Rotary machinery bearing fault real-time diagnosis method, device and system based on deep learning algorithm

Abstract

The invention discloses a real-time fault diagnosis method, device and system for a rotating machinery bearing based on a deep learning algorithm, and the method comprises the steps: designing a data collection board card, and collecting the vibration data and sound data of a bearing in different fault states on an electromechanical fault simulation platform; carrying out normalization processing on the collected signals, and constructing a data set by adopting a sliding window method; performing multi-scale decomposition on the collected bearing vibration signal and sound signal by using a signal time-frequency analysis method to obtain a time-frequency signal with a multi-scale fault feature; taking the decomposed signals as input, training a lightweight deep neural network model and storing optimal model parameters; deploying the trained lightweight deep neural network model based on time-frequency analysis to an edge computing embedded board card; a display screen of the edge computing embedded board card outputs a bearing fault diagnosis result in real time, and an emergency alarm is given through an acousto-optic signal when a fault is detected. According to the method, the advantages of signal time-frequency analysis in processing non-stationary signals and the powerful capability of the deep neural network in feature extraction are fully combined, the bearing fault can be accurately diagnosed in real time in a complex industrial environment, and the problems of poor real-time performance, high calculation complexity and the like existing in bearing fault diagnosis in the prior art are solved.

Description

Technical Field

[0001] The present invention relates to the field of mechanical fault diagnosis, in particular to a method, device and system for fault diagnosis of rotating mechanical bearings. Specifically, the present invention proposes a method, device and system for real-time diagnosis of rotating mechanical bearing faults based on a deep learning algorithm, which can efficiently and accurately identify the type of bearing faults in a complex industrial environment, and has the advantages of strong real-time performance, high diagnostic accuracy, and low consumption of computing resources. Background Art

[0002] Rotating mechanical bearings are widely used in various types of industrial equipment, and their fault diagnosis is of great significance for ensuring the stable operation of equipment and reducing maintenance costs. Traditional bearing fault diagnosis methods usually rely on time domain analysis, frequency domain analysis, or a combination of feature extraction and classification. These methods are often limited in diagnostic performance when faced with non-stationary, nonlinear, and strongly noise-interfered vibration signals, making it difficult to achieve efficient and accurate fault identification. Therefore, with the rapid development of deep learning technology, bearing fault diagnosis methods based on deep neural networks have gradually received widespread attention.

[0003] For example, patent CN118885858A proposes a rolling bearing fault diagnosis method based on convolutional neural network (CNN) and multi-head self-attention (MHSA) mechanism. This method first divides the original vibration data into multiple samples according to a fixed length, then extracts the features in the samples through CNN, and then further learns these features through the MHSA module, and finally inputs the obtained feature vector into the prototype network-based classifier for fault diagnosis. This method effectively improves the accuracy of fault diagnosis through deep learning technology. In addition, patent CN118734190A proposes a bearing fault diagnosis method based on multi-input multi-channel parallel convolutional neural network. This method combines the envelope signal characteristics, statistical parameter characteristics and time-frequency image characteristics of the time domain vibration waveform, and performs comprehensive analysis based on multi-dimensional fault characteristics through the attention mechanism module, fully connected layer and Softmax classifier, thereby improving the accuracy and reliability of fault identification. In this way, the model can effectively integrate information from different signal dimensions, thereby improving the robustness of diagnosis.

[0004] However, although these deep learning methods have performed well in improving the accuracy of fault diagnosis, they usually require large computing resources and a large amount of training data. In addition, these methods also have certain limitations in real-time fault diagnosis and are difficult to meet the requirements of real-time and high efficiency in industrial environments. Especially in production environments, rapid detection and accurate diagnosis of bearing faults are crucial to avoid equipment downtime and reduce maintenance costs. Therefore, how to improve the computational efficiency of deep learning models, reduce dependence on data volume, and achieve real-time diagnosis is still a major challenge facing the current field of bearing fault diagnosis. Summary of the invention

[0005] The present invention aims to overcome the above-mentioned shortcomings of the prior art and provide a real-time diagnosis method, device and system for rotating machinery bearing faults based on deep learning algorithms. The present invention fully combines the advantages of signal time-frequency analysis in processing non-stationary signals with the powerful ability of deep neural networks in feature extraction, and can diagnose bearing faults in real time and accurately in complex industrial environments, solving the problems of poor real-time performance and high computational complexity in bearing fault diagnosis in the prior art.

[0006] To achieve the above objectives, the first aspect of the present invention provides a real-time diagnosis method for rotating machinery bearing faults based on a deep learning algorithm, which specifically includes the following steps:

[0007] S1: Design data acquisition board and collect vibration and sound data of bearings under different fault conditions on the electromechanical fault simulation platform;

[0008] S2: normalize the collected signals and construct the data set using the sliding window method;

[0009] S3: Use the signal time-frequency analysis method to perform multi-scale decomposition on the collected bearing vibration signal and sound signal to obtain a time-frequency signal with multi-scale fault characteristics;

[0010] S4: Using the decomposed signal as input, train a lightweight deep neural network model and save the optimal model parameters;

[0011] S5: Deploy the trained lightweight deep neural network model based on time-frequency analysis to the edge computing embedded board;

[0012] S6: The display screen of the edge computing embedded board outputs the bearing fault diagnosis results in real time, and issues an emergency alarm through sound and light signals when a fault is detected.

