Motor test and diagnosis system and method based on multi-data acquisition

Abstract

The invention relates to the technical field of motor state monitoring and intelligent fault diagnosis, in particular to a motor test and diagnosis system and method based on multi-data acquisition, and the system comprises a multi-mode sensing array module, a heterogeneous data fusion module, a coupling feature mining module and a migration diagnosis decision module. Wherein the multi-mode sensing array module is used for collecting mechanical vibration time domain signals, temperature field space distribution data and near-field electromagnetic radiation spectrums of a motor; the heterogeneous data fusion module is used for generating a time-frequency matrix, a thermal field characteristic matrix and fundamental frequency harmonic energy distribution characteristics; the coupling feature mining module is used for generating a comprehensive diagnosis factor; and the migration diagnosis decision module is used for constructing a feature mapping network based on meta-learning and outputting a fault type and a confidence score. According to the invention, through multi-modal feature fusion and a metalearning-based migration diagnosis mechanism, high-precision identification and cross-model adaptive diagnosis of multi-source fault information under a complex working condition of the motor are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of motor state monitoring and fault diagnosis, and in particular to a motor testing and diagnosis system and method based on multi-data acquisition. Background Art

[0002] As motor systems are widely used in industrial fields such as manufacturing, energy, and transportation, the reliability of their operating status directly affects the safety and stability of key equipment. Traditional motor fault diagnosis methods mostly rely on a single physical quantity for feature analysis. Such methods have certain effects under specific single-modal conditions, but often exhibit problems such as low diagnostic accuracy and insufficient feature response in complex working conditions, early weak faults, or cross-model commonality scenarios. In recent years, with the development of multimodal sensing technology and artificial intelligence, multi-source heterogeneous data fusion and intelligent diagnostic algorithms have gradually become important directions for improving the accuracy of motor fault identification.

[0003] However, existing multimodal diagnostic systems still face the following key technical bottlenecks: First, the inconsistencies in dimensionality, scale, and physical properties between different modal signals lack effective feature coupling mechanisms, making it difficult to construct a unified, high-quality comprehensive diagnostic factor. Second, existing diagnostic models generally rely on large amounts of labeled data, which severely limits their generalization capabilities when there are large differences in device types or a small number of new motor samples, making cross-platform diagnostic migration impossible. Therefore, a motor testing and diagnostic system and method based on multi-data acquisition is urgently needed to address these issues. Summary of the Invention

[0004] Based on the above objectives, the present invention provides a motor testing and diagnosis system and method based on multiple data acquisition.

[0005] The motor testing and diagnosis system based on multi-data acquisition includes a multi-modal sensor array module, a heterogeneous data fusion module, a coupling feature mining module, and a migration diagnosis decision module; wherein:

[0006] Multimodal sensing array module: Arranges three-dimensional orthogonal vibration sensors, distributed infrared thermal imagers, and electromagnetic radiation probes to collect the motor's mechanical vibration time domain signal, temperature field spatial distribution data, and near-field electromagnetic radiation spectrum;

[0007] Heterogeneous data fusion module: used to perform multi-layer time-frequency decomposition on the mechanical vibration time domain signal to generate a time-frequency matrix, perform thermal gradient spatial convolution on the temperature field spatial distribution data to generate a thermal field feature matrix, and extract the fundamental frequency harmonic energy distribution characteristics in the near-field electromagnetic radiation spectrum;

[0008] Coupling feature mining module: used to calculate the nonlinear correlation coefficient between the time-frequency matrix and the thermal field feature matrix, and fuse the harmonic energy distribution characteristics with the nonlinear correlation results to generate a comprehensive diagnostic factor;

[0009] Migration diagnosis decision module: used to build a meta-learning-based feature mapping network, map comprehensive diagnostic factors to a shared feature space, and perform cross-model migration diagnosis in combination with a priori fault case library, outputting fault types and confidence scores.

[0010] Optionally, the multimodal sensing array module includes a vibration sensing unit, a temperature field reconstruction unit, and an electromagnetic spectrum acquisition unit; wherein:

[0011] Vibration sensing unit: Consists of three piezoelectric accelerometers. The first sensor is installed along the motor's axial direction, while the second and third sensors are symmetrically fixed at 45-degree angles on both sides of the motor's radial direction. The sampling frequency of the first sensor is set to above 10kHz to capture rotor dynamic characteristics, while the sampling frequencies of the second and third sensors are set to below 1kHz to collect bearing vibration signals.

[0012] Temperature field reconstruction unit: It includes six infrared thermal imagers, each with a 60-degree field of view and 15-degree overlap between adjacent fields of view. Each thermal imager has a built-in uncooled microbolometer, which generates three-dimensional temperature field distribution data through multi-view data fusion.

