a preset concentration threshold are designated as coarse edge points. If the number of coarse edge points is less than the preset lower limit threshold, the visual feature extraction is deemed to have failed, and the static feature vector of the bearing to be detected is a vector of all zeros. If the number of coarse edge points is greater than or equal to the preset lower limit threshold, then for each coarse edge point, calculate the Zernike orthogonal moment, and use the Zernike orthogonal moment to solve for the normal distance and edge angle corresponding to the pixel. The sub-pixel offsets of the pixel coordinates of the coarse edge point in the horizontal and vertical directions are calculated using trigonometric functions and then superimposed onto the pixel coordinates of the coarse edge point to obtain the sub-pixel coordinates of the coarse edge point. Construct an objective function that minimizes the sum of squared errors, fit a circular equation to the sub-pixel coordinates of coarse edge points, solve for the parameters that minimize the objective function using the least squares method, and output a static feature vector. The static feature vector and the evolution feature vector of each sample in the second type of bearing sample are concatenated to generate the initial sample feature vector.

4. The method for analyzing micro-bearing manufacturing and testing data based on big data as described in claim 3, characterized in that, Step S2, the process of obtaining the final sample feature vector, specifically includes: If the sum of the absolute values ​​of the first three elements of the initial sample feature vector is equal to zero, the current second type of bearing sample is marked as a locally visually missing sample, and the initial sample feature vector of the locally visually missing sample is corrected to obtain the final sample feature vector of the locally visually missing sample. If the sum of the absolute values ​​of the first three elements of the initial sample feature vector is not equal to zero, then the current second type of bearing sample is marked as a visually complete sample. The initial sample feature vector of the visually complete sample is multiplied by the feature transformation weight matrix, and after mapping through the Sigmoid activation function, the final sample feature vector of the visually complete sample is obtained. The process of obtaining the final sample feature vector for any locally visually missing sample specifically includes: The arithmetic mean of the evolution feature vectors of all normal second-class bearing samples is used as the baseline feature vector. The sum of squared Euclidean distances between the baseline feature vector and the evolution feature vectors of locally visually missing samples is calculated. The sum of squared Euclidean distances is input into the exponential decay function to calculate the prior probability of damage. Collect the timestamp of the end of processing of the local visual missing sample. For the local visual missing sample, obtain the second type of bearing sample with adjacent relationship based on the timestamp of the end of processing, and calculate the initial attention correlation score between the local visual missing sample and its adjacent second type of bearing sample. Calculate the prior damage probability of adjacent second-type bearing samples. Adjust the initial attention relevance score using the prior damage probabilities of locally visually missing samples and adjacent second-type bearing samples, obtaining the adjustment weight coefficients. The adjustment rule is as follows: If the prior probability of damage of a local visual missing sample and the prior probability of damage of an adjacent second-class bearing sample are both greater than or equal to a preset probability threshold, or both are less than a preset probability threshold, then the adjustment weight coefficient of the adjacent second-class bearing sample is 1. If the prior probability of damage of a locally visually missing sample and the prior probability of damage of an adjacent second-class bearing sample are both greater than or equal to a preset probability threshold and the other is less than a preset probability threshold, then the adjustment weight coefficient of the adjacent second-class bearing sample is 1 minus the absolute value of the difference between the prior probability of damage of the locally visually missing sample and the adjacent second-class bearing sample. The product of the adjustment weight coefficient of each adjacent second-class bearing sample and the initial attention relevance score is used as the modulation attention relevance score. Softmax normalization is performed on all modulation attention relevance scores to obtain the aggregate weight corresponding to each adjacent second-class bearing sample. Weighted summation is performed on each adjacent second-class bearing sample of the local visual missing sample to obtain the final sample feature vector of the local visual missing sample.

5. The method for analyzing micro-bearing manufacturing and testing data based on big data as described in claim 4, characterized in that, The process of constructing the reconstructed kernel feature matrix in step S3 specifically includes: Extract the final sample feature vectors of all second-class bearing samples, reassemble them by row vectors according to the processing order to form a corrected sample feature matrix, and use the radial basis kernel function to construct the initial estimated kernel matrix; Based on the processing timestamps of the second type of bearing samples, obtain pairs of second type bearing samples with adjacent relationships, calculate the processing time difference between the bearing sample pairs, and calculate the average and variance of all processing time differences; Calculate fractional parameters using the inverse proportional mapping function; Perform orthogonal eigenvalue decomposition on the initial estimated kernel matrix, and perform exponential shrinkage mapping on each main diagonal element of the diagonal matrix obtained after orthogonal eigenvalue decomposition to obtain the reconstructed diagonal matrix; The reconstructed diagonal matrix, orthogonal eigenvector matrix, and transpose of the orthogonal eigenvector matrix are used to restore the original matrix and generate the reconstructed kernel feature matrix.

6. The method for analyzing micro-bearing manufacturing and testing data based on big data as described in claim 5, characterized in that, The bidirectional alternating optimization process in step S3 specifically includes: The second type of bearing samples are divided into a support set and a query set according to a preset ratio; The feature transformation weight matrix and mapping matrix are obtained through random initialization using a normal distribution; The weight matrix of the locked feature transformation does not participate in the differentiation. From the reconstructed kernel feature matrix, all row vectors of the second type of bearing samples whose row index belongs to the support set are extracted and concatenated into the support kernel matrix. Extract the state labels of the second type of bearing samples within the support set, construct a support label matrix, and calculate the first loss function. The mathematical expression of the first loss function is: ； in, For the first loss function, To support the set, To support the first in the kernel matrix row vectors The mapping matrix to be updated. To support the first in the tag matrix row vectors This is the regularization penalty coefficient; To support the total number of second-class bearing samples in the set; The gradient descent algorithm is used to update all elements in the mapping matrix along the negative gradient direction of the first loss function with a preset inner learning rate, so as to obtain the mapping matrix of the current iteration step.