[0013] Step S1 describes designing a data acquisition board and collecting vibration and sound data of the bearing under different fault states on an electromechanical fault simulation platform, specifically including the following steps:

[0014] S11: Select an accelerometer and a microphone sensor to collect vibration signals and sound signals respectively. Connect the accelerometer and the microphone sensor to the AD acquisition module, and use the analog output signal of the sensor as the input of the AD acquisition module;

[0015] S12: Use the single-chip microcomputer as the control core, and establish communication with the AD acquisition module through the parallel interface of the single-chip microcomputer. In order to ensure that the high-frequency harmonics in the bearing fault are not lost, the sampling rate is set to 20KHz. When processing the sensor signal, the AD acquisition module converts the analog signal into a digital signal between -32768 and +32767, and the voltage range corresponds to -5V to +5V;

[0016] S13: In order to ensure the consistency of vibration data and sound data at the same time point, a synchronous clock and a timestamp are set in the microcontroller control program;

[0017] S14: Perform different types of bearing fault simulations on the electromechanical fault simulation platform, set two speeds, and simulate the outer ring fault, inner ring fault, rolling element fault, cage fault, and complex fault of the bearing at the same time. Collect the vibration data and sound data of these faults and transfer them to the local PC through the USB virtual port.

[0018] Step S2 normalizes the collected signals and constructs a data set using a sliding window method, which specifically includes the following steps:

[0019] S21: Perform Z-Score normalization processing on the collected vibration signal and sound signal, which makes the signal have zero mean and unit variance, thus avoiding instability in the training process due to signal amplitude differences.

[0020] S22: The normalized signal is usually used for further feature extraction. A sliding window operation is applied to the signal to decompose the continuous signal into multiple time windows, which will be used as the input of the model. The sliding window operation can capture the local features of the signal.

[0021] Step S3 uses a signal time-frequency analysis method to perform multi-scale decomposition on the collected bearing vibration signal and sound signal to obtain a time-frequency signal with multi-scale fault characteristics, which specifically includes the following steps:

[0022] S31: Initial wavelet decomposition (mother wavelet selection)

[0023] First, select the appropriate mother wavelet function and the corresponding scale function. These two functions determine the way the signal is decomposed and the frequency band division. Through wavelet decomposition, the signal is divided into a low-frequency part (approximate coefficient) and a high-frequency part (detail coefficient). This step uses a low-pass filter and a high-pass filter to perform convolution operations to extract different frequency band information of the signal.

[0024] S32: Recursive Multiscale Decomposition

[0025] Different from the traditional wavelet decomposition method, wavelet packet decomposition not only further decomposes the low-frequency part, but also decomposes the high-frequency part, thus providing a more detailed multi-scale decomposition structure. When decomposing the signal at each level, not only the low-frequency part of the sub-signal is processed, but also the high-frequency part of the sub-signal is recursively decomposed until the required number of decomposition levels is reached. This can capture the time-frequency characteristics of the signal more accurately.

[0026] S33: Decomposition tree structure

[0027] After multiple layers of recursive decomposition, the signal is decomposed into multiple sub-signals, each of which corresponds to a frequency band and covers the entire spectrum. The decomposition results of each layer are obtained by establishing a decomposition tree structure, in which each node represents different frequency band information, and the range of the frequency band is determined by the sampling frequency and the number of decomposition layers. The frequency band range of each sub-signal is gradually subdivided according to the number of decomposition layers, thus providing rich time-frequency information.

[0028] Step S4 uses the decomposed signal as input to train a lightweight deep neural network model and save the optimal model parameters, which specifically includes the following steps:

[0029] S41: Wavelet packet decomposition decomposes the original signal into multiple frequency band sub-signals. These decomposed signals are concatenated into a large input feature vector as the input data of the lightweight residual network model. This step ensures the comprehensiveness and diversity of the input features.

[0030] S42: The lightweight residual network model is a lightweight version of the residual network model, designed for real-time and embedded systems. The lightweight residual network model introduces a residual module to help gradient propagation and reduce the number of model parameters to reduce computational complexity. The residual module includes multiple convolutional layers, batch normalization, and activation functions to help optimize the network training process. The convolutional layer performs feature extraction, batch normalization accelerates the training process and improves stability, and the activation function is used to introduce nonlinearity.

[0031] S43: The lightweight design goal of the lightweight residual network model is to reduce computational complexity, especially for embedded devices. To this end, the lightweight residual network model uses depthwise separable convolution and bottleneck structure to reduce the computational burden. Depthwise separable convolution decomposes the traditional convolution operation into two steps: depthwise convolution and pointwise convolution. Depthwise convolution is a separate convolution for each channel, and pointwise convolution is a fusion of channel information. In addition, the bottleneck structure reduces the amount of computation by compressing high-dimensional features into low dimensions and then restoring them to high dimensions through pointwise convolution.

[0032] S44: When training the lightweight residual network model, a loss function is used to measure the difference between the model's prediction and the true label. During the training process, an optimization algorithm is used to minimize the loss function and update the model parameters. After training, the classification accuracy of the model is evaluated using the validation set, and the best model parameters are saved. The model parameters are stored in the .pth file format to ensure that they can be loaded and used on embedded devices.

[0033] Step S5 deploys the trained lightweight deep neural network model based on time-frequency analysis to the edge computing embedded board, specifically including the following steps:

[0034] S51: The lightweight deep neural network model based on time-frequency analysis trained on the computer is exported to ONNX (Open Neural Network Exchange) format. The network weights and network structure are saved when the model is exported, which ensures that the reasoning process can be reproduced on the edge computing embedded board.