[0013] Electromagnetic spectrum acquisition unit: It consists of an axial loop antenna, a radial loop antenna and a stator gap probe. The axial antenna is installed parallel to the end face of the motor to detect the main frequency component of electromagnetic radiation. The radial antenna vertically surrounds the motor casing to collect leakage magnetic harmonics. The stator gap probe is embedded in the motor slot to directionally receive partial discharge signals. The output signals of the three acquisition channels are processed by frequency band selective fusion to form a near-field electromagnetic radiation spectrum.

[0014] Optionally, the electromagnetic spectrum acquisition unit includes:

[0015] The filter preprocessing subunit is used to perform bandpass filtering on the three signals. Specifically, the axial loop antenna signal passes through a 1MHz to 30MHz passband filter to extract the main frequency radiation component; the radial loop antenna signal passes through a 100kHz to 5MHz passband filter to retain the harmonic components generated by the motor operation; and the stator gap probe signal passes through a 500kHz to 20MHz passband filter to extract the high-frequency pulse components in the partial discharge signal.

[0016] Normalization and compression subunit: used to perform amplitude normalization and logarithmic quantization compression processing on each filtered signal to obtain a compressed signal;

[0017] Spectrum fusion construction subunit: used to convert the normalized three-way compressed signal into the frequency domain and construct the composite near-field electromagnetic radiation spectrum F based on the weighted superposition strategy c (f), its expression is: F c(f)=w1·F1(f)+w2·F2(f)+w3·F3(f), where F1(f), F2(f), and F3(f) represent the spectra of the three signals after fast Fourier transform, respectively; w1, w2, and w3 are spectrum fusion weight coefficients that satisfy the constraint condition of w1+w2+w3=1; and f is a frequency variable.

[0018] Optionally, the heterogeneous data fusion module includes a time-frequency decomposition unit, a thermal field extraction unit, an electromagnetic spectrum analysis unit, and a feature fusion unit; wherein:

[0019] Time-frequency decomposition unit: used to perform multi-layer time-frequency decomposition processing on the mechanical vibration time domain signal from the vibration sensing unit. It uses a step-by-step deconstruction method to reconstruct and decompose the signal at different time scales and frequency scales, and extract a two-dimensional time-frequency matrix containing time locality and frequency distribution;

[0020] Thermal field extraction unit: used to perform thermal gradient convolution calculation on the three-dimensional temperature field spatial distribution data generated by the temperature field reconstruction unit. By constructing a multi-directional thermal gradient template in the spatial domain and performing directional convolution operations along each dimension of the temperature field tensor, it extracts local thermal slope change information and overall thermal diffusion trend, and finally generates a thermal field feature matrix.

[0021] Electromagnetic spectrum analysis unit: used to extract the fundamental frequency harmonic energy distribution from the near-field electromagnetic radiation spectrum from the electromagnetic spectrum acquisition unit. By locating the main frequency position and its integer multiple frequency point area in the frequency domain, the spectrum energy of each harmonic bandwidth interval is integrated and calculated, and the characteristic vector containing multiple harmonic component energy indicators is output;

[0022] Feature fusion unit: used to map the time-frequency matrix, thermal field feature matrix and harmonic energy vector to a unified dimensional structure and construct a third-order feature tensor.

[0023] Optionally, the thermal field extraction unit includes:

[0024] Gradient template construction subunit: used to construct spatial gradient templates in three directions, defining the first-order difference kernel along the X, Y and Z spatial dimensions of the temperature field tensor respectively;

[0025] Spatial convolution calculation subunit: used to perform gradient convolution on the input three-dimensional temperature field tensor in the X, Y and Z directions respectively to obtain the gradient distribution tensor in each direction;

[0026] Thermal field feature generation subunit: used to integrate the gradient distribution tensors in each direction into the gradient mode field; then perform projection averaging along a main observation surface to finally obtain a two-dimensional thermal field feature matrix.

[0027] Optionally, the electromagnetic spectrum analysis unit includes:

[0028] Main frequency positioning subunit: used to scan the composite near-field electromagnetic radiation spectrum, identify and locate the frequency position of the global energy peak, and use it as the reference main frequency point;

[0029] Harmonic interval division sub-unit: Based on the determined reference main frequency point, a series of frequency band areas centered on integer multiples of the main frequency are generated to provide clear interval boundaries for energy calculation;

[0030] Energy product operator unit: used to accumulate the spectrum power in each harmonic interval. By summing the power values of each frequency band, the overall energy value of the corresponding order harmonics is obtained, forming an energy index that can directly reflect the intensity of electromagnetic disturbance;

[0031] Feature vector output subunit: used to arrange and combine the harmonic energy indicators of each order in the order of frequency order from low to high, and output the harmonic energy feature vector.

