Non-contact noise and current combined speed reducer bearing fault detection system and method

Abstract

The invention discloses a non-contact noise and current combined speed reducer bearing fault detection system and method. The system is characterized in that a signal acquisition module, a multi-channel data acquisition module, a data processing module and a cloud collaborative analysis module are connected in an electric loop mode to form a circuit. The signal acquisition module adopts a Hall sensor and a noise sensor to acquire current and noise signals respectively, and synchronous processing is realized through a multi-channel data acquisition card; the data processing module is based on an STM32F767IGT6 chip and supports real-time data acquisition, conversion, transmission and uploading; the fault analysis module adopts a deep learning model to carry out cross-modal feature fusion on noise and current signals, realizes multi-level adaptive fusion of time-frequency features, and carries out fault identification. The method is used for equipment state identification and fault diagnosis, forms an efficient closed-loop monitoring system, is widely applied to fault detection of complex industrial equipment such as a large speed reducer, improves the operation and maintenance management level of the industrial equipment, and has the characteristics of high precision, strong real-time performance and wide adaptability.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial equipment fault diagnosis, and specifically relates to a non-contact noise current combined reducer bearing fault detection system and method. Background Art

[0002] Since the 1980s, domestic and foreign scholars have carried out research on bearing fault diagnosis by means of noise signal analysis, developing from early simple signal decomposition to current signal envelope analysis. The mutation of bearing noise signals usually indicates faults, and signals are collected by acceleration sensors and methods such as Fourier transform (FT) and wavelet transform (WT) are used to extract features. However, due to the non-stationary and non-linear characteristics of noise signals, and the fault signals being time-domain sudden spikes with a wide energy distribution, the fault characteristic frequencies are easily submerged by noise and low-frequency signals. For this reason, researchers have proposed time-frequency analysis methods such as wavelet packet transform (WPT), short-time Fourier transform (STFT), and Hilbert-Huang transform (HHT) to provide more detailed time-frequency characteristics, but these methods have insufficient analysis of the high-frequency part, which may lead to loss of fault information. Existing research is mostly based on simulation data, lacking a complete solution for actual application scenarios. Especially in complex equipment such as large reducers, multiple bearings, multiple noise sources, and strong noise interference pose great challenges to fault detection, and there is no domestic institution that has specifically carried out in-depth research on the bearing fault noise model of such equipment. Summary of the Invention

[0003] The purpose of the present invention is to propose a non-contact noise current combined reducer bearing fault detection system and method, which can distinguish the true from the false under the conditions of multiple bearings, multiple noise sources, and strong noise interference signals, timely discover fault information, and improve the safety and reliability of the operation of large reducer equipment.

[0004] The technical solution of the present invention is: a non-contact noise current combined reducer bearing fault detection system, which includes: a signal acquisition module, a multi-channel data acquisition module, a data processing module, and a cloud collaborative analysis module, and the signal acquisition module, the multi-channel data acquisition module, the data processing module, and the cloud collaborative analysis module are connected into a circuit in an electrical loop manner.

[0005] A non-contact noise current combined reducer bearing fault detection method, the signal acquisition module uses a Hall sensor and a noise sensor to collect current and noise signals respectively, filters and reduces noise by collecting the noise signals and current signals of the reducer bearing under different operating states, and the denoised signal x f(t) Perform the Hilbert-Huang Transform (HHT) to generate time-frequency features under different operating states, and the multi-channel data acquisition card realizes synchronous processing; the data processing module is based on the STM32F767IGT6 data processing chip, supporting real-time data acquisition, conversion, transmission and upload; the cloud collaborative analysis module uses a deep learning model to perform cross-modal feature fusion on noise and current signals, constructs a time-frequency dual-stream feature extraction network by combining time-domain feature processing with depthwise separable convolution and temporal dilated convolution, and introduces a pyramid cross-attention module with a dynamic gating mechanism (DG-PyCA) to achieve multi-level adaptive fusion of time-frequency features for fault diagnosis; the system extracts band energy, statistical features and spectral centroid / bandwidth through the Short-Time Fourier Transform (STFT), and combines the Hilbert-Huang Transform (HHT) to generate time-frequency features for equipment status recognition and fault diagnosis; in addition, the system provides an intelligent visualization interface and maintenance suggestions, supports remote data transmission and predictive maintenance, and forms an efficient closed-loop monitoring system.