[0035] S52: Optimize the exported model to adapt to the hardware architecture of the edge computing embedded board. First, quantize the high floating point model weights to low floating point model weights, thereby reducing the computational complexity and memory usage, and improving the inference speed. Secondly, perform lightweight trimming on the model structure, remove redundant layers or parameters, and reduce the model size while ensuring accuracy. Finally, use the built-in neural network processing unit (NPU) of the edge computing embedded board to accelerate inference, load the calculation from the CPU to the NPU, and make full use of the parallel computing capabilities of the board.

[0036] S53: Convert the optimized model into a format that can be run on edge computing embedded boards through the tool chain, and verify the accuracy and performance of the model locally to ensure that the inference results after deployment are consistent with the original training model.

[0037] S54: Package the model and the inference engine program into an executable file and transmit it to the memory of the edge computing embedded board through the serial port. When the embedded program runs on the edge computing embedded system, it will call the model file and receive the real-time collected vibration and sound signals through the input interface.

[0038] S55: On the edge computing embedded board, the pre-loaded inference engine performs forward reasoning of a lightweight deep neural network model based on time-frequency analysis in real time to output bearing fault diagnosis results. This process is accelerated by the NPU to achieve fast reasoning and provide real-time fault diagnosis information.

[0039] Step S6: The display screen of the edge computing embedded board outputs the bearing fault diagnosis result in real time, and issues an emergency alarm through sound and light signals when a fault is detected, which specifically includes the following steps:

[0040] S61: The edge computing embedded board parses the output of the model and displays text information in real time on the onboard screen, such as "Bearing status: normal" or "Bearing status: rolling element fault", etc. In addition, combined with real-time diagnostic data trend charts, the status of the fault can be intuitively displayed to help operators better understand the operation of the equipment.

[0041] S62: Set the fault threshold and determine whether a fault occurs based on the category probability output by the model. When the category probability exceeds the set threshold, it is judged as a fault, otherwise it is normal. The edge computing embedded board monitors the output results of the model in real time. If a fault is detected, the alarm mechanism is triggered immediately. The buzzer sounds an alarm and the LED light flashes to prompt maintenance personnel to handle it in time.

[0042] S63: The edge computing embedded board can record key information such as the time and type of fault, and store this information in a local storage device. This provides important data support for subsequent fault analysis and maintenance, and facilitates maintenance personnel to conduct subsequent inspections and repairs.

[0043] The second aspect of the present invention relates to a real-time diagnosis device for rotating machinery bearing faults based on a deep learning algorithm, which specifically includes a data acquisition module, a signal preprocessing module, a time-frequency analysis feature extraction module, a lightweight deep neural network diagnosis module, a model deployment module, a fault display and alarm module, and a data storage module.

[0044] The data acquisition module is responsible for acquiring the vibration and sound signals of the bearing during operation. It uses acceleration sensors and microphone sensors to convert analog signals into digital signals through AD data acquisition boards, and is controlled and synchronized by a single-chip microcomputer to ensure high sampling rate and time consistency, and transmit data to embedded devices and storage systems in real time.

[0045] The signal preprocessing module normalizes the collected signals to eliminate amplitude differences and noise interference. Through sliding window operation, the signal is divided into multiple time windows to form a data set suitable for subsequent model input, effectively capturing the local characteristics of the signal and improving the accuracy of fault diagnosis.

[0046] The time-frequency analysis feature extraction module performs multi-scale decomposition of signals through time-frequency analysis, decomposing complex vibration signals and sound signals into sub-signals of multiple frequency bands, extracting time-frequency features, and helping to identify fault features in non-stationary and nonlinear signals. In particular, it provides higher resolution in the high-frequency part, providing support for subsequent fault diagnosis.

[0047] The lightweight deep neural network diagnosis module effectively reduces the computational complexity and model parameter quantity through deep separable convolution and bottleneck structure, and adapts to the computing power of embedded devices. The module uses the extracted time-frequency features to perform deep feature extraction on the signal, detect and identify fault types, such as outer ring fault, inner ring fault, rolling element fault and cage fault.

[0048] The model deployment module optimizes and quantizes the trained lightweight deep neural network model based on time-frequency analysis, and deploys it to the edge computing embedded board. It also accelerates network reasoning through the NPU, supports real-time data input and model reasoning, and ensures efficient real-time diagnosis.

[0049] The fault display and alarm module is responsible for displaying the working status of the bearing in real time and showing the real-time monitoring results through the built-in display. When the diagnosis result reaches the set fault probability threshold, the buzzer alarm is triggered and the LED light flashes to notify relevant personnel to handle it in time.

[0050] The data storage module records historical data during the fault diagnosis process, including the time and type of the fault, as well as the original signal and processing results. These data are stored in the local storage of the edge computing embedded board to provide support for subsequent fault analysis, maintenance management and model optimization.

[0051] The third aspect of the present invention relates to a real-time diagnosis system for rotating machinery bearing faults based on a deep learning algorithm, which is specifically composed of a data acquisition device, a signal processing unit, a diagnostic computing platform, a display and alarm device, and a data storage system. The modules work together to achieve real-time monitoring and diagnosis of bearing faults.

[0052] The data acquisition equipment includes an accelerometer, a microphone sensor, and a data acquisition board, which is responsible for real-time acquisition of the vibration and sound signals of the bearing. The accelerometer captures vibration information, and the microphone sensor collects sound signals. Through high sampling rate and synchronous clock, accurate signal acquisition is ensured in various fault modes, and the analog signal is converted into a digital signal and transmitted to the signal processing unit.

