Motor abnormal sound detection method and system based on contact acquisition and small sample learning

Abstract

The invention discloses a motor abnormal sound detection method and system based on contact acquisition and small sample learning, and the method comprises the steps: directly coupling a motor housing through a contact vibration sensor, collecting an original vibration signal, and generating an anti-interference vibration signal; inputting the anti-interference vibration signal into a nonlinear resonance enhancement module to generate an enhanced sound signal; performing wavelet packet decomposition on the enhanced sound signal, extracting a multi-scale frequency band energy entropy, and generating a motor abnormal sound feature matrix by combining singular value decomposition dimension reduction; on the basis of a dynamic weight distribution element learning algorithm, a small number of normal samples and abnormal samples are utilized to construct an abnormal sound classification model; and inputting the motor abnormal sound characteristic matrix into an abnormal sound classification model, detecting transient abnormality through a sliding window time sequence matching algorithm, and outputting an abnormal sound judgment result. According to the embodiment of the invention, high-precision and low-false-alarm motor abnormal sound detection under complex working conditions can be realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of motor detection, and in particular to a method and system for detecting abnormal noise of a motor based on contact acquisition and small sample learning. Background Art

[0002] In the field of motor operating status monitoring, abnormal noise detection is an important means of identifying early mechanical failures. Traditional detection methods mainly rely on acoustic microphones to collect sound signals, but they are easily interfered by environmental noise and have difficulty capturing low-frequency vibration characteristics. Although existing contact vibration sensors can directly couple with the motor housing to obtain vibration signals, they still face three key technical bottlenecks in complex industrial environments: first, the mutual coupling of environmental noise and the fundamental frequency vibration of the motor makes it difficult to extract weak abnormal noise characteristics; second, traditional signal processing methods are insensitive to transient abnormal responses and cannot effectively identify short-term impact-type abnormal noises; third, deep learning models usually require a large number of abnormal sample training, while the scarcity of fault samples in actual production leads to insufficient model generalization capabilities. Summary of the Invention

[0003] The purpose of the present invention is to provide a motor abnormal noise detection method and system based on contact acquisition and small sample learning to address the deficiencies in the prior art and to achieve high-precision, low-false-alarm motor abnormal noise detection under complex working conditions.

[0004] One embodiment of the present application provides a method for detecting abnormal noise in a motor based on contact acquisition and small sample learning, the method comprising:

[0005] The contact vibration sensor is directly coupled to the motor housing to collect the original vibration signal, and the sensor sampling parameters are dynamically adjusted based on the ambient noise spectrum to generate an anti-interference vibration signal.

[0006] The anti-interference vibration signal is input into a nonlinear resonance enhancement module, and an enhanced sound signal is generated by resonance amplification and fundamental frequency suppression processing of a preset abnormal sound characteristic frequency band;

[0007] Performing wavelet packet decomposition on the enhanced sound signal to extract multi-scale frequency band energy entropy, combining singular value decomposition with dimensionality reduction to generate a motor abnormal sound feature matrix;

[0008] Based on a dynamic weight allocation meta-learning algorithm, a noise classification model is constructed using a small number of normal and abnormal samples. The dynamic weight allocation meta-learning algorithm initializes network parameters through transfer learning and optimizes inter-class separability to generate a robust feature space mapping relationship.

[0009] The motor abnormal noise feature matrix is ​​input into the abnormal noise classification model, transient anomalies are detected through the sliding window timing matching algorithm, and the abnormal noise judgment result is output.

[0010] Optionally, the method of directly coupling a contact vibration sensor to the motor housing to collect an original vibration signal, dynamically adjusting a sensor sampling parameter based on an ambient noise spectrum, and generating an anti-interference vibration signal includes:

[0011] Deploy a three-axis vibration sensor array based on the geometric characteristics of the motor housing, synchronously collect multi-channel raw vibration signals, and generate initial time domain waveform data;

[0012] Perform fast Fourier transform on the initial time domain waveform data to extract the energy distribution of the dominant frequency band of the ambient noise and generate a noise spectrum feature map;

[0013] Based on the noise spectrum feature map, an adaptive notch filter bank is dynamically constructed, and the stopband frequency and bandwidth parameters are adjusted according to the real-time noise energy peak to generate a noise suppression filter;

[0014] The initial time domain waveform data is input into the noise suppression filter to filter out the environmental interference component, and then the sensor coupling delay is compensated through the phase calibration module to output the anti-interference vibration signal.

[0015] Optionally, the step of inputting the anti-interference vibration signal into a nonlinear resonance enhancement module and generating an enhanced sound signal through resonance amplification and fundamental frequency suppression processing in a preset abnormal sound characteristic frequency band includes:

[0016] Based on the motor model library, the fundamental frequency range parameters are loaded and a tunable band-stop filter is designed to filter out the 50-200Hz fundamental frequency component from the anti-interference vibration signal and generate a fundamental frequency suppression signal.

[0017] Perform continuous wavelet transform on the fundamental frequency suppression signal to locate the energy mutation area in the 2-5kHz frequency band and generate the energy distribution map of the abnormal sound characteristic frequency band;

[0018] Dynamically generate a nonlinear gain curve based on the energy distribution diagram, exponentially amplify the target frequency band signal, and generate a resonance enhancement signal;

[0019] The resonance enhancement signal and the fundamental frequency suppression signal are aligned and superimposed in the time domain, and the amplitude of the synthesized signal is constrained by the dynamic range control module to generate a preliminary enhancement signal;

[0020] Harmonic distortion detection and compensation are performed on the preliminary enhanced signal to eliminate artifacts introduced by nonlinear processing and output an enhanced sound signal with optimized signal-to-noise ratio.

[0021] Optionally, performing wavelet packet decomposition on the enhanced sound signal to extract multi-scale frequency band energy entropy, combining singular value decomposition with dimensionality reduction to generate a motor abnormal sound feature matrix, includes:

[0022] The enhanced sound signal is decomposed into 5 layers of wavelet packets using the db8 wavelet basis to generate a set of 32 orthogonal frequency band sub-signals.

[0023] Calculate the sliding window energy entropy of each frequency band sub-signal, count its probability distribution and quantify the uncertainty, and generate a multi-scale energy entropy sequence;

[0024] The multi-scale energy entropy sequence is arranged into an initial feature matrix by frequency band level, and the first k principal components are extracted by sliding window singular value decomposition to generate a reduced dimension feature matrix;

[0025] The reduced-dimensional feature matrix is ​​normalized to eliminate dimensional differences and output the standardized motor abnormal noise feature matrix.

[0026] Optionally, the dynamic weight allocation meta-learning algorithm is based on a small number of normal samples and abnormal samples to construct an abnormal noise classification model, wherein the dynamic weight allocation meta-learning algorithm initializes network parameters and optimizes inter-class separability through transfer learning to generate a robust feature space mapping relationship, including:

[0027] Based on the network structure of the pre-trained motor health status classification model, its convolutional layer and pooling layer parameters are loaded as feature extractors to generate the backbone network for transfer initialization.

[0028] Based on the feature vectors of normal and abnormal samples extracted by the backbone network, the inter-class cosine distance matrix is ​​calculated, the dynamic weight distribution coefficient is generated by entropy weighting, and the initial weight distribution matrix is ​​output;

[0029] The initial weight distribution matrix is ​​injected into the fully connected layer, and the weights of abnormal samples are optimized through the contrast loss function, forcing the abnormal features to stay away from the normal cluster centers in the weight space, thus generating a classifier prototype with a decision boundary.

[0030] Based on the decision boundary of the classifier prototype, a meta-learning optimizer is used to iteratively adjust the weight distribution coefficient on the support set, and the normal sample distribution range is compressed through the intra-class compactness loss function, and the optimized dynamic weight strategy is output;

[0031] The dynamic weight strategy is solidified into the fully connected layer of the classifier, and a small number of abnormal samples are used to fine-tune the output layer threshold parameters to generate the final abnormal noise classification model that can distinguish subtle abnormal noise patterns.

[0032] Optionally, the step of inputting the motor abnormal noise feature matrix into the abnormal noise classification model, detecting transient abnormalities through a sliding window timing matching algorithm, and outputting an abnormal noise determination result includes:

[0033] Adaptively segment the motor abnormal noise feature matrix stream according to the motor speed to generate sliding window data blocks with variable length;

[0034] Calculate the dynamic time-warping distance between each data block and the historical anomaly template to generate a window-level anomaly probability curve;

[0035] Detect the sections in the abnormal probability curve that continuously exceed the preset threshold, perform logical verification based on the judgment results of adjacent windows, and generate a list of candidate abnormal events;

[0036] According to the duration and probability peak intensity of the candidate abnormal event, a multi-level alarm strategy is triggered and the abnormal sound type and confidence score are output.

[0037] Another embodiment of the present application provides a motor abnormal noise detection system based on contact acquisition and small sample learning, the system comprising:

[0038] The acquisition module is used to directly couple the contact vibration sensor to the motor housing to collect the original vibration signal, dynamically adjust the sensor sampling parameters based on the ambient noise spectrum, and generate an anti-interference vibration signal;

[0039] An input module is used to input the anti-interference vibration signal into a nonlinear resonance enhancement module, and generate an enhanced sound signal through resonance amplification and fundamental frequency suppression processing of a preset abnormal sound characteristic frequency band;

[0040] A decomposition module is used to perform wavelet packet decomposition on the enhanced sound signal, extract multi-scale frequency band energy entropy, combine singular value decomposition with dimensionality reduction, and generate a motor abnormal sound feature matrix;

[0041] A construction module is configured to construct an abnormal sound classification model using a small number of normal samples and abnormal samples based on a dynamic weight allocation meta-learning algorithm, wherein the dynamic weight allocation meta-learning algorithm initializes network parameters and optimizes inter-class separability through transfer learning to generate a robust feature space mapping relationship;

[0042] The detection module is used to input the motor abnormal noise feature matrix into the abnormal noise classification model, detect transient anomalies through a sliding window timing matching algorithm, and output an abnormal noise judgment result.

