Bearing vibration anomaly data analysis method and system based on deep learning

Abstract

The application discloses a bearing vibration abnormal data analysis method and system based on deep learning, relates to the technical field of data analysis, and comprises the following steps: collecting bearing vibration data, then performing self-adaptive abnormality identification and pre-classification on the multichannel original vibration data, outputting a deep feature vector and an abnormal data pre-classification result, performing distributed storage on the collected bearing data of various types according to a semantic category division rule, generating a heterogeneous storage board, performing differential abnormality analysis on the heterogeneous storage board through a lightweight deep decoupling network, outputting an abnormality analysis report of the heterogeneous storage board, performing consistency checking on the abnormality analysis report of the heterogeneous storage board, outputting a checking result, then constructing a bearing abnormality causal inference model, tracing an abnormal root cause, outputting an abnormal physical root cause positioning report, and realizing comprehensive visual output of the bearing abnormality analysis result, so that the analysis efficiency and precision of bearing vibration abnormal data are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, specifically to a method and system for analyzing bearing vibration anomaly data based on deep learning. Background Technology

[0002] Bearings are critical supporting components of mechanical equipment, and their operating status directly determines the safety and reliability of the entire machine. During long-term operation, they are prone to various faults, including those affecting the inner ring, outer ring, rolling elements, and cage, due to alternating loads, friction and wear, and fatigue spalling. Conducting abnormal bearing vibration data analysis is a core part of predictive maintenance of equipment. Traditional diagnostic methods rely on signal processing techniques such as Fourier transform, wavelet transform, and empirical mode decomposition to manually extract features. However, the completeness of feature extraction is insufficient under complex operating conditions, the model has weak generalization ability, and it is difficult to adapt to changing operating conditions. Although deep learning is gradually being applied to bearing fault diagnosis, the existing technology has significant shortcomings. Most solutions only output discrete fault category labels and cannot perform fine-grained quantitative classification of the degree of vibration abnormality damage. They generally adopt centralized data storage architectures, resulting in massive amounts of mixed data storage. Retrieving and filtering specific vibration abnormality data is time-consuming and has high transmission overhead, making it difficult to support specialized analysis. These multiple technical pain points restrict the implementation of intelligent bearing operation and maintenance. There is an urgent need for an integrated data analysis solution that incorporates deep learning, is equipped with a distributed storage architecture, and has complete interpretable output capabilities.

[0003] Traditional methods for analyzing massive amounts of bearing vibration data often employ centralized mixed storage or simple time-series storage, resulting in low efficiency, long processing times, and inaccurate analysis of bearing vibration anomaly data. This invention introduces distributed storage technology and combines it with deep learning to sequentially perform initial identification of vibration anomalies, distributed storage, differentiated vibration anomaly analysis and consistency verification, and root cause localization on bearing data, thereby achieving intelligent analysis of bearing vibration anomaly data. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for analyzing bearing vibration anomaly data based on deep learning, so as to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: Firstly, a deep learning-based method for analyzing abnormal bearing vibration data, comprising the following steps: Step 1: Assign numbers to each bearing to collect multi-channel raw vibration data, operational data, parameter data, environmental verification data, and historical vibration anomaly labeled sample data for each bearing, providing a data foundation for subsequent distributed storage and bearing vibration anomaly data analysis; Step 2: Using decoupled contrastive learning technology, with running data and environmental verification data as constraints, supervised fine-tuning is performed using historical vibration anomaly labeled sample data. Adaptive vibration anomaly identification and pre-classification are performed on multi-channel raw vibration data, and deep feature vectors and vibration anomaly data pre-classification results are output, providing a reference for subsequent distributed storage. Step 3: Based on the deep feature vector and the pre-classification results of vibration anomaly data, the collected bearing data of various types are distributed and stored according to the semantic category division rules to generate heterogeneous storage modules, which refines the process of bearing vibration anomaly data analysis and provides technical support for improving the accuracy of bearing vibration anomaly data analysis. Step 4: Perform differentiated vibration anomaly analysis on heterogeneous storage modules using a lightweight deep decoupling network and output a vibration anomaly analysis report for the heterogeneous storage modules. This is a key process in the analysis of bearing vibration anomaly data. Step 5: Perform consistency verification between the vibration anomaly analysis report of the heterogeneous storage module and the pre-classification results of vibration anomaly data and the historical vibration anomaly labeled sample data, and output the verification results. This is an important basis for the credibility of the vibration anomaly analysis report of the heterogeneous storage module. Step 6: Based on the verification results and heterogeneous storage modules, construct a causal inference model for bearing vibration anomaly, trace the root cause of vibration anomaly, and output a report on the physical root cause location of vibration anomaly, thus realizing the tracing of the root cause corresponding to the bearing vibration anomaly data. Step 7: Visualize the output results of the steps to obtain the distributed data topology dashboard for bearings, the distributed thermal matrix of bearing vibration anomalies, the consistency verification and comparison dashboard, and the causal correlation map of bearing vibration anomalies.

[0006] A further improvement to the technical solution of this invention lies in the fact that, in step one, the process of assigning numbers to each bearing to collect multi-channel raw vibration data, operational data, parameter data, environmental verification data, and historical vibration anomaly annotation sample data for each bearing includes: Deploy different types of acquisition devices and combine them with data entry technology to collect multi-channel raw vibration data, operating data, parameter data, environmental verification data, and historical vibration anomaly annotation sample data for each bearing; The data acquisition equipment includes a triaxial piezoelectric vibration acceleration sensor, a multi-channel synchronous data acquisition card, a photoelectric speed sensor, a dynamic torque sensor, a high-precision coordinate measuring machine, a digital display outside micrometer, a laser scanning micrometer, a fiber optic sensor array counting device, a bearing contact angle measuring instrument, a base-mounted piezoelectric acceleration sensor, and an industrial-grade platinum resistance temperature sensor. The multi-channel raw vibration data includes the bearing's raw vibration acceleration signal and its corresponding channel identifier and timestamp; The operating data includes the real-time rotational speed and the real-time load torque borne by the bearing during operation; The parameter data includes the bearing model information, pitch circle diameter, rolling element diameter, number of rolling elements, and contact angle; Environmental verification data includes the background noise signal of the bearing mounting location and the ambient temperature; Historical vibration anomaly labeled sample data includes bearing vibration signal data samples with labeled vibration anomaly types and their corresponding vibration anomaly category labels; Specifically, by combining a triaxial piezoelectric vibration accelerometer with a multi-channel synchronous data acquisition card, the original vibration acceleration signal of the bearing and its corresponding channel identifier and timestamp are acquired; The real-time rotational speed and the real-time load torque of the bearing are collected by using photoelectric speed sensor and dynamic torque sensor respectively during the operation of the bearing. Using data entry technology, the bearing model information is obtained from the bearing manufacturer's database. A high-precision coordinate measuring machine is used to collect the bearing's pitch circle diameter. Combined with a digital display outside micrometer and a laser scanning micrometer, the bearing's rolling element diameter is collected. A fiber optic sensor array counting device is used to collect the number of the bearing's rolling elements. A dedicated bearing contact angle measuring instrument is used to collect the bearing's contact angle. The background noise signal of the bearing mounting location and the ambient temperature were collected by using a base-mounted piezoelectric accelerometer and an industrial-grade platinum resistance temperature sensor, respectively. Using data entry technology, bearing vibration signal data samples with labeled vibration anomaly types and their corresponding vibration anomaly category labels were extracted from the standard bearing vibration anomaly experimental database and the equipment operation and maintenance real-time historical database, respectively. Data cleaning was performed on the collected operating data, parameter data, ambient temperature at the bearing installation location, and historical vibration anomaly labeled sample data. An adaptive wavelet threshold denoising algorithm was used to denoise the original vibration acceleration signal of the bearing and the background noise signal of the machine base vibration at the bearing installation location. The preprocessed multi-channel raw vibration data, operating data, parameter data, and environmental verification data of each bearing are assigned timestamps, the current timestamp is adjusted and recorded, and this timestamp is named the synchronization timestamp to achieve time synchronization of the multi-channel raw vibration data, operating data, parameter data, and environmental verification data of each bearing.

