Bearing fault classification method and system based on intelligent inspection robot

Abstract

The application discloses a bearing fault classification method and system based on an intelligent inspection robot, and belongs to the technical field of sensing technology and digital signal processing technology. The technical problem to be solved by the application is how to use an intelligent inspection robot to classify bearing faults according to sound, and then accurately know the bearing fault position. The technical scheme adopted is as follows: S1, collecting acoustic signals: collecting the fault sound of a bearing inner ring raceway, a bearing outer ring raceway, a bearing rolling element and a bearing assembly based on an intelligent inspection robot's collection card and microphone, and simultaneously collecting the sound information of a normal bearing working; S2, constructing a time-frequency signal: constructing a time-frequency signal to express a data-increased signal feature; S3, feature extraction: extracting features through a self-encoding network; S4, feature classification: classifying features through a convolutional neural network; and S5, optimizing classification: training and testing by using an Adam optimization algorithm, continuously adjusting the parameter settings in the network, and then obtaining more accurate classification results.

Description

Technical Field

[0001] This invention relates to the fields of sensing technology and digital signal processing technology, specifically to a bearing fault classification method and system based on an intelligent inspection robot. Background Technology

[0002] In mechanical fault diagnosis research, vibration signals are typically the primary focus. The connection between the fault source and the vibrations of bearings or motors is used to achieve fault analysis without disassembling the machine. However, in practice, space must be reserved for accelerometers and other sensors, and their use must be strictly regulated; otherwise, the sensors may fly out during bearing operation, posing a danger.

[0003] Analysis reveals that sound is essentially a mechanical wave formed by the vibration of an object propagating outward through the air. The energy generated by the vibration of the sound source propagates outward through the nearby elastic medium, thus forming a sound wave. Therefore, acoustic signals and vibration signals are essentially the same, differing only in their form. Sound signals are collected through microphones. While sound signals undergo superposition and reflection during propagation in the air, this often enhances certain characteristics of the signal, which can be beneficial. Most importantly, acquiring acoustic signals is simpler in practical applications.

[0004] Therefore, how to use intelligent inspection robots to classify bearing faults based on sound and then accurately identify the location of the bearing fault is a technical problem that urgently needs to be solved. Summary of the Invention

[0005] The technical objective of this invention is to provide a bearing fault classification method and system based on an intelligent inspection robot, in order to solve the problem of how to use an intelligent inspection robot to classify bearing faults based on sound, and thus accurately identify the location of the bearing fault.

[0006] The technical objective of this invention is achieved as follows: a bearing fault classification method based on an intelligent inspection robot, the method being as follows:

[0007] S1. Acoustic signal acquisition: Based on the acquisition card and microphone of the intelligent inspection robot, fault sounds of the bearing inner raceway, bearing outer raceway, bearing rolling elements and bearing assembly are acquired, while sound information of the bearing during normal operation is also acquired.

[0008] S2. Constructing Time-Frequency Signals: Constructing time-frequency signals to enhance data dimensionality and express signal features;

[0009] S3. Feature extraction: Feature extraction is performed using an autoencoder network;

[0010] S4. Feature Classification: Feature classification is performed using a convolutional neural network;

[0011] S5. Optimize Classification: Use the Adam optimization algorithm for training and testing, and continuously adjust the parameter settings in the network to obtain more accurate classification results.

[0012] As a preferred embodiment, the acoustic signal acquisition in step S1 is specifically as follows:

[0013] S101. Connect or discover the acquisition device, and the acquisition signal source is a bearing, including bearing inner ring raceway fault signal, bearing outer ring raceway fault signal, bearing rolling element fault signal and bearing assembly fault signal.

[0014] S102. Obtain device status information: Use a GET request to obtain device information, distinguish devices based on IP and MAC address (WebDAQ 504 error), and select the host address as manual input; to facilitate data parsing, the data is converted to .json format;

[0015] S103. Preliminary assessment of equipment availability:

[0016] ① If so, proceed to step S104;

[0017] ② If not, proceed to step S101;

[0018] S104. In order to record the data source and obtain complete system details;

[0019] S105. Configure the audio acquisition task using the acquisition card, including the sampling rate, number of channels, and the sound pressure level settings for the microphone, and then publish the acquisition task.

