A neural network-based rolling bearing fault intelligent diagnosis method and system

Abstract

This invention discloses an intelligent diagnostic method and system for rolling bearing faults based on neural networks. The diagnostic method includes: sampling the vibration signal of the rolling bearing and outputting the sampled signal; performing resonance demodulation processing on the sampled signal using a pre-constructed adaptive Mel filter bank to extract a frame-level feature vector sequence; and using a pre-trained neural network diagnostic model to perform multi-dimensional feature extraction, fusion, and classification operations on the frame-level feature vector sequence to obtain the fault type of the rolling bearing. This invention enhances fault characteristics by performing resonance demodulation processing on the sampled signal using a pre-constructed adaptive Mel filter bank; and achieves feature extraction and classification using a neural network diagnostic model, effectively suppressing background noise interference, improving diagnostic accuracy, and supporting the identification of various bearing faults.

Description

Technical Field

[0001] This invention belongs to the field of mechanical fault diagnosis technology, specifically relating to an intelligent diagnostic method and system for rolling bearing faults based on neural networks. Background Technology

[0002] Rolling bearings are critical components in rotating machinery, and their operating status directly affects the stability and safety of the entire machine. Traditional fault diagnosis methods rely on empirical feature extraction and classification algorithms, which suffer from poor robustness and low generalization ability. In recent years, neural networks have been increasingly used in intelligent diagnosis; however, due to the limited hardware resources of embedded platforms such as FPGAs, directly training and inferring neural networks from raw signals often leads to resource overflow or performance bottlenecks.

[0003] In addition, noise interference remains a significant factor affecting the accuracy of fault identification. Existing methods have certain limitations in signal preprocessing. Therefore, there is an urgent need for a method that combines efficient preprocessing and lightweight neural networks, and can be deployed on edge hardware platforms to achieve real-time fault identification. Summary of the Invention

[0004] To address the aforementioned problems in the existing technology, this invention provides a method and system for intelligent diagnosis of rolling bearing faults based on neural networks. The technical problem to be solved by this invention is achieved through the following technical solution: This invention provides an intelligent fault diagnosis method for rolling bearings based on neural networks, the method comprising: The vibration signal of the rolling bearing is sampled and the sampled signal is output. The sampled signal is subjected to resonant demodulation processing using a pre-constructed adaptive Mel filter bank to extract a frame-level feature vector sequence. A pre-trained neural network diagnostic model is used to perform multi-dimensional feature extraction, fusion, and classification operations on the frame-level feature vector sequence to obtain the fault type of the rolling bearing; wherein, The adaptive Mel filter bank is constructed by automatically selecting the resonant bandwidth parameters based on the fast spectral kurtosis algorithm.

[0005] In one embodiment of the present invention, the sampled signal is subjected to resonant demodulation processing using a pre-constructed filter bank to extract a frame-level feature vector sequence, including: The sampled signal is subjected to frame preprocessing to obtain several adjacent frame signals; The adjacent frame signals are weighted by a window function to obtain the windowed frame signal. Perform a Fast Fourier Transform on the windowed frame signal to obtain the corresponding frequency domain representation; Taking the modulus of the frequency domain representation and squaring it yields the corresponding energy spectrum; The energy spectrum is subjected to feature filtering and dimensionality reduction using an adaptive Mel filter bank, and a frame-level feature vector sequence is output.

[0006] In one embodiment of the present invention, the adaptive Mel filter bank consists of a plurality of triangular bandpass filters; the center frequency and bandwidth of each triangular bandpass filter are adaptively determined according to preset fault-related frequency band parameters.

[0007] In one embodiment of the present invention, the dimension of the frame-level feature vector sequence is the same as the number of filters in the adaptive Mel filter bank.

[0008] In one embodiment of the present invention, the neural network diagnostic model includes: The input layer, the first convolutional layer, the multi-level residual network, and the fully connected layer are arranged sequentially; among them, The input layer transmits the frame-level feature vector sequence to the first convolutional layer; The first convolutional layer extracts the local relevant features corresponding to the received frame-level feature vector sequence in the time dimension and expands the feature channel dimension; The multi-level residual network performs multi-scale feature extraction and deep representation learning on locally related features, and outputs a convolutional feature matrix. The fully connected layer performs feature fusion and classification operations on the convolutional feature matrix to output the fault type of the rolling bearing.

