A fault pattern recognition method and system based on adaptive sliding window

Abstract

A fault mode recognition method and system based on an adaptive sliding window, according to different fault modes of mechanical equipment corresponding data anomaly characteristics, through Fourier transform, frequency domain information of a kind of mechanical equipment test data is obtained, mechanical equipment test data containing normal state and various fault labels is used for training, the characteristic frequency band of various faults is locked and the data characteristics of various faults are learned, so as to construct a mechanical equipment fault mode discrimination method, through the fault mode recognition system and method, the rapid discrimination of the fault mode of the equipment can be realized, and the characteristic frequency band information corresponding to various fault modes is provided for subsequent research.

Description

Technical Field

[0001] This invention relates to a fault mode recognition method and system based on an adaptive sliding window, belonging to the field of mechanical equipment fault mode discrimination. Background Technology

[0002] With the increasing sophistication of advanced sensor, data bus, and network communication technologies, the data elements available for acquiring the operating status of mechanical equipment are now quite comprehensive. High-frequency vibration data can be sampled at frequencies on the order of megahertz, giving the data a strong sensitivity to anomalies and providing a prerequisite for data-driven fault detection and diagnosis. Furthermore, the development of high-performance computing, machine learning, and artificial intelligence technologies is also stimulating in-depth research into data-driven fault detection and diagnosis techniques.

[0003] Compared to traditional methods based on sensor signals or physical models, data-driven fault detection and identification methods possess both the simplicity and ease of implementation of sensor signal-based methods and the in-depth analytical capabilities of physical model-based methods. Furthermore, they are more generalizable, capable of extracting information beyond the modeling experience of technical personnel. Currently, research on data-driven algorithms is receiving increasing attention, and related research and engineering products have been widely applied in various fields such as aviation, aerospace, automotive, and shipbuilding.

[0004] Currently, after the testing of this type of equipment is completed, technicians need to manually analyze the test data using commercial software (Origin, Matlab, Excel, etc.), and judge the characteristics of specific data (such as the harmonic region) based on experience to initially determine the equipment status. For equipment with abnormal status, further disassembly inspection is required to locate and identify the fault. This test analysis method has the following problems: 1. The equipment test data volume is large, and the data processing method of technicians involves a large number of repetitive operations, resulting in a low overall level of automation in data analysis; 2. Equipment startup inspection has high requirements for manpower and time costs, and the experience and knowledge of different technicians vary, making it impossible to guarantee a unified quantitative standard for fault identification. Summary of the Invention

[0005] The technical problem solved by this invention is to address the shortcomings of insufficient automation and excessive reliance on the experience of technicians in fault identification in existing technologies. This invention proposes a fault mode recognition method and system based on an adaptive sliding window.

[0006] The present invention solves the above-mentioned technical problem through the following technical solution:

[0007] A fault mode recognition method based on an adaptive sliding window includes:

[0008] Perform status tests on a specified number of devices, acquire test data for each device, determine the health status label of the devices based on the test results, disassemble and inspect devices in abnormal states, determine the fault mode label of the devices, and summarize and determine the device test parameters.

[0009] The total dataset consists of all test data obtained from the preset number of tests for each test device status test. Based on the test results of each test in the total dataset, health status labels including normal and abnormal status are provided. For specific abnormal test results, the device corresponding to this test is disassembled for inspection to identify the fault mode of this type of abnormality. The fault mode label is confirmed based on the fault mode. The total dataset, health status labels, and fault mode labels are summarized into device test parameters.

[0010] The equipment test parameters are screened, and the importance of the equipment test parameters is sorted by random forest method. The spectrum of the fault characterization parameters in the equipment test parameters is obtained by Fourier transform.

[0011] Select the spectrum corresponding to the fault characterization parameters of the pre-defined equipment involving known fault modes and the equipment with a health status label of healthy as the training set, construct the equipment fault mode recognition algorithm containing various types of fault modes and train the fault mode recognition model.

[0012] Select a test set to test the fault mode recognition capability of the trained equipment fault mode recognition algorithm.

[0013] The fault mode identification is performed on the test data of the equipment to be identified using the equipment fault identification algorithm that has passed the test.

[0014] Based on the characteristics of various fault characterization parameters, high-frequency vibration parameters are selected as fault characterization parameters in the equipment test parameters. A random forest model is constructed using the equipment's health status label and high-frequency vibration parameters to calculate the out-of-bag error rate of various fault characterization parameters. The fault characterization parameters of the current equipment are ranked according to their fault characterization capabilities based on the out-of-bag error rate. The top three ranked fault characterization parameters are selected to characterize the current equipment's working status and are used as input for subsequent algorithms.