[0020] S106. Obtain the device working list status;

[0021] S107. Determine if any tasks are in progress or paused:

[0022] ① If so, stop the existing task or wait in the queue, and proceed to step S108.

[0023] ② If not, proceed to step S108;

[0024] S108. Obtain the equipment operating status and execute tasks;

[0025] S109. Convert and output the relevant sound data of the corresponding acquisition device.

[0026] As a preferred approach, while obtaining the device's work list status, it is also necessary to determine whether the current task list directory is valid.

[0027] If valid, print the list of tasks. The list of tasks includes the task name, number of channels, and task progress. Alternatively, you can select an existing task to view.

[0028] As a preferred option, the data acquisition card uses the WebDAQ 504 IoT sound and vibration data logger;

[0029] The microphone uses two acoustic sensors, SKC AD7199 and SKC AD7018, for dual-channel acquisition.

[0030] As a preferred embodiment, the construction of the time-frequency signal in step S2 is as follows:

[0031] Spatiotemporal frequency domain transformation was employed, and fast Fourier transform was used to construct multi-scale frequency domain information;

[0032] The time-series signal is expanded in two dimensions using a windowing method at the separated frequency domain scale, forming a two-dimensional signal representation under multiple scales. The separated time-series signal is then resampled to obtain a time series, which is then spatially arranged to form a preprocessed signal, as shown in the following formula:

[0033]

[0034] Where L(i) represents the length of the original signal; P(j,k) represents the representation of a single pixel in the image; i = 1,…,M2; j = 1,…..,M; k = 1,……,M;

[0035] The purpose of rearranging the data in two dimensions and rounding down is to bring the data between 0 and 255. The one-dimensional vibration data is then normalized into an image. In fact, the data has been transformed from the original 64 bits to the current 8 bits, which greatly reduces the data storage space of a single point. The method of windowing and resampling the data actually increases the amount of effective analysis data. In this way, the data can be effectively analyzed and the correlation between data can be extracted without occupying too much data storage space.

[0036] As a preferred embodiment, the feature extraction using an autoencoder network in step S3 is as follows:

[0037] After dimensionality augmentation, data denoising and dimensionality reduction are performed using an autoencoder approach: The Tensorflow end-to-end open-source machine learning platform is used, and the Keras deep learning framework is used for data denoising and dimensionality reduction.

[0038] In the autoencoder, the input feature dimension is defined as 2400×1, and the number of channels is 1. Two-dimensional convolution is used to construct 10 filters, and a 128×1 sliding window is used. In order to retain as much data information as possible, the filtering mode is set to "seam", and the activation function is the double tangent function tanh.

[0039] Downsampling is performed using more max pooling, with a sampling factor of 8×1 and a stride of 2 to obtain a feature map of 1200×1×10.

[0040] Upsampling the 1200×1×10 feature map, although it appears to restore the data dimension to the original 2400×1×10, is actually an interpolation operation performed without introducing other parameters to enlarge the feature map and refine the feature information. This involves adding encoding and decoding operations, followed by a standard autoencoder operation. The autoencoder operation is characterized by having the same number of features in both the input and output. It is defined as a reconstruction of the input data after compression and extraction using a neural network. Through multiple stacked feature extraction layers, the convolutional layers and subsampling layers continuously shrink the feature map while the number of features continuously increases.

[0041] At the end of the feature extractor, all feature maps are expanded and arranged into a vector called the feature vector, which serves as the input to the subsequent classifier.

[0042] The signals are fed into a neural network classifier for training to classify bearing fault sound signals.

[0043] More preferably, the feature classification using a convolutional neural network in step S4 is as follows:

[0044] The learning process of a convolutional neural network aims to obtain the minimum error: the actual output label y. ik and predicted label d ik The difference between them should be minimized as much as possible, which depends on the continuous updating of weights ω and bias b during training; among them, the loss function adopts mean squared error, which reflects the closeness between the probability distribution of various classifications of softmax and the actual classification. If the difference between the model prediction result and the actual classification is small, the calculated mean squared error will be small.

[0045] The trained network is used to classify and test the resulting column vectors, proceeding step by step, with the number of classes successively set to 400, 300, and so on.