[0053] The signal processing unit is responsible for preprocessing the collected signals, including normalization and sliding window operations, to ensure that the signal quality meets the requirements of subsequent analysis. The signal processing unit formats the signal into a form suitable for deep learning model input to facilitate accurate extraction of fault features. The unit consists of an embedded control system and computing resources for data processing and formatting.

[0054] The diagnostic computing platform is responsible for the operation and real-time reasoning of lightweight deep neural network models based on time-frequency analysis. The platform runs the trained model through the edge computing embedded board, quickly infers the collected signals, and outputs the fault diagnosis results. Combining time-frequency feature extraction and deep learning capabilities, the platform can accurately identify various types of bearing faults, such as outer ring, inner ring, rolling element and cage faults.

[0055] The display and alarm device displays the operating status and fault type of the bearing in real time and provides a fault alarm function. The display device shows the bearing health status (such as "normal" or "rolling element fault") and fault trend diagram to help operators evaluate the health of the equipment. The alarm device uses a buzzer and LED light to sound and light signal alarms when a fault is detected, reminding relevant personnel to take maintenance measures.

[0056] The data storage system stores historical fault diagnosis records, including fault occurrence time, type, signal characteristics, and diagnostic analysis results. These data are used for subsequent fault analysis, trend prediction, and equipment maintenance management. The storage system supports regular data backup and query to ensure the effective preservation of key fault information and system operation records.

[0057] The beneficial effects of the present invention are:

[0058] 1. The bearing fault diagnosis method provided by the present invention can extract fault features efficiently: The time-frequency analysis method of the present invention can accurately extract local features in the signal through multi-scale decomposition, and is particularly suitable for non-stationary and nonlinear fault components in bearing vibration signals. In addition, it can also finely decompose signals of different frequency bands, making the bearing fault features more prominent in the time-frequency domain, thereby improving the accuracy of subsequent classification.

[0059] 2. The fault diagnosis model provided by the present invention has a strong classification capability: the lightweight deep neural network model of the present invention can effectively deal with the gradient vanishing problem in deep learning and accelerate the training process through residual connections. Compared with the traditional convolutional neural network, the lightweight deep neural network model of the present invention has fewer parameters and lower computational complexity, which improves classification efficiency and accuracy. It can efficiently identify different types of fault modes by automatically learning the advanced features of the signal.

[0060] 3. The bearing fault diagnosis method provided by the present invention combines the multi-scale feature extraction of signal time-frequency analysis and the deep learning classification capability of deep neural networks, making the diagnosis process not only accurate but also efficient, and capable of performing real-time fault detection and diagnosis at industrial sites, meeting the high requirements of the industrial Internet of Things for real-time and accuracy.

[0061] 4. The bearing fault diagnosis model provided by the present invention has the characteristics of low computing resource consumption: the lightweight deep neural network model reduces the network's computing workload through lightweight design and is suitable for low-power embedded platforms. It can run on low-power edge computing embedded boards while maintaining high diagnostic accuracy, which is crucial for industrial applications that require real-time feedback.

[0062] 5. The bearing fault diagnosis system provided by the present invention can be deployed on an edge computing embedded platform to realize on-site real-time reasoning and fault alarm, support effective fault diagnosis and maintenance management at industrial sites, and improve the operation and maintenance efficiency of equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Attached Figure 1 It is a schematic diagram of the process of the present invention;

[0064] Attached Figure 2 It is a flow chart of the signal time-frequency analysis method in the method of the present invention;

[0065] Attached Figure 3 It is a structural schematic diagram of the core part of the lightweight deep neural network model in the method of the present invention;

[0066] Attached Figure 4 It is a structural schematic diagram of the device of the present invention;

[0067] Attached Figure 5 It is a schematic diagram of the structure of the system of the present invention. DETAILED DESCRIPTION

[0068] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0069] Embodiment 1

[0070] Reference Figure 1 , which is a flow chart of a real-time diagnosis method for rotating machinery bearing faults based on a deep learning algorithm provided by the present invention, specifically comprising the following steps:

[0071] S1: Design a data acquisition board and collect the vibration and sound data of the bearing under different fault conditions on the electromechanical fault simulation platform;

[0072] Specifically, S11: select an accelerometer and a microphone sensor to collect vibration signals and sound signals respectively. Connect four accelerometers and one microphone sensor to the AD7606 module, and use the analog output signals of the sensors as inputs of the AD7606 module;

[0073] S12: Use the STM32F407 microcontroller as the control core and communicate with the AD7606 module through the 8080 parallel interface. In order to ensure that the high-frequency harmonics in the bearing fault are not lost, the sampling rate is set to 20KHz. When processing the sensor signal, the AD7606 module converts the analog signal into a digital signal between -32768 and +32767, and the voltage range corresponds to -5V to +5V;

[0074] S13: In order to ensure the consistency of vibration data and sound data at the same time point, a synchronous clock and a timestamp are set in the microcontroller control program;

[0075] S14: Perform different types of bearing fault simulations on the electromechanical fault simulation platform, set the speed to 1800rpm and 1200rpm, and simulate the outer ring fault, inner ring fault, rolling element fault, cage fault, and complex fault of the bearing at the same time. Collect the vibration data and sound data of these faults and transfer them to the local PC through the USB virtual port.

[0076] S2: normalize the collected signals and construct the data set using the sliding window method;

[0077] Specifically, S21: performing Z-Score normalization processing on the collected vibration signal and sound signal, which makes the signal have zero mean and unit variance, thus avoiding instability in the training process due to signal amplitude differences.