[0015] The equipment fault mode identification algorithm is constructed based on an adaptive sliding window.

[0016] The fault characterization parameters of the devices containing the corresponding fault modes and several healthy devices are selected as the training set. Fault identification models are trained for different types of fault modes. The output of the fault identification model is a state value, which is used to characterize the existence of each fault mode.

[0017] In the aforementioned equipment fault mode recognition algorithm, different fault modes are identified through unique frequency domain features, specifically:

[0018] The fault characterization parameters of equipment with a specific fault mode in normal or abnormal state are used as input, and the state value representing whether the corresponding fault mode exists is used as output. By constructing a frequency domain sliding window to exhaustively search the frequency domain interval, cluster analysis is performed in the frequency domain interval corresponding to each fault mode to determine the frequency domain interval corresponding to the fault state and health state of each equipment after cluster analysis. The frequency domain interval corresponding to the fault mode of the current fault model is obtained according to the frequency domain window information, which serves as the basis for fault judgment in the fault mode recognition algorithm.

[0019] The specific steps for testing the fault identification model's fault mode recognition capability are as follows:

[0020] After training, the fault representation parameters of each test are fed into the trained fault identification model. The model outputs a representation that matches the state value of the corresponding fault mode. If the state value matches the fault mode label of the test, the model is considered to be accurate. Otherwise, the model is considered to be inaccurate. The accuracy of the test results is statistically analyzed to determine the accuracy of the model.

[0021] A fault mode recognition system based on an adaptive sliding window includes a data acquisition and classification module, a network model training and testing module, and a fault mode recognition module.

[0022] The data acquisition and classification module receives the equipment test parameter datasets from all devices, including health status labels and some fault mode labels for each device. High-frequency vibration parameters are selected from the dataset. The importance of each high-frequency vibration parameter for health status assessment is analyzed using the random forest method. Based on empirical analysis of the equipment's physical mechanisms and the contribution of the random forest model's permutation parameters to fault characterization, the range of equipment test fault state characterization parameters is determined. The spectrum of the fault state characterization parameters is calculated using Fourier transform. Based on the health status labels and fault mode labels of each device's test dataset, training and test sets are divided.

[0023] The network model training and testing module, based on the adaptive sliding window fault identification algorithm, constructs fault identification models and uses the training set processed by the data acquisition and classification module to train each fault identification model, with one fault mode corresponding to one fault identification model. At the same time, the test set is used to test the fault identification models, and the effectiveness of the fault identification models is judged based on the match between the test results and the equipment fault mode labels.

[0024] The fault mode recognition module extracts fault state characterization parameters from the test dataset of new equipment and calculates the parameter spectrum through Fourier transform. It then uses various fault recognition models that have passed the test to perform fault mode recognition on the test data of the equipment to be identified. If the model outputs a state value corresponding to the fault mode, it determines that the equipment has that type of fault mode.

[0025] The fault characterization parameters used to train the model are determined by the test datasets of each device, health status labels, and fault mode labels; the health status labels and fault mode labels of each device are determined by disassembly inspection after the device test; among them, the fault mode labels include cracks at different locations.

[0026] After analysis and processing by the data acquisition and classification module, the fault characterization parameters are determined;

[0027] An adaptive sliding window fault identification algorithm trains corresponding fault identification models for various fault modes, wherein:

[0028] The fault characterization parameters of the devices containing the corresponding fault modes and the fault characterization parameters of some healthy devices are selected as the training set. The fault identification model for each type of fault is trained respectively, and the output of the model is the state value of whether the corresponding fault mode exists.

[0029] In equipment fault mode recognition algorithms, different fault modes are identified through unique frequency domain features, specifically:

[0030] The fault characterization parameters of equipment with a specific fault mode in normal or abnormal state are used as input, and the state value representing whether the corresponding fault mode exists is used as output. By constructing a frequency domain sliding window to exhaustively search the frequency domain interval, cluster analysis is performed in the frequency domain interval corresponding to each fault mode to determine the frequency domain interval corresponding to the fault state and health state of each equipment after cluster analysis. The frequency domain interval corresponding to the fault mode of the current fault model is obtained according to the frequency domain window information, which serves as the basis for fault judgment in the fault mode recognition algorithm.