[0046] Based on the type of data collected, it is ultimately divided into five categories, corresponding to five different states of the collected signals;

[0047] The evaluation metrics for classification include accuracy, precision, and recall; a parameter F1 is also introduced to evaluate the model's performance, as shown in the following formula:

[0048] F1=2*(Precision*Recall) / (Precision+Recall).

[0049] A bearing fault classification system based on an intelligent inspection robot, the system comprising,

[0050] The signal acquisition module is used to collect fault sounds from the bearing inner raceway, bearing outer raceway, bearing rolling elements, and bearing assembly based on the acquisition card and microphone of the intelligent inspection robot, while also collecting sound information when the bearing is working normally.

[0051] The time-frequency signal construction module is used to construct time-frequency signals for data augmentation and signal feature representation.

[0052] The feature extraction module is used to extract features through an autoencoder network;

[0053] The feature classification module is used to classify features using a convolutional neural network.

[0054] The classification optimization module is used to train and test the Adam optimization algorithm, continuously adjusting the parameter settings in the network to obtain more accurate classification results.

[0055] An electronic device includes: a memory and at least one processor;

[0056] The memory contains computer programs;

[0057] The at least one processor executes the computer program stored in the memory, causing the at least one processor to perform the bearing fault classification method based on an intelligent inspection robot as described in any one of claims 1 to 7.

[0058] A computer-readable storage medium storing a computer program that can be executed by a processor to implement the bearing fault classification method based on an intelligent inspection robot as described above.

[0059] The bearing fault classification method and system based on intelligent inspection robots of the present invention have the following advantages:

[0060] (I) This invention classifies rolling bearings under the same rotational speed to identify five different types of faults. The process involves constructing time-frequency signals to enhance signal feature representation, designing an autoencoder network to extract features, and finally using a convolutional algorithm to classify features. The Adam optimization algorithm is used for training and testing. Throughout the process, the parameters in the network are continuously adjusted to achieve a more accurate classification result.

[0061] (II) Based on the current application of intelligent inspection robots, this invention uses a data acquisition card to design a sound acquisition program to complete the acquisition of sound signals, performs feature extraction by designing an autoencoder network, and adopts a signal classification method under a convolutional neural network to realize the status diagnosis of the equipment based on the acquired sound data. Attached Figure Description

[0062] The invention will be further described below with reference to the accompanying drawings.

[0063] Appendix Figure 1 The flowchart is shown below for a bearing fault classification method based on an intelligent inspection robot.

[0064] Appendix Figure 2 A flowchart for acquiring acoustic signals;

[0065] Appendix Figure 3 Waveform diagram of rolling element fault signal;

[0066] Appendix Figure 4 The waveform diagram of the fault signal of the assembly;

[0067] Appendix Figure 5 Waveform diagram of inner raceway fault signal;

[0068] Appendix Figure 6 The waveform diagram of the bearing under normal conditions;

[0069] Appendix Figure 7 This is a waveform diagram of a fault signal on the outer raceway. Detailed Implementation

[0070] The bearing fault classification method and system based on intelligent inspection robot of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0071] Example 1:

[0072] As attached Figure 1 As shown in the figure, this embodiment provides a bearing fault classification method based on an intelligent inspection robot. The method is as follows:

[0073] S1. Acoustic Signal Acquisition: The intelligent inspection robot's acquisition card and microphone collect fault sounds from the bearing's inner raceway, outer raceway, rolling elements, and bearing assembly. Simultaneously, it collects sound information from normal bearing operation. Using Matlab, the signal amplitude is represented by selecting 12,000 data points. Clearly, the differences in frequency and amplitude are most pronounced in the five cases, demonstrating the feasibility of using acoustic signals for bearing fault classification. Waveform diagrams are attached. Figure 3-7As shown; data acquisition mainly involves obtaining device information and system operating status to facilitate data recording and traceability. The specific acquisition implementation is based on the documentation provided by the acquisition device, in order to achieve autonomous acquisition tasks. This is because the operation of the acquisition card is already quite comprehensive.

[0074] S2. Constructing Time-Frequency Signals: Constructing time-frequency signals to enhance data dimensionality and express signal features;

[0075] S3. Feature extraction: Feature extraction is performed using an autoencoder network;

[0076] S4. Feature Classification: Feature classification is performed using a convolutional neural network;

[0077] S5. Optimize Classification: Use the Adam optimization algorithm for training and testing, and continuously adjust the parameter settings in the network to obtain more accurate classification results.