[0006] Further, the data acquisition and data processing fusion steps are as follows:

[0007] Step1: Use Hall sensors and noise sensors to be responsible for collecting current signals and noise signals respectively; the current signal, as one of the key parameters for monitoring the power system, is converted into a digital signal by the Hall sensor for subsequent processing; at the same time, the noise signal is collected by the noise sensor, and the sensor converts the mechanical noise into digital data that can be processed;

[0008] Step2: After collecting the two signals, enter the multi-channel data acquisition card; the core function of this module is to synchronously collect signals from multiple sensors and achieve efficient data acquisition and management through a multi-channel structure, thereby improving the system's adaptability to complex working conditions;

[0009] Step3: The data processing unit is processed based on the STM32F767IGT6, and the entire system has functions of data acquisition, data conversion, data transmission and data upload; the system obtains data from the Hall sensor and the noise sensor in real time through the data acquisition module; next, the data conversion module performs necessary format conversion on the original signal to make it conform to the standard format for cloud processing; the data transmission module is responsible for uploading the converted data to the cloud through the industrial Ethernet or wireless communication module for further analysis and processing;

[0010] Step 4: The cloud collaborative analysis module adopts a hierarchical processing architecture, including a transmission control layer, a computing engine layer, and a visualization interaction layer; the transmission control layer manages multi-terminal data access through an adaptive traffic scheduling algorithm to achieve load balancing and link fault tolerance; the computing engine layer deploys a fault diagnosis model based on deep learning, and uses a convolutional neural network to perform cross-modal feature fusion on the noise spectrum and current harmonics to identify typical fault modes such as inner ring spalling, wear, and cracking of the bearing; the visualization interaction layer provides a multi-dimensional data display interface, supports real-time visualization of noise waveforms, current spectra, and fault probabilities, and generates a maintenance recommendation report through a human-computer interaction interface; the system realizes high-precision diagnosis while ensuring real-time performance through an edge-cloud collaborative computing strategy, providing reliable technical support for predictive maintenance of industrial equipment.

[0011] Furthermore, the steps for extracting the deep features of the noise data are as follows:

[0012] Step 1: The time-domain feature processing combines depthwise separable convolution and time-domain dilated convolution to construct a time-frequency dual-stream feature extraction network, and introduces a pyramid cross-attention module with a dynamic gating mechanism (DG-PyCA) to achieve multi-level adaptive fusion of time-frequency features. Among them, the depthwise separable convolution decomposes the traditional convolution into depth convolution (channel-wise) and pointwise convolution (1×1 convolution):

[0013]

[0014] where C is the number of input channels, C′ is the number of output channels, K is the convolution kernel size, and d is the dilation rate (a parameter of dilated convolution used to expand the receptive field);

[0015] Step 2: The hierarchical dilated convolution group uses dilated convolutions with different dilation rates in parallel to extract multi-scale time series features:

[0016]

[0017] where the dilation rate d ∈ {1, 2, 4}, l is the index of the convolutional layer, and W l (d) is the weight of the convolution kernel with a dilation rate of d;

[0018] Step 3: The wavelet-FFT joint transform extracts hierarchical frequency-domain features. The sliding window FFT dynamically adjusts the window length for multi-resolution analysis, and the wavelet decomposition is performed on the low frequency (approximation coefficient):

[0019]

[0020] where h is the low-pass filter, j is the decomposition level, and K is the convolution kernel size;

[0021] High frequency (detail coefficient):

[0022]

[0023] where g is the high-pass filter, j is the number of decomposition layers, and K is the convolution kernel size;

[0024] Step4: For adaptive frequency-domain feature fusion, the frequency-domain channel attention (FcaNet) is used to calculate the channel weights for the frequency-domain feature maps:

[0025] α C = σ(W · GAP(F C )) (5)

[0026] where σ is the Sigmoid function and W is the weight of the fully connected layer;

[0027] Perform multi-scale feature weighting to assign weights to feature maps of different resolutions:

[0028]

[0029] β j = Softmax.W j · GAP(F j ) / (7)

[0030] where J is the number of decomposition layers and W j is the learnable weight matrix;

[0031] Step5: Enhance the temporal modeling ability through TCN, and adopt frequency-domain adversarial training and time-frequency joint loss function. Ensure temporal causality through causal convolution to avoid leakage of future information, and then perform residual connection:

[0032] H out = LayerNorm(H conv + H in ) (8)

[0033] where H out is the output residual, H conv is the 1x1 convolution residual, and H in is the input residual;

[0034] Furthermore, the steps for current feature extraction are as follows:

[0035] Step1: Collect the current signal I(t) of the device through a current sensor, and then preprocess the signal to remove noise and DC offset. The process of removing the offset can be completed by calculating the mean value of the signal:

[0036]

[0037] where N is the number of sampling points; I(ti ) is the current value at the i-th time point; I dc is the DC component (mean value) of the signal, and then the DC component is removed to obtain the signal I′(t) after de-offset;

[0038] Step2: Perform short-time Fourier transform (STFT) on the preprocessed current signal I′(t); divide the signal into a spectrum of time-frequency joint distribution;

[0039] Framing and windowing: Divide the signal into frames of length L, with adjacent frames overlapping by M points, and apply a Hamming window to reduce spectral leakage:

[0040] x m [n] = I′(t)·w[n - mR], n = 0, 1, …, L - 1 (10)

[0041] where w[n] is the window function and R is the frame shift (R = L - M);

[0042] Fourier transform: Perform FFT on each frame of the signal to obtain the time-frequency matrix S(m, k):

[0043]

[0044] where m is the frame index and k is the frequency index, corresponding to the frequency is (f s is the sampling rate);

[0045] Step3: After obtaining the time-frequency matrix S(m, k), extract features; first perform frequency band energy division, divide the spectrum into multiple frequency bands (such as 0 - 100Hz, 100 - 200Hz, etc.), and calculate the energy mean value of each frequency band within the time range:

[0046]

[0047] where T is the total number of frames and b is the frequency band number;

[0048] Step4: Perform time-domain statistical features to extract statistics such as mean value, variance, and peak-to-peak value for the amplitude sequence of each frequency band; among them, the spectral centroid and bandwidth in characterizing the main frequency region where energy is concentrated are:

[0049]

[0050] Step5: Combine the energy of each frequency band, statistical features, and spectral centroid / bandwidth into a high-dimensional feature vector F. Use the principal component analysis (PCA) method for dimensionality reduction and maintain the main information of the data; the feature vector after PCA dimensionality reduction is F reduced .

[0051] Furthermore, the implementation process of the noise and current multi-feature fusion fault detection method is as follows:

[0052] Step1: Noise signal acquisition and preprocessing At the initial stage of the monitoring terminal deployment, it is necessary to collect the noise signals and current signals of the reducer bearings under different operating conditions, including: healthy state, bottom crack, broken tooth, tooth root crack, surface crack. The collected original signal x(t) may contain strong high-frequency noise. Therefore, a Butterworth low-pass filter is used for signal denoising, and the filtered signal x f (t) is calculated as follows:

[0053]

[0054] where, w c is the cut-off frequency, and n is the filter order;

[0055] Perform the Hilbert-Huang transform (HHT) on the denoised signal x f (t). Among them, HHT consists of Empirical Mode Decomposition (EMD) and Hilbert Transform (HT);

[0056] Step2: Decompose the signal into several Intrinsic Mode Functions (IMFs) through EMD:

[0057]

[0058] where, IMF i (t) is the i-th component, and r n (t) is the residual component;

[0059] Perform the Hilbert transform on the IMF components to calculate the instantaneous frequency:

[0060]

[0061] where, P.V. represents the Cauchy principal value integral;

[0062] Calculate the instantaneous amplitude A(t) and instantaneous frequency w(t) of the signal:

[0063]

[0064] Step3: Obtain the time-frequency spectrum HHT(w,t) of the signal through the Hilbert transform, and finally generate the time-frequency feature maps under five operating conditions;

[0065] The spectrogram in five states has strong distinguishability. These spectrograms are marked and used as the input of the deep convolutional network model. The input, fully connected layer, and output layer in the pre-trained model ResNet18 are replaced, and the main levels of the model are migrated. Finally, the training of the model is completed.

[0066] Step4: After the training is completed, the model is deployed to the host computer to receive the noise signal and current signal collected by the sensor in real time and convert them into HHT spectrograms. Classification and inference are performed through ResNet18 to obtain the real-time operating state S(t):

[0067] S(t) = argmaxP(y|X) (19)

[0068] where P(y|X) is the class probability distribution predicted by the model;

[0069] Step5: Finally, the system judges the current operating state of the device according to the classification result and provides fault warnings or maintenance suggestions, thus realizing intelligent bearing fault diagnosis.