7. The method for analyzing micro-bearing manufacturing and testing data based on big data as described in claim 6, characterized in that, In step S3 The inner updated mapping matrix is ​​locked and not differentiated. From the reconstructed kernel feature matrix, all row vectors whose row indices belong to the query set are extracted to form the query kernel matrix. The query kernel matrix is ​​multiplied by the updated mapping matrix and then subjected to Softmax normalization to obtain the prediction probability matrix of the query set. Calculate the second loss function, whose mathematical expression is: ； in, For the second loss function, For query set, For the query set The probability that a sample is in a normal state. For the query set The probability that a sample state is abnormal. and These are extracted from the prediction probability matrix. The probability that a sample state is predicted to be normal or abnormal; To query the total number of bearing samples in the set; Using the chain rule, backpropagation is performed to the feature transformation weight matrix and the attention projection vector, and the elements of the feature transformation weight matrix and the attention projection vector are updated with a preset outer learning rate; Calculate the absolute value of the difference between the first loss function and the second loss function. If the absolute value of the difference is greater than or equal to the preset convergence threshold, then regenerate the corrected sample feature matrix and reconstruct the kernel feature matrix using the updated feature transformation weight matrix, and solve the mapping matrix again. If the absolute value of the difference is less than the preset convergence threshold or the total number of alternating iterations reaches the set number, bidirectional convergence is determined, training is terminated, and the feature transformation weight matrix and mapping matrix are solidified.

8. The method for analyzing micro-bearing manufacturing and testing data based on big data as described in claim 7, characterized in that, The process of obtaining the test results of the bearing to be tested specifically includes: Collect the cumulative processing time and end face visual images of the bearing to be inspected to obtain the static feature vector of the bearing to be inspected; The vibration acceleration signal of the bearing to be detected is collected and input into a pre-trained recursive temporal network. After processing through the feature extraction layer and outputting the first hidden state sequence matrix, the sliding interception operation of the transient noise suppression operator is performed on the first hidden state sequence matrix to obtain the filter sequence matrix. After the filtering matrix is ​​processed through the feature transition layer and feature output layer of the recursive temporal network, the evolution feature vector of the bearing to be detected is obtained. The static feature vector and the evolved feature vector of the bearing to be tested are concatenated to form the initial feature vector to be tested; The initial feature vector to be detected is processed by the solidified feature transformation weight matrix to obtain the feature vector to be detected. The exponential mapping value of the Euclidean distance between the feature vector to be detected and the sample feature row vector of each second type of bearing sample in the bearing status database is calculated by using the Gaussian radial basis kernel function. The initial similarity vector is obtained by concatenating the two vectors. The mathematical expression for the exponential mapping value is: ； in, For exponential mapping values, For the bearing status database The feature row vector of each sample This is the radial basis kernel width parameter; The feature vector to be detected; All the index mapping values ​​are concatenated sequentially to form the initial similarity vector; Retrieve the orthogonal feature vector matrix, project the initial similarity vector onto the orthogonal feature space, and perform the same exponential shrinkage mapping in combination with the fractional parameter. Then, perform an inverse transformation to generate a reconstructed similarity vector. Multiply the reconstructed similarity vector by the solidified mapping matrix to obtain the judgment result of the bearing to be detected.

9. A big data-based micro-bearing manufacturing and testing data analysis system, applied in the big data-based micro-bearing manufacturing and testing data analysis method as described in any one of claims 1-8, characterized in that, It includes a pre-training module, a sample feature extraction module, a database construction module, a feature acquisition module, and a detection module; The pre-training module is used to collect vibration acceleration signals of the first type of bearing samples and pre-train the recursive time series network. The sample feature extraction module is used to collect vibration acceleration signals of the second type of bearing samples, obtain the evolution feature vector of the second type of bearing samples through the sliding interception operation of the pre-trained recursive temporal network combined with the transient noise suppression operator, collect the end face visual image and cumulative processing time of the second type of bearing samples, obtain the static feature vector, concatenate the static feature vector and the evolution feature vector into the initial sample feature vector, perform feature correction on samples with local visual defects to obtain the final sample feature vector, and construct the reconstructed kernel feature matrix; The database construction module is used to construct an objective function based on the reconstructed kernel feature matrix and the true labels of the second type of bearing samples, perform bidirectional alternating optimization, solidify the feature transformation weight matrix and the mapping matrix, recalculate the sample feature vectors of all second type of bearing samples using the solidified feature transformation weight matrix, and construct a bearing state database. The module for acquiring the features to be detected is used to collect the cumulative processing time, end face visual image and vibration acceleration signal of the micro bearing to be detected, obtain the static feature vector and evolution feature vector of the bearing to be detected, and splice them into an initial feature vector to be detected. The detection module is used to process the initial feature vector to be detected based on the solidified feature transformation weight matrix, calculate the similarity between the processed initial feature vector to be detected and the sample feature vectors in the bearing status database, and obtain the final detection result of the bearing to be detected by combining the mapping matrix.

Images