[0043] Yet another embodiment of the present application provides a storage medium, wherein the storage medium stores a computer program, wherein the computer program is configured to execute any of the above methods when run.

[0044] Yet another embodiment of the present application provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute any of the above methods.

[0045] Compared with the existing technology, the present invention provides a motor abnormal noise detection method based on contact acquisition and small sample learning. The method directly couples the motor housing through a contact vibration sensor to collect the original vibration signal and generate an anti-interference vibration signal; the anti-interference vibration signal is input into a nonlinear resonance enhancement module to generate an enhanced sound signal; the enhanced sound signal is decomposed by wavelet packets to extract the multi-scale frequency band energy entropy, and combined with singular value decomposition dimensionality reduction, a motor abnormal noise feature matrix is ​​generated; based on the dynamic weight allocation meta-learning algorithm, a small number of normal samples and abnormal samples are used to construct an abnormal noise classification model; the motor abnormal noise feature matrix is ​​input into the abnormal noise classification model, and transient abnormalities are detected through a sliding window timing matching algorithm, and the abnormal noise judgment result is output, thereby realizing high-precision and low-false-alarm motor abnormal noise detection under complex working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A hardware structure block diagram of a computer terminal for a motor abnormal noise detection method based on contact acquisition and small sample learning provided by an embodiment of the present invention;

[0047] Figure 2 A schematic flow chart of a method for detecting abnormal noise in a motor based on contact acquisition and small sample learning provided by an embodiment of the present invention;

[0048] Figure 3 A schematic structural diagram of a motor abnormal noise detection system based on contact acquisition and small sample learning provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limiting the present invention.

[0050] The embodiment of the present invention first provides a method for detecting abnormal noise of a motor based on contact acquisition and small sample learning. The method can be applied to electronic devices such as computer terminals, specifically ordinary computers.

[0051] The following describes it in detail by taking running on a computer terminal as an example. Figure 1 The hardware structure block diagram of a computer terminal for a motor abnormal noise detection method based on contact acquisition and small sample learning provided by an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, a memory, and a network interface connected via a system bus, wherein the memory may include a non-volatile storage medium and an internal memory.

[0052] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions, which, when executed, can cause the processor to execute any one of the motor abnormal noise detection methods based on contact acquisition and small sample learning.

[0053] The processor is used to provide computing and control capabilities and support the operation of the entire computer equipment.

[0054] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor, the processor can execute any motor abnormal noise detection method based on contact acquisition and small sample learning.

[0055] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art will understand that Figure 1 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0056] It should be understood that the processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0057] See also Figure 2 The embodiment of the present invention provides a method for detecting abnormal noise of a motor based on contact acquisition and small sample learning, which may include the following steps:

[0058] S201, directly coupling the contact vibration sensor to the motor housing to collect the original vibration signal, dynamically adjusting the sensor sampling parameters based on the ambient noise spectrum, and generating an anti-interference vibration signal;

[0059] Specifically, a three-axis vibration sensor array can be deployed according to the geometric characteristics of the motor housing to synchronously collect multi-channel raw vibration signals and generate initial time domain waveform data;

[0060] The deployment of triaxial vibration sensors requires consideration of the motor housing's geometry and vibration propagation characteristics. For example, a cylindrical motor's housing typically consists of three parts: the front cover, the main cylinder, and the rear cover. Vibration energy propagates along different paths and intensities at different locations, so sensors must be deployed at key measurement points.

[0061] Sensor placement strategy

[0062] Front cover: Located near the motor shaft output, it primarily detects radial vibration caused by bearing wear or rotor imbalance. A three-axis sensor (labeled SensorA) is placed, with its X-axis (radial), Y-axis (axial), and Z-axis (tangential) aligned with the motor coordinate system.

[0063] In the middle of the main cylinder: Detect vibrations caused by the motor's overall housing resonance (resonance) and electromagnetic force. A second sensor (Sensor B) is placed with its three axes aligned with Sensor A.

[0064] Rear end cover: Detects abnormal vibration of the cooling fan or tail bearing. Place the third sensor (Sensor C) aligned with the first two.

[0065] Example parameters: sensor sensitivity is 100mV / g (g is the acceleration due to gravity), the range is ±50g, and the sampling rate is set to 10kHz (that is, 10,000 data points are collected per second).

[0066] Synchronous acquisition and signal preprocessing

[0067] Hardware synchronization: The three sensors are connected to a data acquisition (DAQ) card via synchronization cables to ensure time stamp alignment. For example, the NI-9234 acquisition card supports multi-channel synchronous sampling with a time base error of less than 1 microsecond.

[0068] Anti-aliasing filtering: Before sampling, an analog low-pass filter (Cutoff Frequency = 5kHz) is used to eliminate high-frequency noise and avoid aliasing effects.

[0069] Initial time-domain waveform data: Collect 9 channels of raw signals from the three sensors' X, Y, and Z axes and store them as a time-amplitude sequence. For example, the format for a single channel is Time = [0ms, 0.1ms, ..., 1000ms], Amplitude = [0.1V, 0.15V, ..., -0.2V].

[0070] Perform fast Fourier transform on the initial time domain waveform data to extract the energy distribution of the dominant frequency band of the ambient noise and generate a noise spectrum feature map;

[0071] Fast Fourier Transform (FFT) converts time domain signals into frequency domain energy distribution, which is used to identify the main frequency components of environmental noise.

[0072] FFT parameter setting and calculation

[0073] Segmentation processing: Split the continuous time domain signal into multiple time windows (windows). For example, a Hanning window is used with a window length of N = 1024 points (corresponding to a time length of 1024 / 10kHz = 102.4ms) and an overlap rate of 50%.

[0074] Spectral calculation: Perform an FFT on each window to obtain a complex spectrum. The power spectral density (PSD) is generated by squared amplitude. For example, the amplitude of the kth frequency component is |X[k]|², and the frequency resolution is Δf = 10kHz / 1024≈9.77Hz.

[0075] Ambient noise feature extraction: When the motor is not running, background noise signals are collected and the peak frequency band of the PSD is analyzed. For example, in a factory, the energy of the 800-1200Hz frequency band accounts for more than 60%, which is determined to be the dominant frequency band.

[0076] Noise spectrum feature map generation

[0077] Energy Accumulation Statistics: Averages multiple FFT results to plot a frequency-energy curve. For example, the frequency axis is 0-5kHz (Nyquist frequency), and the energy axis is a logarithmic scale (in dB).

[0078] Peak Marking: Detects frequencies where energy is significantly above the average level. For example, if a 30dB energy peak is detected at 950Hz, it is identified as the dominant frequency of the ambient noise. The frequency spectrum feature graph format is Frequency = [0Hz, 9.77Hz, ..., 5000Hz], Energy = [-50dB, -45dB, ..., 30dB].

[0079] Based on the noise spectrum feature map, an adaptive notch filter bank is dynamically constructed, and the stopband frequency and bandwidth parameters are adjusted according to the real-time noise energy peak to generate a noise suppression filter;

[0080] The adaptive notch filter suppresses the main frequency component of the noise by dynamically adjusting the stopband parameters.

[0081] Filter Bank Design and Parameter Adjustment

[0082] Stopband frequency setting: According to the peak frequency in the noise spectrum characteristic diagram (such as 950 Hz), the center frequency of the notch filter is set to f_center = 950 Hz.

[0083] Bandwidth adjustment: Define the stopband range based on the peak width. For example, if the peak is significant in the 900-1000Hz range (energy drops by 3dB), set the bandwidth BW to 100Hz and the stopband range to f_center ± BW / 2 = 900-1000Hz.

[0084] Filter order selection: Use a second-order IIR (infinite impulse response) filter with a transfer function of H(z) = (1-2cosθ·z -1 +z -2 ) / (1-2r·cosθ·z -1 +r2·z -2 ), where θ represents the digital angular frequency (dimensionless) corresponding to the center frequency, θ = 2πf_center / fs (fs is the sampling rate of 10 kHz), the damping coefficient r controls the bandwidth (the closer r is to 1, the narrower the bandwidth), and z represents the Z-transform complex variable (complex domain).

[0085] Real-time parameter update mechanism

[0086] Peak tracking algorithm: Recalculates the FFT every 10 seconds to detect the frequency point with the maximum energy in the current noise spectrum. For example, if the main frequency is newly detected to be offset to 980Hz, the value f_center = 980Hz is dynamically updated.

[0087] Bandwidth Adaptive Rule: If the energy of multiple adjacent frequencies exceeds a threshold (e.g., 25dB) within the same time period, the bandwidth is expanded to cover all exceeding frequencies. For example, if all frequencies between 970 and 990 Hz exceed the threshold, the bandwidth is set to 20 Hz (i.e., the stopband is 970-990 Hz).

[0088] The initial time domain waveform data is input into the noise suppression filter to filter out the environmental interference component, and then the sensor coupling delay is compensated through the phase calibration module to output the anti-interference vibration signal.

[0089] After the noise suppression filter removes interference in a specific frequency band, it is necessary to compensate for the signal phase delay to restore the true vibration waveform.

[0090] Filter execution and effect verification

[0091] Filter application: Sensor A's original X-axis signal is fed into a notch filter to remove the 950Hz noise component in real time. For example, the PSD of the filtered signal at 950Hz is reduced to -10dB (originally 30dB).

[0092] Multi-channel parallel processing: Independent filters are applied to the nine channels of the three sensors to ensure pure signals on each axis.

[0093] Phase calibration module design

[0094] Delay measurement: IIR filters introduce nonlinear phase delay, which must be compensated by an allpass filter. For example, if measurements show that the filter introduces a 2ms delay at 100Hz, an allpass filter can be designed to provide an opposite phase response.

[0095] Calibration execution: The filtered signal is passed through the phase calibration module to align the time domain characteristics of the original signal. For example, the peak time error of the calibrated signal is less than 0.05ms.

[0096] Output anti-interference vibration signal

[0097] Signal format: Stored as a time-amplitude sequence, with the same sampling rate as the original signal (10kHz). For example, after filtering, the X-axis signal amplitude range of SensorA is ±0.1V (originally ±0.3V), effectively eliminating noise components.