[0007] A further improvement to the technical solution of this invention lies in the fact that, in step two, the process of supervised fine-tuning using decoupled contrastive learning technology, with operational data and environmental verification data as constraints, and using historical vibration anomaly labeled sample data includes: Set the time window length and window step, and slice the original vibration acceleration signal of the bearing after noise reduction according to the time window length and window step to generate several bearing multi-channel original vibration data samples. Each sample contains a three-channel identified time-series vibration acceleration signal sequence. Input the original vibration data samples of the bearing multi-channel into the deep learning model, and build the content encoder and the working condition encoder with the running data and environmental verification data as decoupling constraints. A self-supervised contrastive learning strategy is adopted to compare and adjust the feature value similarity of multi-channel original vibration data samples of the same bearing under different working conditions. Unsupervised pre-training is performed on the deep learning model so that the deep learning model learns the vibration anomaly feature representation under unchanged working conditions. After pre-training, the classification head is then fine-tuned in a supervised manner using historical vibration anomaly labeled samples to obtain the weights of the content encoder and the working condition encoder.

[0008] A further improvement to the technical solution of this invention lies in the fact that, in step two, the process of adaptive vibration anomaly identification and pre-classification of multi-channel raw vibration data, and outputting depth feature vectors and vibration anomaly data pre-classification results includes: The pre-classification results of vibration anomaly data include the vibration anomaly existence discrimination results (binary classification logical values) and primary vibration anomaly category labels of the bearing multi-channel raw vibration data samples; Load the content encoder and the condition encoder, freeze all parameters of the condition encoder, freeze the low-level convolution parameters of the content encoder, and unfreeze and fine-tune the high-level convolution parameters of the content encoder. A fully connected classification head is added at the end of the network, and historical vibration anomaly labeled sample data is imported. Supervised training is performed using a labeled smooth cross-entropy loss function and an adaptive moment estimation optimizer with weight decay. After supervised training is completed, vibration anomaly identification and pre-classification of the multi-channel raw vibration data of the bearing are adaptively completed, and the deep feature vector and vibration anomaly data pre-classification results are obtained.

[0009] A further improvement to the technical solution of this invention lies in the fact that, in step three, the process of distributing and storing the collected bearing data of various types according to semantic category classification rules based on the deep feature vector and the pre-classification results of vibration anomaly data to generate heterogeneous storage modules includes: Based on the deep feature vector and vibration anomaly data pre-classification results, the collected bearing data are divided into the original signal acquisition layer, the dynamic environment layer, the equipment attribute layer and the historical sample library layer according to semantic categories. A distributed storage architecture is adopted, and data is distributed by segment according to semantic category. Data with different semantic dimensions is stored in storage node shards. At the same time, a unified index is established across semantic categories, and the associated primary key of the collected data corresponding to the bearing with the same number is bound. Metadata service is used to maintain the unified index to ensure that various types of data can be retrieved in a linked manner. The collected bearing data is adapted to the heterogeneous storage engine to form multiple independent yet interconnected storage sub-blocks, ultimately generating a heterogeneous storage block.

[0010] A further improvement to the technical solution of this invention lies in the fact that, in step four, the process of performing differentiated vibration anomaly analysis on heterogeneous storage modules through a lightweight deep decoupling network and outputting a vibration anomaly analysis report for the heterogeneous storage modules includes: The vibration anomaly analysis report for heterogeneous storage modules includes a vibration anomaly severity score for each heterogeneous storage module, a vibration anomaly evolution prediction curve, a characteristic frequency energy spectrum, and comprehensive health indicators. The system reads various types of bearing data stored in the signal raw acquisition layer, working condition dynamic environment layer, equipment attribute layer and historical sample library layer in the heterogeneous storage module, respectively, and performs format conversion and dimension alignment on the read data to adapt to the input requirements of the lightweight deep decoupled network. A lightweight deep decoupling network is used to build a multi-branch structure. Vibration anomaly differentiation analysis is performed on various bearing data stored in the signal raw acquisition layer, working condition dynamic environment layer, equipment attribute layer and historical sample library layer in the heterogeneous storage module through the multi-branch structure to obtain the output characteristics of each branch structure. The output features of each branch structure are integrated into a bearing vibration anomaly analysis dataset, and the bearing vibration anomaly analysis dataset is divided into a training set and a test set in a 7:3 ratio. The neural network algorithm is used as the training set data as the input data, and the vibration anomaly severity scores of each heterogeneous storage board, the numerical sequence of the discrete vibration anomaly prediction of the bearing, the energy of each frequency band of the original vibration acceleration signal of the bearing, and the comprehensive health index are used as the output data to obtain the trained model. The test set data is input into the trained model, the model parameters are adjusted, the model performance is optimized, and the final neural network model is obtained. Combining the output features of each branch structure, the vibration anomaly severity score of each heterogeneous storage block, the numerical sequence of the bearing discrete vibration anomaly prediction, the energy of each frequency band of the bearing's original vibration acceleration signal, and the comprehensive health index are output. By using the numerical sequence of discrete vibration anomaly prediction for bearings, a vibration anomaly evolution prediction curve is established, and the characteristic frequency energy spectrum is obtained by using the energy of each frequency band of the original vibration acceleration signal of the bearing. By integrating the vibration anomaly severity scores, vibration anomaly evolution prediction curves, characteristic frequency energy spectra, and comprehensive health indicators of various heterogeneous storage modules, a vibration anomaly analysis report of the heterogeneous storage modules is obtained.

[0011] A further improvement to the technical solution of this invention lies in the fact that, in step five, the process of verifying the consistency between the vibration anomaly analysis report of the heterogeneous storage module and the pre-classification results of vibration anomaly data and the historical vibration anomaly labeled sample data, and outputting the verification results, includes: The verification results include the vibration anomaly analysis reports of each heterogeneous storage module and the corresponding vibration anomaly confidence consistency score; By using timestamps and the IDs of the original vibration data samples from the bearing multi-channel, the pre-classification results of vibration anomaly data, the historical vibration anomaly labeled sample data, and the vibration anomaly analysis report of the heterogeneous storage module are aligned in terms of time axis and sample dimension. Based on the pre-classification results of vibration anomaly data, the timestamp and the original vibration data sample ID of the bearing multi-channel are extracted. The vibration anomaly existence discrimination results and primary vibration anomaly category labels of the corresponding original vibration data samples of the bearing multi-channel under the current timestamp and data sample ID are retrieved. According to the extracted timestamp and data sample ID, the vibration anomaly severity score, vibration anomaly evolution prediction curve, characteristic frequency energy spectrum and comprehensive health index of the corresponding heterogeneous storage module are retrieved in the vibration anomaly analysis report of the heterogeneous storage module. At the same time, the corresponding data samples and their vibration anomaly category labels are retrieved in the historical vibration anomaly labeled sample data. Using deep learning algorithms, the retrieved data is comprehensively analyzed, and a three-dimensional consistency check is performed sequentially to obtain a consistency check score. The consistency check score includes the consistency score of vibration anomaly category, the consistency score of severity, the consistency score of historical reference, and the consistency score of physical characteristics. The three-dimensional consistency check of results includes the consistency check of vibration anomaly category, the consistency check of severity, the consistency check of historical reference, and the physical consistency check of characteristic frequency. Weights were assigned to the vibration anomaly category consistency score, severity consistency score, historical reference consistency score, and physical characteristic consistency score, respectively. The weighted summation method was used to obtain the vibration anomaly confidence consistency score corresponding to the vibration anomaly analysis report of each heterogeneous storage module. The weighted summation method involves summing the products of the vibration anomaly category consistency score and its weight, the severity consistency score and its weight, the historical reference consistency score and its weight, and the physical characteristic consistency score and its weight.