[0031] The specific steps for testing the fault identification model are as follows:

[0032] After training, the fault characterization parameters from several tests are fed into the fault identification model for each type of fault. The fault identification model outputs a state value indicating whether the characterization corresponds to the type of fault. If the state value matches the fault mode label of the test, the judgment is considered accurate; otherwise, the judgment is considered inaccurate. The accuracy of the test results is statistically analyzed to determine the accuracy of the model.

[0033] The advantages of this invention compared to the prior art are:

[0034] (1) The present invention provides a fault mode recognition method and system based on an adaptive sliding window, which can realize the rapid identification of fault modes of a certain type of mechanical equipment. The method learns the frequency domain information of the fault characterization parameters of known fault modes through training, and mines the linear or nonlinear relationship between data features and potential faults to realize data-driven mechanical equipment fault mode recognition. It solves the problem of insufficient automation in the current manual disassembly inspection and confirmation of fault modes. The method can be used as a general mechanical equipment fault mode recognition process. For other types of mechanical equipment, the corresponding data can be used to train the fault recognition model.

[0035] (2) The present invention proposes a fault mode identification method based on an adaptive sliding window for mechanical equipment. This method can output the parameter spectrum frequency band that represents the corresponding fault mode based on the identification of mechanical equipment fault modes, which helps to improve the technical personnel’s understanding of the data manifestation of specific fault modes and the depth of their understanding of the physical mechanism of fault modes. Attached Figure Description

[0036] Figure 1 A fault mode recognition framework diagram based on an adaptive sliding frequency domain window provided for the invention;

[0037] Figure 2 A schematic diagram of the 19 original signals and their frequency domain characteristics provided for the invention; Detailed Implementation

[0038] A fault mode recognition method and system based on an adaptive sliding window addresses the problem of equipment fault identification from a data-driven perspective. It outputs fault diagnosis results simply by inputting test data, thereby reducing costs and increasing efficiency in equipment test data analysis. Based on the typical anomaly characteristics of different fault modes in mechanical equipment, Fourier transform is used to obtain the frequency domain information of test data for a specific type of mechanical equipment. Training is performed using test data containing normal status and various fault labels to identify the parameter feature frequency bands corresponding to various faults and learn the data features of each fault type. This constructs a mechanical equipment fault mode recognition method. The fault mode recognition system and method enable rapid identification of fault modes for this type of equipment and provide the spectral frequency band information of fault characterization parameters corresponding to various fault modes for subsequent research.

[0039] Fault mode identification based on adaptive sliding window specifically includes:

[0040] Perform status tests on a specified number of devices, acquire test data for each device, determine the health status label of the devices based on the test results, disassemble and inspect devices in abnormal states, determine the fault mode label of the devices, and summarize and determine the device test parameters.

[0041] The total dataset consists of all test data obtained from the preset number of tests for each test device status test. Based on the test results of each test in the total dataset, health status labels including normal and abnormal status are provided. For specific abnormal test results, the device corresponding to this test is disassembled for inspection to identify the fault mode of this type of abnormality. The fault mode label is confirmed based on the fault mode. The total dataset, health status labels, and fault mode labels are summarized into device test parameters.

[0042] The equipment test parameters are screened, and the importance of the equipment test parameters is sorted by random forest method. The spectrum of the fault characterization parameters in the equipment test parameters is obtained by Fourier transform.

[0043] Select the spectrum corresponding to the fault characterization parameters of the pre-defined equipment involving known fault modes and the equipment with a health status label of healthy as the training set, construct the equipment fault mode recognition algorithm containing various types of fault modes and train the fault mode recognition model.

[0044] Select a test set to test the fault mode recognition capability of the trained equipment fault mode recognition algorithm.

[0045] The fault mode identification is performed on the test data of the equipment to be identified using the equipment fault identification algorithm that has passed the test.

[0046] Based on the characteristics of various fault characterization parameters, high-frequency vibration parameters are selected as fault characterization parameters in the equipment test parameters. A random forest model is constructed using the equipment's health status label and high-frequency vibration parameters to calculate the out-of-bag error rate of various fault characterization parameters. The fault characterization parameters of the current equipment are ranked according to their fault characterization capabilities based on the out-of-bag error rate. The top three ranked fault characterization parameters are selected to characterize the current equipment's working status and are used as input for subsequent algorithms.