[0098] Quality assessment: The signal-to-noise ratio (SNR) was calculated. After filtering, the SNR increased from 15dB to 35dB, meeting the requirements for subsequent processing.

[0099] Key technology examples and parameter descriptions

[0100] Three-axis vibration sensor: can simultaneously measure vibration in three directions: X-axis (radial), Y-axis (axial), and Z-axis (tangential).

[0101] FFT (Fast Fourier Transform): Converts the time domain signal into the frequency domain energy distribution. The window length N=1024 means that 1024 data points are analyzed each time.

[0102] Notch filter stopband frequency (f_center): The main frequency of the noise to be suppressed, such as 950Hz.

[0103] Damping coefficient (r): controls the filter bandwidth. When r = 0.9, the bandwidth is narrow, and when r = 0.8, the bandwidth is wide.

[0104] Phase calibration: Eliminate the time delay introduced by the filter to ensure accurate signal time domain characteristics.

[0105] This step captures the time domain waveform data caused by mechanical vibration in real time by physically attaching a highly sensitive three-axis vibration sensor to the surface of the motor housing. During the signal acquisition process, the system continuously monitors the spectral distribution characteristics of the surrounding environmental noise (such as low-frequency electromagnetic interference or high-frequency mechanical resonance noise), and uses an adaptive algorithm to dynamically adjust the sensor's gain, sampling frequency, and filtering parameters, thereby effectively suppressing the interference of environmental noise on the core vibration signal. Phase calibration technology is used to compensate for the coupling delay between the sensor and the motor housing to ensure the high fidelity of the vibration signal. Direct contact acquisition avoids the problem of non-contact sensors being susceptible to air-conducted noise contamination, significantly improving the signal-to-noise ratio. The dynamic parameter adjustment mechanism can adaptively optimize the signal capture quality according to different working conditions, providing a high-precision data basis for subsequent abnormal sound feature extraction.

[0106] S202, inputting the anti-interference vibration signal into a nonlinear resonance enhancement module, generating an enhanced sound signal through resonance amplification and fundamental frequency suppression processing in a preset abnormal sound characteristic frequency band;

[0107] Specifically, the fundamental frequency range parameters can be loaded based on the motor model library to design a tunable band-stop filter to filter out the 50-200 Hz fundamental frequency component from the anti-interference vibration signal and generate a fundamental frequency suppression signal.

[0108] The base frequency is the dominant vibration frequency generated by the rotor's rotation and the periodic changes in the electromagnetic field during normal motor operation. It typically ranges from 50Hz to 200Hz (for example, a four-pole motor with a speed of 1500RPM has a base frequency of 50Hz). Filtering out the base frequency prevents it from masking high-frequency noise characteristics.

[0109] Motor model library matching and parameter loading

[0110] Model library structure: This stores parameters such as the model, number of poles, and rated speed of different motors. For example, a motor with four poles and a rated speed of 1500 RPM uses the formula for calculating the fundamental frequency (Hz) = speed (RPM) / 60 × number of pole pairs, or 1500 / 60 × 2 = 50 Hz.

[0111] Parameter call: Automatically load the base frequency range according to the motor model. For example, the base frequency range of one motor model is 50-120Hz, while that of another model is 80-200Hz.

[0112] Design of tunable band-stop filter

[0113] Center Frequency: Dynamically set based on the base frequency range. For example, for a motor with a base frequency of 50 Hz, set the filter center frequency to 50 Hz and the stopband range to 45-55 Hz (±5 Hz).

[0114] Filter Type: A finite impulse response (FIR) filter was selected because its linear phase characteristic prevents signal distortion. The filter order was set to 128 (i.e., 128 tap coefficients), ensuring a stopband attenuation of -40 dB (i.e., the fundamental frequency component energy was reduced to 1 / 10,000).

[0115] Dynamic tuning mechanism: If motor speed fluctuations cause the fundamental frequency to shift (e.g., ±5%), the filter center frequency automatically tracks and adjusts. For example, if the actual fundamental frequency is detected to be 52Hz, the stopband is updated to 49.4-54.6Hz.

[0116] Filter execution and effect verification

[0117] Input signal processing: The anti-interference vibration signal (sampling rate 10kHz) is input into the filter to remove the fundamental frequency component. For example, the energy of the original signal at 50Hz is 0.5V2 / Hz, which is reduced to 0.005V2 / Hz after filtering.

[0118] Baseband suppressed signal output: stored as a time domain waveform, retaining high-frequency components (such as 2-5kHz).

[0119] Perform continuous wavelet transform on the fundamental frequency suppression signal to locate the energy mutation area in the 2-5kHz frequency band and generate the energy distribution map of the abnormal sound characteristic frequency band;

[0120] Continuous Wavelet Transform (CWT) locates high-frequency transient abnormal noises through multi-scale analysis. Its advantage is that it has high local resolution in time and frequency, and is suitable for detecting non-stationary signal characteristics.

[0121] Wavelet basis selection and parameter setting

[0122] Mother Wavelet: Morlet wavelet (center frequency is 0.849 Hz) is selected because it is suitable for analyzing transient impact signals due to its balance in the time-frequency domain.

[0123] Scale range: corresponding to the target frequency band of 2-5kHz, according to the wavelet center frequency formula frequency (Hz) = center frequency / (scale × sampling interval), set the scale range to 5 to 20 (corresponding to 2kHz to 5kHz).

[0124] Step Size: The scale is divided into 100 levels with logarithmic intervals to ensure that the frequency domain resolution is sufficient to capture energy mutations.

[0125] Energy mutation detection and map generation

[0126] Time-Frequency Energy Calculation: Take the modulus square of the wavelet coefficients at each scale to obtain the time-frequency energy matrix. For example, if the modulus square of the wavelet coefficients at scale 10 (corresponding to 3kHz) at a certain time t is 0.2, it means that the energy in the 3kHz band is high at that time.

[0127] Energy mutation annotation: The energy variance of each frequency band is calculated using a sliding window (window length 100ms). Frequency bands with a variance exceeding a threshold (such as 0.1) are marked as energy mutation areas. For example, a variance of 0.15 in the 3.5kHz frequency band detected at timestamp 1.2s is identified as an abnormal energy area.

[0128] Energy distribution diagram output: the horizontal axis is time (0-10s), the vertical axis is frequency (2-5kHz), and the color depth represents the energy intensity.

[0129] Dynamically generate a nonlinear gain curve based on the energy distribution diagram, exponentially amplify the target frequency band signal, and generate a resonance enhancement signal;

[0130] The nonlinear gain curve dynamically adjusts the amplification factor according to the energy distribution, giving priority to enhancing high-frequency abnormal sound components while avoiding over-amplification of background noise.

[0131] Gain curve generation logic

[0132] Energy threshold setting: Based on historical normal sample statistics, set the energy baseline (Baseline Energy). For example, the average energy of a normal sample in the 3kHz frequency band is 0.05V2 / Hz, and the threshold is set to 3 times the baseline energy (0.15V 2 / Hz).

[0133] Gain calculation rules: If the energy at a frequency exceeds the threshold, the gain is amplified exponentially. For example, if the energy exceeds the threshold by 1 times, the gain is doubled (amplified to 200%), and if it exceeds the threshold by 2 times, the gain is quadrupled (amplified to 400%).

[0134] Smoothing: Perform mean filtering on the gains of adjacent frequency points (with a window width of 3 frequency points) to avoid artifacts introduced by sudden gain changes.

[0135] Signal amplification and synthesis

[0136] Frequency domain filtering and amplification: Apply a gain curve to the wavelet coefficient matrix to amplify the wavelet coefficients in the target frequency band (such as 3.5kHz). For example, the original coefficient of a certain frequency point is 0.1, and after amplification, it becomes 0.4.

[0137] Time-domain signal reconstruction: The amplified wavelet coefficients are converted to a time-domain signal using an inverse continuous wavelet transform (Inverse CWT), generating a resonance-enhanced signal. For example, the amplitude of the reconstructed signal at 3.5kHz is increased from 0.1V to 0.4V.

[0138] The resonance enhancement signal and the fundamental frequency suppression signal are aligned and superimposed in the time domain, and the amplitude of the synthesized signal is constrained by the dynamic range control module to generate a preliminary enhancement signal;

[0139] Time domain superposition ensures that the two sets of signals are phase-aligned, and dynamic range control (DRC) prevents the amplitude of the synthesized signal from exceeding the hardware processing range.

[0140] Signal alignment and overlay

[0141] Delay compensation: Calculate the cross-correlation function of the two signals and find the maximum correlation point to align the time difference. For example, if the resonance-enhanced signal is delayed by 2ms relative to the baseband-suppressed signal, shift it forward by 2ms to align it.

[0142] Overlay ratio: Allocate the overlay coefficient according to the energy weight. For example, the fundamental frequency suppression signal accounts for 60%, and the resonance enhancement signal accounts for 40% to prevent high-frequency components from being overloaded.

[0143] Dynamic Range Control Module

[0144] Threshold setting: Set the maximum allowable amplitude to ±1V (based on the sensor range of ±5V to retain a safety margin).

[0145] Compression Ratio: The signal exceeding the threshold is compressed at a ratio of 4:1 (i.e., for every 4dB increase in the input exceeding the threshold, the output only increases by 1dB).

[0146] Automatic Gain Control (AGC): Dynamically adjusts the overall gain based on the signal average. For example, if the signal average is below 0.2V, the gain is increased by 20%.

[0147] Harmonic distortion detection and compensation are performed on the preliminary enhanced signal to eliminate artifacts introduced by nonlinear processing and output an enhanced sound signal with optimized signal-to-noise ratio.

[0148] Nonlinear amplification may introduce harmonic distortion, and the extra frequency components need to be detected and suppressed.

[0149] Harmonic distortion detection

[0150] Spectral analysis: Perform an FFT on the initially enhanced signal to detect energy anomalies at integer multiples of the fundamental frequency (e.g., harmonics of a 3kHz signal at 6kHz and 9kHz). For example, if the energy at 6kHz is 30dB higher than the original signal, it is considered harmonic distortion.

