Motor fault diagnosis method and motor controller

Abstract

The application discloses a motor fault diagnosis method and a motor controller, and the method comprises the following steps: collecting near-field electromagnetic noise audio signals and vibration acceleration signals of a motor shell when the motor is running; performing short-time Fourier transform on the near-field electromagnetic noise audio signals to extract original electromagnetic acoustic fingerprint characteristic vectors representing electromagnetic states inside the motor; taking the vibration acceleration signals as a modulation source to perform amplitude modulation processing on the original electromagnetic acoustic fingerprint characteristic vectors to generate modulated electromagnetic acoustic fingerprint characteristic vectors; performing time-domain coherence calculation on the original electromagnetic acoustic fingerprint characteristic vectors and the modulated electromagnetic acoustic fingerprint characteristic vectors to generate two-dimensional electromagnetic-mechanical acoustic fingerprint interference maps; and inputting the electromagnetic-mechanical acoustic fingerprint interference maps into a convolutional neural network (CNN) to output final motor fault diagnosis results. According to the embodiment of the application, the distinguishing degree of fault features can be improved, and high-precision diagnosis of motor faults under complex working conditions can be realized.

Description

Technical Field

[0001] This invention belongs to the field of motor technology, specifically a method for diagnosing motor faults and a motor controller. Background Technology

[0002] Traditional motor fault diagnosis methods mainly rely on a single signal source (such as vibration or current signals), which makes it difficult to comprehensively reflect the multi-physics coupled fault characteristics of the motor. Vibration signals are easily affected by environmental noise, and while electromagnetic signals can reflect the internal electromagnetic state, they lack coupling analysis of mechanical vibration. Existing technologies typically employ signal separation processing, ignoring the electromagnetic-mechanical intermodulation effect, resulting in insufficient extraction of fault features. Summary of the Invention

[0003] The purpose of this invention is to provide a motor fault diagnosis method and a motor controller to overcome the shortcomings of the prior art, improve the distinguishability of fault characteristics, and achieve high-precision diagnosis of motor faults under complex working conditions.

[0004] One embodiment of this application provides a method for diagnosing motor faults, the method comprising:

[0005] The system collects near-field electromagnetic noise audio signals and vibration acceleration signals of the motor housing during operation, and ensures signal synchronization through hardware timestamps.

[0006] A short-time Fourier transform is performed on the near-field electromagnetic noise audio signal to extract the instantaneous amplitude envelope sequence of the fundamental frequency and harmonic components, which serves as the original electromagnetic acoustic signature feature vector characterizing the internal electromagnetic state of the motor.

[0007] The vibration acceleration signal is used as a modulation source to perform amplitude modulation processing on the original electromagnetic acoustic pattern feature vector in order to simulate the physical modulation effect of shell vibration on near-field electromagnetic noise and generate a modulated electromagnetic acoustic pattern feature vector.

[0008] The original electromagnetic acoustic signature feature vector and the modulated electromagnetic acoustic signature feature vector are subjected to time-domain coherence calculation to generate a two-dimensional electromagnetic-mechanical acoustic signature interferogram, in which the horizontal axis of the spectrum is time, the vertical axis is frequency, and the pixel value represents the coherence intensity.

[0009] The electromagnetic-mechanical acoustic interferogram is input into a pre-trained convolutional neural network (CNN), which is trained to identify the correspondence between specific coherent fringe patterns in the spectrum and motor fault types, so as to output the final motor fault diagnosis result.

[0010] Optionally, the acquisition of near-field electromagnetic noise audio signals and vibration acceleration signals of the motor housing during operation, and the use of hardware timestamps to ensure signal synchronization, includes:

[0011] An electromagnetic noise sensor is deployed close to the motor surface to capture the raw analog signal of near-field electromagnetic noise with a nanosecond-level response, and outputs an unprocessed analog stream of electromagnetic noise.

[0012] Piezoelectric triaxial accelerometers are installed at key vibration points of the motor housing. Vibration signals are preprocessed by an anti-aliasing filter, and the filtered vibration acceleration analog stream is output.

[0013] The electromagnetic noise analog stream and the vibration acceleration analog stream are input into a multi-channel ADC converter, and a hardware timestamp module driven by an atomic clock is embedded to add a precise nanosecond-level time stamp to each sampling point, and output a pair of digital signals with timestamps.

[0014] A timestamp alignment algorithm is applied to dynamically time-normalize digital signal pairs based on hardware timestamps, eliminating sampling jitter and generating synchronized electromagnetic noise digital signals and vibration acceleration digital signals.

[0015] Optionally, the step of performing a short-time Fourier transform on the near-field electromagnetic noise audio signal to extract the instantaneous amplitude envelope sequence of the fundamental frequency and harmonic components as the original electromagnetic acoustic signature feature vector characterizing the internal electromagnetic state of the motor includes:

[0016] A variable-window-length short-time Fourier transform is applied to synchronous electromagnetic noise digital signals. The window size is adaptively adjusted according to the instantaneous energy of the signal, and a high-resolution time-frequency energy matrix is ​​output.

[0017] Using the time-frequency energy matrix as input, a fundamental frequency detector based on a convolutional neural network is run to identify the dominant fundamental frequency component and its time-varying trajectory, and output the instantaneous position sequence of the fundamental frequency.

[0018] Based on the fundamental frequency instantaneous position sequence, the harmonic bands of integer multiples are extracted using the harmonic tracking algorithm, the instantaneous amplitude of each harmonic component is calculated, and the harmonic amplitude sequence set is output.

[0019] Apply Hilbert transform to the set of fundamental frequency instantaneous position sequence and harmonic amplitude sequence to extract the instantaneous envelope of each component and output the fundamental frequency envelope sequence and harmonic envelope sequence.

[0020] By fusing the fundamental frequency envelope sequence and harmonic envelope sequence, and then splicing and normalizing in the time domain, the original electromagnetic acoustic signature feature vector is generated.

[0021] Optionally, the step of using the vibration acceleration signal as a modulation source to perform amplitude modulation processing on the original electromagnetic acoustic signature feature vector to simulate the physical modulation effect of shell vibration on near-field electromagnetic noise, and generating a modulated electromagnetic acoustic signature feature vector, includes:

[0022] Wavelet packet decomposition is performed on the synchronous vibration acceleration digital signal to extract the frequency band components related to the mechanical resonance of the motor and output the key vibration frequency band signal.

[0023] The key vibration frequency band signal is input into a nonlinear normalizer and converted into a modulation index sequence. Its value range is dynamically mapped to the amplitude range of the original electromagnetic acoustic feature vector, and the modulation index vector is output.

[0024] Using the original electromagnetic acoustic signature feature vector as the carrier and the modulation index vector as the modulation source, the amplitude modulation equation is applied to simulate the physical modulation effect of shell vibration on electromagnetic features, and output the modulation feature sequence.

[0025] The modulation feature sequence is subjected to anti-aliasing smoothing processing and integrated into a temporally continuous modulated electromagnetic acoustic feature vector.

[0026] Optionally, the step of performing time-domain coherence calculation on the original electromagnetic acoustic signature feature vector and the modulated electromagnetic acoustic signature feature vector to generate a two-dimensional electromagnetic-mechanical acoustic signature interferogram, wherein the horizontal axis of the spectrum represents time, the vertical axis represents frequency, and the pixel value represents the coherence intensity, includes:

[0027] The original electromagnetic acoustic signature feature vector and the modulated electromagnetic acoustic signature feature vector are input into the time domain alignment module, and sample-level matching is performed based on the shared time reference to output aligned feature vector pairs.

[0028] Based on the aligned eigenvector pairs, the cross spectral density at each frequency point is calculated, and the original coherence coefficient matrix is ​​generated through the coherence function;

[0029] A multi-scale sliding window is applied to adaptively smooth the original coherence coefficient matrix, suppressing noise and enhancing significant fringes, and outputting a smooth coherence matrix.

[0030] The smooth coherence matrix is ​​mapped to a two-dimensional time-frequency grid, where the horizontal axis represents time and the vertical axis represents frequency. Pixel values ​​are quantized into coherence intensity values, and the initial interferogram is output.

[0031] The initial interferogram is subjected to contrast enhancement processing, and the pixel distribution is optimized by histogram equalization to generate an electromagnetic-mechanical acoustic interferogram with high dynamic range.

[0032] Optionally, the step of inputting the electromagnetic-mechanical acoustic interferogram into a pre-trained convolutional neural network (CNN), which is trained to recognize the correspondence between specific coherent fringe patterns in the spectrum and motor fault types, to output the final motor fault diagnosis result, includes:

[0033] The electromagnetic-mechanical acoustic interferogram is preprocessed, including normalizing pixel values ​​to the [0,1] range and data enhancement, and the enhanced interferogram is output.

[0034] The enhanced interferogram is input into a pre-trained multi-scale attention CNN model, which outputs a fault feature map. The CNN model uses adversarial example enhancement to learn the stripe pattern during the training phase.

[0035] Global average pooling and fully connected layers are applied to the fault feature map to generate a fault probability distribution vector to represent the confidence level of various faults.

[0036] Based on the fault probability distribution vector, the final motor fault diagnosis result, including fault type and severity level, is output through Bayesian decision rules and threshold filtering.

[0037] Another embodiment of this application provides a motor controller, the motor controller comprising:

[0038] The acquisition module is used to acquire near-field electromagnetic noise audio signals and vibration acceleration signals of the motor housing during operation, and ensures signal synchronization through hardware timestamps;

[0039] The extraction module is used to perform a short-time Fourier transform on the near-field electromagnetic noise audio signal to extract the instantaneous amplitude envelope sequence of the fundamental frequency and harmonic components, which serves as the original electromagnetic acoustic signature feature vector characterizing the internal electromagnetic state of the motor.

[0040] The modulation module is used to use the vibration acceleration signal as a modulation source to perform amplitude modulation processing on the original electromagnetic acoustic pattern feature vector in order to simulate the physical modulation effect of shell vibration on near-field electromagnetic noise and generate the modulated electromagnetic acoustic pattern feature vector.

[0041] The calculation module is used to perform time-domain coherence calculation on the original electromagnetic acoustic signature feature vector and the modulated electromagnetic acoustic signature feature vector to generate a two-dimensional electromagnetic-mechanical acoustic signature interferogram, wherein the horizontal axis of the spectrum is time, the vertical axis is frequency, and the pixel value represents the coherence intensity.

[0042] The output module is used to input the electromagnetic-mechanical acoustic interferogram into a pre-trained convolutional neural network (CNN), which is trained to identify the correspondence between specific coherent fringe patterns in the spectrum and motor fault types, so as to output the final motor fault diagnosis result.

[0043] Another embodiment of this application provides a storage medium storing a computer program, wherein the computer program is configured to execute the method described in any of the preceding claims when running.

[0044] Another embodiment of this application provides an electronic device including 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 method described in any of the preceding claims.