[0012] A further improvement to the technical solution of this invention lies in the following step: Step six, based on the verification results and heterogeneous storage modules, constructs a causal inference model for bearing vibration anomaly, traces the root cause of vibration anomaly, and outputs a report on the physical root cause location of vibration anomaly, including: The vibration anomaly analysis reports of each heterogeneous storage module are weighted and aggregated according to the confidence consistency score of each heterogeneous storage module to generate a weighted vibration anomaly evidence matrix. Among them, the analysis results of heterogeneous storage modules with high confidence consistency scores have higher decision weight in subsequent inferences, while the evidence weight of heterogeneous storage modules with low confidence consistency scores is greatly weakened, and their contribution to root cause inference can be ignored or only used as an auxiliary reference. By using historical vibration anomaly labeled sample data, the root cause component of the vibration anomaly was identified, and the weighted vibration anomaly evidence combination pattern of each heterogeneous storage module was determined by using a weighted vibration anomaly evidence matrix. The weighted vibration anomaly evidence matrix and historical vibration anomaly labeled sample data are used as training data to train a structural causal model. The structural causal model learns the correspondence between the evidence combination patterns of each block in the weighted vibration anomaly evidence matrix and the confirmed vibration anomaly root cause components in the historical labeled sample data, establishes a mapping function from weighted vibration anomaly evidence to a single vibration anomaly root cause component, and outputs the vibration anomaly root cause component of each heterogeneous storage block. Using bearing numbers as units, integrate the root cause components of vibration anomalies and the weighted vibration anomaly evidence matrix of all heterogeneous storage modules within each bearing, and output a vibration anomaly physical root cause location report.

[0013] A further improvement to the technical solution of this invention lies in the fact that, in step seven, the process of visualizing the output results of the step to obtain the bearing distributed data topology dashboard, the distributed bearing vibration anomaly thermal matrix, the consistency verification comparison dashboard, and the bearing vibration anomaly causal tracing orientation correlation map includes: The visualization output process of the bearing distributed data topology dashboard includes: referring to each bearing number, scanning all heterogeneous storage modules within the bearing, reading the data stored in each module and the total number of entries, displaying each heterogeneous storage module in the form of topology nodes, with the node size representing the total number of data entries stored in the corresponding heterogeneous storage module, embedding parameter cards inside each topology node, presenting the multi-channel raw vibration data, operating data, parameter data, environmental verification data, historical vibration anomaly annotation sample data, synchronization timestamps, and unified index of each bearing, thus obtaining the bearing distributed data topology dashboard; The visualization output process of the distributed bearing vibration anomaly thermal matrix includes: building a two-dimensional matrix basic structure with rows representing each heterogeneous storage module and columns representing bearing numbers; scheduling the vibration anomaly analysis report of the heterogeneous storage module corresponding to each matrix cell; classifying the vibration anomaly severity level of each heterogeneous storage module based on the vibration anomaly severity score of each heterogeneous storage module; matching thermal color schemes to the vibration anomaly severity level of each heterogeneous storage module; color filling and data embedding processing of each matrix cell; intuitively distinguishing the severity of vibration anomalies corresponding to each heterogeneous storage module; unifying the layout style; completing the overall rendering of the matrix; and obtaining the distributed bearing vibration anomaly thermal matrix. The visualization output process of the consistency verification dashboard includes: dividing the display area into independent display areas, placing the pre-classification results of vibration anomaly data, the vibration anomaly analysis report of the heterogeneous storage module, and the historical vibration anomaly labeled sample data respectively, completing the basic layout, setting the scale and pointer for the heterogeneous storage module, mapping the vibration anomaly confidence consistency score corresponding to the vibration anomaly analysis report of each heterogeneous storage module to the scale range of the scale, marking the specific value with the pointer, setting the scoring threshold, highlighting the heterogeneous storage module that exceeds the scoring threshold, summarizing the independent display areas, pointer scale and highlighted markings, unifying the interface rendering, and generating the consistency verification dashboard; The visualization output process of the bearing vibration anomaly causal directional correlation map includes: taking the bearing number as the central node, and linking the operating data, parameter data and environmental verification data of each bearing to the multi-level inducing factor nodes in a hierarchical manner. Arrows of different thicknesses are used to represent the confidence consistency scores of each heterogeneous storage module. Clicking the central node can expand the corresponding vibration anomaly physical root cause location report, and the bearing vibration anomaly causal directional correlation map is obtained.

[0014] Secondly, a deep learning-based bearing vibration anomaly data analysis system is used to implement the aforementioned deep learning-based bearing vibration anomaly data analysis method. It includes a bearing vibration data acquisition module, a preliminary vibration anomaly analysis module, a distributed storage module, a deep vibration anomaly analysis module, a consistency verification module, a root cause localization module, and a visualization module, wherein each module is interconnected. The bearing vibration data acquisition module assigns a number to each bearing to collect multi-channel raw vibration data, operating data, parameter data, environmental verification data, and historical vibration anomaly annotation sample data for each bearing. The preliminary vibration anomaly analysis module utilizes decoupled contrastive learning technology, with operational data and environmental verification data as constraints, and uses historical vibration anomaly labeled sample data for supervised fine-tuning. It performs adaptive vibration anomaly identification and pre-classification on multi-channel raw vibration data, and outputs deep feature vectors and vibration anomaly data pre-classification results. The distributed storage module, based on deep feature vectors and pre-classification results of vibration anomaly data, performs distributed storage of various types of collected bearing data according to semantic category division rules, generating heterogeneous storage modules; The deep vibration anomaly analysis module uses a lightweight deep decoupling network to perform differentiated vibration anomaly analysis on heterogeneous storage modules and outputs a vibration anomaly analysis report for the heterogeneous storage modules. The consistency verification module performs consistency verification between the vibration anomaly analysis report of the heterogeneous storage module and the pre-classification results of vibration anomaly data and the historical vibration anomaly labeled sample data, and outputs the verification results. The root cause localization module, based on the verification results and heterogeneous storage modules, constructs a causal inference model for bearing vibration anomalies, tracks the root cause of vibration anomalies, and outputs a report on the physical root cause localization of vibration anomalies. The visualization module visualizes the output results of the steps, resulting in a distributed bearing data topology dashboard, a distributed bearing vibration anomaly thermal matrix, a consistency verification comparison dashboard, and a causal correlation map of bearing vibration anomalies.

[0015] The beneficial effects of this invention are as follows: Compared with traditional bearing vibration anomaly data analysis techniques, the bearing vibration anomaly data analysis method and system based on deep learning in this invention closely integrates decoupled comparative learning, lightweight deep decoupled networks, structural causal inference models, and distributed heterogeneous storage technology with the bearing vibration anomaly data analysis process. This achieves real-time and comprehensive monitoring of the vibration state of bearings throughout their entire life cycle based on deep learning. It sequentially completes the entire process of automated intelligent judgment, including vibration anomaly pre-classification, hierarchical storage, differentiated vibration anomaly analysis, consistency verification, vibration anomaly root cause detection, and interpretable deduction with visualized output. This solves the problems of poor generalization of feature extraction, low efficiency of centralized storage and retrieval of massive vibration data, bias in multi-source vibration anomaly analysis results, and inability to trace the causes of vibration anomalies in traditional solutions. It ensures that the method in this invention can refine the dynamic monitoring standards for bearing vibration anomaly data analysis within a more precise range, making the monitored bearing vibration anomaly data a more accurate judgment indicator under the same conditions. The development and application of this method significantly enhances the intelligence, interpretability, and practicality of the deep learning-based bearing vibration anomaly data analysis process. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0017] Figure 1 This is a flowchart of the bearing vibration anomaly data analysis method based on deep learning according to the present invention; Figure 2 This is a block diagram of the bearing vibration anomaly data analysis system based on deep learning according to the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1, as Figure 1 As shown, this invention provides a deep learning-based method for analyzing bearing vibration anomaly data, comprising the following steps: S101. Assign numbers to each bearing to collect multi-channel raw vibration data, operation data, parameter data, environmental verification data and historical vibration anomaly annotation sample data of each bearing, and complete the collection of raw bearing data in all dimensions. S102. Using decoupled contrastive learning technology, with running data and environmental verification data as constraints, supervised fine-tuning is performed using historical vibration anomaly labeled sample data. Adaptive vibration anomaly identification and pre-classification are performed on multi-channel raw vibration data, and the deep feature vector and vibration anomaly data pre-classification results are efficiently output. S103. Based on the deep feature vector and the pre-classification results of vibration anomaly data, the collected bearing data of various types are distributed and stored according to the semantic category division rules to generate heterogeneous storage modules. S104. Introduce a lightweight deep decoupling network to perform differentiated vibration anomaly analysis on heterogeneous storage modules and output a unified and complete vibration anomaly analysis report for heterogeneous storage modules. S105. Conduct consistency verification of the vibration anomaly analysis report of the linked heterogeneous storage module, the pre-classification results of vibration anomaly data, and the historical vibration anomaly labeled sample data, and output the verification results. S106. Based on the verification results and heterogeneous storage modules, construct a causal inference model for bearing vibration anomaly, accurately reverse-track the root cause of vibration anomaly, and output a report on the physical root cause location of vibration anomaly. S107. Visualize the output results of the steps to obtain the bearing distributed data topology dashboard, the distributed bearing vibration anomaly thermal matrix, the consistency verification comparison dashboard, and the bearing vibration anomaly causal tracing orientation correlation map.