[0047] The equipment fault mode identification algorithm is constructed based on an adaptive sliding window.

[0048] The fault characterization parameters of the devices containing the corresponding fault modes and several healthy devices are selected as the training set. Fault identification models are trained for different types of fault modes. The output of the fault identification model is a state value, which is used to characterize the existence of each fault mode.

[0049] In equipment fault mode recognition algorithms, different fault modes are identified through unique frequency domain features, specifically:

[0050] The fault characterization parameters of equipment with a specific fault mode in normal or abnormal state are used as input, and the state value representing whether the corresponding fault mode exists is used as output. By constructing a frequency domain sliding window to exhaustively search the frequency domain interval, cluster analysis is performed in the frequency domain interval corresponding to each fault mode to determine the frequency domain interval corresponding to the fault state and health state of each equipment after cluster analysis. The frequency domain interval corresponding to the fault mode of the current fault model is obtained according to the frequency domain window information, which serves as the basis for fault judgment in the fault mode recognition algorithm.

[0051] The specific steps for testing the fault identification model's fault mode recognition capability are as follows:

[0052] After training, the fault representation parameters of each test are fed into the trained fault identification model. The model outputs a representation that matches the state value of the corresponding fault mode. If the state value matches the fault mode label of the test, the model is considered to be accurate. Otherwise, the model is considered to be inaccurate. The accuracy of the test results is statistically analyzed to determine the accuracy of the model.

[0053] The fault mode recognition system based on adaptive sliding window includes a data acquisition and classification module, a network model training and testing module, and a fault mode recognition module.

[0054] The data acquisition and classification module connects multiple equipment test datasets with health status labels and some fault mode labels, and selects high-frequency vibration parameters; it analyzes the importance of each high-frequency vibration parameter for health status judgment using the random forest method, narrowing down the range of fault status characterization parameters for equipment testing; it calculates the spectrum of fault status characterization parameters through Fourier transform; and it divides the training set and test set according to the health status labels and fault mode labels of each equipment test dataset.

[0055] The network model training and testing module, based on the adaptive sliding window fault identification algorithm, constructs a fault identification model. It trains the model using the training set processed by the data acquisition and classification module, training a corresponding fault identification model for each fault mode. At the same time, it tests the fault identification model using the test set, and judges the effectiveness of the fault identification model based on the match between the test results and the equipment fault mode labels.

[0056] The fault mode recognition module extracts fault state characterization parameters from the test dataset of new equipment and calculates the parameter spectrum through Fourier transform. It uses various fault recognition models that have passed the test to identify fault modes in the test data input by the equipment. If the model outputs a state value corresponding to the fault mode, it determines that the equipment has that type of fault mode.

[0057] The fault characterization parameters used to train the model are determined by the test datasets of each device, health status labels, and fault mode labels; the range of fault characterization parameters is determined by empirical analysis of the physical mechanism of the device and the contribution of the random forest model permutation parameters to the fault characterization; the health status labels and fault mode labels of each device are determined by disassembly inspection after the device is tested.

[0058] After analysis and processing by the data acquisition and classification module, the fault characterization parameters are determined;

[0059] The adaptive sliding window fault identification algorithm requires training a corresponding fault identification model for each type of fault mode, therefore it is necessary to determine the training set of the fault identification model for each type of fault mode.

[0060] Select the fault characterization parameters of the device containing the corresponding fault mode and the fault characterization parameters of several healthy devices as the training set, and train the fault identification model for each type of fault. The output of the model is the state value of whether the corresponding fault mode exists.

[0061] The adaptive sliding window fault mode recognition algorithm is based on the following assumptions: different fault modes can be reflected in the frequency domain features of vibration data, and the features of different fault modes should have unique frequency domain intervals.

[0062] The algorithm takes the fault characterization parameters of multiple normal / faulty devices with specific fault modes as input and the state value representing whether the fault exists as output. It constructs a frequency domain sliding window to exhaustively search the frequency domain intervals, performs cluster analysis in each frequency domain interval, finds each frequency domain interval that can distinguish the fault / health status of the group of devices through clustering, and integrates the frequency domain window information and trains the corresponding fault mode recognition algorithm for each type of fault mode.

[0063] After training, the fault characterization parameters from several tests are fed into the fault identification model for each type of fault. The fault identification model outputs a state value indicating whether the characterization corresponds to the type of fault. If the state value matches the fault mode label of the test, the judgment is considered accurate; otherwise, the judgment is considered inaccurate. The accuracy of the test results is statistically analyzed to determine the accuracy of the model.