[0009] Secondly, the present invention provides an intelligent fault diagnosis system for rolling bearings based on neural networks, comprising: The module consists of a sampling module, an adaptive Mel-filter preprocessing module, and a convolutional neural network diagnostic module; among which, The sampling module samples the vibration signal of the rolling bearing and outputs a sampling signal; The adaptive Mel filter preprocessing module performs resonance demodulation processing on the sampled signal to extract the frame-level feature vector sequence; The convolutional neural network diagnostic module performs multidimensional feature extraction, fusion, and classification operations on the frame-level feature vector sequence to obtain the fault type of the rolling bearing. The convolutional neural network diagnostic module stores the trained model parameters in BRAM in a fixed-point manner. The programmable logic unit (PL) implements the timing and parallel logic of the convolution and pooling layers, and the processing system unit (PS) implements the computation of the fully connected layers. The programmable logic unit (PL) and the processing system unit (PS) are interconnected through the AXI4 bus protocol to realize the interactive transmission of control commands and data streams. The intelligent diagnostic system for rolling bearing faults completes data acquisition and neural network front-end calculation tasks through the programmable logic terminal (PL), and uses a DMA core to transfer data to the DDR for subsequent calculations; the processing system terminal (PS) completes the calculation of the fully connected layer of the neural network, fault classification processing, and overall control management.

[0010] In one embodiment of the present invention, the adaptive Mel filter preprocessing module includes: The module includes a frame-splitting and windowing module, an FFT submodule, and a Mel-filtering submodule; among which... The frame segmentation and windowing module performs frame segmentation preprocessing on the sampled signal to obtain several adjacent frame signals, and performs window function weighting processing on each adjacent frame signal to obtain the windowed frame signal. The FFT submodule performs a fast Fourier transform on the windowed frame signal to obtain the corresponding frequency domain representation; The Mel filtering submodule performs feature filtering and dimensionality reduction on the energy spectrum obtained by taking the modulus and squaring the frequency domain representation, and outputs a frame-level feature vector sequence.

[0011] Thirdly, the present invention provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; The memory is used to store computer programs; When the processor executes the program stored in the memory, it implements the steps of the intelligent diagnosis method for rolling bearing faults based on neural networks provided by the present invention.

[0012] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the steps of a neural network-based intelligent diagnosis method for rolling bearing faults provided in the embodiments of the present invention.

[0013] The beneficial effects of this invention are: The solution provided by this invention uses a pre-built adaptive Mel filter bank to perform resonant demodulation processing on the sampled signal to enhance fault characteristics; it utilizes a neural network diagnostic model to achieve feature extraction and classification, effectively suppressing background noise interference, improving diagnostic accuracy, and supporting the identification of various bearing faults. This invention completes data acquisition and neural network front-end computation tasks through the programmable logic unit (PL), and uses a DMA core to move data to DDR for subsequent computation. The processing system (PS) performs computation of the fully connected layer of the neural network, fault classification processing, and overall control management. The overall system architecture achieves low power consumption and high accuracy edge diagnostic capabilities, significantly reduces neural network inference latency, supports edge deployment, and is suitable for industrial equipment operation status monitoring scenarios. Attached Figure Description

[0014] Figure 1 This is a schematic diagram illustrating the steps of a neural network-based intelligent fault diagnosis method for rolling bearings provided in an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the principle of a neural network-based intelligent fault diagnosis system for rolling bearings provided in an embodiment of the present invention. Figure 3 A flowchart of resonance demodulation processing in a neural network-based intelligent diagnostic system for rolling bearing faults, provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of a convolutional neural network diagnostic module in a rolling bearing fault intelligent diagnostic system based on a neural network, provided in an embodiment of the present invention. Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0015] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0016] This invention provides a method and system for intelligent diagnosis of rolling bearing faults based on neural networks.

[0017] It should be noted that the executing entity of the intelligent diagnosis method for rolling bearing faults based on neural networks provided in this embodiment of the invention can be an intelligent diagnosis system for rolling bearing faults based on neural networks, and the system can run in an electronic device. This electronic device can be a server or a terminal device, but is not limited to these.

[0018] Below, we will first introduce a method for intelligent diagnosis of rolling bearing faults based on neural networks provided in the embodiments of the present invention.

[0019] The present invention provides an intelligent fault diagnosis method for rolling bearings based on neural networks, such as... Figure 1 As shown, it may include the following steps: S1 samples the vibration signal of the rolling bearing and outputs the sampled signal.