[0020] In S101, the process of assigning numbers to each bearing to collect multi-channel raw vibration data, operational data, parameter data, environmental verification data, and historical vibration anomaly annotation sample data for each bearing includes: Deploy different types of acquisition devices and combine them with data entry technology to collect multi-channel raw vibration data, operating data, parameter data, environmental verification data, and historical vibration anomaly annotation sample data for each bearing; The data acquisition equipment includes a triaxial piezoelectric vibration acceleration sensor, a multi-channel synchronous data acquisition card, a photoelectric speed sensor, a dynamic torque sensor, a high-precision coordinate measuring machine, a digital display outside micrometer, a laser scanning micrometer, a fiber optic sensor array counting device, a bearing contact angle measuring instrument, a base-mounted piezoelectric acceleration sensor, and an industrial-grade platinum resistance temperature sensor. The multi-channel raw vibration data includes the bearing's raw vibration acceleration signal and its corresponding channel identifier (which includes the axial direction, horizontal radial direction, and vertical radial direction) and timestamp; The operating data includes the real-time rotational speed and the real-time load torque borne by the bearing during operation; The parameter data includes the bearing model information, pitch circle diameter, rolling element diameter, number of rolling elements, and contact angle; Environmental verification data includes the background noise signal of the bearing mounting location and the ambient temperature; Historical vibration anomaly labeled sample data includes bearing vibration signal data samples with labeled vibration anomaly types and their corresponding vibration anomaly category labels; Specifically, by combining a triaxial piezoelectric vibration accelerometer with a multi-channel synchronous data acquisition card, the original vibration acceleration signal of the bearing and its corresponding channel identifier and timestamp are acquired; The real-time rotational speed and real-time load torque of the bearing during operation are collected by using photoelectric speed sensor and dynamic torque sensor respectively. Using data entry technology, the bearing model information is obtained from the bearing manufacturer's database. A high-precision coordinate measuring machine is used to collect the bearing's pitch circle diameter. Combined with a digital display outside micrometer and a laser scanning micrometer, the bearing's rolling element diameter is collected. A fiber optic sensor array counting device is used to collect the number of the bearing's rolling elements. A dedicated bearing contact angle measuring instrument is used to collect the bearing's contact angle. The background noise signal of the bearing mounting location and the ambient temperature were collected by using a base-mounted piezoelectric accelerometer and an industrial-grade platinum resistance temperature sensor, respectively. Using data entry technology, bearing vibration signal data samples with labeled vibration anomaly types and their corresponding vibration anomaly category labels were extracted from the standard bearing vibration anomaly experimental database and the equipment operation and maintenance real-time historical database, respectively. Data cleaning was performed on the collected operating data, parameter data, ambient temperature at the bearing installation location, and historical vibration anomaly labeled sample data. An adaptive wavelet threshold denoising algorithm was used to denoise the original vibration acceleration signal of the bearing and the background noise signal of the machine base vibration at the bearing installation location. The preprocessed multi-channel raw vibration data, operating data, parameter data, and environmental verification data of each bearing are assigned timestamps, the current timestamp is adjusted and recorded, and this timestamp is named the synchronization timestamp to achieve time synchronization of the multi-channel raw vibration data, operating data, parameter data, and environmental verification data of each bearing.

[0021] In S102, the process of supervised fine-tuning using decoupled contrastive learning techniques, with operational data and environmental verification data as constraints, and using historical vibration anomaly labeled sample data, includes: Set the time window length and window step, and slice the original vibration acceleration signal of the bearing after noise reduction according to the time window length and window step to generate several bearing multi-channel original vibration data samples. Each sample contains a time-series vibration acceleration signal sequence with three channel identifiers (specifically including axial direction, horizontal radial direction, and vertical radial direction). Input the original vibration data samples of the bearing multi-channel into the deep learning model, and build the content encoder and the working condition encoder with the running data and environmental verification data as decoupling constraints. A self-supervised contrastive learning strategy is adopted to compare and adjust the feature value similarity of multi-channel original vibration data samples of the same bearing under different working conditions. Unsupervised pre-training is performed on the deep learning model so that the deep learning model learns the vibration anomaly feature representation under unchanged working conditions. After pre-training, the classification head is then fine-tuned in a supervised manner using historical vibration anomaly labeled samples to obtain the weights of the content encoder and the working condition encoder.

[0022] In S102, the process of adaptive vibration anomaly identification and pre-classification of multi-channel raw vibration data, and outputting depth feature vectors and vibration anomaly data pre-classification results, includes: The pre-classification results of vibration anomaly data include the vibration anomaly existence discrimination results (binary classification logical values) and primary vibration anomaly category labels of the bearing multi-channel raw vibration data samples; Load the content encoder and the condition encoder, freeze all parameters of the condition encoder, freeze the low-level convolution parameters of the content encoder, and unfreeze and fine-tune the high-level convolution parameters of the content encoder. A fully connected classification head is added at the end of the network, historical vibration anomaly labeled sample data is imported, and supervised training is performed using a labeled smooth cross-entropy loss function and an adaptive moment estimation optimizer with weight decay. After supervised training is completed, vibration anomaly identification and pre-classification of multi-channel raw vibration data of bearings are adaptively completed, and deep feature vectors and vibration anomaly data pre-classification results are obtained. It should be added that after the above supervised training is completed, the output of the layer before the classifier is taken as the deep feature vector during forward inference, and the Softmax is replaced with K independent Sigmoids. The probability of existence of each primary vibration anomaly category in the bearing multi-channel original vibration data sample is output respectively. In the scenario of bearing composite vibration anomaly analysis, the probability of existence of different primary vibration anomaly categories in the multi-channel original vibration data sample is different. The above K is the total number of primary vibration anomaly categories in the bearing multi-channel original vibration data sample. A probability threshold is set for each primary vibration anomaly category in the multi-channel raw vibration data sample of the bearing. The probability thresholds for different primary vibration anomaly categories are different. When the probability of the existence of a primary vibration anomaly category in the multi-channel raw vibration data sample of the bearing is less than or equal to the corresponding probability threshold, the primary vibration anomaly of that category does not exist in the corresponding multi-channel raw vibration data sample of the bearing. When the probability of the existence of a primary vibration anomaly category in the multi-channel raw vibration data sample of the bearing is greater than the corresponding probability threshold, the primary vibration anomaly of that category exists in the corresponding multi-channel raw vibration data sample of the bearing. The corresponding label is then assigned to the category, and the primary vibration anomaly label is obtained.

[0023] In S103, based on the deep feature vector and the pre-classification results of vibration anomaly data, and according to semantic category division rules, the process of distributing and storing the collected bearing data of various types to generate heterogeneous storage modules includes: Based on the deep feature vector and pre-classification results of vibration anomaly data, the collected bearing data are divided into four layers according to semantic categories: raw signal acquisition layer, dynamic environment layer, equipment attribute layer, and historical sample library layer. The raw signal acquisition layer stores the raw vibration acceleration signals of the bearings, their corresponding channel identifiers, and the aforementioned synchronization timestamps, partitioned by time. The dynamic environment layer stores the real-time rotational speed and real-time load torque of the bearing during operation, the background noise signal of the bearing mounting location, and the ambient temperature, along with the aforementioned synchronization timestamps. The equipment attribute layer stores the bearing's model information, pitch circle diameter, rolling element diameter, number of rolling elements, and contact angle, along with the aforementioned synchronization timestamps. The historical sample library layer stores bearing vibration signal data samples labeled with vibration anomaly types and their corresponding vibration anomaly category labels. These are the data storage rules for each layer and serve as the basis for subsequent distributed storage sharding routing. A distributed storage architecture is adopted, and data is distributed by segment according to semantic category. Data with different semantic dimensions is stored in storage node shards. At the same time, a unified index is established across semantic categories, and the associated primary key of the collected data corresponding to the bearing with the same number is bound. Metadata service is used to maintain the unified index to ensure that various types of data can be retrieved in a linked manner. The collected bearing data is adapted to the heterogeneous storage engine format, dividing it into multiple independent yet interconnected storage sub-blocks, ultimately generating a heterogeneous storage block. The format adaptation of the collected bearing data to the heterogeneous storage engine mainly includes: adapting the data specified for storage in the original signal acquisition layer to a columnar storage engine, adapting the data specified for storage in the dynamic environment layer to a time-series database, adapting the data specified for storage in the equipment attribute layer to a key-value storage database, and adapting the data specified for storage in the historical sample library layer to a sample database, ultimately generating heterogeneous storage blocks oriented towards different data characteristics.