[0070] The beneficial effects of the present invention are:

[0071] 1. By designing non-contact noise-current combined reducer bearing fault detection, the present invention strengthens the fast and real-time data detection of reducer bearing faults. The data of current and noise signals collected by the information acquisition module are synchronously processed through a multi-channel data acquisition card. Based on the STM32F767IGT6 data processing unit, real-time data acquisition, conversion, transmission, and uploading are carried out. At the same time, combined with cloud collaborative analysis, real-time visualization of noise waveforms, current spectra, and fault probabilities is provided to form a closed-loop monitoring system. Through the edge-cloud collaborative computing strategy, the system realizes high-precision diagnosis while ensuring real-time performance, providing reliable technical support for predictive maintenance of industrial equipment.

[0072] 2. The present invention adopts non-contact detection, avoiding the complexity of traditional contact sensor installation, reducing the interference to the operation of the equipment, and being applicable to harsh working conditions.

[0073] 3. The present invention designs deep extraction of noise data, which can capture weak signals of early bearing faults, such as local damage, cracks, etc., to avoid the expansion of faults. Non-contact detection does not require installing sensors on the equipment, avoiding interference with the operation of the equipment.

[0074] 4. The present invention designs a fault detection method that fuses multiple features of noise and current. The original signal is denoised using a Butterworth low-pass filter, and then the denoised signal is subjected to the Hilbert transform to obtain the time-frequency spectrum HHT(ω, t) of the signal, generating a time-frequency feature map, which is trained through a deep convolutional network model. The current and noise signals are received in real time and converted into HHT spectrograms, which can more accurately identify the fault type. The system determines the current operating state of the device based on the classification result and provides fault warnings or maintenance suggestions. It can more accurately identify the fault type. Through early fault warnings and accurate diagnoses, sudden failures leading to downtime can be avoided, production efficiency can be improved, and intelligent bearing fault diagnosis can be realized.

[0075] 5. In response to the current market demand, the present invention designs a non-contact noise-current combined fault detection method and system for a reduction gearbox bearing, aiming to enhance the ability to quickly and real-time detect faults in the reduction gearbox bearing. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 It is a schematic diagram of the hardware process of the present invention.

[0077] Figure 2 It is a schematic diagram of the deep feature extraction of the noise data of the present invention.

[0078] Figure 3 It is a schematic diagram of the current extraction of the present invention.

[0079] Figure 4 It is a schematic diagram of the fusion of noise and current features of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0080] The present invention will be further described in detail below with reference to the accompanying drawings.

[0081] As Figure 1 shown, the steps of data acquisition and data processing fusion are as follows:

[0082] Step1: A Hall sensor and a noise sensor are respectively responsible for collecting the current signal and the noise signal; the current signal, as one of the key parameters for monitoring the power system, is converted into a digital signal by the Hall sensor for subsequent processing; at the same time, the noise signal is collected by the noise sensor, and the sensor converts the mechanical noise into digital data that can be processed.

[0083] Step2: After the two signals are collected, they enter a multi-channel data acquisition card; the core function of this module is to synchronously collect signals from multiple sensors and achieve efficient data acquisition and management through a multi-channel structure, thereby improving the adaptability of the system to complex working conditions.

[0084] Step 3: The data processing unit is based on STM32F767IGT6 for processing. The entire system has functions of data acquisition, data conversion, data transmission, and data upload. The system obtains data from Hall sensors and noise sensors in real time through the data acquisition module. Next, the data conversion module performs necessary format conversions on the original signals to make them conform to the standard format for cloud processing. The data transmission module is responsible for uploading the converted data to the cloud through industrial Ethernet or wireless communication modules for further analysis and processing.

[0085] Step 4: The cloud collaborative analysis module adopts a hierarchical processing architecture, including a transmission control layer, a computing engine layer, and a visualization interaction layer. The transmission control layer manages multi-terminal data access through an adaptive traffic scheduling algorithm to achieve load balancing and link fault tolerance. The computing engine layer deploys a fault diagnosis model based on deep learning, and uses a convolutional neural network to perform cross-modal feature fusion on noise spectra and current harmonics to identify typical fault modes such as inner ring spalling, wear, and cracking of bearings. The visualization interaction layer provides a multi-dimensional data display interface, supports real-time visualization of noise waveforms, current spectra, and fault probabilities, and generates maintenance recommendation reports through a human-computer interaction interface. The system realizes high-precision diagnosis while ensuring real-time performance through an edge-cloud collaborative computing strategy, providing reliable technical support for predictive maintenance of industrial equipment.