[0032] Optionally, the coupling feature mining module includes a correlation coefficient calculation unit, a dynamic weight allocation unit, and a diagnostic factor fusion unit; wherein:

[0033] Correlation coefficient calculation unit: used to flatten the time-frequency matrix and the thermal field feature matrix into a data sequence of the same dimension according to the corresponding time-space index, and calculate the nonlinear correlation coefficient ρ between the two sets of data sequences based on the mutual information algorithm MI ;

[0034] Dynamic weight allocation unit: used to allocate weights according to the nonlinear correlation coefficient ρ MI The relative size of each order harmonic energy index dynamically generates a set of fusion weights {w v , w e}, where w v represents the weight of the vibration-thermal field coupling feature; w e Represents the overall weight of the harmonic energy characteristics;

[0035] Diagnostic factor fusion unit: used to perform weighted fusion processing on the harmonic energy vector and the nonlinear correlation coefficient according to the dynamic weight to generate the final comprehensive diagnostic factor vector D. The calculation formula is:

[0036] Among them, D is the comprehensive diagnostic factor of the final output; Indicates the average value of harmonic energy index.

[0037] Optionally, the migration diagnosis decision module includes a feature mapping unit, a shared space alignment unit, a fault matching judgment unit, and a confidence score output unit; wherein:

[0038] Feature Mapping Unit: This unit is used to construct a lightweight feature mapping network based on a meta-learning framework. It uses a few-shot support set training approach and uses diagnostic factor and label pairs from historical motor samples to quickly initialize model parameters and obtain a generalizable embedded feature extractor.

[0039] Shared space alignment unit: used to convert the input comprehensive diagnostic factor into a shared embedding space through the mapping network;

[0040] Fault matching and discrimination unit: used to perform distance measurement and category matching between the target feature vector in the shared space and the labeled feature vector in the prior fault case library;

[0041] Confidence score output unit: used to calculate the final fault type confidence score based on the similarity distribution between the target feature and the case library samples.

[0042] Optionally, the fault matching determination unit includes:

[0043] Feature normalization subunit: used to perform unified normalization processing on the embedding vector obtained after the target comprehensive diagnostic factor is input into the feature mapping network and the embedding vector in the case library;

[0044] Vector distance calculation subunit: used to calculate the Euclidean distance between the target sample embedding vector and each labeled sample vector in the case library. Let the target vector be z t , the i-th case vector is z i , then the Euclidean distance d between the two i The expression is: i =||z t -z i ||2;

[0045] Similarity sorting subunit: used to sort all the calculated distance values d i Sort from small to large and select the first K neighboring cases with the smallest distance as the candidate reference sample set;

[0046] Category determination subunit: It is used to count the fault category with the highest frequency among the first K nearest neighbor samples as the final diagnosis result of the target sample and output the final fault type label.

[0047] The motor testing and diagnosis method based on multiple data acquisition is implemented by the above-mentioned motor testing and diagnosis system based on multiple data acquisition, and includes the following steps:

[0048] S1: Deploy three-dimensional orthogonal vibration sensors, distributed infrared thermal imagers, axial loop antennas, radial loop antennas, and stator gap probes to obtain mechanical vibration time domain signals, temperature field spatial distribution data, and near-field electromagnetic radiation spectra.

[0049] S2: Perform multi-layer short-time Fourier decomposition on the mechanical vibration time domain signal to generate a two-dimensional time-frequency matrix. Perform three-dimensional thermal gradient convolution on the temperature field spatial distribution data to generate a thermal field feature matrix. At the same time, locate the main frequency and each order harmonic interval in the electromagnetic radiation spectrum, extract the corresponding energy eigenvectors, and construct the three types of features into a multidimensional tensor according to a unified indexing rule.

[0050] S3: Based on the mutual information algorithm, the nonlinear correlation coefficient between the time-frequency matrix and the thermal field characteristic matrix is calculated, and the comprehensive diagnostic factor vector is formed through the dynamic weight allocation mechanism in combination with the harmonic energy characteristics;

[0051] S4: Utilizes a feature mapping network pre-trained using a meta-learning strategy and projects the comprehensive diagnostic factor vector into a shared embedding space of fixed dimensionality;

[0052] S5: In the shared embedding space, the Euclidean distance between the target embedding vector and the labeled embedding vector in the prior fault case library is calculated. Several neighboring samples with the smallest distance are selected to filter out the fault category, and the inverse distance normalization method is used to output the corresponding confidence score.