[0020] The embodiments of the present invention can perform synchronous AD acquisition based on AD7606 to achieve multi-channel synchronous sampling of vibration signals.

[0021] S2, through a pre-constructed adaptive Mel filter bank, performs resonant demodulation processing on the sampled signal to extract a frame-level feature vector sequence, which may include: S21, perform frame preprocessing on the sampled signal to obtain several adjacent frame signals; S22, perform window function weighting on each adjacent frame signal to obtain the windowed frame signal; S23, Perform a fast Fourier transform on the windowed frame signal to obtain the corresponding frequency domain representation; S24, take the modulus of the frequency domain representation and square it to obtain the corresponding energy spectrum; S25 uses an adaptive Mel filter bank to perform feature filtering and dimensionality reduction on the energy spectrum, and outputs a frame-level feature vector sequence.

[0022] Specifically, regarding steps S21 and S22, the intelligent diagnostic method for rolling bearing faults proposed in this embodiment of the invention analyzes the acquired vibration signals of the noisy rolling bearing. Frame preprocessing is performed to divide the continuous time signal into several adjacent frames, with a partial overlap set between each pair of adjacent frames, preferably 50%.

[0023] For the The adjacent frame signals are weighted by applying a window function to reduce the inter-frame truncation effect, resulting in the windowed frame signal. : ; in, For the selected window function, the Hanning window can be chosen, with its domain being... .

[0024] The above windowing and framing process can effectively suppress spectral leakage and provide a stable frame-level signal input for subsequent feature filtering and dimensionality reduction analysis.

[0025] For steps S23 and S24, the windowed frame signal Perform a Fast Fourier Transform to obtain the frequency domain representation of the signal in that frame. : ; in, Represents a spectrum signal in complex form.

[0026] Subsequently, the energy spectrum is obtained by taking the modulus square of the frequency domain representation. : ; in, It represents the energy distribution of the signal across its frequency components.

[0027] This step converts the time-domain signal into frequency-domain energy features, providing input data for subsequent adaptive filtering and feature extraction.

[0028] Regarding step S25, the energy spectrum obtained in step S24 is... The input is fed into an adaptive Mel filter bank for feature filtering and dimensionality reduction.

[0029] Specifically, the intelligent fault diagnosis method for rolling bearings proposed in this embodiment of the invention performs a fast spectral kurtosis algorithm on the vibration signal of the rolling bearing to achieve rapid estimation of the spectral kurtosis of the signal, which is used to select the optimal fault feature frequency band and provide filtering parameters for the subsequent adaptive Mel filter bank.

[0030] The specific steps are as follows: First, a low-pass analysis filter is constructed, with a cutoff frequency of... It can be set as follows: .

[0031] Based on this, a quasi-analytical low-pass filter and a quasi-analytical high-pass filter are further constructed to process the frequency band. and This completes a basic low-pass / high-pass band decomposition.

[0032] The low-pass-high-pass filter decomposition employs a pyramid structure with iterative steps. At each decomposition level, the input signal is processed simultaneously by both a low-pass and a high-pass filter, followed by a 2x downsampling operation on the filtered output, progressively dividing the signal into sub-bands with different center frequencies and bandwidths. Through multi-level iterative decomposition, a complete tree-like filter bank structure can be constructed, with each node in the tree corresponding to an independent frequency band sub-signal.

[0033] For each sub-band signal in the tree, calculate its analytic complex envelope. And based on the envelope, the spectral kurtosis is defined as follows: ; in, This represents the spectral kurtosis value corresponding to the sub-band signal, used to measure the significance of impulse components in that frequency band. By calculating the spectral kurtosis index for all sub-bands in the filter tree and comparing the magnitudes of the spectral kurtosis of each sub-band, the sub-band node with the largest spectral kurtosis value is selected, and the position of this maximum value is chosen. ; The corresponding center frequency and bandwidth are determined as the optimal fault characteristic frequency band of the vibration signal, providing filtering parameters for the subsequent adaptive Mel filter bank.

[0034] The adaptive Mel filter bank consists of several triangular bandpass filters; the center frequency and bandwidth of each triangular bandpass filter are adaptively determined based on preset fault-related frequency band parameters. The dimension of the frame-level feature vector sequence is the same as the number of filters in the adaptive Mel filter bank.

[0035] The frequency distribution density function is defined as follows: ; In the above formula, This indicates the center frequency of the frequency band with the highest spectral kurtosis value. Indicates bandwidth control parameters. This represents the fault frequency band density coefficient.