[0064] Specifically, when constructing the adaptive sliding window fault mode recognition algorithm, we believed that the relationship between equipment faults and equipment test data features should meet the following conditions: high-frequency vibration parameters have a fast state response and contain a lot of information; therefore, the spectral information of fault state characterization parameters strongly correlated with equipment fault states should include features corresponding to various fault modes. Different fault modes correspond to different crack modes or crack locations; therefore, their related features should be reflected in different frequency bands in the frequency domain of fault characterization parameters. Thus, identifying fault modes means distinguishing the frequency bands of the fault characterization parameters corresponding to each type of fault mode. Based on the above ideas, a sliding window is constructed to determine whether there are abnormal data features related to specific fault modes in each frequency band. The fault frequency band and abnormal feature form of this type of fault mode are learned. Different classification models are trained for different fault types, and a model for identifying this type of fault mode is formed through cross-device joint training and validation, thereby realizing the identification of the corresponding fault mode.

[0065] Since this algorithm analyzes the spectral characteristics of equipment fault characterization parameters, we need to perform a Fourier transform on the original signal.

[0066]

[0067] The frequency domain signal is obtained.

[0068] To capture the representative frequency bands of specific fault modes in the parametric spectrum, a frequency domain sliding window with a width of 625Hz is constructed to scan the entire frequency band (0-6.25kHz). For each frequency domain window, clustering is used to attempt to distinguish the health labels of each device in each training sample. If the frequency band can correctly distinguish the health labels of each device in the training samples, then the frequency band is considered to be related to the fault mode, and the identification of the fault mode is achieved by recognizing the features of the frequency band.

[0069] The specific training steps are as follows:

[0070] 1. Input multiple test data sets containing known pattern failures and healthy states; randomly select two cluster centers from the test data and begin iteratively calculating the clustering results. The calculation process minimizes the objective function J. m :

[0071]

[0072] Where m is the membership factor, u ij Indicates sample x i Membership degree of class j, x i Let x represent the j-th sample, and c be a sample with d-dimensional features. j It is the center of the j-cluster with d-dimensional features, where ||*|| can be any metric representing distance.

[0073] The membership degree u is calculated iteratively. ij and cluster center c j The process continues until the objective function reaches its optimum.

[0074]

[0075] when When k is the number of iterations and ε is the error threshold, the iteration terminates.

[0076] By comparing the values ​​of the membership functions assigned to each sample for each cluster, the samples are classified to obtain the clustering results;

[0077] If the clustering results match the fault / health labels of the sample group, it is considered that the frequency band and the corresponding clustering model can support fault identification and then select the next frequency band; otherwise, the next frequency domain interval is selected directly.

[0078] By using a sliding window scanning method, the frequency domain parameters are exhaustively enumerated across the entire frequency domain to obtain all frequency bands that can characterize the fault mode and the corresponding classification model.

[0079] After training is completed, inputting the test data of the new transmission will allow us to directly determine whether there is a fault in the data of that transmission in each corresponding frequency band. If there is, the result will be confirmed by majority voting. If the majority of the results indicate a fault, then the device is determined to have the fault mode corresponding to the algorithm.

[0080] The following description, in conjunction with the accompanying drawings and preferred embodiments, provides further details: In the current embodiment, for a certain type of mechanical equipment, the adaptive sliding frequency domain window fault mode recognition algorithm constructs a model using three vibration parameters.

[0081] The result of one of the vibration parameters obtained by Fast Fourier Transform is shown below. The left side is the original vibration signal, and the right side is the frequency domain feature extracted from the smoothed vibration signal. Figure 2 As shown, taking device number 19 as an example:

[0082] A fault mode recognition algorithm based on an adaptive sliding window is constructed, with the sliding window width set to 625Hz and the full frequency band width set to 0-6.25kHz. A fault mode recognition model is trained for a specific crack fault mode.

[0083] The test results of the model are shown in the table below.

[0084] Table 1 Fault Mode Identification Results

[0085] 19-1 normal normal 19-2 normal normal 19-3 normal normal 19-4 normal normal 19-5 Unable to determine normal 19-6 Specific faults Specific fault exists

[0086] The test results indicate that the model can correctly identify specific failure modes.

[0087] In summary, this invention is highly applicable to fault diagnosis and fault classification of mechanical equipment, and has positive significance for the modernization of the national defense industry.