[0151] Distortion calculation: The total harmonic distortion (THD) formula is THD (%) = (total harmonic energy / fundamental wave energy) × 100%, and THD is required to be ≤ 5%.

[0152] Distortion compensation and optimization

[0153] Adaptive notch filtering: Design a narrowband notch filter (bandwidth ±50Hz) for the detected harmonic frequency (such as 6kHz) to filter out distortion components.

[0154] Phase compensation: Adjust the signal phase through an all-pass filter to eliminate the delay introduced by filtering.

[0155] Output signal evaluation: The final signal THD dropped to 3%, and the signal-to-noise ratio (SNR) increased from 30dB to 45dB, meeting the requirements for abnormal sound detection.

[0156] Key technology examples and parameter descriptions

[0157] Base Frequency: The natural vibration frequency of the motor during normal operation, measured in Hertz (Hz).

[0158] Tunable Band-Stop Filter: A filter whose center frequency can be dynamically adjusted. The stopband range is usually ±5Hz of the center frequency.

[0159] Continuous Wavelet Transform (CWT): A multi-resolution time-frequency analysis tool based on scale and translation parameters, with the mother wavelet being Morlet.

[0160] Dynamic Range Control (DRC): compresses the signal amplitude to fit within the hardware processing range, with the threshold set to ±1V and a compression ratio of 4:1.

[0161] Total harmonic distortion (THD): A measure of signal distortion, required to be ≤5%.

[0162] This step uses a tunable band-stop filter to filter out the fundamental frequency components of the motor operation (such as the main frequency of rotor rotation) and eliminate background vibration interference under normal operating conditions. Subsequently, the characteristic frequency band of abnormal sound (such as the 2-5kHz high-frequency component unique to bearing damage) is identified through continuous wavelet transform, and a nonlinear gain curve is applied to exponentially amplify the target frequency band. Time domain superposition and dynamic range control technology ensure that the signal amplitude is within the resolvable range. At the same time, the harmonic distortion compensation module eliminates the artifact noise introduced by the amplification process, and the resonance enhancement module focuses on signal enhancement in the frequency band sensitive to abnormal sound, effectively highlighting the characteristics of early weak faults; the fundamental frequency suppression process avoids the interference of normal vibration signals on abnormality detection, greatly improving the recognition of abnormal sound components.

[0163] S203, performing wavelet packet decomposition on the enhanced sound signal, extracting multi-scale frequency band energy entropy, combining singular value decomposition with dimensionality reduction, and generating a motor abnormal sound feature matrix;

[0164] Specifically, the enhanced sound signal may be subjected to a 5-layer wavelet packet decomposition using the db8 wavelet basis to generate a set of 32 orthogonal frequency band sub-signals;

[0165] The db8 wavelet basis (Daubechies Wavelet of Order 8) is a compactly supported and orthogonal wavelet function. The "db" in its name represents the Daubechies wavelet family, and the "8" indicates that the filter coefficient length is 8th order. The db8 wavelet basis is suitable for processing non-stationary signals (such as motor noise) because its high-frequency resolution effectively captures transient characteristics.

[0166] Wavelet packet decomposition parameter settings

[0167] Decomposition level: 5-level decomposition will divide the signal band into 2 5 = 32 orthogonal subbands. For example, if the sampling rate of the enhanced audio signal is 10 kHz (Nyquist frequency is 5 kHz), the frequency band is divided into two levels at each decomposition level. After the fifth level of decomposition, the bandwidth of each subband is 5 kHz / 32 ≈ 156.25 Hz.

[0168] Decomposition tree structure: Full tree decomposition is used, meaning all nodes at each layer are further decomposed to ensure that all frequency bands are evenly covered. For example, the first layer of decomposition separates the signal into low frequencies (0-2.5kHz) and high frequencies (2.5-5kHz). The second layer further separates the low frequencies into 0-1.25kHz and 1.25-2.5kHz, and so on, until the fifth layer generates 32 frequency bands.

[0169] Decomposition process and output

[0170] Input signal preprocessing: The enhanced sound signal (e.g., time domain waveform length 10 seconds, number of sampling points 100,000) is normalized (amplitude range ±1V) to avoid numerical overflow.

[0171] Decomposition execution: Call the wavelet packet decomposition algorithm (such as the wp.destroy() function in the PyWavelets library) to calculate the wavelet coefficients layer by layer. For example, the first sub-band of the fifth layer corresponds to the lowest frequency range of 0-156.25Hz, and the 32nd sub-band corresponds to the highest frequency range of 4843.75-5000Hz.

[0172] Sub-signal reconstruction: Perform an inverse transform on the wavelet coefficients of each sub-band to generate 32 time-domain sub-signals. For example, the amplitude range of the time-domain waveform of sub-signal 1 is ±0.2V, and the amplitude range of sub-signal 32 is ±0.05V (high-frequency energy is low).

[0173] Calculate the sliding window energy entropy of each frequency band sub-signal, count its probability distribution and quantify the uncertainty, and generate a multi-scale energy entropy sequence;

[0174] Energy entropy quantifies the degree of disorder in a signal's energy distribution. Unusual noise often manifests as a sudden change in energy within a specific frequency band, resulting in an increase in entropy. A sliding window is used to capture transient characteristics.

[0175] Sliding window parameter setting

[0176] Window Length: 256 sampling points (corresponding to 256 / 10kHz = 25.6ms), covering the typical duration of abnormal motor noise (for example, the transient impact caused by bearing damage is about 20ms).

[0177] Step Size: 128 samples (12.8ms) with a 50% overlap to ensure that transient events are not truncated.

[0178] Total number of windows: For example, a 10-second signal contains 100,000 sampling points, and the total number of windows is (100,000 - 256) / 128 + 1 ≈ 780 windows.

[0179] Energy entropy calculation process

[0180] Energy calculation: Calculate the total energy of each sub-signal in the window. For example, the window energy of sub-signal 1 is E1 = Σ(amplitude 2). If the amplitude in the window is [0.1, -0.2, 0.05], then E1 = 0.1 2 +(-0.2) 2 +0.05 2 =0.0475V 2 .

[0181] Probability distribution generation: Normalize the sub-signal energies to probability values. For example, if the energies of the 32 sub-bands in a window are E1, E2, ..., E32, respectively, then the probability of the i-th sub-band is Pi = Ei / (E1+E2+...+E32).

[0182] Entropy calculation: Based on the Shannon entropy formula, Entropy = -Σ(Pi * log2(Pi)), with the unit being bits. For example, if all energy is evenly distributed (Pi = 1 / 32), the entropy is -32 * (1 / 32) * log2 (1 / 32) = 5 bits (maximum value). If the energy is concentrated in a single subband (Pi = 1), the entropy is 0.

[0183] Multi-scale energy entropy sequence generation

[0184] Sequence structure: Each subband generates an entropy sequence of length 780, for a total of 32 sequences. For example, the entropy sequence of subband 5 has values ​​of 1.2 bits, 1.5 bits, ..., and 0.8 bits in windows 1 to 780, respectively.

[0185] Uncertainty quantification: The variance of each entropy sequence is calculated. For example, the entropy variance of sub-band 10 is 0.1 under normal operating conditions, but rises to 0.5 under abnormal operating conditions, indicating that the energy distribution fluctuation is increasing.

[0186] The multi-scale energy entropy sequence is arranged into an initial feature matrix by frequency band level, and the first k principal components are extracted by sliding window singular value decomposition to generate a reduced dimension feature matrix;

[0187] Singular Value Decomposition (SVD) is used to extract key features of data and reduce redundant information. Sliding window SVD extracts principal components segment by segment based on the characteristics of time series data.

[0188] Initial feature matrix construction

[0189] Matrix dimensions: 32 bands × 780 windows = 32 × 780 matrix. Each column represents the 32-dimensional entropy value of a window. For example, the data in column 1 is [1.2, 1.5, ..., 0.8] (32 entropy values).

[0190] Frequency band hierarchy: Arrange the rows by frequency band from low to high (row 1 is 0-156.25Hz, row 32 is 4843.75-5000Hz).

[0191] Sliding window SVD parameter setting

[0192] Window length: 50 consecutive windows (corresponding to a 50×12.8ms=640ms time period) to capture the temporal correlation of abnormal sound events.

[0193] Step size: 25 windows (320ms), overlap rate 50%. For example, the total number of windows 780 can be divided into (780-50) / 25+1=30 sliding blocks.

[0194] The number of principal components k is determined by the Cumulative Variance Contribution Rate. For example, if the contribution rate is set to ≥ 85%, and the variance of the first five principal components accounts for 85%, then k = 5.

[0195] Dimensionality reduction execution and output

[0196] Local SVD calculation: SVD is performed on each sliding block (50 windows × 32 bands), decomposing it into U·S·V^T, where S is the singular value matrix.

[0197] Principal component extraction: Take the right singular vectors corresponding to the first k singular values ​​(the first k columns of the V matrix) and generate a reduced-dimensional feature matrix block. For example, a single sliding block is reduced from a 32×50 matrix to a 5×50 matrix.

[0198] Matrix concatenation: All sliding blocks are concatenated into the final reduced-dimensional feature matrix (5×780). For example, the feature dimension of each window is reduced from 32 to 5, reducing computational complexity.

[0199] The reduced-dimensional feature matrix is ​​normalized to eliminate dimensional differences and output the standardized motor abnormal noise feature matrix.

[0200] Normalization eliminates the dimensional differences of different principal components and ensures balanced weights in the classification model.

[0201] Normalization method selection

[0202] Z-Score normalization: Calculate the mean and standard deviation for each feature dimension (principal component) using the conversion formula (x-μ) / σ. For example, if the mean of principal component 1 is 0.5 and the standard deviation is 0.2, the original value of 0.7 is converted to (0.7-0.5) / 0.2=1.0.

[0203] Min-Max Normalization: Scales eigenvalues ​​to the range [0, 1] using the formula (x - x_min) / (x_max - x_min). For example, if a principal component ranges from [-1.5, 2.0], then -1.5 is mapped to 0 and 2.0 is mapped to 1.