[0078] As attached Figure 2 As shown, the acquisition of acoustic signals in step S1 of this embodiment is specifically as follows:

[0079] S101. Connect or discover the acquisition device, and the acquisition signal source is a bearing, including bearing inner ring raceway fault signal, bearing outer ring raceway fault signal, bearing rolling element fault signal and bearing assembly fault signal.

[0080] S102. Obtain device status information: Use a GET request to obtain device information, distinguish devices based on IP and MAC address (WebDAQ 504 error), and select the host address as manual input; to facilitate data parsing, the data is converted to .json format;

[0081] S103. Preliminary assessment of equipment availability:

[0082] ① If so, proceed to step S104;

[0083] ② If not, proceed to step S101;

[0084] S104. In order to record the data source and obtain complete system details;

[0085] S105. Configure the audio acquisition task using the acquisition card, including the sampling rate, number of channels, and the sound pressure level settings for the microphone, and then publish the acquisition task.

[0086] S106. Obtain the device working list status;

[0087] S107. Determine if any tasks are in progress or paused:

[0088] ① If so, stop the existing task or wait in the queue, and proceed to step S108.

[0089] ② If not, proceed to step S108;

[0090] S108. Obtain the equipment operating status and execute tasks;

[0091] S109. Convert and output the relevant sound data of the corresponding acquisition device.

[0092] In step S106 of this embodiment, while obtaining the device work list status, it is determined whether the current task list directory is valid:

[0093] If valid, print the list of tasks. The list of tasks includes the task name, number of channels, and task progress. Alternatively, you can select an existing task to view.

[0094] In this embodiment, the acquisition card is the WebDAQ504 IoT sound and vibration data recorder developed by MCC.

[0095] To achieve the highest possible complexity of the audio signal, the microphone in this embodiment uses two acoustic sensors, SKC AD7199 and SKC AD7018, for dual-channel acquisition.

[0096] The construction of the time-frequency signal in step S2 of this embodiment is as follows:

[0097] The effective information in the raw signals measured by ordinary sensors cannot be deeply explored. Data that has been preprocessed can often highlight the features expressed by the original data.

[0098] In one-dimensional time-domain signals, preprocessing through resampling augments the data and uncovers the signal's implicit features. The preprocessed signal often highlights the features expressed by the original data. Extending traditional time-domain varying signals to the two-dimensional spatial domain through windowing and resampling not only enriches the signal feature representation but also allows for analysis using existing network weights. Essentially, the correlation and variation indices between sampled data points are key metrics in network classification.

[0099] To balance the computational complexity of edge devices with the sufficiency of data preprocessing, a spatiotemporal frequency domain transformation was adopted, and fast Fourier transform was used to construct multi-scale frequency domain information.

[0100] The time-series signal is expanded in two dimensions using a windowing method at the separated frequency domain scale, forming a two-dimensional signal representation under multiple scales. The separated time-series signal is then resampled to obtain a time series, which is then spatially arranged to form a preprocessed signal, as shown in the following formula:

[0101]

[0102] Where L(i) represents the length of the original signal; P(j,k) represents the representation of a single pixel in the image; i = 1,…,M2; j = 1,…..,M; k = 1,……,M;

[0103] The purpose of rearranging the data in two dimensions and rounding down is to bring the data between 0 and 255. The one-dimensional vibration data is then normalized into an image. In fact, the data has been transformed from the original 64 bits to the current 8 bits, which greatly reduces the data storage space of a single point. The method of windowing and resampling the data actually increases the amount of effective analysis data. In this way, the data can be effectively analyzed and the correlation between data can be extracted without occupying too much data storage space.

[0104] This step essentially applies the algorithmic thinking of image extraction, without needing to perform separate noise reduction, and directly uses the acquired original signal to enhance its features by adding dimensionality.

[0105] In this embodiment, the feature extraction via an autoencoder network in step S3 is as follows:

[0106] After dimensionality augmentation, data denoising and dimensionality reduction are performed using an autoencoder approach: The Tensorflow end-to-end open-source machine learning platform is used, and the Keras deep learning framework is used for data denoising and dimensionality reduction.