[0024] In S104, the process of performing differentiated vibration anomaly analysis on heterogeneous storage modules and outputting a vibration anomaly analysis report for the heterogeneous storage modules through a lightweight deep decoupling network includes: The vibration anomaly analysis report for heterogeneous storage modules includes a vibration anomaly severity score for each heterogeneous storage module, a vibration anomaly evolution prediction curve, a characteristic frequency energy spectrum, and comprehensive health indicators. The system reads various types of bearing data stored in the signal raw acquisition layer, working condition dynamic environment layer, equipment attribute layer and historical sample library layer in the heterogeneous storage module, respectively, and performs format conversion and dimension alignment on the read data to adapt to the input requirements of the lightweight deep decoupled network. A lightweight deep decoupled network is used to build a multi-branch structure. The multi-branch structure is used to perform vibration anomaly differential analysis on various bearing data stored in the signal raw acquisition layer, working condition dynamic environment layer, equipment attribute layer and historical sample library layer in the heterogeneous storage module. The output characteristics of each branch structure are obtained. The multi-branch structure includes vibration signal branch, working condition environment branch, equipment attribute branch and historical sample branch, which are used to perform vibration anomaly differential analysis on various bearing data stored in the signal raw acquisition layer, working condition dynamic environment layer, equipment attribute layer and historical sample library layer in the heterogeneous storage module, respectively. The output features of each branch structure are integrated into a bearing vibration anomaly analysis dataset, and the bearing vibration anomaly analysis dataset is divided into a training set and a test set in a 7:3 ratio. The neural network algorithm is used as the training set data as the input data, and the vibration anomaly severity scores of each heterogeneous storage board, the numerical sequence of the discrete vibration anomaly prediction of the bearing, the energy of each frequency band of the original vibration acceleration signal of the bearing, and the comprehensive health index are used as the output data to obtain the trained model. The test set data is input into the trained model, the model parameters are adjusted, the model performance is optimized, and the final neural network model is obtained. Combining the output features of each branch structure, the vibration anomaly severity score of each heterogeneous storage block, the numerical sequence of the bearing discrete vibration anomaly prediction, the energy of each frequency band of the bearing's original vibration acceleration signal, and the comprehensive health index are output. By using the numerical sequence of discrete vibration anomaly prediction for bearings, a vibration anomaly evolution prediction curve is established, and the characteristic frequency energy spectrum is obtained by using the energy of each frequency band of the original vibration acceleration signal of the bearing. By integrating the vibration anomaly severity scores, vibration anomaly evolution prediction curves, characteristic frequency energy spectra, and comprehensive health indicators of various heterogeneous storage modules, a vibration anomaly analysis report of the heterogeneous storage modules is obtained. The specific process of establishing the vibration anomaly evolution prediction curve includes: establishing a two-dimensional coordinate system, dividing the discrete vibration anomaly prediction numerical sequence of the bearing into several discrete time points at equal intervals, mapping each discrete time point to the horizontal axis in ascending order from left to right, mapping the corresponding discrete vibration anomaly prediction numerical values ​​of the bearing to the vertical axis, completing the coordinate scale configuration, connecting all discrete data points with a line graph, and combining smoothing fitting technology to obtain the vibration anomaly evolution prediction curve. The specific process of establishing the characteristic frequency energy spectrum includes: establishing a two-dimensional coordinate system, setting a fixed length, dividing the current horizontal axis into equal parts with the fixed length as the interval, mapping the frequency bands corresponding to the fixed length to the horizontal axis in ascending order from left to right, mapping the energy of each frequency band of the original vibration acceleration signal of the bearing to the vertical axis corresponding to each frequency band, with each frequency band corresponding to a column, and visually comparing the energy strength of different frequency bands to generate the characteristic frequency energy spectrum.

[0025] In S105, the process of verifying the consistency between the vibration anomaly analysis report of the heterogeneous storage module and the pre-classification results of vibration anomaly data and the historical vibration anomaly labeled sample data, and outputting the verification results includes: The verification results include the vibration anomaly analysis reports of each heterogeneous storage module and the corresponding vibration anomaly confidence consistency score; By using timestamps and the IDs of the original vibration data samples from the bearing multi-channel, the pre-classification results of vibration anomaly data, the historical vibration anomaly labeled sample data, and the vibration anomaly analysis report of the heterogeneous storage module are aligned in terms of time axis and sample dimension. Based on the pre-classification results of vibration anomaly data, the timestamp and the original vibration data sample ID of the bearing multi-channel are extracted. The vibration anomaly existence discrimination results and primary vibration anomaly category labels of the corresponding original vibration data samples of the bearing multi-channel under the current timestamp and data sample ID are retrieved. According to the extracted timestamp and data sample ID, the vibration anomaly severity score, vibration anomaly evolution prediction curve, characteristic frequency energy spectrum and comprehensive health index of the corresponding heterogeneous storage module are retrieved in the vibration anomaly analysis report of the heterogeneous storage module. At the same time, the corresponding data samples and their vibration anomaly category labels are retrieved in the historical vibration anomaly labeled sample data. Using deep learning algorithms, the retrieved data is comprehensively analyzed, and a three-dimensional consistency check is performed sequentially to obtain a consistency check score. The consistency check score includes the consistency score of vibration anomaly category, the consistency score of severity, the consistency score of historical reference, and the consistency score of physical characteristics. The three-dimensional consistency check of results includes the consistency check of vibration anomaly category, the consistency check of severity, the consistency check of historical reference, and the physical consistency check of characteristic frequency. Specifically, the retrieved data is integrated to obtain a consistency verification dataset, which is then divided into a consistency verification training set and a consistency verification test set in an 8:2 ratio. The retrieved data includes the vibration anomaly existence discrimination results and primary vibration anomaly category labels of the corresponding bearing multi-channel original vibration data samples under the current timestamp and data sample ID, the vibration anomaly severity score of the corresponding heterogeneous storage module, the vibration anomaly evolution prediction curve, characteristic frequency energy spectrum and comprehensive health index, as well as the corresponding data samples and their vibration anomaly category labels in the historical vibration anomaly labeled sample data. Using a deep learning algorithm, the consistency verification training set is used as input data, and the vibration anomaly category consistency score, severity consistency score, historical reference consistency score, and physical feature consistency score are used as output data. Under the conditions of the extracted timestamps and data sample IDs, the nonlinear relationship between the retrieved data is learned to verify the consistency of vibration anomaly category, severity consistency, historical reference consistency, and physical consistency of characteristic frequency in the vibration anomaly analysis report of heterogeneous storage modules, thus obtaining a trained deep learning model. Input the consistency verification test set into the currently trained deep learning model, adjust the deep learning model parameters, optimize the deep learning parameters, and obtain the final vibration anomaly consistency verification model. Based on the currently retrieved data, output the vibration anomaly category consistency score, severity consistency score, historical reference consistency score, and physical characteristic consistency score; Weights were assigned to the vibration anomaly category consistency score, severity consistency score, historical reference consistency score, and physical characteristic consistency score, respectively. The weighted summation method was used to obtain the vibration anomaly confidence consistency score corresponding to the vibration anomaly analysis report of each heterogeneous storage module. The weighted summation method involves summing the products of the vibration anomaly category consistency score and its weight, the severity consistency score and its weight, the historical reference consistency score and its weight, and the physical characteristic consistency score and its weight.