[0204] Normalized execution process

[0205] Parameter calculation: Calculate the mean, standard deviation, or extreme value by column (feature dimension). For example, the mean of principal component 2 is 0.1 and the standard deviation is 0.05.

[0206] Data transformation: Apply the normalization formula column by column. For example, convert the 5×780 matrix of the reduced feature matrix into a standardized matrix with each column having a mean of 0 (Z-Score) or a range of [0, 1] (Min-Max).

[0207] Outlier handling: Clipping is performed on outliers that exceed ±3σ (Z-Score) or [0,1] (Min-Max). For example, if a principal component value is 3.5σ, it is forced to be 3.0σ.

[0208] Normalized feature matrix output

[0209] Matrix structure: 5×780 normalized matrix, each row corresponds to a principal component, and each column corresponds to a window. For example, the value of the principal component in row 1 is 0.8 in window 1 and 1.2 in window 2.

[0210] Application scenario: The standardized matrix can be directly input into the classification model (such as the abnormal noise classification model of claim 5) to avoid the model being biased towards high-level features due to dimensional differences.

[0211] Key technology examples and parameter descriptions

[0212] db8 wavelet basis: Daubechies 8th-order wavelet, filter coefficient length is 8, suitable for analyzing transient signals.

[0213] Orthogonal Subbands: The decomposed subbands do not overlap and the energy is conserved.

[0214] Sliding window energy entropy: window length 25.6ms, step size 12.8ms, quantifies the disorder of local energy distribution.

[0215] Singular value decomposition (SVD): k = 5 means retaining the first five principal components, and the cumulative variance contribution rate ≥ 85%.

[0216] Z-Score standardization: mean μ = 0, standard deviation σ = 1, eliminating dimensional differences.

[0217] This step uses the db8 wavelet basis function to perform multi-level wavelet packet decomposition on the signal, subdividing the time-frequency domain signal into 32 orthogonal subbands and calculating the sliding window energy entropy of each subband (reflecting signal complexity and uncertainty). Singular value decomposition is used to extract the principal component characteristics of the energy entropy sequence, eliminating redundant information and compressing the data dimension, ultimately generating a multidimensional feature matrix representing the abnormal noise pattern. Normalization ensures balanced weighting of features of different physical dimensions. The combination of wavelet packet decomposition and energy entropy enables multi-scale characterization of the signal, accurately capturing the time-frequency characteristics of transient abnormal noise. Dimensionality reduction reduces the model's computational complexity while retaining key discriminant information, providing high-quality input for the classification model.

[0218] S204, constructing an abnormal sound classification model using a small number of normal samples and abnormal samples based on a dynamic weight allocation meta-learning algorithm, wherein the dynamic weight allocation meta-learning algorithm initializes network parameters and optimizes inter-class separability through transfer learning to generate a robust feature space mapping relationship;

[0219] Specifically, based on the network structure of the pre-trained motor health status classification model, its convolutional layer and pooling layer parameters can be loaded as feature extractors to generate a backbone network for migration initialization.

[0220] A pre-trained model is a neural network pre-trained on a large-scale dataset (such as ImageNet or the Motor Health Database). Its underlying feature extraction capabilities can be directly transferred to the new task. Here, we use a pre-trained model optimized for motor vibration data (such as ResNet-18, an 18-layer residual network). Its convolutional and pooling layers have learned universal vibration feature representations.

[0221] Model structure adjustment and parameter loading

[0222] Input layer adaptation: The original input of the pre-trained model is a three-channel image (RGB), which needs to be adjusted to adapt to the input of the motor feature matrix. For example, the number of input channels is changed to 5 (corresponding to the 5 principal components of the normalized feature matrix output in claim 4), and the first-layer convolution kernel is changed from 3×3×64 (3-channel input, 64 filters) to 5×1×64 (5-channel input, filter width covering a single time step).

[0223] Parameter Freeze: Freezes the weights of the convolutional and pooling layers, allowing only the fully connected layers to be updated during subsequent training. For example, the parameters of the first 17 layers of ResNet-18 are fixed, and only the last fully connected layer is adjustable.

[0224] Example parameters: The loaded convolution kernel size is 3×3 (time×feature), the stride is 1, the padding is 1, the pooling window is 2×2, and the maximum pooling strategy is used.

[0225] Feature Extractor Validation

[0226] Test set inference: Input normal vibration samples (e.g., a 5×780 feature matrix with 100 windows) and observe the activation patterns of the feature maps output by the convolutional layers. For example, the output feature map of the third convolutional layer (size 64×195) for normal samples shows a uniform energy distribution, while abnormal samples show localized high activation areas.

[0227] Transfer effect evaluation: Comparing the feature discrimination of random initialization and transfer initialization. The transfer initialization model improves the inter-class distance between normal and abnormal samples by 30% (e.g., from 1.2 to 1.56, using the cosine distance metric) within the same training epoch.

[0228] Based on the feature vectors of normal and abnormal samples extracted by the backbone network, the inter-class cosine distance matrix is ​​calculated, the dynamic weight distribution coefficient is generated by entropy weighting, and the initial weight distribution matrix is ​​output;

[0229] Cosine distance measures the directional difference between two feature vectors in space. It is calculated as 1-(A·B) / (||A||·||B||) and ranges from 0 to 2. Smaller values ​​indicate greater similarity. Entropy weighting dynamically adjusts sample weights based on the degree of class mixing to improve the clarity of classification boundaries.

[0230] Feature vector extraction and distance calculation

[0231] Feature vector dimension: The output dimension of the backbone network's fully connected layer is set to 256 (i.e., each sample is mapped to a 256-dimensional vector). For example, the mean feature vector of a normal sample is [0.1, -0.2, ..., 0.05], while that of an abnormal sample is [0.8, -0.3, ..., 0.6].

[0232] Inter-class distance matrix: Calculate the cosine distance between all normal samples and abnormal samples to generate an N×M matrix (N is the number of normal samples, M is the number of abnormal samples). For example, 100 normal samples and 20 abnormal samples generate a 100×20 matrix, where each element represents the distance between a pair of samples.

[0233] Entropy weighted strategy

[0234] Entropy calculation: For each normal sample, calculate the distribution uncertainty of its distance from all abnormal samples. For example, the distance between normal sample A and 20 abnormal samples is [0.3, 0.5, ..., 1.2]. Its probability distribution P = [0.3 / total, 0.5 / total, ..., 1.2 / total], and the entropy is calculated as -Σ(P_i*log2(P_i)).

[0235] Weight allocation coefficient: The higher the entropy value, the lower the ability to distinguish between normal and abnormal samples. A higher weight should be assigned to strengthen learning of such samples. For example, a sample with an entropy value of 1.5 has a weight of 0.8, and a sample with an entropy value of 0.5 has a weight of 0.2.

[0236] Weight matrix generation: Arrange the weight coefficients into an N×1 matrix (normal sample weight) and a 1×M matrix (abnormal sample weight) according to the sample index, and multiply them to generate an N×M dynamic weight matrix.

[0237] The initial weight distribution matrix is ​​injected into the fully connected layer, and the weights of abnormal samples are optimized through the contrast loss function, forcing the abnormal features to stay away from the normal cluster centers in the weight space, thus generating a classifier prototype with a decision boundary.

[0238] The contrastive loss function optimizes the feature space distribution by bringing similar samples closer together and pushing different samples further apart. The cluster center refers to the mean position of the feature vectors of similar samples.

[0239] Weight injection and fully connected layer adjustment

[0240] Fully connected layer structure: The last fully connected layer of the backbone network (original output dimension 1000) is replaced with a two-layer structure adapted to the binary classification task: 256-dimensional input → 128-dimensional hidden layer → 2-dimensional output (normal / abnormal).

[0241] Weight injection mechanism: The dynamic weight matrix (N×M) and the fully connected layer weight (128×2) are weighted by sample pair. For example, if the weight coefficient of normal sample i and abnormal sample j is 0.8, the gradient update of the fully connected layer for this pair of samples will be amplified by 0.8 times.

[0242] Contrastive loss function design

[0243] Triplet Sampling: Randomly select an anchor, a positive sample (same category), and a negative sample (different category). For example, the anchor is a normal sample A, the positive sample is a normal sample B, and the negative sample is an abnormal sample C.

[0244] Loss calculation: The loss function formula is max(d(A,P)-d(A,N)+margin,0), where d is the cosine distance and margin is a preset threshold (such as 0.5). If the distance between the anchor point and the negative sample does not exceed the distance between the positive sample and the margin, a loss is triggered.

[0245] Optimization goal: Update the fully connected layer parameters through backpropagation to move the abnormal sample features away from the normal cluster center (for example, the normal class center coordinates are [0.2, -0.1], and the abnormal sample features are pushed to [-0.5, 0.3]).

[0246] Based on the decision boundary of the classifier prototype, a meta-learning optimizer is used to iteratively adjust the weight distribution coefficient on the support set, and the normal sample distribution range is compressed through the intra-class compactness loss function, and the optimized dynamic weight strategy is output;

[0247] Meta-learning optimizers, such as MAML (Model-Agnostic Meta-Learning), improve model generalization by rapidly adapting across multiple tasks. Intra-class compactness loss enforces clustering of features from similar samples.

[0248] Meta-learning task construction

[0249] Support Set: A small batch of samples (such as 10 normal samples and 5 abnormal samples) is randomly selected from the training set to form a meta-task.

[0250] Task goal: Fine-tune the dynamic weight strategy on the support set to ensure that the model maintains high accuracy on unseen query sets.

[0251] Compactness Loss and Meta-Optimization

[0252] Compactness calculation: Calculate the mean Euclidean distance between the normal sample feature vector and the cluster center. For example, if the normal cluster center is [0.1, -0.2], the distance between sample A and sample B is 0.3, and the mean is 0.35.

[0253] Loss function superposition: The total loss is the weighted sum of contrast loss and compactness loss (for example, contrast loss weight is 0.7, compactness loss weight is 0.3).