[0078] Specifically, the signal is normalized by the mean and variance, and converted into a standard normal distribution. The formula is:

[0079]

[0080] Where μ is the mean of the signal, σ is the standard deviation of the signal, x is the original signal, and x norm is the normalized signal.

[0081] S22: The normalized signal is used for further feature extraction. A sliding window operation is applied to the signal to decompose the continuous signal into multiple time windows, which will be used as the input of the model. The sliding window operation can capture the local features of the signal.

[0082] Specifically, let the original signal be x(t), and the sliding window operation divides the signal into multiple sub-blocks (i.e., data blocks). Assuming the window size is N and the step size is S, the signal of each window is:

[0083] X i =[x(i),x(i+1),...,x(i+N-1)] (1.2)

[0084] Among them, i is the starting position of the window, N is the window size, X i is the i-th data block.

[0085] Furthermore, the window size N and step size S determine the accuracy of the sliding window operation. After the sliding window processing, the original signal x(t) is divided into multiple signal segments X1, X2, …, X n , these data blocks will serve as the model input feature set.

[0086] S3: See attached Figure 2 , wavelet packet decomposition (WPD) is used to perform multi-scale decomposition on the collected bearing vibration signal and sound signal to obtain a time-frequency signal with multi-scale fault characteristics;

[0087] Specifically, S31: initial wavelet decomposition (mother wavelet selection)

[0088] Specifically, given a signal x(t), first select the mother wavelet function ψ(t) and its corresponding scaling function φ(t). These two functions determine the decomposition method and the division of frequency bands. The mother wavelet function has the characteristics of time-frequency localization and is suitable for analyzing non-stationary signals.

[0089] In the first layer of the initial wavelet decomposition, the signal x(t) is decomposed into two parts: low-frequency part (approximation coefficient, A1) and high-frequency part (detail coefficient, D1). This process is achieved through convolution operation, using low-pass filter and high-pass filter to extract low-frequency and high-frequency parts respectively. The formula is:

[0090]

[0091] Among them, h[k] and g[k] are the filter coefficients of the low-pass and high-pass filters respectively, and n is an index on the time scale, indicating different time positions of the signal.

[0092] S32: Recursive Multiscale Decomposition

[0093] Specifically, in the traditional wavelet decomposition method, only the low-frequency part A1 is further decomposed, but in WPD, not only the low-frequency part is decomposed, but also the high-frequency part D1 is decomposed. Therefore, WPD can provide a more refined multi-scale decomposition structure.

[0094] In the j-th level decomposition, signal Aj is decomposed into 2 j sub-signals, each corresponding to a frequency band. The decomposition of these sub-signals depends on the sub-signals of the previous layer (whether low frequency or high frequency). Specifically, the decomposition formula for each node (j, k) in the jth layer is as follows:

[0095]

[0096] Among them: A j+1,2k is the low-frequency component of node (j, k), A j+1,2k+1 is the high frequency component of node (j,k).

[0097] These operations are performed recursively in each layer (from layer 1 to layer J), and each time the decomposition signal A j and D j , until the desired level of decomposition is reached.

[0098] S33: Decomposition tree structure

[0099] After J-layer WPD decomposition, the signal x(t) is decomposed into 2 J The frequency bands of the signal are divided into several sub-signals, which cover the entire spectrum. Each frequency band corresponds to a sub-signal, and the frequency band range of the signal is determined by the sampling frequency f. s and decomposition level J to determine.

[0100] At the jth layer, the frequency band range of the kth sub-signal is:

[0101]

[0102] Among them, f s is the sampling frequency of the signal; j is the number of decomposition layers, which determines the degree of subdivision of the frequency band; k is the sub-signal index of the jth layer, and the value range is k = 0, 1, ..., 2 j -1.

[0103] S4: Take the WPD decomposition signal as input, train the LResNet model and save the optimal model parameters;

[0104] Specifically, S41: WPD can decompose the original signal x(t) into multiple frequency band sub-signals, assuming that x LF1 ,x HF1 ,x LF2 ,x HF2 ,…, concatenate these signals into a large input feature vector:

[0105] X input =[x LF1 ,x HF1 ,x LF2 ,x HF2,...](1.8)

[0106] These concatenated feature vectors X input Will be used as input data for the LResNet model.

[0107] S42: LResNet is a lightweight version of ResNet, suitable for real-time and embedded systems. The design idea of ​​LResNet is to introduce a residual block to help gradient propagation, while reducing the number of model parameters to reduce computational complexity.

[0108] Reference Figure 3 , the core part of LResNet is the residual module, and its basic formula is as follows:

[0109] y=F(x,{W i})+x(1.9)

[0110] Among them, x is the input feature, F(x,{W i}) is the convolutional layer, batch normalization, and activation function in the residual module, and y is the output of the module.

[0111] Furthermore, each residual module includes multiple convolutional layers, and the formula for the convolution operation is:

[0112]

[0113] Among them, x i is the input, w i is the convolution kernel, b is the bias, and * indicates the convolution operation.

[0114] Furthermore, batch normalization is used to speed up network training and stabilize the model. In the batch normalization layer, the input x is standardized to a distribution with a mean of 0 and a variance of 1:

[0115]

[0116] Among them, μ and σ 2 are the mean and variance of the batch data, respectively, and ∈ is a small constant to prevent division by zero errors.

[0117] Furthermore, the ReLU function is a commonly used nonlinear activation function, and the formula is:

[0118] ReLU(x)=max(0,x) (1.12)

[0119] S43: In LResNet, the main goal of lightweight design is to reduce computational complexity and parameter quantity, which is particularly suitable for embedded devices. This embodiment uses the method of depthwise separable convolution and bottleneck structure to lightweight the model.