[0086] As Figure 2 shown, the steps for deep feature extraction of the noise data are as follows:

[0087] Step 1: Time-domain feature processing combines depthwise separable convolution and temporal dilated convolution to construct a time-frequency dual-stream feature extraction network, and introduces a pyramid cross-attention module with a dynamic gating mechanism (DG-PyCA) to achieve multi-level adaptive fusion of time-frequency features. Among them, depthwise separable convolution decomposes traditional convolution into depth convolution (channel-wise) and pointwise convolution (1×1 convolution):

[0088]

[0089] where C is the number of input channels, C′ is the number of output channels, K is the convolution kernel size, and d is the dilation rate (a parameter of dilated convolution used to expand the receptive field);

[0090] Step 2: The hierarchical dilated convolution group uses dilated convolutions with different dilation rates in parallel to extract multi-scale temporal features:

[0091]

[0092] where the dilation rate d ∈ {1, 2, 4}, l is the convolution layer index, and W l (d) is the weight of the convolution kernel with a dilation rate of d.

[0093] Step 3: The wavelet-FFT joint transform is used to extract hierarchical frequency-domain features. The sliding window FFT dynamically adjusts the window length for multi-resolution analysis. Wavelet decomposition of the low frequency (approximation coefficient):

[0094]

[0095] where h is the low-pass filter, j is the decomposition level, and K is the convolution kernel size;

[0096] High frequency (detail coefficient):

[0097]

[0098] where g is the high-pass filter, j is the decomposition level, and K is the convolution kernel size;

[0099] Step 4: Adaptive frequency-domain feature fusion uses frequency-domain channel attention (FcaNet) to calculate the channel weights for the frequency-domain feature map:

[0100] α C = σ(W·GAP(F C )) (5)

[0101] where σ is the Sigmoid function and W is the weight of the fully connected layer;

[0102] Perform multi-scale feature weighting to assign weights to feature maps of different resolutions:

[0103]

[0104] β j = Softmax.W j ·GAP(F j ) / (7)

[0105] where J is the decomposition level and W j is the learnable weight matrix;

[0106] Step 5: Enhance the temporal modeling ability through TCN, and adopt frequency-domain adversarial training and time-frequency joint loss function. Ensure temporal causality through causal convolution to avoid leakage of future information, and then perform residual connection:

[0107] H out = LayerNorm(H conv + H in ) (8)

[0108] where H out is the output residual, H conv is the 1x1 convolution residual, and H in is the input residual.

[0109] As shown Figure 3 below, the steps for extracting the current characteristics are as follows:

[0110] Step1: Collect the current signal I(t) of the device through a current sensor and then preprocess the signal to remove noise and DC offset; the de-offset process can be completed by calculating the mean value of the signal:

[0111]

[0112] where N is the number of sampling points; I(t i ) is the current value at the i-th time point; I dc is the DC component (mean value) of the signal, and then the DC component is removed to obtain the de-offset signal I′(t);

[0113] Step2: Perform a short-time Fourier transform (STFT) on the preprocessed current signal I′(t). Divide the signal into a spectrum with a time-frequency joint distribution;

[0114] Frame division and windowing: Divide the signal into frames of length L, with adjacent frames overlapping by M points, and apply a Hamming window to reduce spectral leakage:

[0115] x m [n] = I′(t)·w[n - mR], n = 0, 1, …, L - 1 (10)

[0116] where w[n] is the window function and R is the frame shift (R = L - M);

[0117] Fourier transform: Perform an FFT on each frame of the signal to obtain the time-frequency matrix S(m, k):

[0118]

[0119] where m is the frame index and k is the frequency index, corresponding to the frequency is (f s is the sampling rate);

[0120] Step3: After obtaining the time-frequency matrix S(m, k), extract the features. First, perform frequency band energy division, divide the spectrum into multiple frequency bands (such as 0 - 100Hz, 100 - 200Hz, etc.), and calculate the energy mean value of each frequency band within the time range:

[0121]

[0122] where T is the total number of frames and b is the frequency band number;

[0123] Step 4: Perform time domain statistical features to extract the mean, variance, peak-to-peak value and other statistics of the amplitude sequence of each frequency band. The centroid and bandwidth of the spectrum in the main frequency region where the energy is concentrated are:

[0124]

[0125] Step 5: Combine the energy, statistical characteristics and spectrum centroid / bandwidth of each frequency band into a high-dimensional feature vector F. Use the principal component analysis (PCA) method to reduce the dimension and retain the main information of the data. The feature vector after PCA dimension reduction is F reduced .

[0126] like Figure 4 As shown, the implementation process of the noise and current multi-feature fusion fault detection method is as follows:

[0127] Step 1: Noise signal collection and preprocessing In the early stage of monitoring terminal deployment, it is necessary to collect noise signals and current signals of reducer bearings in different operating states, including: healthy state, bottom cracks, broken teeth, tooth root cracks, and surface cracks. The collected original signal x(t) may contain strong high-frequency noise, so a Butterworth low-pass filter is used to reduce the signal noise. The filtered signal x f (t) is calculated as follows:

[0128]

[0129] Among them, w c is the cutoff frequency, n is the filter order;

[0130] The denoised signal x f (t) Performing Hilber-Huang transform (HHT), where HHT consists of Empirical Mode Decomposition (EMD) and Hilbert Transform (HT);

[0131] Step 2: Decompose the signal into several intrinsic mode functions (IMFs) through EMD:

[0132]

[0133] Among them, the IMF i (t) is the i-th component, r n (t) is the residual component;

[0134] Perform Hilbert transform on the IMF component and calculate the instantaneous frequency:

[0135]

[0136] Among them, P.V represents the Cauchy principal value integral;

[0137] Calculate the instantaneous amplitude A(t) and instantaneous frequency w(t) of the signal:

[0138]

[0139] Step3: Obtain the time-frequency spectrum HHT(w, t) of the signal through the Hilbert transform, and finally generate the time-frequency characteristic diagrams under five operating states;

[0140] The time-frequency spectrum diagrams under the five states have strong distinguishability. Mark these time-frequency spectrum diagrams as the input of the deep convolutional network model; replace the input, fully connected layer, and output layer in the pre-trained model ResNet18, and transfer the main levels of the model. Finally, complete the training of the model;

[0141] Step4: After the training is completed, deploy the model to the host computer, receive the noise signal and current signal collected by the sensor in real time, and convert them into HHT spectrograms. Perform classification inference through ResNet18 to obtain the real-time operating state S(t):

[0142] S(t) = argmax P(y|X) (19)

[0143] where P(y|X) is the class probability distribution predicted by the model;

[0144] Step5: Finally, the system judges the current operating state of the device according to the classification result, and provides fault warnings or maintenance suggestions, so as to realize intelligent bearing fault diagnosis.

Claims

1. A non-contact noise current combined reducer bearing fault detection system, characterized in that The detection system includes: a signal acquisition module, a multi-channel data acquisition module, a data processing module, and a cloud collaborative analysis module. The signal acquisition module, multi-channel data acquisition module, data processing module, and cloud collaborative analysis module are connected into a circuit in an electrical loop manner.

2. A non-contact noise current combined reduction gear bearing fault detection method, characterized in that: The signal acquisition module uses a Hall sensor and a noise sensor to collect current and noise signals respectively. By collecting the noise signals and current signals of the reducer bearing under different operating states and performing filtering and noise reduction, the denoised signal x f (t) is subjected to the Hilbert-Huang transform (HHT) to generate time-frequency features under different operating states, and synchronous processing is realized by a multi-channel data acquisition card; the data processing module is based on the STM32F767IGT6 data processing chip, supporting real-time data acquisition, conversion, transmission and upload; The cloud collaborative analysis module uses a deep learning model to perform cross-modal feature fusion on noise and current signals. It constructs a time-frequency two-stream feature extraction network by combining time-domain feature processing with depthwise separable convolution and time-domain dilated convolution, and introduces a pyramid cross-attention module with a dynamic gating mechanism (DG-PyCA) to achieve multi-level adaptive fusion of time-frequency features for fault diagnosis. The system extracts band energy, statistical features, and spectral centroid / bandwidth through short-time Fourier transform (STFT), and generates time-frequency features in combination with Hilbert-Huang transform (HHT) for equipment status recognition and fault diagnosis. In addition, the system provides an intelligent visualization interface and maintenance suggestions, supports remote data transmission and predictive maintenance, and forms an efficient closed-loop monitoring system.