[0053] Beneficial effects of the present invention:

[0054] The present invention uses multimodal sensors to collaboratively collect vibration, temperature field and electromagnetic spectrum signals, and constructs a time-frequency matrix, a thermal field feature matrix and a harmonic energy vector respectively, to achieve a multi-level physical quantity characterization of the motor's operating state; then, it performs feature coupling measurement based on the mutual information algorithm, and introduces a dynamic weighting mechanism to fuse multi-source features to generate a unified comprehensive diagnostic factor, which effectively improves the recognition ability and diagnostic accuracy of weak signals under complex fault modes.

[0055] The present invention projects comprehensive diagnostic factors into a shared embedding space by adopting a meta-learning-driven feature mapping network, and performs nearest neighbor weighted matching and confidence assessment in combination with a priori fault case library, thereby achieving adaptive elimination of data distribution differences between different models and knowledge transfer; the diagnostic process has the ability of rapid generalization with a small number of samples and high interpretability, significantly enhancing the fault identification efficiency and reliability of motor systems in cross-platform deployment scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0057] Figure 1 A schematic diagram of a motor testing and diagnostic system according to an embodiment of the present invention;

[0058] Figure 2 Schematic diagram of a motor testing and diagnosis method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0059] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. It is also noted that, to provide a more detailed description, the following embodiments are best and preferred embodiments, and those skilled in the art may employ alternative methods for implementing certain known technologies. Furthermore, the accompanying drawings are intended only to provide a more detailed description of the embodiments and are not intended to limit the present invention.

[0060] like Figure 1 As shown in FIG, the motor testing and diagnosis system based on multi-data acquisition includes a multi-modal sensor array module, a heterogeneous data fusion module, a coupling feature mining module, and a migration diagnosis decision module; wherein:

[0061] Multimodal sensing array module: Arranges three-dimensional orthogonal vibration sensors, distributed infrared thermal imagers, and electromagnetic radiation probes to collect the motor's mechanical vibration time domain signal, temperature field spatial distribution data, and near-field electromagnetic radiation spectrum;

[0062] Heterogeneous data fusion module: used to perform multi-layer time-frequency decomposition on the mechanical vibration time domain signal to generate a time-frequency matrix, perform thermal gradient spatial convolution on the temperature field spatial distribution data to generate a thermal field feature matrix, and extract the fundamental frequency harmonic energy distribution characteristics in the near-field electromagnetic radiation spectrum;

[0063] Coupling feature mining module: used to calculate the nonlinear correlation coefficient between the time-frequency matrix and the thermal field feature matrix, and fuse the harmonic energy distribution characteristics with the nonlinear correlation results to generate a comprehensive diagnostic factor;

[0064] Migration diagnosis decision module: used to build a meta-learning-based feature mapping network, map comprehensive diagnostic factors to a shared feature space, and perform cross-model migration diagnosis in combination with a priori fault case library, outputting fault types and confidence scores.

[0065] The multimodal sensing array module includes a vibration sensing unit, a temperature field reconstruction unit, and an electromagnetic spectrum acquisition unit; wherein:

[0066] Vibration sensing unit: Consists of three piezoelectric accelerometers. The first sensor is installed along the motor's axial direction, while the second and third sensors are symmetrically fixed at 45-degree angles on both sides of the motor's radial direction. The sampling frequency of the first sensor is set to above 10kHz to capture rotor dynamic characteristics, while the sampling frequencies of the second and third sensors are set to below 1kHz to collect bearing vibration signals.

[0067] Temperature field reconstruction unit: It includes six infrared thermal imagers, each with a 60-degree field of view and 15-degree overlap between adjacent fields of view. Each thermal imager has a built-in uncooled microbolometer, which generates three-dimensional temperature field distribution data through multi-view data fusion.

[0068] Electromagnetic spectrum acquisition unit: consists of an axial loop antenna, a radial loop antenna and a stator gap probe. The axial antenna is installed parallel to the end face of the motor to detect the main frequency component of the electromagnetic radiation, the radial antenna vertically surrounds the motor casing to collect leakage magnetic harmonics, and the stator gap probe is embedded in the motor slot to directionally receive partial discharge signals. The output signals of the three acquisition channels are processed by frequency band selective fusion to form a near-field electromagnetic radiation spectrum; the above-mentioned unit can effectively obtain multi-source heterogeneous physical quantity information of the motor's operating status through the high-precision acquisition of vibration signals, thermal field data and electromagnetic spectrum signals in different channels. The multimodal sensor array module can effectively obtain multi-source heterogeneous physical quantity information of the motor's operating status, and through structured layout and directional sampling design, it can achieve comprehensive coverage of the status of key components, providing a high signal-to-noise ratio and high-dimensional raw data foundation for subsequent feature fusion and diagnostic analysis.