[0120] Specifically, we first perform depthwise separable convolution on ResNet. In traditional convolution, the convolution operation performs convolution calculations on all input channels, which requires a lot of calculations. Depthwise separable convolution decomposes the standard convolution operation into two steps:

[0121] (1) Depthwise convolution: Convolution operation is performed on each input channel separately.

[0122] (2) Point-by-point convolution: Information between channels is fused through 1×1 convolution.

[0123] The formula for depthwise separable convolution is:

[0124] y=(x*ω d )*ω p (1.13)

[0125] Among them, ω d is the depth convolution kernel, ω p is the point-wise convolution kernel.

[0126] Furthermore, a bottleneck structure is used to reduce the computational complexity of the ResNet network. The bottleneck structure compresses the high-dimensional feature map into a low-dimensional feature map by introducing a 1×1 convolution, thereby reducing the number of parameters in the convolution layer. Assume that the input feature map is The convolution operation of the bottleneck structure can be expressed as:

[0127] y=(x*ω1)*ω2*ω3 (1.14)

[0128] Among them, ω1 is a 1×1 convolution kernel, ω2 and ω3 are other convolution kernels. The bottleneck design compresses the high-dimensional feature map to a low dimension, and then restores it to a high dimension through point-by-point convolution.

[0129] S44: When training LResNet, the loss function used is the cross entropy loss function. Assuming that the output of the model is y and the true label is y, the cross entropy loss function can be expressed as:

[0130]

[0131] Among them, C is the number of categories, y i is the true label of the i-th category, y i is the predicted probability output by the model.

[0132] Furthermore, an optimization algorithm is used in the optimization process to minimize the loss function. This embodiment uses the SGD optimization algorithm, and the formula is as follows:

[0133]

[0134] Among them, W iis the current weight, η is the learning rate, is the gradient of the loss function with respect to the weights.

[0135] Furthermore, after training is completed, the classification accuracy of the model is evaluated on the validation set, and the best model weights are saved for subsequent inference. When saving, the model parameters are stored in the .pth file format to ensure that they can be loaded and used on embedded devices.

[0136] S5: Deploy the trained WPD-LResNet model to the RV1126 embedded board;

[0137] Specifically, S51: Export the WPD-LResNet model trained on the computer to ONNX format. The network weights and network structure will be saved when the model is exported, so as to ensure that the reasoning process can be reproduced on the RV1126 device.

[0138] S52: Optimize the exported model to fit the hardware architecture of RV1126.

[0139] Specifically, the 32-bit floating point model weights are quantized to 16-bit floating points to reduce the computational complexity and memory usage of the model and improve the inference speed.

[0140] Furthermore, the model structure is pruned to remove redundant layers or parameters, further reducing the model size while maintaining accuracy.

[0141] Furthermore, the built-in NPU of the RV1126 board is used to accelerate reasoning, adapt the model to the NPU engine, and load the calculation from the CPU to the NPU, giving full play to the parallel computing capability of the board.

[0142] S53: Convert the optimized model into the RKNN format that can be run on RV1126 through the RKNN toolchain provided by Rockchip, and verify the accuracy and performance of the model locally to ensure that the reasoning results after deployment are consistent with the original training model.

[0143] S54: The model and inference engine program are packaged into an executable file and transferred to the memory of the RV1126 board through the serial port. When the embedded program runs on the RV1126 system, it calls the model file and receives the real-time collected vibration and sound signals through the input interface.

[0144] S55: On the RV1126 board, the forward reasoning of the WPD-LResNet model is performed in real time through the pre-loaded RKNN Runtime inference engine to output the bearing fault diagnosis results.

[0145] S6: The display screen of the RV1126 board outputs the bearing fault diagnosis results in real time, and issues an emergency alarm through sound and light signals when a fault is detected;

[0146] Specifically, the S61:RV1126 board analyzes the output of the model and displays text information in real time on the onboard screen, such as "Bearing status: normal" or "Bearing status: rolling element fault", etc. In addition, combined with real-time diagnostic data trend charts, the status of the fault can be intuitively displayed to help operators better understand the operation of the equipment.

[0147] S62: Set the fault threshold and determine whether a fault occurs based on the class probability output by the model. When the class probability exceeds the set threshold, it is judged as a fault, otherwise it is normal. The RV1126 board monitors the output results of the model in real time. If a fault is detected, the alarm mechanism is triggered immediately. The buzzer sounds an alarm and the LED light flashes to prompt maintenance personnel to handle it in time.

[0148] S63: The RV1126 board can record key information such as the time and type of fault, and store this information in a local storage device. This provides important data support for subsequent fault analysis and maintenance, and facilitates maintenance personnel to conduct subsequent inspections and repairs.

[0149] Embodiment 2

[0150] Reference Figure 4 , which is a structural schematic diagram of a real-time diagnosis device for rotating machinery bearing faults based on a deep learning algorithm provided by the present invention, specifically including a data acquisition module, a signal preprocessing module, a wavelet packet decomposition feature extraction module, a lightweight residual network diagnosis module, a model deployment module, a fault display and alarm module, and a data storage module.

[0151] The data acquisition module is responsible for acquiring vibration and sound signals during the operation of the bearing. The module includes four acceleration sensors and one microphone sensor. These sensors convert analog signals into digital signals through the AD7606 data acquisition board, and are controlled and synchronized by the STM32F407 microcontroller to ensure that the collected data has a high sampling rate and time consistency. The data is transmitted to the embedded device and storage system in real time through the USB interface.