[0036] The normalized mapping relationship is obtained by integrating the density function: ; exist Selected at equal intervals on the axis N sampling points ( N (Number of filters): ; Based on the above sampling points, by adjusting the function... Perform an inverse transform to obtain the frequency value corresponding to each point. : ; This results in a higher density of filter distribution in the fault frequency band.

[0037] Construct the frequency response of the triangular filter, and analyze the frequency response of each filter. Defined as: ; Through this step, the embodiments of the present invention can achieve adaptive enhancement of the energy of the fault frequency band, while mapping the original high-dimensional spectrum data into a low-dimensional feature vector, thereby achieving effective feature filtering and dimensionality reduction.

[0038] S3 utilizes a pre-trained neural network diagnostic model to perform multi-dimensional feature extraction, fusion, and classification operations on the frame-level feature vector sequence to obtain the fault type of the rolling bearing; among which, The adaptive Mel filter bank is constructed by automatically selecting the resonant bandwidth parameters based on the fast spectral kurtosis algorithm.

[0039] The frame-level feature vector sequence obtained in step S2 is used as the output feature of the Mel filter bank. The number of filters can be 30, and the number of frames can be 187. The 30-dimensional features corresponding to each frame are arranged in chronological order to construct a 187×30 two-dimensional feature map, which serves as the input to the neural network diagnostic model. A neural network diagnostic model (1D-CNN) is constructed to achieve deep nonlinear mapping of features and fault mode recognition.

[0040] Neural network diagnostic models may include: The input layer, the first convolutional layer, the multi-level residual network, and the fully connected layer are arranged sequentially; among them, The input layer transmits the frame-level feature vector sequence to the first convolutional layer; The first convolutional layer extracts the local relevant features corresponding to the received frame-level feature vector sequence in the time dimension and expands the feature channel dimension; Multi-level residual networks perform multi-scale feature extraction and deep representation learning on locally correlated features, and output convolutional feature matrices. The fully connected layer performs feature fusion and classification operations on the convolutional feature matrix to output the fault type of the rolling bearing.

[0041] For the input layer, the network input is a sequence of frame-level feature vectors extracted by the adaptive Mel filter bank, and the number of input channels can be 30.

[0042] The first convolutional layer employs a one-dimensional convolution operation with a kernel size of 3×1, a stride of 1, 30 input channels, and 64 output channels. This convolutional layer is used to extract locally relevant features in the temporal dimension and expand the feature channel dimension.

[0043] For multi-level residual networks, the output of the first convolutional layer passes through three residual structures in sequence. Each residual block is composed of two one-dimensional convolutional layers connected sequentially, as follows: In each residual block, the kernel size of the two convolutional layers is 3×3; Each convolutional layer has 64 output channels; The residual block uses identity mapping, which adds the input features to the output of two convolutional layers and then uses the sum as the module output.

[0044] This invention, through its multi-level residual network, enables multi-scale feature extraction and deep representation learning of input features, while avoiding network degradation and improving the robustness and discriminative power of fault features.

[0045] The convolutional feature matrix output by the multi-level residual network is fed into the fully connected layer to achieve fault category discrimination.

[0046] After calculation by a three-layer residual network, the resulting feature output size is 187×64. This two-dimensional feature matrix is ​​then flattened into a one-dimensional feature vector and input into a fully connected layer for feature fusion and classification operations. The number of output nodes in this fully connected layer is set to 10, corresponding to 10 different rolling bearing fault types.

[0047] The intelligent diagnostic method for rolling bearing faults proposed in this invention uses a pre-constructed adaptive Mel filter bank to perform resonance demodulation processing on the sampled signal to enhance fault characteristics; it utilizes a neural network diagnostic model to achieve feature extraction and classification, effectively suppressing background noise interference, improving diagnostic accuracy, and supporting the identification of various bearing faults.