[0045] Compared with existing technologies, the present invention provides a motor fault diagnosis method that collects near-field electromagnetic noise audio signals and vibration acceleration signals of the motor housing during motor operation; performs short-time Fourier transform on the near-field electromagnetic noise audio signals to extract the original electromagnetic acoustic signature feature vector characterizing the internal electromagnetic state of the motor; uses the vibration acceleration signal as a modulation source to perform amplitude modulation processing on the original electromagnetic acoustic signature feature vector to generate a modulated electromagnetic acoustic signature feature vector; performs time-domain coherence calculation on the original electromagnetic acoustic signature feature vector and the modulated electromagnetic acoustic signature feature vector to generate a two-dimensional electromagnetic-mechanical acoustic signature interferogram; inputs the electromagnetic-mechanical acoustic signature interferogram into a convolutional neural network (CNN) to output the final motor fault diagnosis result, thereby improving the distinguishability of fault features and achieving high-precision diagnosis of motor faults under complex operating conditions. Attached Figure Description

[0046] Figure 1 Hardware structure block diagram of a computer terminal for a motor fault diagnosis method provided in an embodiment of the present invention;

[0047] Figure 2 This is a flowchart illustrating a motor fault diagnosis method provided in an embodiment of the present invention;

[0048] Figure 3 This is a schematic diagram of the structure of a motor controller provided in an embodiment of the present invention. Detailed Implementation

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

[0050] This invention first provides a method for diagnosing motor faults, which can be applied to electronic devices, such as computer terminals, specifically ordinary computers.

[0051] The following detailed explanation uses a computer terminal as an example. Figure 1 This is a hardware structure block diagram of a computer terminal for a motor fault diagnosis method provided in an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.

[0052] Non-volatile storage media can store operating systems and computer programs. These computer programs include program instructions that, when executed, cause the processor to perform any motor fault diagnosis method.

[0053] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0054] The internal memory provides an environment for the execution of computer programs stored in non-volatile storage media. When the computer program is executed by the processor, it enables the processor to perform any motor fault diagnosis method.

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

[0056] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0057] See Figure 2 The present invention provides a method for diagnosing motor faults, which may include the following steps:

[0058] S201 collects near-field electromagnetic noise audio signals and vibration acceleration signals of the motor housing during operation, and ensures signal synchronization through hardware timestamps;

[0059] Specifically, an electromagnetic noise sensor can be deployed close to the motor surface to capture the raw analog signal of near-field electromagnetic noise with a nanosecond-level response, and output an unprocessed analog stream of electromagnetic noise.

[0060] A high-sensitivity magnetoelectric sensor (such as a giant magnetoresistive (GMR) sensor or a Hall effect sensor) is selected. Its core sensing element is composed of a thin layer of special alloy (such as cobalt-iron-boron), capable of sensing micro-Tesla-level magnetic field changes. The sensor is encapsulated in an electromagnetically shielded alloy shell, retaining only the probe contact surface made of nanocrystalline soft magnetic material. During installation, the probe is vertically pressed against the surface of the motor stator housing using a vacuum adsorption clamp (contact pressure controlled at 5-10 Newtons). The probe surface is coated with thermally conductive silicone grease to eliminate air gaps. The sensor has a built-in preamplifier (60dB gain, bandwidth 0.1Hz-100kHz). When the motor winding current changes, its radiated leakage magnetic field penetrates the motor housing, inducing a microvolt-level voltage signal in the sensor probe. This signal, after preamplification, is output as a differential analog signal through a twisted-pair shielded cable, forming a raw electromagnetic noise analog stream (RENAS) with a voltage range of ±2V and a signal-to-noise ratio better than 70dB.

[0061] Nanosecond-level response guarantee mechanism: The sensor's response delay mainly comes from probe eddy current losses and circuit group delay. To achieve nanosecond-level response (NRR), the probe uses a nanocrystalline ribbon material (such as Finemet alloy) with a thickness of only 20 micrometers, reducing eddy current losses to the picosecond level; the signal path uses a gallium arsenide (GaAs) operational amplifier, controlling the group delay to within 1.5 nanoseconds. In real-time monitoring, when a sudden current change of hundreds of amperes occurs at the moment of motor startup, the sensor can capture the magnetic field step signal within 15 nanoseconds, outputting a pulse waveform with a rise time of less than 25 nanoseconds. This performance ensures the resolution of the motor's PWM drive switching frequency (e.g., 20kHz) and its higher harmonics (up to 1MHz).

[0062] Interference immunity and signal fidelity: The sensor output employs double-layer shielding: the inner layer is a copper-plated silver braided mesh (shielding effectiveness >100dB@1MHz), and the outer layer is covered with a ferrite magnetic ring (to suppress common-mode interference). The signal transmission line length does not exceed 1 meter, and the end is connected to an impedance matching network (50-ohm terminating resistor). In variable frequency motor applications, to resist electromagnetic pulse interference (EMI) caused by IGBT switching, the probe incorporates an active cancellation coil, which generates a reverse magnetic field through a feedback circuit, suppressing common-mode interference by more than 40dB. The final output Untreated Electromagnetic Noise Analog Stream (UENAS) retains the original characteristics of the motor's electromagnetic state, providing a high-fidelity foundation for subsequent analysis.

[0063] Piezoelectric triaxial accelerometers are installed at key vibration points of the motor housing. Vibration signals are preprocessed by an anti-aliasing filter, and the filtered vibration acceleration analog stream is output.

[0064] Vibration Point Location and Sensor Installation: Based on motor structure dynamics simulation, critical vibration points (CVPs) are identified: typically located directly above the bearing housing, at the cooling fan mounting base, and at the end of the stator core. A piezoelectric triaxial accelerometer (sensitivity 100mV / g, range ±50g) is selected, containing three sets of orthogonally arranged piezoelectric ceramic crystal stacks (e.g., PZT-5H). During installation, a steel magnetic base (with an adsorption force >200 Newtons) is used, and a rigid coupling agent (tungsten powder epoxy resin) is applied between the base and the housing. Each accelerometer is equipped with an independent charge amplifier (conversion rate 10mV / pC) to convert the charge signal generated by the piezoelectric crystal into a voltage signal. Triaxial signals (X / Y / Z directions) are output separately, comprehensively capturing radial, axial, and tangential vibrations.

[0065] Anti-aliasing filtering: Vibration signals contain high-frequency noise (such as electromagnetic interference and structural resonance). The signal first passes through an anti-aliasing filter (AAF), employing an eighth-order Butterworth low-pass filter (with a cutoff frequency set to 1 / 2.56 times the accelerometer range, configured according to the sampling theorem). For example, when the sampling rate is set to 25.6kHz, the cutoff frequency is 10kHz. The filter is implemented using an analog integrated circuit (such as the MAX291 chip), with in-band ripple less than 0.1dB and stopband attenuation of -80dB / decathlon. In motor bearing fault diagnosis, this design can retain the impact characteristic frequency (such as the bearing outer race fault frequency of 3kHz) while filtering out switching noise higher than 10kHz.

[0066] Signal Conditioning and Output: The filtered signal enters a programmable gain amplifier (PGA), and the gain is dynamically adjusted (1-100 times) according to the vibration intensity. For example, when no-load vibration of the motor (0.1g RMS) is detected, the gain is automatically increased; when overloaded (>40g peak), the protection circuit is triggered to limit the amplitude. The final output is a filtered vibration acceleration analog stream (FVAAS), with a voltage range of ±5V and a frequency response of 0.5Hz-10kHz (±3dB). This signal stream is transmitted through a coaxial cable with an impedance of 75 ohms and can transmit vibration modulation information containing fault characteristics (such as a 100Hz sideband caused by a broken rotor bar) without distortion.

[0067] The electromagnetic noise analog stream and the vibration acceleration analog stream are input into a multi-channel ADC converter, and a hardware timestamp module driven by an atomic clock is embedded to add a precise nanosecond-level time stamp to each sampling point, and output a pair of digital signals with timestamps.

[0068] Synchronous Sampling Architecture: A multi-channel synchronous sampling ADC (such as the ADI AD7606) is employed, configured in dual-channel parallel mode: Channel 1 acquires an electromagnetic noise analog stream (UENAS), and Channel 2 acquires a vibration acceleration analog stream (FVAAS). The ADC resolution is 16 bits, and the sampling rate is programmable (typically 51.2kHz). A key design feature is the synchronous triggering of the sample-and-hold circuit (SHA): the SHA for both channels is controlled by the same trigger pulse (CONVST), ensuring a sampling time deviation of less than 100 picoseconds. The input stage is equipped with overvoltage protection (±15V clamping diodes) and RF suppression (π-type filter).

[0069] Atomic clock timestamp mechanism: The core of the Hardware Timestamp Module (HWTS) is a miniature rubidium atomic clock (such as the Microsemi SA.45s), with a frequency stability of 5e-11. The atomic clock outputs a 10MHz square wave as a time base, driving a 32-bit counter (such as the IDT TDC-GPX) with a counting step size of 100 nanoseconds. When the ADC completes each sampling (i.e., the rising edge of CONVST), the following actions are triggered:

[0070] The current value of the counter is latched into the timestamp register;

[0071] The ADC conversion result (digital signal) is linked to a timestamp;

[0072] A timestamped digital signal pair (TDSP) is formed, with the data structure being: {electromagnetic noise sample value (int16), vibration acceleration sample value (int16), timestamp (uint32)}. The timestamp accuracy reaches the nanosecond level (measured jitter <3ns), completely eliminating software timing errors.

[0073] Real-time data stream processing: The TDSP transmits data to the FPGA preprocessing unit via a high-speed parallel bus (such as LVDS). The FPGA implements a ping-pong buffer architecture: dual-port RAM partitions alternately store data packets (512 sampling points per packet). A CRC-32 checksum is added during transmission, resulting in a bit error rate of less than 1e-9. Under motor acceleration conditions (such as 0-3000rpm ramp start), the system can still maintain timestamp continuity, ensuring no loss of 51,200 sampling points per second.

[0074] A timestamp alignment algorithm is applied to dynamically time-normalize digital signal pairs based on hardware timestamps, eliminating sampling jitter and generating synchronized electromagnetic noise digital signals and vibration acceleration digital signals.

[0075] Time base normalization: Convert the timestamps in TDSP to an absolute time axis (unit: nanoseconds). Establish a global time reference point (e.g., set the timestamp of the first sampling point to 0). Since the sampling rate of electromagnetic and vibration signals is the same (e.g., 51.2kHz), the theoretical sampling interval should be 19.53125 microseconds. However, due to clock jitter, the actual interval has a deviation of ±15 nanoseconds. Timestamp Alignment Algorithm (TAA): First, calculate the theoretical time position of each sampling point:

[0076] The theoretical time for the nth point is T_theory(n) = n×19,531,250ns; the actual timestamp T_actual(n) comes from HWTS, and the deviation ΔT(n) = T_actual(n) - T_theory(n) is the sampling jitter.

[0077] Implementation of dynamic time warping:

[0078] Dynamic Time Warping (DTW) algorithm is used to eliminate jitter:

[0079] Construct the cost matrix: Using the theoretical time axis as a reference, calculate the Euclidean distance between the actual sampling points and the theoretical locations;

[0080] Path search: Use dynamic programming to find the minimum cost path (Sakoe-Chiba bandwidth constraint ±5 samples);

[0081] Signal resampling: Cubic spline interpolation is performed on sample points that deviate from the theoretical position.

[0082] For example, if the actual time of a vibration signal sample point is 200 nanoseconds ahead of its theoretical time point, its amplitude is adjusted to the theoretical time point using an interpolation function. This process simultaneously processes electromagnetic and vibration signals, ensuring strict alignment of the two signal time axes.