[0107] In the autoencoder, the input feature dimension is defined as 2400×1, and the number of channels is 1. Two-dimensional convolution is used to construct 10 filters, and a 128×1 sliding window is used. In order to retain as much data information as possible, the filtering mode is set to "seam", and the activation function is the double tangent function tanh.

[0108] Downsampling is performed using more max pooling, with a sampling factor of 8×1 and a stride of 2 to obtain a feature map of 1200×1×10.

[0109] Upsampling the 1200×1×10 feature map, although it appears to restore the data dimension to the original 2400×1×10, is actually an interpolation operation performed without introducing other parameters to enlarge the feature map and refine the feature information. This involves adding encoding and decoding operations, followed by a standard autoencoder operation. The autoencoder operation is characterized by having the same number of features in both the input and output. It is defined as a reconstruction of the input data after compression and extraction using a neural network. Through multiple stacked feature extraction layers, the convolutional layers and subsampling layers continuously shrink the feature map while the number of features continuously increases.

[0110] At the end of the feature extractor, all feature maps are expanded and arranged into a vector called the feature vector, which serves as the input to the subsequent classifier.

[0111] The signals are fed into a neural network classifier for training to classify bearing fault sound signals.

[0112] In this embodiment, step S4, feature classification using a convolutional neural network, is specifically as follows:

[0113] In classification, the Adam optimization algorithm is used. Simply put, Adam uses momentum and an adaptive learning rate to accelerate convergence. It is suitable for non-fixed objectives and is applicable to problems with very noisy or sparse gradients, and can replace the first-order optimization algorithm of the traditional stochastic gradient descent process.

[0114] The learning process of a convolutional neural network aims to obtain the minimum error: the actual output label y. ik and predicted label d ik The difference between them should be minimized as much as possible, which depends on the continuous updating of weights ω and bias b during training; among them, the loss function adopts mean squared error, which reflects the closeness between the probability distribution of various classifications of softmax and the actual classification. If the difference between the model prediction result and the actual classification is small, the calculated mean squared error will be small.

[0115] The trained network is used to classify and test the resulting column vectors, proceeding step by step, with the number of classes successively set to 400, 300, and so on.

[0116] Based on the type of data collected, it is ultimately divided into five categories, corresponding to five different states of the collected signals;

[0117] The evaluation metrics for classification include accuracy, precision, and recall; a parameter F1 is also introduced to evaluate the model's performance, as shown in the following formula:

[0118] F1=2*(Precision*Recall) / (Precision+Recall).

[0119] Example 2:

[0120] This embodiment provides a bearing fault classification system based on an intelligent inspection robot. The system includes:

[0121] The signal acquisition module is used to collect fault sounds from the bearing inner raceway, bearing outer raceway, bearing rolling elements, and bearing assembly based on the acquisition card and microphone of the intelligent inspection robot, while also collecting sound information when the bearing is working normally.

[0122] The time-frequency signal construction module is used to construct time-frequency signals for data augmentation and signal feature representation.

[0123] The feature extraction module is used to extract features through an autoencoder network;

[0124] The feature classification module is used to classify features using a convolutional neural network.

[0125] The classification optimization module is used to train and test the Adam optimization algorithm, continuously adjusting the parameter settings in the network to obtain more accurate classification results.

[0126] The working process of this system is as follows:

[0127] (1) Install the acquisition card and microphone equipment on the intelligent inspection robot, and call the equipment to collect the acoustic signal of bearing failure at the actual working condition location where data collection and analysis are required.

[0128] (2) Continuously adjust parameters to optimize training effect, and finally encapsulate the classification model of convolutional neural network into intelligent inspection robot, and call the model to extract features and classify the collected raw signals.

[0129] It can be set to collect data at regular intervals to achieve fully automatic bearing fault monitoring.

[0130] The feature classification using a convolutional neural network is as follows:

[0131] ① By using Fast Fourier Transform, the signal is decomposed into signals in different frequency bands, and the dataset is expanded by time-domain resampling, which is then unfolded in two dimensions to form a two-dimensional signal representation at multiple scales.

[0132] ② Divide the dataset into training and testing sets, and assign labels to each sample;

[0133] ③ Input the training samples obtained in step ② into the autoencoder network for training to obtain the feature representation of the training and test data of the last fully connected layer. Finally, use the features obtained from the training samples and the original labels to calculate the loss function.