[0088] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications to the technical solutions of the present invention by utilizing the methods and techniques disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention shall fall within the protection scope of the technical solutions of the present invention.

[0089] The contents not described in detail in this specification are common knowledge to those skilled in the art.

Claims

1. A fault mode recognition method based on an adaptive sliding window, characterized in that... include: Perform status tests on a specified number of devices, acquire test data for each device, determine the health status label of the devices based on the test results, disassemble and inspect devices in abnormal states, determine the fault mode label of the devices, and summarize and determine the device test parameters; the device test parameters include the total dataset, health status label, and fault mode label, and the total dataset is all the test data of each test device for a preset number of times; The equipment test parameters are screened, and the importance of the equipment test parameters is sorted by random forest method. The spectrum of the fault characterization parameters in the equipment test parameters is obtained by Fourier transform. Select the spectrum corresponding to the fault characterization parameters of the pre-defined equipment involving known fault modes and the equipment with a health status label of healthy as the training set, construct the equipment fault mode recognition algorithm containing various types of fault modes and train the fault mode recognition model. Select a test set to test the fault mode recognition capability of the trained equipment fault mode recognition algorithm. Fault mode identification is performed on the test data of the equipment to be identified using a tested equipment fault identification algorithm. Based on the characteristics of various fault characterization parameters, high-frequency vibration parameters are selected as fault characterization parameters in the equipment test parameters. A random forest model is constructed using the equipment's health status labels and high-frequency vibration parameters. The out-of-bag error rate (OOB) of various fault characterization parameters is calculated. Based on the OOB, the fault characterization parameters of the current equipment are ranked according to their fault characterization capabilities. The top three ranked fault characterization parameters are selected to characterize the current equipment's operating status and are used as input for subsequent algorithms. The equipment fault mode identification algorithm is constructed based on an adaptive sliding window. The training set is selected from the fault characterization parameters of the corresponding fault modes and several healthy devices. Fault identification models are trained for different types of fault modes. The output of the fault identification model is a state value, which represents the existence of each fault mode. In the aforementioned equipment fault mode recognition algorithm, different fault modes are identified through unique frequency domain features, specifically: The fault characterization parameters of devices in normal or abnormal states with specific fault modes are used as input, and the state value representing whether the corresponding fault mode state exists is used as output. By constructing a frequency domain sliding window to exhaustively search the frequency domain intervals, cluster analysis is performed in the frequency domain intervals corresponding to each fault mode to determine the frequency domain intervals corresponding to the fault state and healthy state of each device after cluster analysis. Based on the frequency domain window information, the frequency domain interval of the fault mode corresponding to the current fault model is obtained, which serves as the basis for fault judgment in the fault mode recognition algorithm. The specific steps for testing the fault identification model's fault mode recognition capability are as follows: After training, the fault representation parameters of each test are fed into the trained fault identification model. The model outputs a representation that matches the state value of the corresponding fault mode. If the state value matches the fault mode label of the test, the model is considered to be accurate. Otherwise, the model is considered to be inaccurate. The accuracy of the test results is statistically analyzed to determine the accuracy of the model. Among them, identifying fault modes involves distinguishing the frequency bands of the fault characterization parameters corresponding to each type of fault mode. The method is as follows: A sliding window is constructed to determine whether there are abnormal data features related to specific fault modes in each frequency band. The fault frequency band and abnormal feature form of such fault modes are learned. Different classification models are trained for different fault types. Through cross-device joint training and verification, a model for recognizing such fault modes is formed, and the corresponding fault mode is recognized. Perform a Fourier transform on the original signal: Obtain the frequency domain signal; To capture the frequency bands representing specific fault modes in the parametric spectrum, a frequency domain sliding window with a width of 625Hz is constructed to scan the entire frequency band, including 0-6.25kHz. For each frequency domain window, clustering is used to attempt to distinguish the health labels of each device in each training sample. If the frequency band can correctly distinguish the health labels of each device in the training sample, then the frequency band is considered to be related to the fault mode, and the identification of the frequency band features is used to identify the fault mode. The specific training steps are as follows: The input consists of multiple test data sets containing known fault modes and health states. Two cluster centers are randomly selected from the test data, and the clustering results are iteratively calculated by minimizing the objective function. : in It is the membership factor. Indicates sample belong Membership degree of a class Indicates the first One sample, It has Samples with 3D features It has dimensional features Cluster center, ||*|| is any metric representing distance; Iterative calculation of membership degree and cluster center The process continues until the objective function reaches its optimum. when When the iteration terminates, where It is the number of iterations. It is the error threshold; By comparing the values ​​of the membership functions assigned to each sample for each cluster, the samples are classified to obtain the clustering results; If the clustering results match the fault / health labels of the sample group, it is considered that the frequency band and the corresponding clustering model can support fault identification and then select the next frequency band; otherwise, the next frequency domain interval is selected directly. By using a sliding window scanning method, the frequency domain parameters are exhaustively enumerated across the entire frequency domain to obtain all frequency bands that can characterize the fault mode and the corresponding classification model. After training is completed, inputting the test data of the new transmission will allow us to directly determine whether there is a fault in the data of that transmission in each corresponding frequency band. If there is, the result will be confirmed by majority voting. If the majority of the results indicate a fault, then the device is determined to have the fault mode corresponding to the algorithm.