[0083] Synchronization signals are generated: After normalization, synchronized electromagnetic noise digital signals (SENDS) and synchronized vibration acceleration digital signals (SAVADS) are output. The signal format is an equally spaced array (strictly spaced at 19.53125 microsecond intervals), with the same length and index numbers directly corresponding to the same time. In the detection of faults in the outer race of motor bearings, this synchronization ensures that the 1kHz carrier wave in the electromagnetic signal and the 3kHz fault characteristics in the vibration signal can be accurately correlated in the time and frequency domain, laying the foundation for subsequent modulation analysis.

[0084] This step involves synchronously acquiring electromagnetic noise and vibration signals from the motor using an electromagnetic noise sensor and a piezoelectric accelerometer, and precisely aligning the temporal relationship between the two types of signals using hardware timestamp technology. Electromagnetic noise reflects the internal electromagnetic state of the motor, while vibration signals characterize the dynamic properties of the mechanical structure. The synchronous acquisition of both lays the foundation for subsequent fusion analysis, ensuring strict synchronization between electromagnetic and mechanical signals, avoiding feature correlation failure due to time deviations, and providing a high-precision data source for multimodal fault feature extraction.

[0085] S202, Perform a short-time Fourier transform on the near-field electromagnetic noise audio signal to extract the instantaneous amplitude envelope sequence of the fundamental frequency and harmonic components, which serves as the original electromagnetic acoustic signature feature vector characterizing the internal electromagnetic state of the motor.

[0086] Specifically, a variable window length short-time Fourier transform can be applied to synchronized electromagnetic noise digital signals, with the window size adaptively adjusted according to the instantaneous energy of the signal, to output a high-resolution time-frequency energy matrix.

[0087] The system first receives a timestamped, time-stamped electromagnetic noise digital signal (ENDS), which is a discrete-time sequence acquired at a fixed sampling rate (e.g., 48 kHz). To achieve high-resolution spectrum analysis, a Variable Window Short-Time Fourier Transform (VW-STFT) is employed. The core mechanism lies in the dynamic adjustment of the window length (WL): the system calculates the instantaneous energy (IE) of the signal within a sliding time frame in real time, which is the mean of the sum of the squares of the amplitudes of all sampling points within that frame. For example, when a sudden increase in instantaneous energy is detected (e.g., a current surge at motor startup causing a 30% increase in noise energy), the system automatically shortens the window length (e.g., from 256 points to 128 points) to improve time-domain resolution and capture rapid transient changes; conversely, during periods of stable energy (e.g., constant speed operation), the window length is increased (e.g., to 512 points) to improve frequency-domain resolution and accurately separate dense harmonics. The Hanning window type is selected to reduce spectral leakage. Each time STFT is performed, the window slides along the time axis with a fixed step size (e.g., 64 points), and a Fast Fourier Transform (FFT) is performed on each window segment to calculate the complex spectrum.

[0088] Adaptive window adjustment relies on a pre-defined Energy-Window Length Mapping Table (EWLMT). This table, trained using historical data, establishes a correspondence between instantaneous energy values ​​(IE) and optimal window lengths (WL). For example, when IE < 0.1 (normalized value), WL = 512 points; when 0.1 ≤ IE < 0.3, WL = 256 points; and when IE ≥ 0.3, WL = 128 points. In real-time processing, the system calculates the IE value for each new time frame received, determines the WL by looking up the table or interpolation, and dynamically calls the corresponding FFT algorithm. The FFT result for each window segment generates a Spectral Vector (SV), containing both the magnitude spectrum and the phase spectrum. Arrange the amplitude spectra of all time frames in chronological order to form a two-dimensional matrix. The row indices correspond to frequency points (frequency resolution is determined by the sampling rate and window length), and the column indices correspond to time points (time resolution is determined by the step size). The matrix element value is the energy density (ED) of that time-frequency unit, which is the square of the amplitude. This matrix is ​​the Time-Frequency Energy Matrix (TFEM).

[0089] To ensure the usability of the matrix, the system performs post-processing: First, frequency axis calibration is performed, calculating the actual frequency value (FrequencyValue, FV = k×SR / WL, where k is the frequency point index) based on the sampling rate (SR) and window length (WL). Second, background noise suppression (BNS) is implemented, for example, by calculating the full-time average energy spectrum as the noise basis and subtracting it point by point from the TFEM. Finally, dynamic range compression (DRC) is performed, such as taking the logarithm of the energy value (log(1+ED)) to enhance the visibility of weak components. The output TFEM has the dimension of [number of time frames × number of frequency points], with each element storing the compressed energy value, forming a high-resolution time-frequency representation that can be directly used for subsequent analysis.

[0090] Using the time-frequency energy matrix as input, a fundamental frequency detector based on a convolutional neural network is run to identify the dominant fundamental frequency component and its time-varying trajectory, and output the instantaneous position sequence of the fundamental frequency.

[0091] The Time-Frequency Energy Matrix (TFEM) is fed into a pre-trained Convolutional Neural Network Fundamental Frequency Detector (CNN-FFD). This CNN employs a lightweight architecture: the input layer receives the TFEM (considered a single-channel grayscale image); the first layer is a convolutional layer (Conv) with a 3×3 convolution kernel to extract local time-frequency features (such as energy ridges); the second layer is a 2×2 max pooling layer (MP) to reduce dimensionality and enhance translation invariance; subsequent layers of convolutional and pooling are stacked to gradually abstract high-level features; the network ends with a fully connected layer (FC) and a softmax output layer. The network output is a Fundamental Frequency Probability Distribution Vector (FFPDV) for each time frame, with a length equal to the number of frequency points, and each element representing the probability that the frequency point is the fundamental frequency.

[0092] The training of CNN-FFD is based on massive amounts of labeled data: motor noise TFEM samples are collected under different operating conditions (no load, load, variable speed) and fault states (such as rotor bar breakage, bearing wear), and the true fundamental frequency position (TFFP) of each frame is manually labeled. During training, the cross-entropy loss function (CELF) and the adaptive moment estimation optimizer (AMEO) are used. During the inference phase, the system processes TFEM frame by frame: the matrix of the current frame is input into CNN-FFD and the output is FFPDV; the frequency point with the highest probability is selected as the candidate fundamental frequency (CFF); the temporal continuity constraint (TCC) is applied: if the CFF jump of the adjacent frame exceeds the threshold (e.g., 10Hz), the trajectory smoothing algorithm (TSA) is started, and Kalman filtering (KF) is performed based on the position of the previous frame and the rate of change of motor speed to correct the position, and the corrected instantaneous fundamental frequency position (IFFP) is output.

[0093] To handle harmonic interference or transient signal loss, the system introduces Confidence Verification (CV): the ratio of the highest probability value to the second highest probability value in the FFPDV is calculated as the Confidence Score (CS). If the CS is below a threshold (e.g., 1.5), the detection of that frame is deemed unreliable, and a Redundant Detection Mechanism (RDM) is activated: the theoretical fundamental frequency range is calculated based on the motor's rated speed and number of pole pairs, and the spectral peak with the highest energy within this frequency band is searched as a supplementary estimate. Finally, the IFFPs of all time frames are arranged in chronological order to form a Fundamental Instantaneous Position Sequence (FIPS), where each element is the frequency point number or actual frequency value (in Hz) corresponding to the fundamental frequency, accurately depicting the trajectory of the fundamental frequency over time.

[0094] Based on the fundamental frequency instantaneous position sequence, the harmonic bands of integer multiples are extracted using the harmonic tracking algorithm, the instantaneous amplitude of each harmonic component is calculated, and the harmonic amplitude sequence set is output.

[0095] The Harmonic Tracking Algorithm (HTA) uses the Fundamental Frequency Instantaneous Position Sequence (FIPS) as a reference. For each time frame, based on the current frame's Base Frequency Value (BFV), the theoretical positions of its integer harmonics (e.g., 2nd, 3rd, ... Nth times) are calculated, forming the Harmonic Band Center (HBC). A symmetrical Extraction Bandwidth (EBW) is set around each HBC, for example, ±5% of BFV. In the current frame sequence of the Time-Frequency Energy Matrix (TFEM), the spectral peak with the highest energy within this band is located, and its frequency is the Instantaneous Harmonic Frequency (IHF). For example, if the current BFV = 50Hz, then the 2nd harmonic band is 95-105Hz. The search for the maximum energy peak within this range, assuming it is located at 101Hz, then IHF = 101Hz.

[0096] Harmonic amplitude calculation employs the Adaptive Bandpass Filtering Method (ABFM): a digital bandpass filter (such as an FIR filter) is constructed with IHF as the center frequency, and its bandwidth is dynamically set according to the harmonic sparsity (e.g., narrow near the fundamental frequency and widened in the high-frequency range). This filter is applied to the current frame of the original electromagnetic noise digital signal (ENDS) to obtain the harmonic component time-domain signal (HCTDS). The root mean square (RMS) value of this signal is calculated as the instantaneous amplitude (IA) of the harmonic. For example, after filtering the 3rd harmonic component, the RMS value of its signal near 150Hz is calculated as the amplitude. To ensure noise immunity, a Harmonic Confidence Verification (HCV) is introduced: if the ratio of the highest energy in the frequency band to the energy of the adjacent noise floor is less than a threshold (e.g., 10dB), the harmonic is considered missing, and its amplitude is recorded as 0.

[0097] The algorithm processes all harmonic orders (e.g., 1st to 10th) and all time frames frame by frame. For the k-th harmonic (k=2,3,...,N), the instantaneous amplitudes of all its time frames constitute a Harmonic Amplitude Sequence (HAS_k). The HAS_k values ​​of all harmonics form a Harmonic Amplitude Sequence Set (HASS), which can be represented as {HAS_2, HAS_3, ..., HAS_N}. Each HAS_k is a one-dimensional array of length equal to the number of time frames, storing the amplitude of the corresponding harmonic as it changes over time. The system also outputs the Instantaneous Frequency Sequence (IFS_k) of each harmonic, recording the change of IHF over time, for analyzing frequency modulation phenomena.

[0098] Apply Hilbert transform to the set of fundamental frequency instantaneous position sequence and harmonic amplitude sequence to extract the instantaneous envelope of each component and output the fundamental frequency envelope sequence and harmonic envelope sequence.

[0099] The Hilbert Transform (HT) is used to extract the instantaneous envelope (IE) from oscillating signals. First, the fundamental frequency component is processed: based on the frequency trajectory in the fundamental frequency instantaneous position sequence (FIPS), a dynamic bandpass filter (DBF) is applied to the original electromagnetic noise digital signal (ENDS). The filter's center frequency tracks the value in the FIPS in real time, with a fixed bandwidth (e.g., ±2Hz). The filtered signal yields the fundamental time-domain signal (FTDS). The Hilbert Transform is then applied to this signal: a 90-degree phase-shift network is constructed (approximately achievable using an FIR filter), generating the orthogonal signal (QS) of the FTDS. The FTDS and QS together constitute an analytic signal (AS), whose magnitude is the instantaneous envelope value.

[0100] The harmonic components are processed in parallel: starting from the harmonic amplitude sequence set (HASS) and the corresponding instantaneous frequency sequence (IFS_k), the time-domain signal (Harmonic Time-Domain Signal, HTDS_k) of each harmonic is extracted through dynamic bandpass filtering (center frequency tracking IFS_k). A Hilbert transform is applied independently to each HTDS_k: its orthogonal signal QS_k is calculated, and the analytic signal AS_k = HTDS_k + j·QS_k is constructed. The instantaneous envelope IE_k is |AS_k| (the modulus of the complex number). Since the envelope signal frequency is much lower than the carrier frequency, low-pass filtering (LPF) is required to smooth out glitches; the cutoff frequency is set to twice the highest envelope variation frequency (e.g., 50Hz). Finally, the instantaneous envelope sequence (IES_k) of each harmonic is obtained.