[0134] ④ The test samples are fed into the trained model to verify the effectiveness of the diagnostic model.

[0135] Example 3:

[0136] This embodiment also provides an electronic device, including: a memory and a processor;

[0137] The memory stores the instructions executed by the computer.

[0138] The processor executes computer execution instructions stored in the memory, causing the processor to execute the bearing fault classification method based on an intelligent inspection robot in any embodiment of the present invention.

[0139] The processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor can be a microprocessor or any conventional processor.

[0140] Memory is used to store computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory, and by accessing data stored in the memory. Memory can mainly include a program storage area and a data storage area. The program storage area can store the operating system, at least one application program required for a function, etc.; the data storage area can store data created based on the use of the terminal, etc. In addition, memory can also include high-speed random access memory, and can also include non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart memory cards (SMC), secure digital cards (SD cards), flash memory cards, at least one disk storage device, flash memory devices, or other volatile solid-state storage devices.

[0141] Example 4:

[0142] This invention also provides a computer-readable storage medium storing multiple instructions, which are loaded by a processor to cause the processor to execute the bearing fault classification method based on an intelligent inspection robot according to any embodiment of this invention. Specifically, a system or device equipped with a storage medium may be provided, on which software program code implementing the functions of any of the above embodiments is stored, and the computer (or CPU or MPU) of the system or device may read and execute the program code stored in the storage medium.

[0143] In this case, the program code read from the storage medium can itself implement the function of any of the above embodiments, and therefore the program code and the storage medium storing the program code constitute part of the present invention.

[0144] Storage media embodiments for providing program code include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RYM, DVD-RW, DVD+RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, program code can be downloaded from a server computer via a communication network.

[0145] Furthermore, it should be clear that not only can the program code read by the computer be executed, but also the operating system or other components operating on the computer can be instructed based on the program code to perform some or all of the actual operations, thereby realizing the function of any of the embodiments described above.

[0146] Furthermore, it is understood that the program code read from the storage medium is written to the memory set in the expansion board inserted into the computer or to the memory set in the expansion unit connected to the computer. Then, based on the instructions of the program code, the CPU or other components installed on the expansion board or expansion unit execute some and all of the actual operations, thereby realizing the function of any of the embodiments described above.