[0026] In S106, the process of constructing a causal inference model for bearing vibration anomaly based on the verification results and heterogeneous storage modules, tracing the root cause of vibration anomaly, and outputting a report on the physical source location of vibration anomaly includes: The vibration anomaly analysis reports of each heterogeneous storage module are weighted and aggregated according to the confidence consistency score of each heterogeneous storage module to generate a weighted vibration anomaly evidence matrix. Among them, the analysis results of heterogeneous storage modules with high confidence consistency scores have higher decision weight in subsequent inferences, while the evidence weight of heterogeneous storage modules with low confidence consistency scores is greatly weakened, and their contribution to root cause inference can be ignored or only used as an auxiliary reference. By using historical vibration anomaly labeled sample data, the root cause component of the vibration anomaly was identified, and the weighted vibration anomaly evidence combination pattern of each heterogeneous storage module was determined by using a weighted vibration anomaly evidence matrix. The weighted vibration anomaly evidence matrix and historical vibration anomaly labeled sample data are used as training data to train a structural causal model. The structural causal model learns the correspondence between the evidence combination patterns of each block in the weighted vibration anomaly evidence matrix and the confirmed vibration anomaly root cause components in the historical labeled sample data, establishes a mapping function from weighted vibration anomaly evidence to a single vibration anomaly root cause component, and outputs the vibration anomaly root cause component of each heterogeneous storage block. During the learning process, the structural causal model automatically retrieves the historical cases in the historical vibration anomaly labeled sample data that are most similar to the current evidence pattern, uses the fault category label in the most similar historical case as the main basis for the current root cause inference, and outputs the component corresponding to the vibration anomaly root cause. Using bearing numbers as units, integrate the root cause components of vibration anomalies and the weighted vibration anomaly evidence matrix of all heterogeneous storage modules within each bearing, and output a vibration anomaly physical root cause location report.

[0027] In S107, the process of visualizing the output results of the step to obtain the distributed bearing data topology dashboard, the distributed bearing vibration anomaly thermal matrix, the consistency verification comparison dashboard, and the bearing vibration anomaly causal tracing orientation correlation map includes: The visualization output process of the bearing distributed data topology dashboard includes: referring to each bearing number, scanning all heterogeneous storage modules within the bearing, reading the data stored in each module and the total number of entries, displaying each heterogeneous storage module in the form of topology nodes, with the node size representing the total number of data entries stored in the corresponding heterogeneous storage module, embedding parameter cards inside each topology node, presenting the multi-channel raw vibration data, operating data, parameter data, environmental verification data, historical vibration anomaly annotation sample data, synchronization timestamps, and unified index of each bearing, thus obtaining the bearing distributed data topology dashboard. The card information is presented as a concise key-value pair list and is displayed in the topology layout along with the nodes. To avoid information overload, the system displays summary information by default, and the complete parameter card can be expanded when the user hovers over or clicks on the node. The visualization output process of the distributed bearing vibration anomaly thermal matrix includes: constructing a two-dimensional matrix basic structure with rows representing each heterogeneous storage module and columns representing bearing numbers; scheduling vibration anomaly analysis reports for each heterogeneous storage module corresponding to each matrix cell; setting a first threshold and a second threshold for vibration anomaly severity scoring; determining whether the first threshold is less than the second threshold; and further classifying the severity scores as below or equal to the first threshold, between the first and second thresholds, or above or equal to the second threshold. The gradient of values ​​is used to classify the severity of vibration anomalies of each heterogeneous storage module based on the severity score of each heterogeneous storage module. A thermal color system is matched to the severity level of vibration anomalies of each heterogeneous storage module. Color filling and data embedding are performed on each matrix cell to intuitively distinguish the severity of vibration anomalies of each heterogeneous storage module. The layout style is unified and the matrix is ​​rendered as a whole to obtain the distributed bearing vibration anomaly thermal matrix. The specific data embedding process is as follows: the severity score of vibration anomalies of the corresponding heterogeneous storage module, the fault evolution prediction curve, the characteristic frequency energy spectrum and the comprehensive health index are embedded in each matrix cell. The visualization output process of the consistency verification dashboard includes: dividing the display area into independent display areas, placing the pre-classification results of vibration anomaly data, the vibration anomaly analysis report of heterogeneous storage modules, and the historical vibration anomaly labeled sample data respectively, completing the basic layout, setting the scale and pointer for the heterogeneous storage modules, mapping the vibration anomaly confidence consistency score corresponding to the vibration anomaly analysis report of each heterogeneous storage module to the scale range of the scale, marking the specific value with the pointer, setting the scoring threshold, comparing the vibration anomaly confidence consistency score corresponding to the vibration anomaly analysis report of each heterogeneous storage module with the scoring threshold one by one, highlighting the heterogeneous storage modules that exceed the scoring threshold, summarizing the independent display areas, pointer scale, and highlighted markings, unifying the interface rendering, and generating the consistency verification dashboard; The visualization output process of the bearing vibration anomaly causal source directional correlation map includes: taking the bearing number as the central node, and outwardly linking the multi-level inducing factor nodes corresponding to the operating data, parameter data, and environmental verification data of each bearing. Arrows of different thicknesses represent the confidence consistency scores corresponding to each heterogeneous storage module. Thicker arrows correspond to higher confidence consistency scores for each heterogeneous storage module, while thinner arrows correspond to lower confidence consistency scores. Clicking on the central node allows you to expand the corresponding vibration anomaly physical root cause location report, thus obtaining the bearing vibration anomaly causal source directional correlation map.

[0028] Example 2, as Figure 2 As shown, based on Embodiment 1, the present invention provides a technical solution: a bearing vibration anomaly data analysis system based on deep learning, used to implement the above-mentioned bearing vibration anomaly data analysis method based on deep learning, including a bearing vibration data acquisition module, a preliminary vibration anomaly analysis module, a distributed storage module, a deep vibration anomaly analysis module, a consistency verification module, a root cause localization module, and a visualization module, wherein each module is communicatively connected; The bearing vibration data acquisition module assigns a number to each bearing to collect multi-channel raw vibration data, operating data, parameter data, environmental verification data, and historical vibration anomaly annotation sample data for each bearing. The preliminary vibration anomaly analysis module utilizes decoupled contrastive learning technology, with operational data and environmental verification data as constraints, and uses historical vibration anomaly labeled sample data for supervised fine-tuning. It performs adaptive vibration anomaly identification and pre-classification on multi-channel raw vibration data, and outputs deep feature vectors and vibration anomaly data pre-classification results. The distributed storage module, based on deep feature vectors and pre-classification results of vibration anomaly data, performs distributed storage of various types of collected bearing data according to semantic category division rules, generating heterogeneous storage modules; The deep vibration anomaly analysis module uses a lightweight deep decoupling network to perform differentiated vibration anomaly analysis on heterogeneous storage modules and outputs a vibration anomaly analysis report for the heterogeneous storage modules. The consistency verification module performs consistency verification between the vibration anomaly analysis report of the heterogeneous storage module and the pre-classification results of vibration anomaly data and the historical vibration anomaly labeled sample data, and outputs the verification results. The root cause localization module, based on the verification results and heterogeneous storage modules, constructs a causal inference model for bearing vibration anomalies, tracks the root cause of vibration anomalies, and outputs a report on the physical root cause localization of vibration anomalies. The visualization module visualizes the output results of the steps, resulting in a distributed bearing data topology dashboard, a distributed bearing vibration anomaly thermal matrix, a consistency verification comparison dashboard, and a causal correlation map of bearing vibration anomalies.