[0101] To eliminate differences in amplitude dimensions, all envelope sequences are normalized (NP): using the long-term mean of the fundamental frequency envelope as a benchmark, the ratio of each harmonic envelope to this benchmark is calculated. The output includes:

[0102] Fundamental Envelope Sequence (FES): A sequence of envelope values ​​extracted from a fundamental frequency signal.

[0103] Harmonic Envelope Sequence Set (HESS): Contains envelope sequences {HES_2, HES_3, ..., HES_N} for each harmonic (e.g., from the 2nd to the 10th harmonic).

[0104] Each sequence has a length consistent with the number of time frames and stores the envelope amplitude (dimensionless relative value), which characterizes the slow change of amplitude of each frequency component over time, providing key modulation information for subsequent fault diagnosis.

[0105] By fusing the fundamental frequency envelope sequence and harmonic envelope sequence, and then splicing and normalizing in the time domain, the original electromagnetic acoustic signature feature vector is generated.

[0106] The goal of the fusion operation is to integrate the scattered envelope sequences into a unified feature representation. First, temporal alignment (TA) is performed: since the fundamental frequency and harmonic envelope sequences are generated based on the same time frame, their time axes are intrinsically aligned. Feature vectors are organized by time frame: the feature of each time frame consists of the fundamental envelope value (FEV) and all harmonic envelope values ​​(HEV_k) for that frame. For example, if tracking 10 harmonics, a single frame feature contains 11 elements: [FEV, HEV_2, HEV_3, ..., HEV_10]. The features of all time frames are arranged vertically in chronological order to form an initial feature matrix (FM), with dimensions of [number of time frames × feature dimension (1 + number of harmonics)].

[0107] To eliminate dimensional differences and numerical range fluctuations, normalization is performed:

[0108] Intra-frame Normalization (IFN): For all feature elements (FEV and HEV_k) of a single frame, divide by the fundamental frequency envelope value FEV of that frame to make the fundamental frequency envelope always 1. The harmonic envelope represents the relative intensity with respect to the fundamental frequency.

[0109] Inter-frame Normalization (IFN): Calculate the mean and standard deviation of each feature dimension (such as HEV_3) on the entire training set. Perform z-score normalization on each column of FM (i.e. each feature dimension), that is, subtract the mean of the column and divide by the standard deviation, so that the mean of each dimension is 0 and the variance is 1.

[0110] This process preserves the relative change patterns of the envelope while improving feature stability.

[0111] The final result is the raw electromagnetic voiceprint feature vector (REVFV). This "vector" is actually a time-varying multidimensional sequence, which can be viewed as a sequence of feature vectors arranged in chronological order. Its data structure is as follows:

[0112] Time dimension: Length equals the total number of time frames (e.g., T-frames).

[0113] Feature dimension: Length equals 1 (fundamental frequency envelope) + N (harmonic envelope number, such as 10), totaling D dimensions.

[0114] The output is a two-dimensional array [time index × feature index], where each element stores the normalized envelope value. This feature vector fully characterizes the dynamic behavior of the fundamental frequency and harmonic amplitude of the motor's electromagnetic noise as a function of time, providing core input for subsequent vibration modulation processing and fault diagnosis.

[0115] By using a short-time Fourier transform with an adaptive window length, the electromagnetic noise signal is converted into a time-frequency energy matrix. Then, the instantaneous amplitude envelope of key frequency components is extracted using a fundamental frequency detection and harmonic tracing algorithm to form an electromagnetic acoustic signature vector. This separates the fundamental frequency and harmonic features directly related to the electromagnetic state of the motor from the electromagnetic noise, quantifies the dynamic change law of electromagnetic anomalies, and provides an interpretable electromagnetic feature expression for fault diagnosis.

[0116] S203, using the vibration acceleration signal as a modulation source, the original electromagnetic acoustic signature feature vector is subjected to amplitude modulation processing to simulate the physical modulation effect of shell vibration on near-field electromagnetic noise, and a modulated electromagnetic acoustic signature feature vector is generated.

[0117] Specifically, wavelet packet decomposition can be performed on the synchronous vibration acceleration digital signal to extract the frequency band components related to the mechanical resonance of the motor and output the key vibration frequency band signal.

[0118] Initialization and parameter configuration of wavelet packet decomposition: First, the system receives the synchronized vibration acceleration digital signal (SVADS) output from the above steps. This signal has been aligned with hardware timestamps to ensure strict synchronization with the electromagnetic noise signal. Wavelet packet decomposition (WPD) uses the Db10 wavelet basis function (Daubechies 10, a compactly supported orthogonal wavelet with a 10th-order vanishing moment), as it can effectively preserve transient impact characteristics in mechanical vibration signal analysis. The decomposition level is set to 6 levels (Decomposition Level, DL), and the optimal sub-band tree structure is adaptively selected using the Shannon Entropy Criterion (SEC). At the hardware level, the embedded DSP (Digital Signal Processor) performs real-time decomposition: SVADS is divided into blocks with a frame length of 32 milliseconds (ms). Each frame of signal is divided into frequency bands through a polyphase filter bank (PFB) to generate 64 wavelet packet nodes (WPNs) covering the entire frequency band from 0 to 5 kHz. Each node corresponds to a specific sub-band frequency range (SFR).

[0119] Feature extraction of mechanical resonance frequency bands: To identify key frequency bands related to motor mechanical resonance, the system employs the Envelope Entropy Algorithm (EEA) to quantify the fault sensitivity of each sub-frequency band. The specific process is as follows:

[0120] Perform a Hilbert Transform (HT) on each wavelet packet node signal to extract its instantaneous envelope (IE).

[0121] Calculate the Shannon Entropy Value (SEV) of the envelope signal. The lower the entropy value, the more concentrated the signal energy (potential resonant band).

[0122] By combining the Prior Fault Spectrum Library (PFSL), sub-bands with entropy values ​​below the threshold of 0.8 (an empirical value) and located within the motor's natural frequency range (such as the bearing fault characteristic band of 1-3 kHz) are selected. Finally, 3-5 key vibration band signals (KVBS) are output, such as the bearing outer ring fault characteristic band (BPFO band) and the rotor unbalance band.

[0123] Real-time processing and resource optimization: To meet the needs of online diagnosis, the system adopts a sliding window incremental update strategy (SWIUS): full decomposition is performed only on signal variation areas (identified by a differential energy detector (DED)), while historical decomposition results are reused for other areas. Key vibration frequency band signals (KVBS) are output to the next module in the form of floating-point arrays. Each array contains a timestamp (TS), a band identifier (BID), and a normalized amplitude sequence (NAS).

[0124] The key vibration frequency band signal is input into a nonlinear normalizer and converted into a modulation index sequence. Its value range is dynamically mapped to the amplitude range of the original electromagnetic acoustic feature vector, and the modulation index vector is output.

[0125] Design principle of nonlinear normalizer: The core function of nonlinear normalizer (NLN) is to dynamically compress the amplitude of KVBS to a range compatible with the original electromagnetic acoustic feature vector (OEMAFV) (e.g., [0, 1]). This is achieved using a piecewise sigmoid function (PSF).

[0126] Low amplitude region (amplitude < threshold TH_low=0.2): Use a near-linear mapping with slope K1=5.0 to enhance sensitivity to weak vibrations.

[0127] In the mid-amplitude range (0.2 ≤ amplitude ≤ 0.8): a gradually varying Sigmoid function (SigmoidFunction, SF) in the saturation range is used to avoid overmodulation.

[0128] High amplitude region (amplitude > threshold TH_high=0.8): The output is limited to no more than 1.0 by upper clipping (UC) to suppress impulse noise interference.

[0129] Dynamic range mapping mechanism: To adapt to the amplitude fluctuations of OEMAFV under different operating conditions, NLN introduces a dynamic range adaptation module (DRAM):

[0130] Real-time monitoring of the sliding window amplitude range (SWAR) of OEMAFV, with a window length of 1 second.

[0131] Adjust the PSF parameters based on SWAR: when SWAR increases, increase TH_low to expand the linear region; when SWAR decreases, decrease K1 to enhance sensitivity.

[0132] The final output is a Modulation Index Sequence (MIS), where each element value (e.g., 0.35) represents the modulation intensity of the vibration on the electromagnetic characteristics at the current moment.

[0133] Generation and verification of the modulation index vector: The MIS is subjected to a temporal consistency check (TCC) to remove outliers (such as amplitude abrupt changes caused by sensor interruptions) and then smoothed by a Kalman filter (KF). The final output is a modulation index vector (MIV), whose dimensions are strictly aligned with the OEMAFV time points and whose value range is always [0,1], which can be directly used in the amplitude modulation equation.

[0134] Using the original electromagnetic acoustic signature feature vector as the carrier and the modulation index vector as the modulation source, the amplitude modulation equation is applied to simulate the physical modulation effect of shell vibration on electromagnetic features, and output the modulation feature sequence.

[0135] Physical Modeling of the Amplitude Modulation Equation: The Amplitude Modulation Equation (AME) is constructed based on Electromechanical Coupling Theory (ECT) to simulate how vibration periodically disturbs the electromagnetic field distribution by changing the stator-rotor gap or magnetic reluctance of the motor. The mathematical formula is: Modulated Amplitude = Original Amplitude × (1 + Modulation Index × Carrier Frequency Coupling Factor).

[0136] Among them, the carrier frequency coupling factor (CFCF) is determined by the intermodulation coefficient (IMC) between the fundamental frequency (e.g., 50Hz) of the electromagnetic characteristic component and the center frequency of the vibration band (e.g., 2kHz), and is obtained through offline calibration experiments (e.g., when the fundamental frequency is 50Hz and the vibration band is 2kHz, CFCF=0.07).

[0137] Real-time modulation process execution: The system implements modulation in hardware using a point-wise multiplier (PWM).

[0138] For each component of OEMAFV (such as the fundamental frequency amplitude A_k at time k), the modulation index M_k at the corresponding time is read from MIV.

[0139] Based on the frequency identifier of this component (provided by the instantaneous position sequence of the fundamental frequency output from the above steps), retrieve the preset CFCF value (stored in the Coupling Factor Lookup Table (CFLT)). Calculate the modulated amplitude: A'_k = A_k×[1+ M_k×CFCF].

[0140] Repeat the above operation on the harmonic components to generate a Modulated Feature Sequence (MFS).

[0141] Accurate simulation of physical effects: To enhance realism, the system introduces a multi-path modulation compensation (MPMC) mechanism.

[0142] Path 1: Direct modulation (main path), as described above; Path 2: Inter-harmonic intermodulation caused by vibration (such as 2nd harmonic and 3rd harmonic generating 5th harmonic), sideband components are added through the Intermodulation Term Injection Module (ITIM).

[0143] Ultimately, the MFS includes the main modulation signal and compensation sidebands, fully restoring the physical modulation effect of shell vibration.

[0144] The modulation feature sequence is subjected to anti-aliasing smoothing processing and integrated into a temporally continuous modulated electromagnetic acoustic feature vector.