[0069] The electromagnetic spectrum acquisition unit includes:

[0070] The filter preprocessing subunit is used to perform bandpass filtering on the three signals. Specifically, the axial loop antenna signal passes through a 1MHz to 30MHz passband filter to extract the main frequency radiation component; the radial loop antenna signal passes through a 100kHz to 5MHz passband filter to retain the harmonic components generated by the motor operation; and the stator gap probe signal passes through a 500kHz to 20MHz passband filter to extract the high-frequency pulse components in the partial discharge signal.

[0071] Normalization compression subunit: used to perform amplitude normalization and logarithmic quantization compression on each filtered signal to obtain a compressed signal, which is expressed as: in, represents the result of logarithmic quantization compression of the i-th signal; S i (t) represents the time domain signal output by the i-th acquisition channel; A i Indicates the reference amplitude value of the normalized amplitude of the i-th signal; log 10 (·) represents a logarithmic function with base 10, which is used to compress the signal amplitude range, improve numerical stability, and maintain unit consistency; i∈{1,2,3}, corresponding to the axial loop antenna, radial loop antenna, and stator gap probe channel, respectively;

[0072] Spectrum fusion construction subunit: used to convert the normalized three-way compressed signal into the frequency domain and construct the composite near-field electromagnetic radiation spectrum F based on the weighted superposition strategy c (f), its expression is: F c(f)=w1·F1(f)+w2·F2(f)+w3·F3(f), where F1(f), F2(f), and F3(f) represent the spectra of the three signals after fast Fourier transform (FFT), respectively; w1, w2, and w3 are spectrum fusion weight coefficients, satisfying the constraint of w1+w2+w3=1, and the weights are dynamically set according to the signal-to-noise ratio of each channel; f is a frequency variable in Hz. The above subunit constructs a step-by-step processing chain with filtering, normalization, spectrum weighted fusion, and spectrum shape correction as the core. The frequency band fusion processing subunit realizes the unified and standardized fusion of three electromagnetic signals from different sources, effectively suppressing the problems of inconsistent amplitude scales and frequency feature conflicts between channels, ensuring that the generated near-field electromagnetic radiation spectrum has a clear spectrum structure and fault identifiable characteristics, providing high-quality input basis for subsequent diagnostic algorithms.

[0073] The heterogeneous data fusion module includes a time-frequency decomposition unit, a thermal field extraction unit, an electromagnetic spectrum analysis unit, and a feature fusion unit; among which:

[0074] Time-frequency decomposition unit: This unit is used to perform multi-layer time-frequency decomposition processing on the mechanical vibration time-domain signal from the vibration sensing unit. It uses a step-by-step deconstruction method to reconstruct and decompose the signal at different time and frequency scales, extracting a two-dimensional time-frequency matrix containing time locality and frequency distribution to characterize the instantaneous dynamic response characteristics of the motor during operation.

[0075] The specific operation steps of the time-frequency decomposition unit are as follows:

[0076] Signal preprocessing: The input mechanical vibration time domain signal s(t) is windowed, and the entire signal sequence is framed using a sliding window of fixed length. Each frame is multiplied by a weighted window function (such as a Hamming window) to obtain a time domain sub-signal s k (t), expressed as: s k (t) = s(t) w(t-kΔt), where s(t) is the original mechanical vibration time domain signal; w(t) is the window function, using the Hamming window; k is the kth frame number; Δt is the time interval between adjacent frames, in seconds;

[0077] Time-frequency transformation: for each frame signal s k (t) Perform short-time Fourier transform (STFT) to obtain the energy distribution of each frame in the frequency domain. The specific formula is: Where: S(k, f) is the complex spectrum value at the kth frame and frequency f; f is the frequency variable in Hz; j is the imaginary unit, j 2 =-1;

[0078] Power spectrum calculation and matrix construction: Calculate the energy density corresponding to each frequency point to form a two-dimensional time-frequency matrix, whose value is the square normalization result of the spectrum modulus, and the expression is: Where |S(k, f)| represents the modulus of the complex spectrum; N is the number of sampling points per frame, which is used for energy normalization; P(k, f) is a two-dimensional time-frequency matrix, each element of which represents the energy of the vibration signal at the kth frame and frequency f;

[0079] Finally, all P(k, f) are arranged according to the time frame k as the row index and the frequency point f as the column index to construct a complete two-dimensional time-frequency matrix.