[0254] Meta-optimization step: Use the MAML algorithm to perform five gradient updates on the support set in the inner loop, and the outer loop to update the global model parameters. For example, the learning rate is set to 0.001 and the inner loop step size is 0.01.

[0255] The dynamic weight strategy is solidified into the fully connected layer of the classifier, and a small number of abnormal samples are used to fine-tune the output layer threshold parameters to generate the final abnormal noise classification model that can distinguish subtle abnormal noise patterns.

[0256] Threshold fine-tuning balances recall and false alarm rates by adjusting the classification decision boundary. The Sigmoid function maps outputs to probability values ​​between 0 and 1. The default threshold is 0.5 and can be adjusted based on the sample distribution.

[0257] Threshold optimization strategy

[0258] ROC curve analysis: Draw a receiver operating characteristic (ROC) curve on the validation set and calculate the true positive rate (TPR) and false positive rate (FPR) at different thresholds. For example, at a threshold of 0.6, TPR = 85%, FPR = 5%; at a threshold of 0.4, TPR = 95%, FPR = 20%.

[0259] Optimal threshold selection: Based on business requirements (such as prioritizing reducing false positives), select the highest threshold corresponding to FPR ≤ 10% (such as 0.65).

[0260] Output layer fine-tuning

[0261] Parameter update: Fix the parameters of the fully connected layer and only train the bias term of the output layer. For example, the initial bias is 0 and is adjusted to -0.3 through gradient descent to make the Sigmoid output more inclined to abnormality judgment.

[0262] Example parameters: learning rate is set to 0.0001, number of iterations is 50, and loss function is cross-entropy.

[0263] Key technology examples and parameter descriptions

[0264] ResNet-18: An 18-layer residual network consisting of 17 convolutional / pooling layers and 1 fully connected layer, suitable for temporal feature extraction.

[0265] Cosine distance: The value range is 0-2, where 0 means the two vectors are completely in the same direction, and 2 means they are completely in opposite directions.

[0266] Contrastive loss function: In triplet sampling, margin = 0.5 means that the distance between negative samples must be at least 0.5 farther than that between positive samples.

[0267] MAML: Model-agnostic meta-learning algorithm with inner loop learning rate of 0.01 and outer loop learning rate of 0.001.

[0268] Sigmoid threshold: The default value is 0.5. Adjusting it to 0.6 can reduce false positives, but may miss some abnormal sounds.

[0269] This step uses the parameters of a pre-trained health status classification model to initialize the feature extraction network and utilizes transfer learning to address the problem of model overfitting in small sample scenarios. By calculating the inter-class cosine distance between the feature vectors of normal and abnormal samples, weight coefficients are dynamically assigned to strengthen the weight of abnormal samples. The contrastive loss function forces abnormal features to move away from the normal cluster center, and the meta-learning optimizer further compresses the intra-class distribution differences, ultimately constructing a classifier with a clear decision boundary. The dynamic weight allocation mechanism alleviates the class imbalance problem in small sample learning. The combination of transfer learning and meta-learning significantly improves the model's ability to generalize unknown abnormal sound patterns, ensuring high-precision classification even with limited labeled data.

[0270] S205 , inputting the motor abnormal noise feature matrix into the abnormal noise classification model, detecting transient abnormalities through a sliding window time series matching algorithm, and outputting an abnormal noise determination result.

[0271] Specifically, the motor abnormal noise feature matrix stream can be adaptively segmented according to the motor speed to generate sliding window data blocks with variable length;

[0272] Adaptive motor speed segmentation dynamically adjusts the length of the sliding window based on the real-time speed, ensuring that each window covers a fixed mechanical period (such as the vibration period generated by a single rotation or multiple rotations). Its core goal is to align the motor's physical motion with the signal analysis window to avoid feature truncation or phase shifts caused by speed fluctuations.

[0273] Speed ​​synchronization and window length calculation

[0274] Speed ​​sensor data access: The real-time motor speed (unit: RPM, Revolutions Per Minute) is obtained through a Hall effect sensor or encoder. For example, a motor with a rated speed of 1500 RPM may fluctuate between 1480-1520 RPM (±2%) during actual operation.

[0275] Window length formula: Each window is set to cover an integer multiple of the rotation period. For example, if the current speed is 1500 RPM (25 revolutions per second), the period of a single revolution is 40 milliseconds (ms). If the window is set to cover 5 periods (5 × 40ms = 200ms), the window length is 200ms. If the speed drops to 1440 RPM (41.67ms period), the window length is adjusted to 5 × 41.67 ≈ 208ms.

[0276] Sliding step size setting: The step size is half the window length (i.e. 50% overlap). For example, the step size of a 200ms window is 100ms, ensuring that adjacent windows have 50% overlap to avoid missing transient features.

[0277] Adaptive segmentation algorithm execution

[0278] Real-time speed tracking: The speed sensor data is read every 100ms and the window length is updated. For example, if the current window length is 208ms, the next window will be shortened to 200ms if the speed is detected to rise to 1500RPM.

[0279] Data block generation: The input feature matrix stream (e.g., a normalized 5×780 matrix with rows representing principal components and columns representing time windows) is split according to the new window length. For example, a 10-second data stream (780 windows) is split into 785 windows of variable length when the speed changes.

[0280] Boundary processing: Zero pad the last incomplete window or truncate and discard it to ensure that all windows have the same length. For example, the remaining 15ms of data is padded with zeros to 208ms.

[0281] Verification and fault tolerance mechanism

[0282] Sudden speed fluctuation detection: Set a speed fluctuation threshold (e.g., ±5%). If the threshold is exceeded, an abnormal flag is triggered, segmentation is paused, and the speed stabilizes. For example, if the speed suddenly drops from 1500 RPM to 1300 RPM (exceeding 13%), segmentation is paused and the event log is recorded.

[0283] Window length upper and lower limits: Limit the window length to between 100ms (minimum) and 500ms (maximum) to prevent extreme speeds from causing the window to be too short or too long. For example, at 600 RPM, a single cycle of 100ms results in a window length of 5 × 100 = 500ms (reaching the upper limit).

[0284] Calculate the dynamic time-warping distance between each data block and the historical anomaly template to generate a window-level anomaly probability curve;

[0285] Dynamic Time Warping (DTW) is an algorithm that measures the similarity between two time series. It adapts to the expansion and contraction of the time series by elastically aligning the time axis. Historical anomaly templates are typical feature patterns extracted from known anomaly cases (such as feature matrix fragments corresponding to bearing wear and rotor imbalance).

[0286] Template library construction and matching strategy

[0287] Template type: Each abnormal noise type (such as bearing damage and winding looseness) corresponds to one or more templates. For example, the bearing damage template is a 5×200 matrix (5 principal components, 200 time windows), which captures the periodic impact characteristics caused by damage.

[0288] Multi-template matching: Calculate the DTW distance between each data block and all templates, and take the minimum value as the final matching distance. For example, if the DTW distance between a data block and bearing damage template A is 1.2 and with template B is 1.5, then 1.2 is taken as the anomaly distance for this data block.

[0289] DTW distance calculation process

[0290] Accumulated Cost Matrix: Construct a two-dimensional matrix with rows and columns representing the time points of the data block and template, respectively. The matrix elements are the cumulative minimum values ​​of the Euclidean distances (the square root of the sum of the squared differences between two vectors in each dimension) between the corresponding points. For example, if the Euclidean distance between point i in the data block and point j in the template is 0.8, the cumulative path is the minimum value of the left, above, or diagonal neighbors plus 0.8.

[0291] Warping Path: Tracing back from the lower right corner of the matrix to the upper left corner, select the path with the lowest cumulative cost. The total cost of the path is the DTW distance. For example, the total cost of the warping path between a data block and the template is 1.5.

[0292] Normalize to probability: Convert the DTW distance to an anomaly probability between 0 and 1 using Min-Max Normalization. For example, if the maximum DTW distance for a normal sample is 2.0 and the minimum distance for an anomaly template is 0.5, then the probability corresponding to a distance of 1.2 is (2.0 - 1.2) / (2.0 - 0.5) = 0.53.

[0293] Abnormal probability curve generation

[0294] Time series alignment: Probability values ​​are arranged by window timestamp to generate a continuous curve. For example, the probability of window 1 is 0.1, the probability of window 2 is 0.3, ..., and the probability of window 785 is 0.6.

[0295] Smoothing: Apply a sliding average filter (window length 3, i.e., taking the average of the current point, the previous point, and the next point) to suppress random fluctuations. For example, after smoothing, the probability of window 2 changes from 0.3 to (0.1 + 0.3 + 0.4) / 3 = 0.27.

[0296] Detect the sections in the abnormal probability curve that continuously exceed the preset threshold, perform logical verification based on the judgment results of adjacent windows, and generate a list of candidate abnormal events;

[0297] The preset threshold is determined based on historical data statistics, typically the 95% quantile of the normal sample probability distribution. For example, if the maximum normal sample probability is 0.2, the threshold is set to 0.2. Logical verification reduces false positives through temporal continuity constraints and neighborhood consistency checks.

[0298] Threshold setting and super-threshold segment detection

[0299] Adaptive threshold: If the probability distribution of normal samples changes over time (e.g., baseline shift due to device aging), the threshold is updated using rolling window statistics. For example, the 95% quantile is calculated every 100 windows and the threshold is adjusted dynamically.

[0300] Segment marking: Scan the probability curve and record the window intervals that continuously exceed the threshold. For example, if the probability of windows 50-58 is ≥ 0.2, mark them as candidate segments.

[0301] Logical validation rules

[0302] Minimum duration: Set the minimum duration of candidate events (e.g., 3 consecutive windows, corresponding to 3 × window step). For example, if windows 50-52 last for 3 windows (total duration 300ms), they meet the criteria; while windows 50-51 only last for 2 windows and are filtered out.

[0303] Neighborhood consistency: Check the probability trend of the windows before and after the candidate segment. For example, if the probability of window 49 is 0.18 (slightly below the threshold), windows 50-58 are ≥ 0.2, and window 59 is 0.19, then windows 50-58 are considered independent events.