[0145] Anti-aliasing filter design: The modulation process may introduce spurious high-frequency components above the Nyquist frequency (NF, determined by the original sampling rate of 48kHz, NF=24kHz). The system employs a Finite Impulse Response Low-pass Filter (FIR-LPF) for anti-aliasing:

[0146] Filter order: 128 taps, to ensure a steep transition band;

[0147] Cutoff frequency: 0.8 × NF = 19.2 kHz;

[0148] Window function: Blackman window (BW), which suppresses spectral leakage.

[0149] FIR-LPF processes MFS in real time using convolution operation (CO) and outputs a filtered modulated sequence (FMS).

[0150] Temporal smoothing and continuity assurance: To avoid boundary discontinuities caused by frame processing, the overlap-save method (OLSM) is adopted.

[0151] FMS is divided into 50% overlapping frames (frame length 256 points).

[0152] Apply a Hanning Window (HW) to the edges of each frame to smooth them.

[0153] Continuous waveforms are reconstructed by weighted overlap-add (WOLA) to generate smoothed modulated sequences (SMS).

[0154] Feature vector integration and output: The SMS is recombined in chronological order and amplitude normalization (AN) is performed to eliminate scale differences: normalized value = (current amplitude - minimum value of sliding window) / (maximum value of sliding window - minimum value of sliding window).

[0155] The window length is 1 second and it updates dynamically. The final output is a modulated electromagnetic acoustic feature vector (MEAFV), whose dimensions and time reference are completely consistent with the original vector (OEMAFV), and it can be directly used for subsequent coherence calculations.

[0156] Based on the key frequency band components of the vibration signal, a modulation index is generated to modulate the amplitude of the electromagnetic acoustic signature, simulating the actual physical coupling effect of mechanical vibration on electromagnetic noise. Through physical modeling, the correlation between electromagnetic features and mechanical vibration is enhanced, revealing the abnormal mode of electromechanical coupling mechanism under faults and improving the sensitivity of features to complex faults.

[0157] S204, perform time-domain coherence calculation on the original electromagnetic acoustic signature feature vector and the modulated electromagnetic acoustic signature feature vector to generate a two-dimensional electromagnetic-mechanical acoustic signature interferogram, wherein the horizontal axis of the spectrum is time, the vertical axis is frequency, and the pixel value represents the coherence intensity.

[0158] Specifically, the original electromagnetic acoustic signature feature vector and the modulated electromagnetic acoustic signature feature vector can be input into the time domain alignment module, and sample-level matching can be performed based on the shared time reference to output aligned feature vector pairs.

[0159] Time reference synchronization mechanism: Although the original electromagnetic acoustic signature feature vector (representing the internal electromagnetic state of the motor) and the modulated electromagnetic acoustic signature feature vector (reflecting the vibration modulation effect) have been macroscopically synchronized through hardware timestamps, further elimination of microscopic sampling deviations is needed. The time-domain alignment module first reads the nanosecond-level hardware timestamp shared by the two feature vectors (generated by a timing module driven by an atomic clock, with an accuracy of 10 nanoseconds). This module incorporates a Dynamic Time Warping (DTW) algorithm, which calculates the local distance matrix of the two vector time axes (using Euclidean distance as the distance metric) to find the optimal warping path, forcibly aligning the sampling points (samples) describing the same physical moment in the two vectors. For example, if the electromagnetic vector index is T1 and the vibration modulation vector index is T2 at a certain moment, and DTW determines that the time deviation between T1 and T2 is less than 20 microseconds (the allowable error threshold for motor fault diagnosis), then T1 and T2 are marked as matching points. Finally, a strictly aligned feature vector pair (FVP) is generated, ensuring that subsequent coherence calculations are based on data from the same physical moment.

[0160] Sample-level interpolation compensation technology: When two vectors experience non-integer multiple sample point offsets due to sensor response delays or ADC sampling rate differences, the alignment module activates a cubic spline interpolator. Using the electromagnetic vector as the reference time axis, the modulation vector is resampled. For example, if the electromagnetic vector sampling rate is 48 kHz and the modulation vector is 44.1 kHz, the interpolator calculates new sample point values ​​for the modulation vector at 48 kHz intervals, fitting a smooth curve based on four adjacent original samples (curve parameters are determined by solving a third-order polynomial coefficient). After interpolation, the two vectors have the same length and their time axes completely overlap. Simultaneously, the module detects and removes invalid segments caused by signal loss (such as sensor momentary interruptions), retaining only continuous valid data segments as the FVP output.

[0161] Real-time drift correction strategy: To address the cumulative clock drift that may occur during long-term operation, the alignment module periodically calls the Least Squares Linear Fit (LSLF) algorithm. A timestamp sequence is collected every 10 seconds to fit the time offset function of the electromagnetic and vibration modulation signals (in the form Δt = a×t + b, where t is time, a is the drift rate, and b is the initial offset). The fitting results are used to dynamically correct the time stamp of the modulation vector (e.g., the modulation vector sample at the 5th second is forward-compensated by a×5 + b milliseconds), ensuring that the sample-level synchronization accuracy of the entire data segment is better than 50 microseconds (the minimum time resolution for the electromagnetic-vibration coupling effect of the motor).

[0162] Based on the aligned eigenvector pairs, the cross spectral density at each frequency point is calculated, and the original coherence coefficient matrix is ​​generated through the coherence function;

[0163] Segmented spectrum estimation method: The aligned feature vector pairs (FVP) are segmented into overlapping time windows, the window length of which is dynamically adjusted according to the motor speed (e.g., a window length of 100 milliseconds (ms) corresponds to a rated speed of 1500 rpm). A Hanning window function (a cosine-weighted window) is applied to each window to suppress spectral leakage. The complex spectra of the original electromagnetic vector segment (denoted as X) and the modulated vector segment (denoted as Y) are calculated using a Fast Fourier Transform (FFT) to obtain X(f) and Y(f), where f is the frequency point (frequency resolution of 10 Hz).

[0164] Cross Spectral Density Calculation: For each frequency point f, calculate the cross spectral density (CSD) of the two vectors. The formula is conceptualized as: CSD(f) = X(f) × conj(Y(f)), where conj represents the complex conjugate operation, which reflects the phase relationship.

[0165] For example, at a fundamental frequency of 50 Hz, if X(f) has an amplitude of 0.8 and a phase of 30 degrees, and Y(f) has an amplitude of 0.75 and a phase of 45 degrees, then CSD(f) has an amplitude of 0.6 (0.8 × 0.75) and a phase of -15 degrees (30° - 45°). CSD reflects the energy coupling strength and phase difference between the two signals in the frequency domain and is the basis for coherence calculation.

[0166] Construction of the coherence coefficient matrix: Based on CSD(f), the magnitude squared coherence (MSC) is further calculated, and its mathematical formula is: MSC(f) = |CSD(f)| 2 / [PXX(f) × PYY(f)].

[0167] Where PXX(f) is the autopower spectrum of X (X(f)×conj(X(f))) and PYY(f) is the autopower spectrum of Y. The MSC(f) value range is 0 to 1 (1 indicates complete coherence, 0 indicates no correlation). For each time window and frequency point, an MSC value is generated. The MSC values ​​of all time windows are arranged in chronological order to form the raw coherence matrix (RCM), where the row index is the time window number (time dimension), the column index is the frequency point (frequency dimension), and the element value is MSC(f).

[0168] A multi-scale sliding window is applied to adaptively smooth the original coherence coefficient matrix, suppressing noise and enhancing significant fringes, and outputting a smooth coherence matrix.

[0169] Multi-scale window design: To address potential random noise (such as sensor thermal noise) and local interference in RCM, a three-level sliding window is used for smoothing.

[0170] Micro-scale: A 3×3 pixel window used to filter out single-point impulse noise (such as electromagnetic interference spikes).

[0171] Meso-scale: A 5×5 pixel window covering the width of a typical fault stripe (such as the 100 Hz sideband of a bearing fault).

[0172] Macro-scale: 15-pixel time axis × 7-pixel frequency axis window, matching the motor load fluctuation cycle (e.g., 150 time windows corresponding to a 10-second cycle).

[0173] Adaptive weighted smoothing algorithm: Gaussian weighted average (GWA) is used within each window. Pixels at the center of the window have the highest weight (e.g., Gaussian function standard deviation σ=0.5), while edge pixels have lower weight. For meso- and macro-scale windows, a Local Variance Detector (LVD) is introduced: the variance Var of pixel values ​​within the window is calculated. If Var is higher than a threshold (e.g., 0.1), a valid fault stripe is identified, and only micro-scale smoothing is initiated; if Var is lower than the threshold, it is identified as a background noise region, and three-level joint smoothing is initiated. For example, for a 5×5 window with Var=0.05 (low variance), three-level Gaussian smoothing is applied to it, significantly suppressing uniform noise.

[0174] Stripe Enhancement: After smoothing, Directional Gradient Enhancement (DGE) is used to enhance fault-related coherent stripes (typically appearing as diagonal lines or curves). Using the stripe orientation as a reference (angle θ estimated via Hough Transform), the gradient perpendicular to θ is calculated (using the Sobel operator). The gradient magnitude is added to the original matrix (weighting coefficient β = 0.3) to sharpen the stripe edges. The final output is a Smoothed Coherence Matrix (SCM), which improves the signal-to-noise ratio by at least 10 dB.

[0175] The smooth coherence matrix is ​​mapped to a two-dimensional time-frequency grid, where the horizontal axis represents time and the vertical axis represents frequency. Pixel values ​​are quantized into coherence intensity values, and the initial interferogram is output.

[0176] Mesh parameterization mapping: Define a two-dimensional mesh: horizontal axis (time axis): the length is equal to the number of time windows of the SCM (e.g., 1000 windows), and the unit time interval is the window shift (Hop Size, e.g., 5 milliseconds (ms)).

[0177] Vertical axis (frequency axis): range from 0 to the analysis bandwidth (e.g., 5 kHz), the frequency resolution is determined by the number of FFT points (e.g., 512-point FFT corresponds to 9.76 Hz / pixel).

[0178] Linearly map each element MSC(t,f) in SCM to a grid pixel brightness value (BV): BV(t,f) = floor(MSC(t,f) × 255), where floor represents rounding down, converting MSC values ​​from 0 to 1 into 8-bit grayscale values ​​from 0 to 255.

[0179] Non-uniform data interpolation: If the SCM frequency resolution is inconsistent with the grid (e.g., the SCM resolution is 10 Hz, but the grid requires 9.76 Hz), bilinear interpolation is used. Taking a target grid point (t_target, f_target) as an example:

[0180] Find the four points around it in the SCM: (t1,f1), (t1,f2), (t2,f1), (t2,f2);

[0181] Calculate the weight w_f of f_target between f1 and f2 (w_f = (f_target - f1) / (f2 - f1));

[0182] Calculate the weight w_t of t_target between t1 and t2 (w_t = (t_target - t1) / (t2 - t1));

[0183] The MSC values ​​of the four points are mixed: MSC_target=(1-w_t)(1-w_f)MSC(t1,f1) + (1-w_t)w_fMSC(t1,f2) + w_t(1-w_f) MSC(t2,f1) + w_t w_f MSC(t2,f2).

[0184] Finally, an initial interference map (IIM) with continuous pixel positions is generated.

[0185] The initial interferogram is subjected to contrast enhancement processing, and the pixel distribution is optimized by histogram equalization to generate an electromagnetic-mechanical acoustic interferogram with high dynamic range.