[0080] Thermal field extraction unit: This unit is used to perform thermal gradient convolution calculations on the three-dimensional temperature field spatial distribution data generated by the temperature field reconstruction unit. By constructing a multi-directional thermal gradient template in the spatial domain and performing directional convolution operations along each dimension of the temperature field tensor, it extracts local thermal slope change information and overall thermal diffusion trends, ultimately generating a thermal field feature matrix with spatial structural characteristics to characterize the spatial evolution of abnormal heat conduction on the motor surface.

[0081] Electromagnetic spectrum analysis unit: used to extract the fundamental frequency harmonic energy distribution from the near-field electromagnetic radiation spectrum from the electromagnetic spectrum acquisition unit. By locating the main frequency position and its integer multiple frequency points in the frequency domain, the spectrum energy of each harmonic bandwidth interval is integrated and calculated, and the characteristic vector containing multiple harmonic component energy indicators is output to reflect the periodic disturbance pattern of the electromagnetic interference source in the motor system.

[0082] Feature fusion unit: used to map the time-frequency matrix, thermal field feature matrix and harmonic energy vector to a unified dimensional structure, and construct a third-order feature tensor, forming a unified multimodal data structure through dimensional normalization and alignment. This feature tensor serves as the input for subsequent diagnosis and is used to support the joint analysis of cross-modal information and fault classification. By setting up a four-level feature processing unit, the heterogeneous data fusion module realizes the channel processing and unified structure conversion of multi-source physical signals, and can efficiently integrate the key feature information in the time domain, space domain and frequency domain, improve the fusion expression ability of multimodal data, and provide a unified format and dimension matching data input basis for subsequent fault coupling mining and intelligent diagnosis.

[0083] The thermal field extraction unit includes:

[0084] Gradient template construction subunit: used to construct spatial gradient templates in three directions, defining the first-order difference kernel G along the three spatial dimensions of the temperature field tensor X (horizontal), Y (vertical) and Z (depth) x ,G y ,G z , where the general expression for each core is: G d= [-1, 0, 1], d∈{x, y, z}, this template represents the sensitivity of the temperature at the center point to the temperature changes in the adjacent spatial directions, and will be used in the convolution operation later;

[0085] Spatial convolution calculation subunit: used to perform gradient convolution on the input three-dimensional temperature field tensor in the X, Y and Z directions respectively, and obtain the gradient distribution tensor in each direction. The expressions are as follows:

[0086] Where T(x, y, z) is the three-dimensional temperature field tensor generated by the fusion of infrared thermal imagers, in degrees Celsius; * represents the three-dimensional discrete convolution operation; Represents the spatial gradient value of the temperature field in direction d;

[0087] Thermal field feature generation subunit: used to integrate the gradient distribution tensors in each direction into a gradient mode field, which is used to reflect the degree of thermal inhomogeneity at each spatial point. The expression is:

[0088] Among them, G(x, y, z) is the temperature gradient modulus at the three-dimensional coordinate point (x, y, z), and are the first-order spatial gradient values of temperature in each direction respectively; this formula calculates the thermal gradient size of a point in three-dimensional space by taking the square root of the sum of squares. Its physical meaning is the comprehensive rate of temperature change within a unit distance, and it is widely used to identify the changing boundary or thermal anomaly area in the temperature field; then the projection and averaging are performed along a main observation plane (such as the Y-Z plane), and finally a two-dimensional thermal field characteristic matrix H(y, z) is obtained, in which each element represents the average thermal gradient value of the area, and its expression is:

[0089] Where H(y, z) is the thermal field projection eigenvalue under the fixed YZ plane; L x is the integral length in the X direction, in meters; this operation realizes the projection averaging of the three-dimensional gradient mode field G(x, y, z) along the X axis.

[0090] The electromagnetic spectrum analysis unit includes:

[0091] Main frequency positioning subunit: used to scan the composite near-field electromagnetic radiation spectrum, identify and locate the frequency position of the global energy peak, which serves as the benchmark main frequency point and provides an accurate reference for subsequent harmonic analysis;

[0092] Harmonic interval division sub-unit: Based on the determined reference main frequency point, a series of frequency band areas centered on integer multiples of the main frequency are generated to ensure that each frequency band completely covers the energy distribution range of the corresponding order harmonics, providing clear interval boundaries for energy calculation;

[0093] Energy product operator unit: used to accumulate the spectrum power in each harmonic interval. By summing the power values of each frequency band, the overall energy value of the corresponding order harmonics is obtained, forming an energy index that can directly reflect the intensity of electromagnetic disturbance;

[0094] Eigenvector output subunit: used to arrange and combine the harmonic energy indicators of each order in order from low to high frequency order, and output the harmonic energy eigenvector for input into the subsequent coupling analysis and diagnosis decision module, ensuring that the electromagnetic characteristics under different motor states are comparable and classifiable.