[0304] Multi-level verification: the first level verification is through duration, the second level verification is through neighborhood trends, and the third level verification is through merging with other candidate events that overlap with the segment (such as merging windows 50-58 and 55-63 into 50-63).

[0305] Candidate event list generation

[0306] Event attribute records: Each candidate event records the start time, end time, maximum probability value, and average probability value. For example, event number 1: start time 5.0 seconds, end time 5.8 seconds, maximum probability 0.6, and average probability 0.45.

[0307] False positive filtering: If the probability curve for an event exhibits a "spike" (e.g., a very high probability in a single window but a rapidly decreasing probability around it), it is marked as noise and removed. For example, if the probability in window 100 is 0.9, and both the preceding and following windows are 0.1, it is considered transient interference.

[0308] According to the duration and probability peak intensity of the candidate abnormal event, a multi-level alarm strategy is triggered and the abnormal sound type and confidence score are output.

[0309] The multi-level alarm strategy divides the response level according to the severity of the event, such as "Warning", "Critical", and "Urgent". The confidence score is calculated based on the duration, probability peak, and the degree of match with the template.

[0310] Alarm level classification rules

[0311] Level threshold setting:

[0312] Warning level: duration ≥ 300ms and peak probability ≥ 0.3.

[0313] Severe level: Duration ≥ 500ms or peak probability ≥ 0.5.

[0314] Emergency level: duration ≥ 1 second and peak probability ≥ 0.7.

[0315] Example judgment: Event 1 lasts for 800ms and has a peak value of 0.6, triggering a severe alarm; Event 2 lasts for 1.2 seconds and has a peak value of 0.75, triggering an emergency alarm.

[0316] Abnormal sound type identification

[0317] Template matching score: Calculates the DTW distance between the candidate event and each anomaly template, and takes the type corresponding to the minimum distance as the predicted label. For example, if event 1 has a distance of 1.2 from the bearing damage template and a distance of 2.5 from the winding looseness template, it is judged to be bearing damage.

[0318] Confidence calculation: Confidence = 1 - (minimum DTW distance / maximum possible distance). For example, if the maximum possible distance is 3.0, then the confidence of event 1 = 1 - 1.2 / 3.0 = 0.6 (60%).

[0319] Alarm output and logging

[0320] Alarm information format: Contains event ID, abnormal sound type, alarm level, confidence level, and timestamp. For example: {Event ID: 001, Type: Bearing Damage, Level: Severe, Confidence Level: 60%, Time: 2023-10-01 14:30:25}.

[0321] Log storage: Records raw data blocks, probability curves, and decision basis for subsequent review and analysis. For example, this can be stored as a CSV file or written to a database table.

[0322] Key technology examples and parameter descriptions

[0323] RPM (Revolutions Per Minute): revolutions per minute, used to quantify motor speed.

[0324] DTW (Dynamic Time Warping): Dynamic time warping algorithm that measures sequence similarity by elastically aligning the time axis.

[0325] Sliding window step size: The window sliding interval. For example, 100ms means that the window advances one window every 100ms.

[0326] Normalization (Min-Max Normalization): Linearly map the data to the 0-1 range using the formula (x-min) / (max-min).

[0327] Confidence Score: A value between 0 and 1, indicating the reliability of the judgment result.

[0328] This step dynamically divides the sliding window length of the input feature matrix based on motor speed. A dynamic time warping algorithm is used to calculate the similarity distance between the current window and historical anomaly templates, generating a probabilistic anomaly confidence curve. Multi-threshold logic verification eliminates occasional false alarms, and a graded alarm is triggered based on the duration and probability peak intensity. The final output includes a judgment result including the type of abnormal sound (such as bearing wear and rotor eccentricity) and a confidence score. The sliding window timing matching algorithm adapts to the non-stationary characteristics of the signal under variable motor speed conditions, improving the sensitivity of transient anomaly detection. A multi-level verification mechanism effectively reduces the false alarm rate and ensures that the judgment result is engineering-operable.

[0329] It can be seen that the contact vibration sensor is directly coupled to the motor housing to collect the original vibration signal and generate an anti-interference vibration signal; the anti-interference vibration signal is input into the nonlinear resonance enhancement module to generate an enhanced sound signal; the enhanced sound signal is decomposed by wavelet packets to extract the multi-scale frequency band energy entropy, and combined with the singular value decomposition dimensionality reduction, the motor abnormal noise feature matrix is ​​generated; based on the dynamic weight allocation meta-learning algorithm, a small number of normal samples and abnormal samples are used to construct an abnormal noise classification model; the motor abnormal noise feature matrix is ​​input into the abnormal noise classification model, and the sliding window timing matching algorithm is used to detect transient anomalies, and the abnormal noise judgment result is output, thereby realizing high-precision and low-false-alarm motor abnormal noise detection under complex working conditions.

[0330] Another embodiment of the present invention provides a motor abnormal noise detection system based on contact acquisition and small sample learning, see Figure 3 , the system may include:

[0331] An acquisition module 301 is configured to acquire the original vibration signal by directly coupling the contact vibration sensor to the motor housing, dynamically adjust the sensor sampling parameters based on the ambient noise spectrum, and generate an anti-interference vibration signal;

[0332] An input module 302 is configured to input the anti-interference vibration signal into a nonlinear resonance enhancement module, and generate an enhanced sound signal through resonance amplification and fundamental frequency suppression processing in a preset abnormal sound characteristic frequency band;

[0333] A decomposition module 303 is configured to perform wavelet packet decomposition on the enhanced sound signal, extract multi-scale frequency band energy entropy, and generate a motor abnormal sound feature matrix by combining singular value decomposition with dimensionality reduction;

[0334] A construction module 304 is configured to construct an abnormal sound classification model using a small number of normal samples and abnormal samples based on a dynamic weight allocation meta-learning algorithm, wherein the dynamic weight allocation meta-learning algorithm initializes network parameters and optimizes inter-class separability through transfer learning to generate a robust feature space mapping relationship;

[0335] The detection module 305 is used to input the motor abnormal noise feature matrix into the abnormal noise classification model, detect transient abnormalities through a sliding window time series matching algorithm, and output an abnormal noise determination result.

[0336] It can be seen that the contact vibration sensor is directly coupled to the motor housing to collect the original vibration signal and generate an anti-interference vibration signal; the anti-interference vibration signal is input into the nonlinear resonance enhancement module to generate an enhanced sound signal; the enhanced sound signal is decomposed by wavelet packets to extract the multi-scale frequency band energy entropy, and combined with the singular value decomposition dimensionality reduction, the motor abnormal noise feature matrix is ​​generated; based on the dynamic weight allocation meta-learning algorithm, a small number of normal samples and abnormal samples are used to construct an abnormal noise classification model; the motor abnormal noise feature matrix is ​​input into the abnormal noise classification model, and the sliding window timing matching algorithm is used to detect transient anomalies, and the abnormal noise judgment result is output, thereby realizing high-precision and low-false-alarm motor abnormal noise detection under complex working conditions.

[0337] An embodiment of the present invention further provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps of any one of the above method embodiments when running.

[0338] Specifically, in this embodiment, the above-mentioned storage medium may be configured to store a computer program for performing the following steps:

[0339] S201, directly coupling the contact vibration sensor to the motor housing to collect the original vibration signal, dynamically adjusting the sensor sampling parameters based on the ambient noise spectrum, and generating an anti-interference vibration signal;

[0340] S202, inputting the anti-interference vibration signal into a nonlinear resonance enhancement module, generating an enhanced sound signal through resonance amplification and fundamental frequency suppression processing in a preset abnormal sound characteristic frequency band;

[0341] S203, performing wavelet packet decomposition on the enhanced sound signal, extracting multi-scale frequency band energy entropy, combining singular value decomposition with dimensionality reduction, and generating a motor abnormal sound feature matrix;

[0342] S204, constructing an abnormal sound classification model using a small number of normal samples and abnormal samples based on a dynamic weight allocation meta-learning algorithm, wherein the dynamic weight allocation meta-learning algorithm initializes network parameters and optimizes inter-class separability through transfer learning to generate a robust feature space mapping relationship;

[0343] S205 , inputting the motor abnormal noise feature matrix into the abnormal noise classification model, detecting transient abnormalities through a sliding window time series matching algorithm, and outputting an abnormal noise determination result.

[0344] It can be seen that the contact vibration sensor is directly coupled to the motor housing to collect the original vibration signal and generate an anti-interference vibration signal; the anti-interference vibration signal is input into the nonlinear resonance enhancement module to generate an enhanced sound signal; the enhanced sound signal is decomposed by wavelet packets to extract the multi-scale frequency band energy entropy, and combined with the singular value decomposition dimensionality reduction, the motor abnormal noise feature matrix is ​​generated; based on the dynamic weight allocation meta-learning algorithm, a small number of normal samples and abnormal samples are used to construct an abnormal noise classification model; the motor abnormal noise feature matrix is ​​input into the abnormal noise classification model, and the sliding window timing matching algorithm is used to detect transient anomalies, and the abnormal noise judgment result is output, thereby realizing high-precision and low-false-alarm motor abnormal noise detection under complex working conditions.

[0345] An embodiment of the present invention further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any one of the above method embodiments.

[0346] Specifically, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0347] Specifically, in this embodiment, the processor may be configured to execute the following steps through a computer program:

[0348] S201, directly coupling the contact vibration sensor to the motor housing to collect the original vibration signal, dynamically adjusting the sensor sampling parameters based on the ambient noise spectrum, and generating an anti-interference vibration signal;

[0349] S202, inputting the anti-interference vibration signal into a nonlinear resonance enhancement module, generating an enhanced sound signal through resonance amplification and fundamental frequency suppression processing in a preset abnormal sound characteristic frequency band;

[0350] S203, performing wavelet packet decomposition on the enhanced sound signal, extracting multi-scale frequency band energy entropy, combining singular value decomposition with dimensionality reduction, and generating a motor abnormal sound feature matrix;

[0351] S204, constructing an abnormal sound classification model using a small number of normal samples and abnormal samples based on a dynamic weight allocation meta-learning algorithm, wherein the dynamic weight allocation meta-learning algorithm initializes network parameters and optimizes inter-class separability through transfer learning to generate a robust feature space mapping relationship;

[0352] S205 , inputting the motor abnormal noise feature matrix into the abnormal noise classification model, detecting transient abnormalities through a sliding window time series matching algorithm, and outputting an abnormal noise determination result.