[0186] Histogram Statistics and Cumulative Distribution: Calculate the grayscale histogram of the IIM, recording the number of pixels H(k) for each grayscale level k (0 - 255). Calculate the normalized cumulative distribution function (Cumulative Distribution Function, CDF): CDF(k) = Σ_{i=0}^k H(i) / Total_Pixels, where Total_Pixels is the total number of pixels in the image. For example, if CDF = 0.15 when k = 100, it means that the brightness of 15% of the pixels ≤ 100.

[0187] Adaptive Equalization Mapping: Use the CDF as the transformation function to reassign pixel grayscale values: New_BV = floor(255 × CDF(Original_BV)).

[0188] This operation stretches the original grayscale distribution to the full range (0 - 255). Considering the characteristics of the motor fault pattern (fault stripes usually occupy a small number of pixels), add a contrast limiting mechanism: set the limiting threshold CL = 0.02 (empirical value), merge the grayscale levels in the CDF below CL, and compress the grayscale levels above 1 - CL to prevent excessive amplification of noise. For example, if the pixel with the original BV = 10 accounts for 0.5% (< CL = 2%), it is mapped to the new BV = 5; the pixel with the original BV = 250 accounts for 1.8%, and is mapped to the new BV = 252.

[0189] High Dynamic Range Output: After equalization, use gamma correction (Gamma Correction, γ = 0.8) to fine-tune the contrast in the middle grayscale region (to enhance the visibility of dark stripes). Finally, generate an Electro-Mechanical Acoustic Interference Pattern (EMAIP), whose dynamic range is extended to 8-bit full scale (0 - 255), and the contrast between the fault coherence stripes (such as the "fishbone pattern" of bearing damage and the "oblique grid" of rotor bar breakage) and the background is increased by more than 3 times, meeting the input requirements of the convolutional neural network (CNN).

[0190] By calculating the time-domain coherence between the original features and the modulated features, an interference pattern reflecting the electromagnetic-mechanical coupling strength is generated. The pixel distribution of the interference pattern reveals the dynamic correlation mode between the two types of signals, transforming the abstract electromechanical coupling relationship into a visual interference fringe pattern, providing input features with clear physical meanings for the convolutional neural network, and enhancing the model's ability to identify fault patterns.

[0191] S205, the electromagnetic-mechanical acoustic interferogram is input into a pre-trained convolutional neural network (CNN). The CNN is trained to identify the correspondence between specific coherent fringe patterns in the spectrum and motor fault types, so as to output the final motor fault diagnosis result.

[0192] Specifically, the electromagnetic-mechanical acoustic interferogram can be preprocessed, including standardizing pixel values ​​to the [0,1] range and data enhancement, and the enhanced interferogram can be output.

[0193] After receiving the electromagnetic-mechanical acoustic interferogram (a two-dimensional image with time on the horizontal axis, frequency on the vertical axis, and pixel brightness representing the coherence intensity of the original electromagnetic acoustic pattern and the vibration-modulated acoustic pattern), the primary task of the system is to standardize it. Due to differences in the acquisition environment (such as sensor sensitivity and motor load fluctuations), the dynamic range of pixel values ​​(PV) in the original spectrum may vary greatly (for example, some spectrums have pixel values ​​concentrated in the range of 0-50, while others are distributed in the range of 0-255). Standardization adopts a linear normalization algorithm: the system first scans all pixels in the entire spectrum to find the minimum pixel value (MinPV) and the maximum pixel value (MaxPV) in the current spectrum. Then, the transformation formula is applied to each pixel in the spectrum: new pixel value = (original pixel value - MinPV) / (MaxPV - MinPV). This operation ensures that in the processed spectrum, the darkest pixel value is 0 (representing the lowest coherence intensity), the brightest pixel value is 1 (representing the highest coherence intensity), and all other pixel values ​​are uniformly distributed in the interval [0,1]. This standardization eliminates the absolute brightness differences between different samples, allowing subsequent convolutional neural networks (CNNs) to focus on the relative patterns of coherent stripes (such as stripe shape, orientation, and density) rather than absolute brightness.

[0194] To avoid model overfitting and improve its generalization ability, the system performs data augmentation on the standardized graph. This is not a simple data replication, but rather the application of a series of controllable geometric and physical transformations to generate diverse new samples. Key techniques include:

[0195] Random Time-axis Cropping and Scaling: While ensuring no critical events (such as motor startup and load switching) are lost, 80%-95% of the graph's length is randomly cropped along the time axis and linearly scaled back to the original time length. This simulates the subtle differences in the starting point of the analysis window in actual diagnostics.

[0196] Micro Frequency-axis Shift: The entire spectrum is shifted up or down by 1-3 pixels along the frequency axis (corresponding to an actual frequency shift of approximately a few hertz to tens of hertz). This simulates the slight drift of the fundamental frequency and harmonic frequencies caused by minor fluctuations in motor speed or minute differences in sensor installation during operation.

[0197] Controlled Gaussian Noise Injection: Adds Gaussian random noise (GRN) with a mean of 0 and a standard deviation of 0.02-0.05 to the spectrum. The intensity of this noise is carefully set so as not to overwhelm the real coherence fringes, while simulating sensor electronic noise or environmental interference.

[0198] Elastic Deformation: A mesh-based algorithm for slight non-rigid deformation is applied to induce localized, smooth distortions in the spectrum along the time and frequency directions (amplitude controlled within a few pixels). This simulates the potential impact of nonlinear mechanical vibration transmission paths or small electromagnetic field distortions on fringe continuity. Each enhancement operation randomly selects one or a combination of the above methods to generate a batch (e.g., 5-10 times the original number) of enhanced interferograms (AI).

[0199] The augmented interferograms still require final format unification to ensure they meet CNN input requirements. This includes: 1) adjusting all map sizes to the fixed size specified by the CNN input layer (e.g., 256 pixels x 256 pixels), using a bicubic interpolation algorithm to maintain stripe continuity; 2) organizing the maps into batch tensors (BTs), typically 32 or 64 maps per batch for GPU parallel computation; 3) performing slight brightness-contrast fine-tuning on the data within each batch, randomly adjusting the overall brightness within a factor of [0.9, 1.1] and the contrast within a factor of [0.95, 1.05]. At this point, preprocessing is complete, and the output augmented interferogram (AI) possesses a standardized numerical range, rich morphological diversity, and a consistent input format, laying a high-quality data foundation for subsequent CNN feature extraction.

[0200] The enhanced interferogram is input into a pre-trained multi-scale attention CNN model, which outputs a fault feature map. The CNN model uses adversarial example enhancement to learn the stripe pattern during the training phase.

[0201] The core architecture of the Multi-scale Attention CNN Model (MACM) is designed to effectively capture coherent fringe features at different scales in interferograms. Its "multi-scale" characteristic is reflected in:

[0202] Parallel Multi-branch Convolution: Multiple parallel convolutional branches follow the model's input layer. For example: Branch 1 uses a small kernel (e.g., 3x3) to focus on capturing fine, local stripe edges and texture details (e.g., short, coherent abrupt changes); Branch 2 uses a medium kernel (e.g., 5x5) to identify medium-scale stripe segments and basic patterns (e.g., a slanted stripe); Branch 3 uses a large kernel (e.g., 7x7) or convolutions with an initial layer stride to perceive large-scale stripe orientations and overall distribution patterns (e.g., continuous high-coherence bands spanning the entire time axis). Each branch convolution is followed by a non-linear activation function (e.g., ReLU, Rectified Linear Unit) and batch normalization (BN).

[0203] Feature Map Fusion: This involves concatenating the feature maps (FM) output from different branches along the channel dimension to form a fused feature map that incorporates multi-scale information. This ensures that the model can utilize both local details and global contextual information simultaneously.

[0204] An attention mechanism is integrated into the CNN, giving the model the ability to focus on the regions most relevant to the fault. Here, the Channel-Spatial Dual Attention Module (CSDAM) is used:

[0205] Channel Attention (CA): Global Average Pooling (GAP) is calculated for each channel of the fused feature map, resulting in a Channel Descriptor Vector (CDV). This vector is then fed into a small fully connected network (typically containing a dimensionality reduction layer and an dimensionality increase layer, with ReLU activation in between), outputting the Channel Weight Coefficient (CWC) for each channel. Channels with larger coefficients represent those whose stripe pattern features are more important to the current diagnostic task. Multiplying each channel of the original feature map by its corresponding CWC achieves channel-dimensional feature selection.

[0206] Spatial Attention (SA): Max pooling and average pooling are applied along the channel dimension to the channel-weighted feature map, resulting in two spatial feature maps (SFMs). These two SFMs are concatenated and passed through a standard convolutional layer (e.g., a 7x7 kernel) and a sigmoid activation function to generate a spatial weight map (SWM). Each pixel value in SWM is between [0,1], with larger values ​​indicating greater importance of the spatial location (e.g., dense areas of high-coherence fringes or areas where specific interference patterns appear). The original feature map is then multiplied point-by-point by the SWM in space to achieve spatial feature focusing. Dual attention enables the model to adaptively enhance fault-sensitive multi-scale fringe regions, suppress background noise and irrelevant information, and output a cleaner fault feature map (FFM).

[0207] This MACM model employs an adversarial sample augmentation (ASA) strategy during the training phase to improve its robustness to interference and its ability to resolve subtle stripe patterns. Specific methods:

[0208] Generating Adversarial Perturbation (GAP): This method uses the Fast Gradient Sign Method (FGSM) or a variant thereof. The gradient (G) of the training graph input is calculated relative to the model's current prediction loss function (LF, such as cross-entropy loss). Then, a small perturbation (P) is added along the gradient direction. The magnitude of the perturbation is controlled by a hyperparameter ε (Epsilon, typically ranging from 0.01 to 0.03), i.e., P = ε * sign(G). This perturbation is imperceptible to the human eye but sufficient to cause an unvalidated model to misclassify.

[0209] Adversarial Training (AT): In the regular training batches, a portion of adversarial samples (AS) generated by the methods described above are mixed in. The model not only learns to identify faults from the original augmented interferogram (AI), but must also learn to correctly classify these adversarial samples that are carefully designed to "fool" it. This training forces the model to deeply understand the essential characteristics of coherent fringes (such as specific spatial frequencies, directionality, and continuity), rather than simply memorizing surface pixel patterns, enabling it to maintain high diagnostic accuracy when faced with slight distortions, noise, or unseen but similar fringe patterns that may exist in real-world applications. The MACM model, after sufficient adversarial training, can ultimately perform deep feature extraction on the input augmented interferogram (AI) and output a highly abstract fault feature map (FFM) rich in discriminative information.

[0210] Global average pooling and fully connected layers are applied to the fault feature map to generate a fault probability distribution vector to represent the confidence level of various faults.

[0211] Fault Feature Maps (FFMs) output from MACM models typically have large spatial dimensions (e.g., H pixels high x W pixels wide) and a large number of channels (e.g., C channels, where C might be 256 or 512). Direct processing of these FFMs is computationally intensive and prone to overfitting. Therefore, Global Average Pooling (GAP) is used for dimensionality reduction and information aggregation. Specifically, each channel of the FFM is processed individually. For the feature map of the c-th channel (c=1,2,...,C), the arithmetic mean of the values ​​at all spatial locations (H x W pixels) is calculated. That is, the final output value of this channel is GAP_c = (1 / (H*W)) * Σ (values ​​at all locations). After GAP processing, the original H x W x C dimensional 3D feature map is compressed into a one-dimensional vector containing only C elements, called the Global Feature Vector (GFV). The advantages of GAP are: 1) It significantly reduces the number of parameters and prevents overfitting in subsequent layers; 2) It has inherent robustness to spatial transformations of the input feature map (such as slight translations); 3) The average value of each channel can be regarded as a macroscopic representation of the "response intensity" or "presence" of the features extracted by that channel on the entire map.