[0152] The signal preprocessing module normalizes the collected vibration and sound signals to remove signal amplitude differences and noise interference. This module divides the continuous signal data into multiple time windows through sliding window operations to form a data set suitable for subsequent model input. This step effectively captures the local characteristics of the signal and improves the accuracy of fault diagnosis.

[0153] The wavelet packet decomposition feature extraction module uses the wavelet packet decomposition method to perform multi-scale decomposition of the signal, decomposing the complex vibration signal and sound signal into multiple sub-signals of different frequency bands. These sub-signals cover the time-frequency characteristics of the original signal, which helps to identify the fault characteristics in non-stationary and nonlinear signals, especially providing higher resolution in the high-frequency part. The decomposed time-frequency characteristics are used as the input data of the diagnostic model.

[0154] The lightweight residual network diagnosis module is the core part of this device. Based on the lightweight residual neural network, this module greatly reduces the computational complexity and model parameter quantity through deep separable convolution and bottleneck structure, so as to adapt to the computing power limitation of embedded devices. In this module, the time-frequency features extracted by WPD are input into the LResNet model, which performs deep feature extraction and fault classification on the signal data, and outputs the fault type of the bearing, including outer ring fault, inner ring fault, rolling element fault and cage fault.

[0155] The model deployment module deploys the trained WPD-LResNet model to the RV1126 embedded board. Through model optimization and quantization, the model is converted into the RKNN format suitable for embedded hardware, and the NPU built into the RV1126 board is used to accelerate network reasoning. This module implements real-time data input and model forward reasoning, ensuring the efficiency and real-time performance of the diagnosis process.

[0156] The fault display and alarm module is responsible for real-time output of diagnostic results and triggering the alarm mechanism when a fault is detected. The RV1126 embedded board displays the bearing status through the built-in display screen and intuitively displays the real-time monitoring results in combination with the data trend chart. When the fault probability output by the model exceeds the set threshold, the buzzer is triggered to sound an alarm and the LED light flashes to notify relevant personnel to carry out maintenance in a timely manner.

[0157] The data storage module is used to record the historical data of fault diagnosis, including the time and type of fault occurrence, as well as the original signal and processing results during the diagnosis process. These data are stored in the local storage device of the RV1126 board to facilitate subsequent fault analysis, maintenance management and model optimization.

[0158] Embodiment 3

[0159] Reference Figure 5, is a structural diagram of a real-time diagnosis system for bearing faults of rotating machinery based on a deep learning algorithm provided by the present invention, which is specifically composed of a data acquisition device, a signal processing unit, a diagnostic computing platform, a display and alarm device, and a data storage system. The various devices in the system cooperate with each other to complete the real-time monitoring and diagnosis of bearing faults. The system is suitable for rotating machinery equipment in an industrial environment, and provides reliable equipment operation monitoring and fault warning functions through efficient and accurate fault diagnosis.

[0160] Data acquisition equipment includes accelerometers, microphone sensors, and data acquisition boards, which are responsible for real-time acquisition of vibration and sound signals from rotating machinery bearings. Accelerometers are used to capture the acceleration information of bearing vibration, while microphone sensors are used to collect sound signals generated during the operation of the equipment. The data acquisition board converts these analog signals into digital signals through connection with the sensors and transmits them to the signal processing unit. These signal acquisition devices use technologies such as high sampling rate and synchronous clock to ensure accurate acquisition of signal data under various fault modes during operation.

[0161] The signal processing unit is responsible for preprocessing the collected signals, including normalization and sliding window operations, to ensure that the data quality of the signals meets the requirements of subsequent analysis. Through the data preprocessing module, the signals are converted into a format suitable for deep learning model input, so that the fault characteristics in the bearing operation status can be accurately extracted. The unit includes an embedded control system for the data acquisition board and computing resources for processing and formatting data.

[0162] The diagnostic computing platform is the core computing unit of this system, responsible for the operation and real-time reasoning of the WPD-LResNet model. The platform runs the trained WPD-LResNet model through the RV1126 embedded board, quickly reasons the real-time collected signal data, and outputs the fault diagnosis results of the bearing. The computing platform extracts the time-frequency features of the signal data and combines the deep learning capabilities of LResNet to accurately identify various types of bearing faults (such as outer ring faults, inner ring faults, rolling element faults, cage faults, etc.).

[0163] The display and alarm device is used to display the operating status and fault type of the bearing in real time, and provide a fault alarm function. By connecting to the diagnostic computing platform, the display device can show the health status of the bearing in real time (such as "normal" or "rolling element fault", etc.), and display the fault trend through a chart to help operators judge the health status of the equipment. The alarm device will promptly send out sound and light signals (buzzer alarm, LED flashing prompt) when a fault is detected, reminding relevant personnel to take necessary maintenance measures.

[0164] The data storage system is used to store historical fault diagnosis records, including the bearing failure time, fault type, signal characteristic data, and analysis results during the diagnosis process. These historical data can be used for subsequent fault analysis, trend analysis, and equipment maintenance management. The storage system also supports regular data backup and query to ensure that key fault information and system operation records are effectively retained.