3. A non-contact noise current combined reducer bearing fault detection method according to claim 2, characterized in that The steps of data acquisition and data processing fusion are as follows: Step1: A Hall sensor and a noise sensor are respectively responsible for collecting current signals and noise signals. As one of the key parameters for monitoring the power system, the current signal is converted into a digital signal by the Hall sensor for subsequent processing. At the same time, the noise signal is collected by the noise sensor, and the sensor converts mechanical noise into digital data that can be processed. Step2: After the two signals are collected, they enter the multi-channel data acquisition card. The core function of this module is to synchronously collect signals from multiple sensors and achieve efficient data acquisition and management through a multi-channel structure, thereby improving the system's adaptability to complex working conditions. Step3: The data processing module is based on STM32F767IGT6 for processing. The entire system has functions of data acquisition, data conversion, data transmission, and data uploading. The system obtains data from the Hall sensor and the noise sensor in real time through the data acquisition module. Next, the data conversion module performs necessary format conversion on the original signal to make it conform to the standard format for cloud processing. The data transmission module is responsible for uploading the converted data to the cloud through an industrial Ethernet or a wireless communication module for further analysis and processing. Step4: The cloud collaborative analysis module adopts a hierarchical processing architecture, including a transmission control layer, a computing engine layer, and a visualization interaction layer. The transmission control layer manages multi-terminal data access through an adaptive traffic scheduling algorithm to achieve load balancing and link fault tolerance. The computing engine layer deploys a fault diagnosis model based on deep learning, uses a convolutional neural network to perform cross-modal feature fusion on noise spectra and current harmonics, and identifies typical fault modes such as inner ring spalling, wear, and cracking of bearings. The visualization interaction layer provides a multi-dimensional data display interface, supports real-time visualization of noise waveforms, current spectra, and fault probabilities, and generates a maintenance suggestion report through a human-computer interaction interface. The system uses an edge-cloud collaborative computing strategy to achieve high-precision diagnosis while ensuring real-time performance, providing reliable technical support for predictive maintenance of industrial equipment.

4. A non-contact noise current combined reducer bearing fault detection method according to claim 3, characterized in that The steps for extracting the deep features of the noise data are as follows: Step1: Time-domain feature processing combines depthwise separable convolution and temporal dilated convolution to construct a time-frequency two-stream feature extraction network, and introduces a pyramid cross-attention module with a dynamic gating mechanism (DG-PyCA) to achieve multi-level adaptive fusion of time-frequency features. Among them, depthwise separable convolution decomposes traditional convolution into depthwise convolution (channel-wise) and pointwise convolution (1×1 convolution): Where C is the number of input channels, C′ is the number of output channels, K is the convolution kernel size, and d is the dilation rate (the parameter of dilated convolution, used to expand the receptive field); Step2: The hierarchical dilated convolution group uses dilated convolutions with different dilation rates in parallel to extract multi-scale temporal features: where the dilation rate d ∈ {1, 2, 4}, l is the convolutional layer index, and W l (d) is the convolutional kernel weight with dilation rate d; Step3: The wavelet-FFT joint transform extracts hierarchical frequency-domain features. The sliding window FFT dynamically adjusts the window length for multi-resolution analysis, and the wavelet decomposition is performed on the low-frequency (approximate coefficient): Where h is the low-pass filter, j is the decomposition level, and K is the convolution kernel size; High-frequency (detail coefficient): Where g is the high-pass filter, j is the decomposition level, and K is the convolution kernel size; Step4: Adaptive frequency-domain feature fusion uses frequency-domain channel attention (FcaNet) to calculate the channel weights for the frequency-domain feature map: α C = σ(W·GAP(F C )) (5) Where σ is the Sigmoid function and W is the weight of the fully connected layer; Perform multi-scale feature weighting to assign weights to feature maps with different resolutions: β j = Softmax.W j ·GAP(F j ) / (7) where J is the number of decomposition levels, and W j is a learnable weight matrix; Step5: Enhance the temporal modeling ability through TCN, and adopt frequency-domain adversarial training and time-frequency joint loss function. Ensure temporal causality through causal convolution to avoid leakage of future information, and then perform residual connection: H out = LayerNorm(H conv + H in ) (8) Among them, H out is the output residual, H conv is the 1x1 convolutional residual, H in is the input residual.