[0353] It can be seen that the contact vibration sensor is directly coupled to the motor housing to collect the original vibration signal and generate an anti-interference vibration signal; the anti-interference vibration signal is input into the nonlinear resonance enhancement module to generate an enhanced sound signal; the enhanced sound signal is decomposed by wavelet packets to extract the multi-scale frequency band energy entropy, and combined with the singular value decomposition dimensionality reduction, the motor abnormal noise feature matrix is ​​generated; based on the dynamic weight allocation meta-learning algorithm, a small number of normal samples and abnormal samples are used to construct an abnormal noise classification model; the motor abnormal noise feature matrix is ​​input into the abnormal noise classification model, and the sliding window timing matching algorithm is used to detect transient anomalies, and the abnormal noise judgment result is output, thereby realizing high-precision and low-false-alarm motor abnormal noise detection under complex working conditions.

[0354] The above describes in detail the structure, features and effects of the present invention based on the embodiments shown in the drawings. The above is only a preferred embodiment of the present invention, but the scope of implementation of the present invention is not limited to what is shown in the drawings. Any changes made in accordance with the concept of the present invention, or modifications to equivalent embodiments with equivalent changes, which do not exceed the spirit covered by the description and drawings, should be within the scope of protection of the present invention.

Claims

1. A motor abnormal noise detection method based on contact acquisition and small sample learning, characterized in that: The method comprises: The contact vibration sensor is directly coupled to the motor housing to collect the original vibration signal, and the sensor sampling parameters are dynamically adjusted based on the ambient noise spectrum to generate an anti-interference vibration signal. The anti-interference vibration signal is input into a nonlinear resonance enhancement module, and an enhanced sound signal is generated by resonance amplification and fundamental frequency suppression processing of a preset abnormal sound characteristic frequency band; Performing wavelet packet decomposition on the enhanced sound signal to extract multi-scale frequency band energy entropy, combining singular value decomposition with dimensionality reduction to generate a motor abnormal sound feature matrix; Based on a dynamic weight allocation meta-learning algorithm, a noise classification model is constructed using a small number of normal and abnormal samples. The dynamic weight allocation meta-learning algorithm initializes network parameters through transfer learning and optimizes inter-class separability to generate a robust feature space mapping relationship. The motor abnormal noise feature matrix is ​​input into the abnormal noise classification model, transient anomalies are detected through the sliding window timing matching algorithm, and the abnormal noise judgment result is output.

2. The method according to claim 1, characterized in that The method directly couples the contact vibration sensor to the motor housing to collect the original vibration signal, dynamically adjusts the sensor sampling parameters based on the ambient noise spectrum, and generates an anti-interference vibration signal, including: Deploy a three-axis vibration sensor array based on the geometric characteristics of the motor housing, synchronously collect multi-channel raw vibration signals, and generate initial time domain waveform data; Perform fast Fourier transform on the initial time domain waveform data to extract the energy distribution of the dominant frequency band of the ambient noise and generate a noise spectrum feature map; Based on the noise spectrum feature map, an adaptive notch filter bank is dynamically constructed, and the stopband frequency and bandwidth parameters are adjusted according to the real-time noise energy peak to generate a noise suppression filter; The initial time domain waveform data is input into the noise suppression filter to filter out the environmental interference component, and then the sensor coupling delay is compensated through the phase calibration module to output the anti-interference vibration signal.

3. The method according to claim 2, characterized in that The anti-interference vibration signal is input into a nonlinear resonance enhancement module, and an enhanced sound signal is generated by resonance amplification and fundamental frequency suppression processing of a preset abnormal sound characteristic frequency band, including: Based on the motor model library, the fundamental frequency range parameters are loaded and a tunable band-stop filter is designed to filter out the 50-200Hz fundamental frequency component from the anti-interference vibration signal and generate a fundamental frequency suppression signal. Perform continuous wavelet transform on the fundamental frequency suppression signal to locate the energy mutation area in the 2-5kHz frequency band and generate the energy distribution map of the abnormal sound characteristic frequency band; Dynamically generate a nonlinear gain curve based on the energy distribution diagram, exponentially amplify the target frequency band signal, and generate a resonance enhancement signal; The resonance enhancement signal and the fundamental frequency suppression signal are aligned and superimposed in the time domain, and the amplitude of the synthesized signal is constrained by the dynamic range control module to generate a preliminary enhancement signal; Harmonic distortion detection and compensation are performed on the preliminary enhanced signal to eliminate artifacts introduced by nonlinear processing and output an enhanced sound signal with optimized signal-to-noise ratio.

4. The method according to claim 3, characterized in that The enhanced sound signal is subjected to wavelet packet decomposition to extract multi-scale frequency band energy entropy, and the singular value decomposition is combined with dimensionality reduction to generate a motor abnormal sound feature matrix, including: The enhanced sound signal is decomposed into 5 layers of wavelet packets using the db8 wavelet basis to generate a set of 32 orthogonal frequency band sub-signals. Calculate the sliding window energy entropy of each frequency band sub-signal, count its probability distribution and quantify the uncertainty, and generate a multi-scale energy entropy sequence; The multi-scale energy entropy sequence is arranged into an initial feature matrix by frequency band level, and the first k principal components are extracted by sliding window singular value decomposition to generate a reduced dimension feature matrix; The reduced-dimensional feature matrix is ​​normalized to eliminate dimensional differences and output the standardized motor abnormal noise feature matrix.

5. The method according to claim 4, characterized in that The dynamic weight allocation meta-learning algorithm is based on a small number of normal samples and abnormal samples to build an abnormal noise classification model. The dynamic weight allocation meta-learning algorithm initializes network parameters and optimizes inter-class separability through transfer learning to generate a robust feature space mapping relationship, including: Based on the network structure of the pre-trained motor health status classification model, its convolutional layer and pooling layer parameters are loaded as feature extractors to generate the backbone network for transfer initialization. Based on the feature vectors of normal and abnormal samples extracted by the backbone network, the inter-class cosine distance matrix is ​​calculated, the dynamic weight distribution coefficient is generated by entropy weighting, and the initial weight distribution matrix is ​​output; The initial weight distribution matrix is ​​injected into the fully connected layer, and the weights of abnormal samples are optimized through the contrast loss function, forcing the abnormal features to stay away from the normal cluster centers in the weight space, thus generating a classifier prototype with a decision boundary. Based on the decision boundary of the classifier prototype, a meta-learning optimizer is used to iteratively adjust the weight distribution coefficient on the support set, and the normal sample distribution range is compressed through the intra-class compactness loss function, and the optimized dynamic weight strategy is output; The dynamic weight strategy is solidified into the fully connected layer of the classifier, and a small number of abnormal samples are used to fine-tune the output layer threshold parameters to generate the final abnormal noise classification model that can distinguish subtle abnormal noise patterns.

6. The method according to claim 5, characterized in that The motor abnormal noise feature matrix is ​​input into the abnormal noise classification model, transient abnormalities are detected by a sliding window timing matching algorithm, and an abnormal noise determination result is output, including: Adaptively segment the motor abnormal noise feature matrix stream according to the motor speed to generate sliding window data blocks with variable length; Calculate the dynamic time-warping distance between each data block and the historical anomaly template to generate a window-level anomaly probability curve; Detect the sections in the abnormal probability curve that continuously exceed the preset threshold, perform logical verification based on the judgment results of adjacent windows, and generate a list of candidate abnormal events; According to the duration and probability peak intensity of the candidate abnormal event, a multi-level alarm strategy is triggered and the abnormal sound type and confidence score are output.

7. A motor abnormal noise detection system based on contact acquisition and small sample learning, characterized in that: The system comprises: The acquisition module is used to directly couple the contact vibration sensor to the motor housing to collect the original vibration signal, dynamically adjust the sensor sampling parameters based on the ambient noise spectrum, and generate an anti-interference vibration signal; An input module is used to input the anti-interference vibration signal into a nonlinear resonance enhancement module, and generate an enhanced sound signal through resonance amplification and fundamental frequency suppression processing of a preset abnormal sound characteristic frequency band; A decomposition module is used to perform wavelet packet decomposition on the enhanced sound signal, extract multi-scale frequency band energy entropy, combine singular value decomposition with dimensionality reduction, and generate a motor abnormal sound feature matrix; A construction module is configured to construct an abnormal noise classification model using a small number of normal and abnormal samples based on a dynamic weight allocation meta-learning algorithm, wherein the dynamic weight allocation meta-learning algorithm initializes network parameters and optimizes inter-class separability through transfer learning to generate a robust feature space mapping relationship; The detection module is used to input the motor abnormal noise feature matrix into the abnormal noise classification model, detect transient anomalies through a sliding window timing matching algorithm, and output an abnormal noise judgment result.

8. The system according to claim 7, characterized in that The acquisition module is specifically used to: Deploy a three-axis vibration sensor array based on the geometric characteristics of the motor housing, synchronously collect multi-channel raw vibration signals, and generate initial time domain waveform data; Perform fast Fourier transform on the initial time domain waveform data to extract the energy distribution of the dominant frequency band of the ambient noise and generate a noise spectrum feature map; Based on the noise spectrum feature map, an adaptive notch filter bank is dynamically constructed, and the stopband frequency and bandwidth parameters are adjusted according to the real-time noise energy peak to generate a noise suppression filter; The initial time domain waveform data is input into the noise suppression filter to filter out the environmental interference component, and then the sensor coupling delay is compensated through the phase calibration module to output the anti-interference vibration signal.

9. A storage medium, characterized in that: The storage medium stores a computer program, wherein the computer program is configured to execute the method according to any one of claims 1 to 6 when executed.

10. An electronic device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to perform the method according to any one of claims 1 to 6.

Images