[0095] The operation steps of the electromagnetic spectrum analysis unit are as follows:

[0096] Step 1: Let f0 represent the main frequency point, and find the energy density function |F c (f)| 2 The global maximum point of determines the main frequency position, and we get: Among them, f is the frequency variable, the unit is Hz; f min ,f max To analyze the lower and upper limits of the frequency band; f0 is the main frequency position, which is used as a reference for harmonic interval division;

[0097] Step 2: Based on the main frequency f0, a set of equal bandwidth frequency intervals with nf0 as the center frequency is constructed. The bandwidth is set to Δf to form the harmonic interval I n , whose expression is: n=1,2,…,N1, where N1 is the number of harmonics extracted;

[0098] Step S3: For each harmonic interval I n Integrate the spectrum energy within to obtain the total energy E of the corresponding harmonic order n , the calculation formula is: Among them, E n is the nth order harmonic energy index; |F c (f)| 2 is the power spectral density of the electromagnetic spectrum at frequency f;

[0099] Step 4: Arrange all N1 harmonic energy indices in ascending order of order to form a harmonic energy feature vector.

[0100] The coupling feature mining module includes a correlation coefficient calculation unit, a dynamic weight allocation unit, and a diagnostic factor fusion unit; wherein:

[0101] Correlation coefficient calculation unit: used to flatten the time-frequency matrix and the thermal field feature matrix into a data sequence of the same dimension according to the corresponding time-space index, and calculate the nonlinear correlation coefficient ρ between the two sets of data sequences based on the mutual information algorithm MI, to quantify the coupling strength between vibration characteristics and thermal field characteristics, the calculation formula is:

[0102] Among them, p XY (x i ,y j ) is the joint probability density function, which means x i with y j The probability of simultaneous occurrence; p X (x i ),p Y (y j ) are the marginal probability density functions of X and Y respectively; ρ MI is the nonlinear correlation coefficient, the larger the value, the stronger the coupling between features; i is the i-th time-frequency eigenvalue after flattening; y j is the jth thermal field eigenvalue after flattening;

[0103] Dynamic weight allocation unit: used to allocate weights according to the nonlinear correlation coefficient ρ MI The relative size of each order harmonic energy index dynamically generates a set of fusion weights {w v , w e}, which is calculated as:

[0104] Among them, w v represents the weight of the vibration-thermal field coupling feature; w e Represents the overall weight of the harmonic energy feature; satisfies the normalization constraint w v +w e =1;E i is the spectrum energy value in the i-th order harmonic interval;

[0105] Diagnostic factor fusion unit: used to perform weighted fusion processing on the harmonic energy vector and the nonlinear correlation coefficient according to the dynamic weight to generate the final comprehensive diagnostic factor vector D. The calculation formula is:

[0106] Among them, D is the final output comprehensive diagnostic factor, which serves as the input for the next step of migration diagnosis; It represents the average value of the harmonic energy index and reflects the overall intensity level of the electromagnetic disturbance. This vector also carries the coupling information of multi-source characteristics and their respective fault sensitivities, and is used to support the fault type identification and confidence assessment of the subsequent migration diagnosis decision module.

[0107] The migration diagnosis decision module includes a feature mapping unit, a shared space alignment unit, a fault matching judgment unit, and a confidence score output unit; wherein:

[0108] Feature Mapping Unit: This unit is used to construct a lightweight feature mapping network based on a meta-learning framework. It employs a few-shot support set training approach and uses diagnostic factor and label pairs from historical motor samples to rapidly initialize model parameters, thereby obtaining a generalizable embedded feature extractor. This network structure can be adaptively expanded based on the feature dimensions of the comprehensive diagnostic factor, supporting multimodal feature input from motors with diverse sources.

[0109] Shared Space Alignment Unit: This unit transforms the input comprehensive diagnostic factors into a shared embedding space through a mapping network, maintaining inter-modal structural consistency and comparability of the input data. In this space, the diagnostic factors generated by different motor platforms are uniformly encoded into structurally aligned vector representations to avoid migration drift caused by differences in data distribution.

[0110] Fault matching and discrimination unit: This unit is used to perform distance measurement and category matching between the target feature vector in the shared space and the labeled feature vectors in the prior fault case library. The case library consists of historical samples from multiple motor platforms, including fault type labels and their corresponding feature embedding vectors. Cross-platform fault type identification is achieved through a nearest neighbor matching strategy or a prototype classifier structure.