[0029] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A deep learning-based method for analyzing bearing vibration anomaly data, characterized in that, Includes the following steps: Each bearing is assigned a number to collect multi-channel raw vibration data, operational data, parameter data, environmental verification data, and historical vibration anomaly annotation sample data for each bearing; Using decoupled contrastive learning technology, with the running data and the environmental verification data as constraints, supervised fine-tuning is performed using the historical vibration anomaly labeled sample data, adaptive vibration anomaly identification and pre-classification are performed on the multi-channel raw vibration data, and the deep feature vector and vibration anomaly data pre-classification results are output. Based on the depth feature vector and the pre-classification results of vibration anomaly data, the collected bearing data of various types are distributed and stored according to the semantic category division rules to generate heterogeneous storage modules. A lightweight, deeply decoupled network is used to perform differentiated vibration anomaly analysis on the heterogeneous storage modules, and a vibration anomaly analysis report of the heterogeneous storage modules is output. The consistency of the vibration anomaly analysis report of the heterogeneous storage module with the vibration anomaly data pre-classification results and the historical vibration anomaly labeled sample data is verified, and the verification results are output. Based on the verification results and the heterogeneous storage module, a causal inference model for bearing vibration anomaly is constructed to trace the root cause of vibration anomaly and output a report on the physical root cause location of vibration anomaly. The output results of the above steps are visualized to obtain a distributed data topology dashboard for bearings, a distributed thermal matrix of bearing vibration anomalies, a consistency verification and comparison dashboard, and a causal correlation map of bearing vibration anomalies.

2. The bearing vibration anomaly data analysis method based on deep learning according to claim 1, characterized in that, The process of assigning numbers to each bearing to collect multi-channel raw vibration data, operational data, parameter data, environmental verification data, and historical vibration anomaly annotation sample data for each bearing includes: Deploy different types of acquisition devices and combine them with data entry technology to collect multi-channel raw vibration data, operating data, parameter data, environmental verification data, and historical vibration anomaly annotation sample data for each bearing; The data acquisition equipment includes a triaxial piezoelectric vibration acceleration sensor, a multi-channel synchronous data acquisition card, a photoelectric speed sensor, a dynamic torque sensor, a high-precision coordinate measuring machine, a digital display outside micrometer, a laser scanning micrometer, a fiber optic sensor array counting device, a bearing contact angle measuring instrument, a base-mounted piezoelectric acceleration sensor, and an industrial-grade platinum resistance temperature sensor. The multi-channel raw vibration data includes the bearing's raw vibration acceleration signal and its corresponding channel identifier and timestamp; The operating data includes the real-time rotational speed and the real-time load torque borne by the bearing during operation; The parameter data includes the bearing model information, pitch circle diameter, rolling element diameter, number of rolling elements, and contact angle; The environmental verification data includes the background noise signal of the bearing mounting location and the ambient temperature. The historical vibration anomaly labeled sample data includes bearing vibration signal data samples with labeled vibration anomaly types and their corresponding vibration anomaly category labels; The collected operating data, parameter data, ambient temperature at the bearing installation location, and historical vibration anomaly labeled sample data are cleaned. An adaptive wavelet threshold denoising algorithm is used to denoise the original vibration acceleration signal of the bearing and the background noise signal of the machine base vibration at the bearing installation location. The preprocessed multi-channel raw vibration data, operating data, parameter data, and environmental verification data of each bearing are assigned timestamps, the current timestamp is adjusted and recorded, and the timestamp is named the synchronization timestamp to achieve time synchronization of the multi-channel raw vibration data, operating data, parameter data, and environmental verification data of each bearing.

3. The bearing vibration anomaly data analysis method based on deep learning according to claim 2, characterized in that, The process of using decoupled contrastive learning technology, with the operational data and environmental verification data as constraints, and using the historical vibration anomaly labeled sample data for supervised fine-tuning includes: Set the time window length and window step, and slice the original vibration acceleration signal of the denoised bearing according to the time window length and window step to generate several bearing multi-channel original vibration data samples, each sample containing a time-series vibration acceleration signal sequence; The bearing multi-channel raw vibration data samples are input into a deep learning model, and the content encoder and the working condition encoder are built using the running data and the environmental verification data as decoupling constraints. A self-supervised contrastive learning strategy is adopted to compare and adjust the feature value similarity of multi-channel original vibration data samples of the same bearing under different working conditions. The deep learning model is pre-trained unsupervised, and then the classification head is fine-tuned in a supervised manner using the historical vibration anomaly labeled samples to obtain the weights of the content encoder and the working condition encoder.

4. The bearing vibration anomaly data analysis method based on deep learning according to claim 3, characterized in that, The process of adaptive vibration anomaly identification and pre-classification of the multi-channel raw vibration data, and outputting depth feature vectors and vibration anomaly data pre-classification results includes: The pre-classification results of the vibration anomaly data include the vibration anomaly existence discrimination results and primary vibration anomaly category labels of the bearing multi-channel raw vibration data samples; Load the content encoder and the operating condition encoder, freeze all parameters of the operating condition encoder, freeze the bottom convolution parameters of the content encoder, and unfreeze and fine-tune the high-level convolution parameters of the content encoder. A fully connected classification head is added at the end of the network, and the historical vibration anomaly labeled sample data is imported. Supervised training is performed using a labeled smooth cross-entropy loss function and an adaptive moment estimation optimizer with weight decay. After the supervised training is completed, the vibration anomaly identification and pre-classification of the multi-channel raw vibration data of the bearing is adaptively completed, and the deep feature vector and the pre-classification result of the vibration anomaly data are obtained.

5. The bearing vibration anomaly data analysis method based on deep learning according to claim 4, characterized in that, The process of distributing and storing the collected bearing data of various types according to semantic category division rules based on the deep feature vector and the pre-classification results of vibration anomaly data to generate heterogeneous storage modules includes: Based on the depth feature vector and the pre-classification results of vibration anomaly data, the collected bearing data are divided into the original signal acquisition layer, the dynamic environment layer, the equipment attribute layer, and the historical sample library layer according to semantic categories. A distributed storage architecture is adopted, and data is distributed by segment according to semantic category. Data with different semantic dimensions is stored in storage node shards. At the same time, a unified index is established across semantic categories, and the associated primary key of the collected data corresponding to the bearing with the same number is bound. Metadata service is used to maintain this unified index. The collected bearing data is adapted to the heterogeneous storage engine to form multiple storage sub-blocks, and finally the heterogeneous storage blocks are generated.

6. The bearing vibration anomaly data analysis method based on deep learning according to claim 5, characterized in that, The process of performing differentiated vibration anomaly analysis on the heterogeneous storage modules using a lightweight deep decoupling network and outputting a vibration anomaly analysis report for the heterogeneous storage modules includes: The vibration anomaly analysis report of the heterogeneous storage modules includes a vibration anomaly severity score for each heterogeneous storage module, a vibration anomaly evolution prediction curve, a characteristic frequency energy spectrum, and a comprehensive health index. The various types of bearing data stored in the signal raw acquisition layer, working condition dynamic environment layer, equipment attribute layer and historical sample library layer of the heterogeneous storage module are read respectively. The read data is processed by format conversion and dimension alignment to adapt to the input requirements of the lightweight deep decoupling network. A multi-branch structure is built using a lightweight deep decoupling network. Vibration anomaly differentiation analysis is performed on various types of bearing data stored in the signal raw acquisition layer, working condition dynamic environment layer, equipment attribute layer and historical sample library layer in the heterogeneous storage module through the multi-branch structure to obtain the output characteristics of each branch structure. The output features of each branch structure are integrated into a bearing vibration anomaly analysis dataset, and the bearing vibration anomaly analysis dataset is divided into a training set and a test set. Using a neural network algorithm, the training set data is used as input data, and the vibration anomaly severity scores of each heterogeneous storage module, the numerical sequence of predicted discrete vibration anomalies of the bearing, the energy of each frequency band of the original vibration acceleration signal of the bearing, and the comprehensive health index are used as output data to obtain a trained model. The test set data is input into the trained model, the model parameters are adjusted, the performance of the model is optimized, and the final neural network model is obtained. Combining the output features of each branch structure, the vibration anomaly severity score of each heterogeneous storage module, the numerical sequence of the bearing discrete vibration anomaly prediction, the energy of each frequency band of the bearing's original vibration acceleration signal, and the comprehensive health index are output. Using the numerical sequence of the bearing discrete vibration anomaly prediction, the vibration anomaly evolution prediction curve is established, and the characteristic frequency energy spectrum is obtained through the energy of each frequency band of the bearing's original vibration acceleration signal. By integrating the vibration anomaly severity scores, vibration anomaly evolution prediction curves, characteristic frequency energy spectra, and comprehensive health indicators of each heterogeneous storage module, a vibration anomaly analysis report of the heterogeneous storage module is obtained.