[0212] The Global Feature Vector (GFV) contains highly abstract fault-related information extracted from the original interferogram, after multi-scale analysis and attention focusing, but it is still high-dimensional (C-dimensional) and nonlinear. To map it to specific fault categories, a fully connected layer (FCL, also known as a dense layer) is introduced. A fully connected layer consists of a series of neurons, each connected to all elements of the input vector. Suppose we need to diagnose K types of faults (e.g., bearing outer race damage, bearing inner race damage, rotor bar breakage, stator inter-turn short circuit, imbalance, misalignment, foundation looseness, etc.), then a fully connected output layer with K neurons is typically used. Each neuron is calculated by weighting the input GFV (C-dimensional) and adding a bias term, then passing it through a nonlinear activation function. The mathematical essence is: Output of neuron k = Activation function (Σ (weights_w_k_i * GFV_i) + Bias_b_k), where i ranges from 1 to C. Key points: 1) Weights (w): Learnable parameters connecting each element of GFV to this neuron, determining the importance of different features in identifying the fault type. 2) Bias (b): Learnable parameters unique to each neuron, representing the "baseline confidence" of the fault type. 3) Activation function: The output layer typically uses the Softmax function. The Softmax function operates on the raw output values ​​of all K neurons, transforming them into a probability distribution: Probability of class k = exp(raw output of neuron k) / Σ (exp(raw output of neuron j), j=1 to K). This ensures that the sum of the probabilities of all classes is 1.

[0213] The fully connected layer after Softmax activation outputs a K-dimensional vector called the Fault Probability Distribution Vector (FPDV). Each element of this vector corresponds to a predefined motor fault type (or health state), with a value in the range [0,1]. This value represents the model's confidence level or probability estimate that the input graph belongs to that type of fault. For example, a possible FPDV might be: [0.02, 0.85, 0.10, 0.03, ..., 0.00], where the maximum value of 0.85 appears at the second element position, indicating that the model has a high confidence (85% probability) that the current motor state corresponds to the second fault type (such as "bearing outer race damage"). This vector encapsulates the model's comprehensive judgment of the current input and forms the basis for subsequent decisions. During training, by minimizing the cross-entropy loss (CEL) between the FPDV and the true fault labels (one-hot encoded vectors), the model learns to adjust all its parameters (convolutional kernel weights, attention weights, fully connected layer weights, and biases) to make the predicted probability distribution as close as possible to the true distribution. Thus, the complex interferometric information is compressed into a probability vector that intuitively represents the likelihood of various faults.

[0214] Based on the fault probability distribution vector, the final motor fault diagnosis result, including fault type and severity level, is output through Bayesian decision rules and threshold filtering.

[0215] The obtained Fault Probability Distribution Vector (FPDV) provides the likelihood of various fault types, but directly taking the fault type corresponding to the highest probability as the diagnostic result may not be robust enough, especially when the probability distribution is relatively flat (i.e., the model is not very certain) or when there are low-probability but high-risk faults. Therefore, a Bayesian Decision Rule (BDR) is introduced for more refined decision optimization. The core idea is to combine prior knowledge and the cost of diagnostic errors. Specific implementation:

[0216] Define the prior probability (PP): Based on historical statistics or domain knowledge, assign a prior probability P(ω_k) (where ω_k represents the k-th type of fault) to each type of fault k. For example, in a specific industrial scenario, bearing failure may be more common than rotor bar breakage.

[0217] Define the cost matrix (CM): Construct a K x K matrix where each element C(i|j) represents the cost of diagnosing a true fault as type i when the actual fault is type j. Typically, the cost of a correct diagnosis (i=j) is zero. The cost of a serious misdiagnosis (such as misdiagnosing a severe mechanical fault as healthy) is much higher than that of a minor misdiagnosis (such as misjudging two types of bearing faults).

[0218] Calculating Bayesian Risk (BR): For a fault diagnosis result of type i, the expected risk is: R(i) = Σ [ C(i|j) * P(ω_j |spectral) ], where P(ω_j |spectral) is the j-th probability value in the FPDV (i.e., the posterior probability estimate given by the model). The Bayesian decision rule selects the i-th fault type that minimizes the expected risk R(i) as the final diagnosed fault type. This is equivalent to maximizing the posterior probability and then weighting it according to the cost, making the diagnostic decision more in line with the needs of actual engineering risk control.

[0219] Even after selecting the fault type using Bayesian decision rules, its severity level (SL) still needs to be evaluated, and low-confidence diagnoses need to be filtered out. This is achieved through threshold filtering (TF):

[0220] Set the confidence threshold (CT): This is a preset value (e.g., 0.7 or 0.8). If the posterior probability (the maximum value in FPDV) corresponding to the selected fault type is lower than the CT, the system considers the model confidence insufficient and may not be able to make a reliable diagnosis. In this case, the output may be "Status unknown, further investigation recommended" or "No significant fault detected".

[0221] Severity Level Mapping (SLM): For diagnostic results with a confidence level exceeding that of a CT scan, it is necessary to determine the severity of the fault. This is typically not determined directly by probability values, but rather:

[0222] Utilizing feature map information: Retrospectively analyze the spatial distribution intensity or area of ​​the channels most relevant to the diagnosed fault type in the Fault Feature Map (FFM) output by the MACM model (obtained by analyzing the weights of fully connected layers or the attention map). Higher intensity and larger area generally indicate a more severe fault.

[0223] Analyze the raw features of the interferogram: In the raw or standardized interferogram, quantitatively calculate the energy (E, such as the sum of squares of coherence intensity within the relevant frequency band), continuity (Con, such as fringe length or number of breakpoints), or modulation depth (MD) of specific coherent fringe patterns associated with the diagnosed fault. These feature values ​​are input into a pre-calibrated severity assessment model (based on historical fault cases or simulation data) with a three-level severity rating (mild-moderate-severe) or a continuous numerical rating (such as a simple thresholding or small regression model).

[0224] Health status assessment: If the type selected by the Bayesian decision is "Healthy" (H), and its probability is much higher than other faults (e.g., > 0.95), then the "Healthy" status is output directly without needing a severity level.

[0225] Based on all the above steps, the system generates and outputs the final motor fault diagnosis result (FMFDR). This result is a structured information package, typically containing:

[0226] Diagnosed Fault Type (DFT): The most likely and sufficiently confident specific fault name selected based on Bayesian decision rules (e.g., "pitting on the outer ring of rolling bearing").

[0227] Fault Severity Level (FSL): A qualitative or semi-quantitative description (such as “mild”, “moderate”, “severe”) or a quantitative score (such as 70 / 100) derived from threshold filtering and severity mapping.

[0228] Diagnosis Confidence (DC): The probability value of the corresponding diagnosis type is directly taken from the Fault Probability Distribution Vector (FPDV) (e.g., 0.92).

[0229] The results are clearly presented to the user through a human-machine interface (such as a monitor, mobile app, or industrial SCADA system), or integrated into the equipment maintenance management system to trigger a maintenance work order. All intermediate data in the diagnostic process (original map, enhanced map, feature map, probability vector, decision basis) are usually recorded and stored for subsequent fault tracking, model optimization, and diagnostic report generation.

[0230] By using a pre-trained CNN model to learn fault-related coherent fringe patterns from interferograms, and outputting fault type and severity level through probability distribution, a diagnostic framework combining deep learning and physical feature enhancement is achieved to realize high-precision and interpretable motor fault classification, meeting the dual requirements of real-time performance and reliability in industrial scenarios.

[0231] As can be seen, the process involves collecting near-field electromagnetic noise audio signals and vibration acceleration signals of the motor casing during operation; performing a short-time Fourier transform on the near-field electromagnetic noise audio signals to extract the original electromagnetic acoustic signature feature vector characterizing the internal electromagnetic state of the motor; using the vibration acceleration signal as a modulation source to perform amplitude modulation processing on the original electromagnetic acoustic signature feature vector to generate a modulated electromagnetic acoustic signature feature vector; performing time-domain coherence calculation on the original electromagnetic acoustic signature feature vector and the modulated electromagnetic acoustic signature feature vector to generate a two-dimensional electromagnetic-mechanical acoustic signature interferogram; and inputting the electromagnetic-mechanical acoustic signature interferogram into a convolutional neural network (CNN) to output the final motor fault diagnosis result. This process can improve the distinguishability of fault features and achieve high-precision diagnosis of motor faults under complex operating conditions.

[0232] Another embodiment of the present invention provides a motor controller, see below. Figure 3 The motor controller may include:

[0233] The acquisition module 301 is used to acquire near-field electromagnetic noise audio signals and vibration acceleration signals of the motor housing during motor operation, and to ensure signal synchronization through hardware timestamps.

[0234] Extraction module 302 is used to perform short-time Fourier transform on the near-field electromagnetic noise audio signal to extract the instantaneous amplitude envelope sequence of the fundamental frequency and harmonic components, as the original electromagnetic acoustic signature feature vector characterizing the internal electromagnetic state of the motor.

[0235] The modulation module 303 is used to use the vibration acceleration signal as a modulation source to perform amplitude modulation processing on the original electromagnetic acoustic pattern feature vector in order to simulate the physical modulation effect of shell vibration on near-field electromagnetic noise and generate a modulated electromagnetic acoustic pattern feature vector.

[0236] The calculation module 304 is used to perform time-domain coherence calculation on the original electromagnetic acoustic feature vector and the modulated electromagnetic acoustic feature vector to generate a two-dimensional electromagnetic-mechanical acoustic interference spectrum, wherein the horizontal axis of the spectrum is time, the vertical axis is frequency, and the pixel value represents the coherence intensity.

[0237] The output module 305 is used to input the electromagnetic-mechanical acoustic interferogram into a pre-trained convolutional neural network (CNN), which is trained to identify the correspondence between specific coherent fringe patterns in the spectrum and motor fault types, so as to output the final motor fault diagnosis result.

[0238] This invention also provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above method embodiments when running.

[0239] Specifically, in this embodiment, the storage medium can be configured to store a computer program for performing the following steps:

[0240] S201 collects near-field electromagnetic noise audio signals and vibration acceleration signals of the motor housing during operation, and ensures signal synchronization through hardware timestamps;

[0241] S202, Perform a short-time Fourier transform on the near-field electromagnetic noise audio signal to extract the instantaneous amplitude envelope sequence of the fundamental frequency and harmonic components, which serves as the original electromagnetic acoustic signature feature vector characterizing the internal electromagnetic state of the motor.

[0242] S203, using the vibration acceleration signal as a modulation source, the original electromagnetic acoustic signature feature vector is subjected to amplitude modulation processing to simulate the physical modulation effect of shell vibration on near-field electromagnetic noise, and a modulated electromagnetic acoustic signature feature vector is generated.

[0243] S204, perform time-domain coherence calculation on the original electromagnetic acoustic signature feature vector and the modulated electromagnetic acoustic signature feature vector to generate a two-dimensional electromagnetic-mechanical acoustic signature interferogram, wherein the horizontal axis of the spectrum is time, the vertical axis is frequency, and the pixel value represents the coherence intensity.