5. A non-contact noise current combined reducer bearing fault detection method according to claim 3, characterized in that The steps for extracting the current features are as follows: Step1: Collect the current signal I(t) of the device through a current sensor, and then preprocess the signal to remove noise and DC offset. The process of removing the offset can be completed by calculating the mean value of the signal: where N is the number of sampling points; I(t i ) is the current value at the i-th time point; I dc is the DC component (mean value) of the signal, and then the DC component is removed to obtain the detrended signal I′(t); Step2: Perform a short-time Fourier transform (STFT) on the preprocessed current signal I′(t) to divide the signal into a spectrum of time-frequency joint distribution; Framing and windowing: Divide the signal into frames of length L, with adjacent frames overlapping by M points, and add a Hamming window to reduce spectral leakage: x m [n] = I′(t)·w[n - mR], n = 0, 1, …, L - 1 (10) Where w[n] is the window function and R is the frame shift (R = L - M); Fourier transform: Perform FFT on each frame of the signal to obtain the time-frequency matrix S(m,k): where m is the frame index and k is the frequency index, and the corresponding frequency is (f s is the sampling rate); Step3: After obtaining the time-frequency matrix S(m,k), then extract features; First, perform frequency band energy division, divide the spectrum into multiple frequency bands (such as 0 - 100Hz, 100 - 200Hz, etc.), and calculate the energy mean value of each frequency band within the time range: Where T is the total number of frames and b is the frequency band number; Step4: Perform time-domain statistical features to extract statistics such as mean, variance, and peak-to-peak value for the amplitude sequence of each frequency band; Where the spectral centroid and bandwidth are in the main frequency region representing energy concentration: Step 5: Combine the energy of each frequency band, statistical features, and spectral centroid / bandwidth into a high-dimensional feature vector F; use the principal component analysis (PCA) method for dimensionality reduction while preserving the main information of the data; the feature vector after dimensionality reduction by PCA is F reduced .

6. A non-contact noise current combined reducer bearing fault detection method according to claim 2, characterized in that: The implementation process of the multi-feature fusion fault detection method for noise and current is as follows: Step1: Noise signal acquisition and preprocessing At the initial stage of the monitoring terminal deployment, it is necessary to collect the noise signals and current signals of the reducer bearings under different operating conditions, including: healthy state, bottom crack, broken tooth, tooth root crack, surface crack. The collected original signal x(t) may contain strong high-frequency noise. Therefore, a Butterworth low-pass filter is used to reduce the signal noise, and the filtered signal x f (t) is calculated as follows: where w c is the cut-off frequency and n is the filter order; The denoised signal x f (t) is subjected to the Hilbert-Huang Transform (HHT), where HHT consists of Empirical Mode Decomposition (EMD) and Hilbert Transform (HT); Step 2: Decompose the signal into several Intrinsic Mode Functions (IMFs) through EMD: Among them, IMF i (t) is the i-th component, and r n (t) is the residual component; Perform Hilbert transform on the IMF components to calculate the instantaneous frequency: where P.V. represents the Cauchy principal value integral; Calculate the instantaneous amplitude A(t) and instantaneous frequency ω(t) of the signal: Step 3: Obtain the time-frequency spectrum HHT(ω, t) of the signal through Hilbert transform, and finally generate the time-frequency feature maps under five operating states; The time-frequency spectra under the five states have strong distinguishability. Mark these time-frequency spectra as the input of the deep convolutional network model; replace the input, fully connected layer, and output layer in the pre-trained model ResNet18, and transfer the main levels of the model. Finally, complete the training of the model; Step 4: After the training is completed, deploy the model to the host computer to receive the noise signal and current signal collected by the sensor in real time, and convert them into HHT spectrograms; perform classification inference through ResNet18 to obtain the real-time operating state S(t): S(t) = argmax P(y|X) (19) where P(y|X) is the class probability distribution predicted by the model; Step 5: Finally, the system judges the current operating state of the device according to the classification result and provides fault warnings or maintenance suggestions, so as to realize intelligent bearing fault diagnosis.

Images