[0048] Secondly, corresponding to the above method embodiments, this invention also provides a rolling bearing fault intelligent diagnosis system based on neural networks, such as... Figure 2 As shown, it may include: The module consists of a sampling module, an adaptive Mel-filter preprocessing module, and a convolutional neural network diagnostic module; among which, The sampling module samples the vibration signal of the rolling bearing and outputs the sampled signal; The adaptive Mel filter preprocessing module performs resonance demodulation processing on the sampled signal to extract the frame-level feature vector sequence; The convolutional neural network diagnostic module performs multidimensional feature extraction, fusion, and classification operations on the frame-level feature vector sequence to obtain the fault type of the rolling bearing. The convolutional neural network diagnostic module stores the trained model parameters in BRAM in a fixed-point manner. The programmable logic unit (PL) implements the timing and parallel logic of the convolution and pooling layers, and the processing system unit (PS) implements the computation of the fully connected layers. The PL and PS are interconnected through the AXI4 bus protocol to realize the interactive transmission of control commands and data streams. The intelligent diagnostic system for rolling bearing faults completes data acquisition and neural network front-end calculation tasks through the programmable logic unit (PL), and uses a DMA core to transfer data to the DDR for subsequent calculations; the processing system (PS) completes the calculation of the fully connected layer of the neural network, fault classification and processing, and overall control management.

[0049] This invention allows the deployment of a trained neural network diagnostic model to a Zynq-based hardware and software collaborative platform. This deployment structure includes a processing system (PS, ARM) and a programmable logic (PL, FPGA), interconnected via the AXI4 bus protocol to enable interactive transmission of control commands and data streams. The PS and PL are connected via an AXI4 interface. The AXI-Lite bus transmits control data, including model startup commands, configuration parameters, and data transfer instructions; the AXI-Stream interface handles high-throughput data stream transmission; and the DMA (Direct Memory Access) controller streams input data, convolutional weight parameters, and adaptive Mel filter parameters stored in DDR memory to various computational modules in the PL. The convolutional weight parameters, adaptive Mel filter parameters, and input / output buffers are all stored in the PS's DDR memory. The final fully connected layer computation and classification result output are also completed on the PS.

[0050] Understandably, the PL (Programmer) module is primarily responsible for implementing the adaptive Mel filter preprocessing module and the convolutional neural network diagnostic module; the flowchart for the resonance demodulation process is as follows: Figure 3 As shown, the adaptive Mel filter preprocessing module may include: The module includes a frame-splitting and windowing module, an FFT submodule, and a Mel-filtering submodule; among which... The frame segmentation and windowing module performs frame preprocessing on the sampled signal to obtain several adjacent frame signals. The adjacent frame signals are then weighted by a window function to obtain the windowed frame signal. The FFT submodule performs a fast Fourier transform on the windowed frame signal to obtain the corresponding frequency domain representation; The Mel filtering submodule performs feature filtering and dimensionality reduction on the energy spectrum obtained by taking the modulus and squaring the frequency domain representation, and outputs a frame-level feature vector sequence.

[0051] The framing and windowing submodule implements sliding window framing through a multiplexer (MUX) and register array; the windowing operation is performed by storing Hanning window coefficients in the ROM core and multiplying them with the input signal through a multiplier.

[0052] The FFT submodule performs a Fast Fourier Transform operation on each frame of signal.

[0053] The FFT module can adopt a 256-point pipelined architecture (Radix-2DIF FFT), with windowed time-domain data as input and complex spectrum data as output. The module internally performs butterfly operations through a multiply-accumulate (MAC) array, implemented using Xilinx's FFT IP core.

[0054] The energy spectrum signal output by the FFT is input to the Mel filtering submodule. The Mel filtering submodule stores the odd and even triangular filter coefficients by two sets of ROM cores respectively. The input spectrum data and the filter coefficients are weighted and summed by a multiply-accumulate arithmetic unit array. The data is processed in a streaming manner, and the output is 30-dimensional feature data of each frame after Mel filtering.

[0055] After processing by this module, the 187 frames of signals are generated into 30-dimensional feature vectors, which are stored in a two-dimensional BRAM array to form a 187×30 feature map matrix, which serves as the input to the convolutional neural network diagnostic module.

[0056] Understandably, the fast spectral kurtosis algorithm in MATLAB is used to automatically select the resonant bandwidth parameter, filter the vibration signal of the rolling bearing, perform envelope demodulation processing, and extract the frequency domain envelope signal. The results are then used to train a neural network model.

[0057] The structural diagram of the convolutional neural network diagnostic module is shown below. Figure 4 As shown, it mainly consists of a first-layer convolutional unit and a residual inference unit. To reduce hardware resource consumption and improve the module's versatility, this embodiment of the invention constructs a reusable single-layer convolutional computation module, and achieves temporal reuse of this module through state machine scheduling, thereby completing the sequential execution of multi-layer convolutional operations in the entire network.