[0165] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A real-time diagnosis method for rotating machinery bearing faults based on deep learning algorithm, characterized in that: The following steps are involved: S1: Design data acquisition board and collect vibration and sound data of bearings under different fault conditions on the electromechanical fault simulation platform; S2: normalize the collected signals and construct the data set using the sliding window method; S3: Use the signal time-frequency analysis method to perform multi-scale decomposition on the collected bearing vibration signal and sound signal to obtain a time-frequency signal with multi-scale fault characteristics; S4: Using the decomposed signal as input, train a lightweight deep neural network model and save the optimal model parameters; S5: Deploy the trained lightweight deep neural network model based on time-frequency analysis to the edge computing embedded board; S6: The display screen of the edge computing embedded board outputs the bearing fault diagnosis results in real time, and issues an emergency alarm through sound and light signals when a fault is detected.

2. The real-time diagnosis method for rotating machinery bearing faults based on deep learning algorithm according to claim 1 is characterized in that: The design of the data acquisition board described in step S1 and the collection of the vibration and sound data of the bearing under different fault states on the electromechanical fault simulation platform specifically include: S11: Select an acceleration sensor and a microphone sensor to collect vibration signals and sound signals respectively; S12: Use the single-chip microcomputer as the control core, and establish communication with the AD acquisition module through the parallel interface of the single-chip microcomputer, and set the sampling rate to 20KHz; S13: Setting the synchronous clock and timestamp to ensure the consistency of vibration data and sound data at the same time point; S14: Perform different types of bearing fault simulations on the electromechanical fault simulation platform, and collect vibration data and sound data under different fault types.

3. The real-time diagnosis method for rotating machinery bearing faults based on deep learning algorithm according to claim 1 is characterized in that: The step S2 of normalizing the collected signals and constructing a data set using a sliding window method specifically includes: S21: performing Z-Score standardization processing on the collected vibration signal and sound signal; S22: Apply a sliding window operation to the normalized signal to decompose the signal into multiple time windows as model input.

4. The real-time diagnosis method for rotating machinery bearing fault based on deep learning algorithm according to claim 1 is characterized in that: The signal time-frequency analysis method described in step S3 is used to perform multi-scale decomposition on the collected bearing vibration signal and sound signal to obtain a time-frequency signal with multi-scale fault characteristics. The specific steps include: S31: Select a suitable mother wavelet function and perform initial wavelet decomposition to divide the signal into a low-frequency part and a high-frequency part; S32: using a recursive multi-scale decomposition method to further decompose the high frequency part and gradually refine the frequency band information of the signal; S33: The decomposition result of each layer is represented by a decomposition tree structure, and each node represents different frequency band information.

5. The real-time diagnosis method for rotating machinery bearing fault based on deep learning algorithm according to claim 1 is characterized in that: The step S4 uses the decomposed signal as input to train a lightweight deep neural network model and save the optimal model parameters, and the specific steps include: S41: concatenate the signal after wavelet packet decomposition into an input feature vector as the input of the lightweight residual network; S42: Lightweight residual network introduces residual module for gradient propagation and uses depthwise separable convolution and bottleneck structure to reduce computational complexity; S43: Use an optimization algorithm to minimize the loss function, evaluate the classification accuracy through the validation set, and save the best model parameters.

6. The real-time diagnosis method for rotating machinery bearing fault based on deep learning algorithm according to claim 1 is characterized in that: Step S5 of deploying the trained lightweight deep neural network model based on time-frequency analysis to the edge computing embedded board includes: S51: Export the trained model to ONNX format and optimize it to fit the hardware architecture of the edge computing embedded board; S52: quantize and trim the exported model to reduce the computational complexity and storage requirements of the model; S53: Use the NPU of the edge computing embedded board to accelerate reasoning and ensure real-time diagnosis; S54: Package the model and the inference engine program into an executable file and transmit it to the embedded board through the serial port.

7. The real-time diagnosis method for rotating machinery bearing fault based on deep learning algorithm according to claim 1 is characterized in that: The display screen of the edge computing embedded board described in step S6 outputs the bearing fault diagnosis result in real time, and issues an emergency alarm through sound and light signals when a fault is detected. The specific steps include: S61: Display the bearing status in real time on the display screen of the embedded board, and display the fault changes through the fault trend diagram; S62: Determine whether a fault occurs according to the set threshold value, and trigger a buzzer alarm and LED flashing when a fault occurs; S63: Record key information such as the time and type of the fault, and store the information in a local storage device.

8. A real-time diagnosis device for rotating machinery bearing faults based on deep learning algorithm, characterized in that: include: The data acquisition module is used to collect the vibration signal and sound signal of the bearing, and convert the analog signal into a digital signal through the AD acquisition module; The signal preprocessing module is used to normalize the collected signals and construct a data set through a sliding window method; The time-frequency analysis feature extraction module is used to perform multi-scale decomposition of the signal and extract time-frequency features; A lightweight deep neural network diagnosis module for fault diagnosis based on time-frequency feature extraction; Model deployment module, used to deploy the trained lightweight deep neural network model based on time-frequency analysis to the edge computing embedded board; Fault display and alarm module, used to display fault diagnosis results and alarm through sound and light signals; The data storage module is used to store historical fault diagnosis records.

9. A real-time diagnosis system for rotating machinery bearing faults based on deep learning algorithm, characterized in that: include: Data acquisition equipment, including an acceleration sensor, a microphone sensor and a data acquisition board, for collecting vibration and sound signals of the bearing; A signal processing unit, used for preprocessing the collected signal, including signal normalization and sliding window operation; The diagnostic computing platform is used to run the trained lightweight deep neural network model based on time-frequency analysis and perform real-time reasoning to output fault diagnosis results; Display and alarm equipment, used to display the bearing fault diagnosis results and provide sound and light alarms when a fault occurs; Data storage system, used to store fault diagnosis records and other related information.

Images