2. A fault mode recognition system implementing the fault mode recognition method of claim 1, characterized in that: It includes a data acquisition and classification module, a network model training and testing module, and a fault mode recognition module; The data acquisition and classification module receives the equipment test parameter datasets from all devices, including health status labels and some fault mode labels for each device. High-frequency vibration parameters are selected from the dataset. The importance of each high-frequency vibration parameter for health status assessment is analyzed using the random forest method. Based on empirical analysis of the equipment's physical mechanisms and the contribution of the random forest model's permutation parameters to fault characterization, the range of equipment test fault state characterization parameters is determined. The spectrum of the fault state characterization parameters is calculated using Fourier transform. Based on the health status labels and fault mode labels of each device's test dataset, training and test sets are divided. The network model training and testing module, based on the adaptive sliding window fault identification algorithm, constructs fault identification models and uses the training set processed by the data acquisition and classification module to train each fault identification model, with one fault mode corresponding to one fault identification model. At the same time, the test set is used to test the fault identification models, and the effectiveness of the fault identification models is judged based on the match between the test results and the equipment fault mode labels. The fault mode recognition module extracts fault state characterization parameters from the test dataset of new equipment and calculates the parameter spectrum through Fourier transform. It then uses various fault recognition models that have passed the test to perform fault mode recognition on the test data of the equipment to be identified. If the model outputs a state value corresponding to the fault mode, it determines that the equipment has that type of fault mode.

3. The fault mode recognition system based on an adaptive sliding window according to claim 2, characterized in that: The fault characterization parameters used to train the model are determined by the test datasets of each device, health status labels, and fault mode labels; the health status labels and fault mode labels of each device are determined by disassembly inspection after the device test; among them, the fault mode labels include cracks at different locations.

4. The fault mode recognition system based on an adaptive sliding window according to claim 3, characterized in that: After analysis and processing by the data acquisition and classification module, the fault characterization parameters are determined; An adaptive sliding window fault identification algorithm trains corresponding fault identification models for various fault modes, wherein: The fault characterization parameters of the devices containing the corresponding fault modes and the fault characterization parameters of some healthy devices are selected as the training set. The fault identification model for each type of fault is trained respectively, and the output of the model is the state value of whether the corresponding fault mode exists.

5. The fault mode recognition system based on an adaptive sliding window according to claim 4, characterized in that: In equipment fault mode recognition algorithms, different fault modes are identified through unique frequency domain features, specifically: The fault characterization parameters of equipment with a specific fault mode in normal or abnormal state are used as input, and the state value representing whether the corresponding fault mode exists is used as output. By constructing a frequency domain sliding window to exhaustively search the frequency domain interval, cluster analysis is performed in the frequency domain interval corresponding to each fault mode to determine the frequency domain interval corresponding to the fault state and health state of each equipment after cluster analysis. The frequency domain interval corresponding to the fault mode of the current fault model is obtained according to the frequency domain window information, which serves as the basis for fault judgment in the fault mode recognition algorithm.

6. The fault mode recognition system based on an adaptive sliding window according to claim 5, characterized in that: The specific steps for testing the fault identification model are as follows: After training, the fault characterization parameters from several tests are fed into the fault identification model for each type of fault. The fault identification model outputs a state value indicating whether the characterization corresponds to the type of fault. If the state value matches the fault mode label of the test, the judgment is considered accurate; otherwise, the judgment is considered inaccurate. The accuracy of the test results is statistically analyzed to determine the accuracy of the model.

Images