[0147] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A bearing fault classification method based on an intelligent inspection robot, characterized in that, The method is as follows: S1. Acoustic signal acquisition: Based on the acquisition card and microphone of the intelligent inspection robot, fault sounds of the bearing inner raceway, bearing outer raceway, bearing rolling elements and bearing assembly are acquired, while sound information of the bearing during normal operation is also acquired. S2. Constructing Time-Frequency Signals: Constructing time-frequency signals to enhance data dimensionality and express signal features; S3. Feature extraction: Feature extraction is performed using an autoencoder network; S4. Feature Classification: Feature classification is performed using a convolutional neural network; S5. Optimize classification: Use the Adam optimization algorithm for training and testing, and continuously adjust the parameter settings in the network to obtain more accurate classification results; The specific steps for acquiring acoustic signals in step S1 are as follows: S101. Connect or discover the acquisition device, and the acquisition signal source is a bearing, including bearing inner ring raceway fault signal, bearing outer ring raceway fault signal, bearing rolling element fault signal and bearing assembly fault signal. S102. Obtain device status information: Use the GET request method to obtain device information, distinguish devices based on IP and MAC address, distinguish WebDAQ 504 errors, and select to manually input the host address. Convert the data into .json format. S103. Preliminary assessment of equipment availability: ① If so, proceed to step S104; ② If not, proceed to step S101; S104. Obtain complete system details; S105. Configure the audio acquisition task using the acquisition card, including the sampling rate, number of channels, and the sound pressure level settings for the microphone, and then publish the acquisition task. S106. Obtain the device working list status; S107. Determine if any tasks are in progress or paused: ① If so, stop the existing task or wait in the queue, and proceed to step S108. ② If not, proceed to step S108; S108. Obtain the equipment operating status and execute tasks; S109. Convert and output the relevant sound data of the corresponding acquisition device; While retrieving the device's work list status, determine whether the current task list directory is valid: If valid, print the list of tasks. The list of tasks includes the task name, number of channels, and progress status of the task. Alternatively, you can select an existing task to view. The data acquisition card uses the WebDAQ 504 IoT sound and vibration data logger; The microphone uses two acoustic sensors, SKC AD7199 and SKC AD7018, for dual-channel acquisition. The construction of the time-frequency signal in step S2 is as follows: Spatiotemporal frequency domain transformation was employed, and fast Fourier transform was used to construct multi-scale frequency domain information; The time-series signal is expanded in two dimensions using a windowing method at the separated frequency domain scale, forming a two-dimensional signal representation under multiple scales. The separated time-series signal is then resampled to obtain a time series, which is then spatially arranged to form a preprocessed signal, as shown in the following formula: ； Where L(i) represents the length of the original signal, i = (j-1) × M; i = 1, ..., M 2 j=1,…,M+1; P(j, K) represents the representation of a single pixel in the image; K=1,……,M; The purpose of rearranging the data in two dimensions and rounding down is to bring the data between 0 and 255. The one-dimensional vibration data is then normalized into an image. In fact, the data has been transformed from the original 64 bits to the current 8 bits. The feature extraction using an autoencoder network in step S3 is as follows: After dimensionality augmentation, data denoising and dimensionality reduction are performed using an autoencoder approach: The Tensorflow end-to-end open-source machine learning platform is used, and the Keras deep learning framework is used for data denoising and dimensionality reduction. In the autoencoder, the input feature dimension is defined as 2400×1, the number of channels is 1, the two-dimensional convolution constructs 10 filters, a 128×1 sliding window is used, the filter mode is set to same, and the activation function is the double tangent function tanh. Downsampling was performed using max pooling with a sampling factor of 8×1 and a stride of 2 to obtain a feature map of 1200×1×10. Upsampling is performed on the 1200×1×10 feature map to enlarge it and refine the feature information. After adding the encoding and decoding operations, a standard autoencoder operation is inserted. The autoencoder operation is characterized by the same number of features in the input and output. It is defined as the reconstruction of the input data after compression and extraction by the neural network. Through multiple stacked feature extractions, the convolutional layers and subsampling layers continuously shrink the feature map while the number of feature maps continuously increases. At the end of the feature extractor, all feature maps are expanded and arranged into a vector called the feature vector, which serves as the input to the subsequent classifier. The signals are fed into a neural network classifier for training to classify bearing fault sound signals. The feature classification using a convolutional neural network in step S4 is as follows: The learning process of a multi-neural network aims to obtain the minimum error: the actual output label. and predicted labels The difference between them decreases, which depends on the continuous updating of weights ω and bias b during training; among them, the loss function adopts mean squared error, which reflects the closeness between the probability distribution of various classifications in softmax and the actual classification. If the difference between the model prediction result and the actual classification is small, the calculated mean squared error will be small. The trained network is used to classify and test the resulting column vectors, proceeding step by step, with the number of classes successively set to 400, 300, and so on. Based on the type of data collected, it is ultimately divided into five categories, corresponding to five different states of the collected signals; The evaluation metrics for classification include accuracy, precision, and recall; a parameter F1 is also introduced to evaluate the model's performance, as shown in the following formula: F1=2*(Precision*Recall) / (Precision+Recall).

2. A bearing fault classification system based on an intelligent inspection robot, characterized in that, This system is used to implement the bearing fault classification method based on an intelligent inspection robot as described in claim 1; the system includes, The signal acquisition module is used to collect fault sounds from the bearing inner raceway, bearing outer raceway, bearing rolling elements, and bearing assembly based on the acquisition card and microphone of the intelligent inspection robot, while also collecting sound information when the bearing is working normally. The time-frequency signal construction module is used to construct time-frequency signals for data augmentation and signal feature representation. The feature extraction module is used to extract features through an autoencoder network; The feature classification module is used to classify features using a convolutional neural network. The classification optimization module is used to train and test the Adam optimization algorithm, continuously adjusting the parameter settings in the network to obtain more accurate classification results.

3. An electronic device, characterized in that, include: Memory and at least one processor; The memory contains computer programs; The at least one processor executes the computer program stored in the memory, causing the at least one processor to perform the bearing fault classification method based on the intelligent inspection robot as described in claim 1.

4. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that can be executed by a processor to implement the bearing fault classification method based on an intelligent inspection robot as described in claim 1.

Images