7. The bearing vibration anomaly data analysis method based on deep learning according to claim 6, characterized in that, The process of verifying the consistency between the vibration anomaly analysis report of the heterogeneous storage module, the vibration anomaly data pre-classification results, and the historical vibration anomaly labeled sample data, and outputting the verification results includes: The verification results include the vibration anomaly confidence consistency score corresponding to the vibration anomaly analysis report of each heterogeneous storage module; By using the timestamp and the ID of the original vibration data sample from the bearing multi-channel, the pre-classification results of the vibration anomaly data, the historical vibration anomaly labeled sample data, and the vibration anomaly analysis report of the heterogeneous storage module are aligned in terms of time axis and sample dimension. Based on the pre-classification results of the vibration anomaly data, the timestamp and the original vibration data sample ID of the bearing multi-channel are extracted. The vibration anomaly existence discrimination result and primary vibration anomaly category label of the corresponding original vibration data sample of the bearing multi-channel under the current timestamp and data sample ID are retrieved. According to the extracted timestamp and data sample ID, the vibration anomaly severity score, vibration anomaly evolution prediction curve, characteristic frequency energy spectrum and comprehensive health index of the corresponding heterogeneous storage module are retrieved in the vibration anomaly analysis report of the heterogeneous storage module. At the same time, the corresponding data sample and its vibration anomaly category label are retrieved in the historical vibration anomaly labeled sample data. Using a deep learning algorithm, the retrieved data is integrated, and a three-dimensional consistency check is performed sequentially to obtain a consistency check score. The consistency check score includes a vibration anomaly category consistency score, a severity consistency score, a historical reference consistency score, and a physical feature consistency score. The three-dimensional consistency check includes vibration anomaly category consistency check, severity consistency check, historical reference consistency check, and characteristic frequency physical consistency check. Weights were assigned to the vibration anomaly category consistency score, severity consistency score, historical reference consistency score, and physical characteristic consistency score, respectively. A weighted summation method was used to obtain the vibration anomaly confidence consistency score corresponding to the vibration anomaly analysis report of each heterogeneous storage module.

8. The bearing vibration anomaly data analysis method based on deep learning according to claim 7, characterized in that, The process of constructing a causal inference model for bearing vibration anomaly based on the verification results and the heterogeneous storage module, tracing the root cause of vibration anomaly, and outputting a report on the physical root cause location of vibration anomaly includes: The vibration anomaly analysis reports of each heterogeneous storage module are weighted and aggregated according to the confidence consistency score corresponding to each heterogeneous storage module to generate a weighted vibration anomaly evidence matrix. By using the historical vibration anomaly labeled sample data, the root cause component of the vibration anomaly is identified, and by using the weighted vibration anomaly evidence matrix, the weighted vibration anomaly evidence combination pattern of each heterogeneous storage module is determined. The weighted vibration anomaly evidence matrix and the historical vibration anomaly labeled sample data are used as training data to train a structural causal model. The structural causal model learns the correspondence between the evidence combination patterns of each block in the weighted vibration anomaly evidence matrix and the confirmed vibration anomaly root cause components in the historical labeled sample data, establishes a mapping function from weighted vibration anomaly evidence to a single vibration anomaly root cause component, and outputs the vibration anomaly root cause component of each heterogeneous storage block. Using the bearing number as the unit, integrate the root cause components of vibration anomalies and the weighted vibration anomaly evidence matrix of all heterogeneous storage modules in each bearing, and output the physical root cause location report of the vibration anomaly.

9. The bearing vibration anomaly data analysis method based on deep learning according to claim 8, characterized in that, The process of visualizing the output results of the above steps to obtain the bearing distributed data topology dashboard, the distributed bearing vibration anomaly thermal matrix, the consistency verification comparison dashboard, and the bearing vibration anomaly causal tracing orientation correlation map includes: The visualization output process of the bearing distributed data topology dashboard includes: referring to the bearing numbers, scanning all heterogeneous storage modules within the bearing, reading the data stored in each module and the total number of entries, displaying each heterogeneous storage module in the form of topology nodes, with the node size representing the total number of data entries stored in the corresponding heterogeneous storage module, embedding parameter cards inside each topology node, presenting the multi-channel original vibration data, operating data, parameter data, environmental verification data, historical vibration anomaly annotation sample data, synchronization timestamps, and unified index of each bearing, thus obtaining the bearing distributed data topology dashboard; The visualization output process of the distributed bearing vibration anomaly thermal matrix includes: building a two-dimensional matrix basic structure with rows representing each heterogeneous storage module and columns representing each bearing number; scheduling the vibration anomaly analysis report of each heterogeneous storage module corresponding to each matrix cell; classifying the vibration anomaly severity level of each heterogeneous storage module based on the vibration anomaly severity score of each heterogeneous storage module; matching thermal color schemes to the vibration anomaly severity levels of each heterogeneous storage module; color filling and data embedding processing of each matrix cell; unifying the layout style; completing the overall rendering of the matrix; and obtaining the distributed bearing vibration anomaly thermal matrix. The visualization output process of the consistency verification comparison dashboard includes: dividing the display area into independent display areas, placing the pre-classification results of the vibration anomaly data, the vibration anomaly analysis report of the heterogeneous storage module, and the historical vibration anomaly labeled sample data respectively; setting a scale and pointer for the heterogeneous storage module; mapping the vibration anomaly confidence consistency score corresponding to the vibration anomaly analysis report of each heterogeneous storage module to the scale range of the scale; marking specific values ​​with the pointer; setting a preset scoring threshold; highlighting heterogeneous storage modules that exceed the scoring threshold; summarizing the independent display areas, pointer scale, and highlighted markings; unifying the interface rendering; and generating the consistency verification comparison dashboard. The visualization output process of the bearing vibration anomaly causal directional correlation map includes: taking the bearing number as the central node, and outwardly associating the multi-level inducing factor nodes corresponding to the operating data, parameter data and environmental verification data of each bearing; using arrows of different thicknesses to represent the confidence consistency scores corresponding to each heterogeneous storage module; supporting clicking the central node to expand the corresponding vibration anomaly physical root cause location report, and obtaining the bearing vibration anomaly causal directional correlation map.

10. A deep learning-based bearing vibration anomaly data analysis system, used to implement the deep learning-based bearing vibration anomaly data analysis method according to any one of claims 1-9, comprising a bearing vibration data acquisition module, a preliminary vibration anomaly analysis module, a distributed storage module, a deep vibration anomaly analysis module, a consistency verification module, a root cause localization module, and a visualization module, wherein, The various modules are connected for communication, characterized in that, The bearing vibration data acquisition module assigns a number to each bearing to collect multi-channel raw vibration data, operating data, parameter data, environmental verification data, and historical vibration anomaly annotation sample data for each bearing. The preliminary vibration anomaly analysis module utilizes decoupled contrastive learning technology, with the running data and the environmental verification data as constraints, and uses the historical vibration anomaly labeled sample data for supervised fine-tuning, to perform adaptive vibration anomaly identification and pre-classification on the multi-channel raw vibration data, and outputs a deep feature vector and vibration anomaly data pre-classification results. The distributed storage module, based on the deep feature vector and the pre-classification results of the vibration anomaly data, performs distributed storage of the collected bearing data according to the semantic category division rules, generating heterogeneous storage modules; The deep vibration anomaly analysis module performs differentiated vibration anomaly analysis on the heterogeneous storage module through a lightweight deep decoupling network and outputs a vibration anomaly analysis report of the heterogeneous storage module. The consistency verification module performs consistency verification between the vibration anomaly analysis report of the heterogeneous storage module, the vibration anomaly data pre-classification result, and the historical vibration anomaly labeled sample data, and outputs the verification result. The root cause localization module, based on the verification results and the heterogeneous storage module, constructs a causal inference model for bearing vibration anomaly, tracks the root cause of vibration anomaly, and outputs a physical root cause localization report for vibration anomaly. The visualization module visualizes the output results of the steps to obtain a distributed data topology dashboard for bearings, a distributed thermal matrix of bearing vibration anomalies, a consistency verification and comparison dashboard, and a causal correlation map of bearing vibration anomalies.

Images