[0058] A single-layer convolution computation module may include a weight storage BRAM, a data storage BRAM, and several convolution operators. Each convolution operator performs a single-channel convolution operation. Specifically, the convolution operator first reads a 3×1 convolution kernel weight from the weight storage BRAM and writes it into an internal register. Simultaneously, it reads the input data sequentially from the data storage BRAM and performs a three-level register delay processing on the read data, enabling a convolution computation to be completed in parallel using three multipliers in the fourth clock cycle. After completing one convolution computation, the convolution operator continues to read the next sampling point from the data storage BRAM to achieve continuous sliding of the convolution kernel window. By configuring the convolution operators in multi-instance parallelization, synchronous convolution processing of multi-channel input data can be achieved.

[0059] The output of the convolution operation is written to the buffer BRAM and used as input data for the next convolutional unit. When the computation of a certain convolutional layer is completed, the convolution module sends a convolution completion indication signal to the AXI-Lite interface. After receiving the signal, the PS end writes the weight parameters required for the next convolution to the weight storage BRAM (Block Random Access Memory) of the PL end via DMA to start the subsequent convolution computation.

[0060] After the first convolutional layer calculation and the subsequent three residual convolutional layers are completed sequentially, the final feature mapping data is written into the output FIFO (First In First Out). Subsequently, the PS end uses a DMA kernel to transfer the inference result data in the FIFO to the DDR (Double Data Rate Synchronous Dynamic Random Access Memory) at high speed via AXI-Stream for use in subsequent fully connected layer calculations and fault classification.

[0061] The intelligent rolling bearing fault diagnosis system proposed in this invention operates on the ZYNQ7020 platform. The AD7606 is responsible for analog signal acquisition; the PL (Programmable Logic Unit) handles convolution and pooling operations and sends the results to the DDR (Data Transfer Unit) via DMA; the PS (Power Supply Unit) executes the fully connected layer and classification decisions; the entire system can achieve real-time fault diagnosis at the edge. The PL completes data acquisition and neural network front-end computation tasks, and uses the DMA core to transfer data to the DDR for subsequent computation. The PS at the processing system end performs neural network fully connected layer computation, fault classification processing, and overall control management. The overall system architecture achieves low power consumption and high accuracy edge diagnostic capabilities, significantly reduces neural network inference latency, supports edge deployment, and is suitable for industrial equipment operation status monitoring scenarios.

[0062] Thirdly, embodiments of the present invention also provide an electronic device, such as... Figure 5 As shown, it includes a processor 001, a communication interface 002, a memory 003, and a communication bus 004, wherein the processor 001, the communication interface 002, and the memory 003 communicate with each other through the communication bus 004. The memory is used to store computer programs; When the processor executes the program stored in the memory, it implements the steps of any of the intelligent diagnosis methods for rolling bearing faults based on neural networks provided in the first aspect of the present invention.

[0063] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0064] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0065] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0066] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0067] The method provided in this invention can be applied to electronic devices. Specifically, the electronic device can be a desktop computer, a portable computer, a smart mobile terminal, a server, etc. No limitation is made herein; any electronic device that can implement this invention falls within the protection scope of this invention.

[0068] Fourthly, corresponding to the intelligent diagnosis method for rolling bearing faults based on neural networks provided in the first aspect, this embodiment of the invention also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the steps of any of the intelligent diagnosis methods for rolling bearing faults based on neural networks provided in the first aspect of this invention.

[0069] For system / electronic device / storage medium embodiments, since they are basically similar to method embodiments, the description is relatively simple, and relevant parts can be referred to in the description of the method embodiments.

[0070] It should be noted that the device, electronic device and storage medium in the embodiments of the present invention are respectively the system, electronic device and storage medium applying the above method. Therefore, all embodiments of the above-mentioned intelligent diagnosis method for rolling bearing faults based on neural networks are applicable to the system, electronic device and storage medium, and can achieve the same or similar beneficial effects.

[0071] It should be noted that, in the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0072] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0073] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A method for intelligent diagnosis of rolling bearing faults based on neural networks, characterized in that, include: The vibration signal of the rolling bearing is sampled and the sampled signal is output. The sampled signal is subjected to resonant demodulation processing using a pre-constructed adaptive Mel filter bank to extract a frame-level feature vector sequence. A pre-trained neural network diagnostic model is used to perform multi-dimensional feature extraction, fusion, and classification operations on the frame-level feature vector sequence to obtain the fault type of the rolling bearing; wherein, The adaptive Mel filter bank is constructed by automatically selecting the resonant bandwidth parameters based on the fast spectral kurtosis algorithm.