[0244] S205, the electromagnetic-mechanical acoustic interferogram is input into a pre-trained convolutional neural network (CNN). The CNN is trained to identify the correspondence between specific coherent fringe patterns in the spectrum and motor fault types, so as to output the final motor fault diagnosis result.

[0245] This invention also provides an electronic device, including 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 of the above method embodiments.

[0246] Specifically, the aforementioned electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the aforementioned processor, and the input / output device is connected to the aforementioned processor.

[0247] Specifically, in this embodiment, the processor can be configured to perform the following steps via a computer program:

[0248] S201 collects near-field electromagnetic noise audio signals and vibration acceleration signals of the motor housing during operation, and ensures signal synchronization through hardware timestamps;

[0249] S202, Perform a short-time Fourier transform on the near-field electromagnetic noise audio signal to extract the instantaneous amplitude envelope sequence of the fundamental frequency and harmonic components, which serves as the original electromagnetic acoustic signature feature vector characterizing the internal electromagnetic state of the motor.

[0250] S203, using the vibration acceleration signal as a modulation source, the original electromagnetic acoustic signature feature vector is subjected to amplitude modulation processing to simulate the physical modulation effect of shell vibration on near-field electromagnetic noise, and a modulated electromagnetic acoustic signature feature vector is generated.

[0251] S204, perform time-domain coherence calculation on the original electromagnetic acoustic signature feature vector and the modulated electromagnetic acoustic signature feature vector to generate a two-dimensional electromagnetic-mechanical acoustic signature interferogram, wherein the horizontal axis of the spectrum is time, the vertical axis is frequency, and the pixel value represents the coherence intensity.

[0252] S205, the electromagnetic-mechanical acoustic interferogram is input into a pre-trained convolutional neural network (CNN). The CNN is trained to identify the correspondence between specific coherent fringe patterns in the spectrum and motor fault types, so as to output the final motor fault diagnosis result.

[0253] The above description, based on the embodiments shown in the figures, details the structure, features, and effects of the present invention. The above description is only a preferred embodiment of the present invention, but the present invention is not limited to the scope of implementation shown in the figures. Any changes made in accordance with the concept of the present invention, or equivalent embodiments modified to have equivalent changes, that do not exceed the spirit covered by the specification and figures, should be within the protection scope of the present invention.

Claims

1. A method for diagnosing motor faults, characterized in that, The method includes: The system collects near-field electromagnetic noise audio signals and vibration acceleration signals of the motor housing during operation, and ensures signal synchronization through hardware timestamps. A short-time Fourier transform is performed on the near-field electromagnetic noise audio signal to extract the instantaneous amplitude envelope sequence of the fundamental frequency and harmonic components, which serves as the original electromagnetic acoustic signature feature vector characterizing the internal electromagnetic state of the motor. The vibration acceleration signal is used as a modulation source to perform amplitude modulation processing on the original electromagnetic acoustic pattern feature vector in order to simulate the physical modulation effect of shell vibration on near-field electromagnetic noise and generate a modulated electromagnetic acoustic pattern feature vector. The original electromagnetic acoustic signature feature vector and the modulated electromagnetic acoustic signature feature vector are subjected to time-domain coherence calculation to generate a two-dimensional electromagnetic-mechanical acoustic signature interferogram, in which the horizontal axis of the spectrum is time, the vertical axis is frequency, and the pixel value represents the coherence intensity. The electromagnetic-mechanical acoustic interferogram is input into a pre-trained convolutional neural network (CNN), which is trained to identify the correspondence between specific coherent fringe patterns in the spectrum and motor fault types, so as to output the final motor fault diagnosis result.

2. The method according to claim 1, characterized in that, The acquisition of near-field electromagnetic noise audio signals and vibration acceleration signals of the motor housing during operation, and the use of hardware timestamps to ensure signal synchronization, includes: An electromagnetic noise sensor is deployed close to the motor surface to capture the raw analog signal of near-field electromagnetic noise with a nanosecond-level response, and outputs an unprocessed analog stream of electromagnetic noise. Piezoelectric triaxial accelerometers are installed at key vibration points of the motor housing. Vibration signals are preprocessed by an anti-aliasing filter, and the filtered vibration acceleration analog stream is output. The electromagnetic noise analog stream and the vibration acceleration analog stream are input into a multi-channel ADC converter, and an atomic clock-driven hardware timestamp module is embedded to add a precise nanosecond-level time stamp to each sampling point, and output a pair of timestamped digital signals. A timestamp alignment algorithm is applied to dynamically time-normalize digital signal pairs based on hardware timestamps, eliminating sampling jitter and generating synchronized electromagnetic noise digital signals and vibration acceleration digital signals.

3. The method according to claim 2, characterized in that, The step of performing a short-time Fourier transform on the near-field electromagnetic noise audio signal to extract the instantaneous amplitude envelope sequence of the fundamental frequency and harmonic components, as the original electromagnetic acoustic signature feature vector characterizing the internal electromagnetic state of the motor, includes: A variable-window-length short-time Fourier transform is applied to synchronous electromagnetic noise digital signals. The window size is adaptively adjusted according to the instantaneous energy of the signal, and a high-resolution time-frequency energy matrix is ​​output. Using the time-frequency energy matrix as input, a fundamental frequency detector based on a convolutional neural network is run to identify the dominant fundamental frequency component and its time-varying trajectory, and output the instantaneous position sequence of the fundamental frequency. Based on the fundamental frequency instantaneous position sequence, the harmonic bands of integer multiples are extracted using the harmonic tracking algorithm, the instantaneous amplitude of each harmonic component is calculated, and the harmonic amplitude sequence set is output. Apply Hilbert transform to the set of fundamental frequency instantaneous position sequence and harmonic amplitude sequence to extract the instantaneous envelope of each component and output the fundamental frequency envelope sequence and harmonic envelope sequence. By fusing the fundamental frequency envelope sequence and harmonic envelope sequence, and then splicing and normalizing in the time domain, the original electromagnetic acoustic signature feature vector is generated.

4. The method according to claim 3, characterized in that, The step of using the vibration acceleration signal as a modulation source to perform amplitude modulation processing on the original electromagnetic acoustic signature feature vector to simulate the physical modulation effect of shell vibration on near-field electromagnetic noise, and generating a modulated electromagnetic acoustic signature feature vector, includes: Wavelet packet decomposition is performed on the synchronous vibration acceleration digital signal to extract the frequency band components related to the mechanical resonance of the motor and output the key vibration frequency band signal. The key vibration frequency band signal is input into a nonlinear normalizer and converted into a modulation index sequence. Its value range is dynamically mapped to the amplitude range of the original electromagnetic acoustic feature vector, and the modulation index vector is output. Using the original electromagnetic acoustic signature feature vector as the carrier and the modulation index vector as the modulation source, the amplitude modulation equation is applied to simulate the physical modulation effect of shell vibration on electromagnetic features, and output the modulation feature sequence. The modulation feature sequence is subjected to anti-aliasing smoothing processing and integrated into a temporally continuous modulated electromagnetic acoustic feature vector.

5. The method according to claim 4, characterized in that, The step involves performing time-domain coherence calculations on the original electromagnetic acoustic signature feature vector and the modulated electromagnetic acoustic signature feature vector to generate a two-dimensional electromagnetic-mechanical acoustic signature interferogram. In this interferogram, the horizontal axis represents time, the vertical axis represents frequency, and the pixel values ​​represent coherence intensity. This includes: The original electromagnetic acoustic signature feature vector and the modulated electromagnetic acoustic signature feature vector are input into the time domain alignment module, and sample-level matching is performed based on the shared time reference to output aligned feature vector pairs. Based on the aligned eigenvector pairs, the cross spectral density at each frequency point is calculated, and the original coherence coefficient matrix is ​​generated through the coherence function; A multi-scale sliding window is applied to adaptively smooth the original coherence coefficient matrix, suppressing noise and enhancing significant fringes, and outputting a smooth coherence matrix. The smooth coherence matrix is ​​mapped to a two-dimensional time-frequency grid, where the horizontal axis represents time and the vertical axis represents frequency. Pixel values ​​are quantized into coherence intensity values, and the initial interferogram is output. The initial interferogram is subjected to contrast enhancement processing, and the pixel distribution is optimized by histogram equalization to generate an electromagnetic-mechanical acoustic interferogram with high dynamic range.

6. The method according to claim 5, characterized in that, The electromagnetic-mechanical acoustic interferogram is input into a pre-trained convolutional neural network (CNN). This CNN is trained to recognize the correspondence between specific coherent fringe patterns in the spectrum and motor fault types, in order to output the final motor fault diagnosis result, including: The electromagnetic-mechanical acoustic interferogram is preprocessed, including normalizing pixel values ​​to the [0,1] range and data enhancement, and the enhanced interferogram is output. The enhanced interferogram is input into a pre-trained multi-scale attention CNN model, which outputs a fault feature map. The CNN model uses adversarial example enhancement to learn the stripe pattern during the training phase. Global average pooling and fully connected layers are applied to the fault feature map to generate a fault probability distribution vector to represent the confidence level of various faults. Based on the fault probability distribution vector, the final motor fault diagnosis result, including fault type and severity level, is output through Bayesian decision rules and threshold filtering.

7. A motor controller, characterized in that, The motor controller includes: The acquisition module is used to acquire near-field electromagnetic noise audio signals and vibration acceleration signals of the motor housing during operation, and ensures signal synchronization through hardware timestamps; The extraction module is used to perform a short-time Fourier transform on the near-field electromagnetic noise audio signal to extract the instantaneous amplitude envelope sequence of the fundamental frequency and harmonic components, which serves as the original electromagnetic acoustic signature feature vector characterizing the internal electromagnetic state of the motor. The modulation module is used to use the vibration acceleration signal as a modulation source to perform amplitude modulation processing on the original electromagnetic acoustic pattern feature vector in order to simulate the physical modulation effect of shell vibration on near-field electromagnetic noise and generate the modulated electromagnetic acoustic pattern feature vector. The calculation module is used to perform time-domain coherence calculation on the original electromagnetic acoustic signature feature vector and the modulated electromagnetic acoustic signature feature vector to generate a two-dimensional electromagnetic-mechanical acoustic signature interferogram, wherein the horizontal axis of the spectrum is time, the vertical axis is frequency, and the pixel value represents the coherence intensity. The output module is used to input the electromagnetic-mechanical acoustic interferogram into a pre-trained convolutional neural network (CNN), which is trained to identify the correspondence between specific coherent fringe patterns in the spectrum and motor fault types, so as to output the final motor fault diagnosis result.

8. The motor controller according to claim 7, characterized in that, The acquisition module is specifically used for: An electromagnetic noise sensor is deployed close to the motor surface to capture the raw analog signal of near-field electromagnetic noise with a nanosecond-level response, and outputs an unprocessed analog stream of electromagnetic noise. Piezoelectric triaxial accelerometers are installed at key vibration points of the motor housing. Vibration signals are preprocessed by an anti-aliasing filter, and the filtered vibration acceleration analog stream is output. The electromagnetic noise analog stream and the vibration acceleration analog stream are input into a multi-channel ADC converter, and an atomic clock-driven hardware timestamp module is embedded to add a precise nanosecond-level time stamp to each sampling point, and output a pair of timestamped digital signals. A timestamp alignment algorithm is applied to dynamically time-normalize digital signal pairs based on hardware timestamps, eliminating sampling jitter and generating synchronized electromagnetic noise digital signals and vibration acceleration digital signals.

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

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

Images