2. The intelligent fault diagnosis method for rolling bearings based on neural networks according to claim 1, characterized in that, The step of performing resonant demodulation processing on the sampled signal using a pre-constructed filter bank to extract a frame-level feature vector sequence includes: The sampled signal is subjected to frame preprocessing to obtain several adjacent frame signals; The adjacent frame signals are weighted by a window function to obtain the windowed frame signal. Perform a Fast Fourier Transform on the windowed frame signal to obtain the corresponding frequency domain representation; Taking the modulus of the frequency domain representation and squaring it yields the corresponding energy spectrum; The energy spectrum is subjected to feature filtering and dimensionality reduction using an adaptive Mel filter bank, and a frame-level feature vector sequence is output.

3. The intelligent fault diagnosis method for rolling bearings based on neural networks according to claim 1, characterized in that, The adaptive Mel filter bank consists of several triangular bandpass filters; the center frequency and bandwidth of each triangular bandpass filter are adaptively determined based on preset fault-related frequency band parameters.

4. The intelligent fault diagnosis method for rolling bearings based on neural networks according to claim 1, characterized in that, The dimension of the frame-level feature vector sequence is the same as the number of filters in the adaptive Mel filter bank.

5. The intelligent fault diagnosis method for rolling bearings based on neural networks according to claim 1, characterized in that, The neural network diagnostic model includes: The input layer, the first convolutional layer, the multi-level residual network, and the fully connected layer are arranged sequentially; among them, The input layer transmits the frame-level feature vector sequence to the first convolutional layer; The first convolutional layer extracts the local relevant features corresponding to the received frame-level feature vector sequence in the time dimension and expands the feature channel dimension; The multi-level residual network performs multi-scale feature extraction and deep representation learning on locally related features, and outputs a convolutional feature matrix. The fully connected layer performs feature fusion and classification operations on the convolutional feature matrix to output the fault type of the rolling bearing.

6. A neural network-based intelligent fault diagnosis system for rolling bearings, characterized in that, include: The module consists of a sampling module, an adaptive Mel-filter preprocessing module, and a convolutional neural network diagnostic module; among which, The sampling module samples the vibration signal of the rolling bearing and outputs a sampling signal; The adaptive Mel filter preprocessing module performs resonance demodulation processing on the sampled signal to extract the frame-level feature vector sequence; The convolutional neural network diagnostic module performs multidimensional feature extraction, fusion, and classification operations on the frame-level feature vector sequence to obtain the fault type of the rolling bearing. The convolutional neural network diagnostic module stores the trained model parameters in BRAM in a fixed-point manner. The programmable logic unit (PL) implements the timing and parallel logic of the convolution and pooling layers, and the processing system unit (PS) implements the computation of the fully connected layers. The programmable logic unit (PL) and the processing system unit (PS) are interconnected through the AXI4 bus protocol to realize the interactive transmission of control commands and data streams. The intelligent diagnostic system for rolling bearing faults completes data acquisition and neural network front-end calculation tasks through the programmable logic terminal (PL), and uses a DMA core to transfer data to the DDR for subsequent calculations; the processing system terminal (PS) completes the calculation of the fully connected layer of the neural network, fault classification processing, and overall control management.

7. The intelligent fault diagnosis system for rolling bearings based on neural networks according to claim 6, characterized in that, The adaptive Mel filter preprocessing module includes: The module includes a frame-splitting and windowing module, an FFT submodule, and a Mel-filtering submodule; among which... The frame segmentation and windowing module performs frame segmentation preprocessing on the sampled signal to obtain several adjacent frame signals, and performs window function weighting processing on each adjacent frame signal to obtain the windowed frame signal. The FFT submodule performs a fast Fourier transform on the windowed frame signal to obtain the corresponding frequency domain representation; The Mel filtering submodule performs feature filtering and dimensionality reduction on the energy spectrum obtained by taking the modulus and squaring the frequency domain representation, and outputs a frame-level feature vector sequence.

8. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; The memory is used to store computer programs; When the processor executes the program stored in the memory, it implements the steps of the intelligent diagnosis method for rolling bearing faults based on neural networks as described in any one of claims 1-5.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the intelligent diagnosis method for rolling bearing faults based on neural networks as described in any one of claims 1-5.

Images