[0111] Table 1 Example of the structure of the prior fault case library

[0112]

[0113]

[0114] The case library uses an "embedded feature vector + fault label" structure to facilitate feature space distance calculation or prototype matching between the target motor sample's embedded vector and known cases during the migration diagnosis phase, thereby outputting the most similar fault type and confidence score.

[0115] Confidence score output unit: used to calculate the final fault type confidence score based on the similarity distribution between the target features and the case library samples; the above unit constructs a meta-learning-based feature mapping network and introduces a shared embedding space alignment mechanism. This migration diagnosis decision module can achieve accurate fault identification and classification output under conditions with very few samples or significant differences in models. It has strong generalization ability and good platform adaptability, significantly improving the diagnostic efficiency and stability of the motor system under cross-scenario deployment.

[0116] The fault matching judgment unit includes:

[0117] Feature normalization subunit: This is used to uniformly normalize the embedding vectors of the target comprehensive diagnostic factor obtained after inputting the feature mapping network and the embedding vectors in the case library to eliminate the impact of differences in numerical scales of different dimensions and enhance the fairness of the distance metric. This operation maps each vector to the unit norm range.

[0118] Vector distance calculation subunit: used to calculate the Euclidean distance between the target sample embedding vector and each labeled sample vector in the case library. Let the target vector be z t , the i-th case vector is z i , then the Euclidean distance d between the two i The expression is: i =||z t -z i ||2;

[0119] Similarity sorting subunit: used to sort all the calculated distance values d i Sort from small to large and select the first K neighboring cases with the smallest distance as the candidate reference sample set;

[0120] Category determination subunit: This unit is used to count the fault category with the highest frequency among the first K nearest neighbor samples as the final diagnosis result of the target sample and output the final fault type label. The above-mentioned unit can efficiently complete cross-model and multi-modal fault type mapping by standardizing, measuring and sorting the distance relationship between the target sample and the historical case samples in the shared feature space, thereby improving the diagnostic accuracy under few-sample conditions. It also has good interpretability and computational efficiency, making it suitable for deployment in real-time motor monitoring systems.

[0121] The confidence score output unit includes:

[0122] Distance inverse conversion subunit: used to convert the Euclidean distance values of the first K nearest neighbor samples into similarity weights, using the method of adding a smoothing factor to the inverse distance to suppress the near-zero denominator, as expressed as follows: Among them, s i is the similarity score of the ith neighbor sample; ε is a smoothing constant to prevent division by zero errors, and its value is 10 -6 ;

[0123] Category weight accumulation subunit: It is used to perform weighted accumulation of similarity scores of neighbor samples with the same fault type label to construct the confidence original score of each candidate fault type. Suppose the fault type set is C = {c1, c2, ..., c M}, then the score calculation formula for each type is: Where: S(c M ) indicates fault type c j The cumulative score of Indicates that it belongs to category c M The set of all neighbor sample indexes;

[0124] Score normalization subunit: used to convert the scores of all candidate fault types into a confidence probability distribution in the range of 0–1 and output the final fault confidence score vector.

[0125] like Figure 2 As shown, the motor testing and diagnosis method based on multiple data acquisition is implemented by the above-mentioned motor testing and diagnosis system based on multiple data acquisition, and includes the following steps:

[0126] S1: Deploy three-dimensional orthogonal vibration sensors, distributed infrared thermal imagers, axial loop antennas, radial loop antennas, and stator gap probes to obtain mechanical vibration time domain signals, temperature field spatial distribution data, and near-field electromagnetic radiation spectra.

[0127] S2: Perform multi-layer short-time Fourier decomposition on the mechanical vibration time domain signal to generate a two-dimensional time-frequency matrix. Perform three-dimensional thermal gradient convolution on the temperature field spatial distribution data to generate a thermal field feature matrix. At the same time, locate the main frequency and each order harmonic interval in the electromagnetic radiation spectrum, extract the corresponding energy eigenvectors, and construct the three types of features into a multidimensional tensor according to a unified indexing rule.

[0128] S3: Based on the mutual information algorithm, the nonlinear correlation coefficient between the time-frequency matrix and the thermal field characteristic matrix is calculated. Combined with the harmonic energy characteristics, a dynamic weight distribution mechanism is used to form a comprehensive diagnostic factor vector to describe the multimodal coupling response of the motor operating state.

[0129] S4: Utilizes a pre-trained feature mapping network using a meta-learning strategy and projects the comprehensive diagnostic factor vector into a shared embedding space of fixed dimension to maintain feature alignment and comparability across samples of different models;

[0130] S5: In the shared embedding space, the Euclidean distance between the target embedding vector and the labeled embedding vector in the prior fault case library is calculated. Several neighboring samples with the smallest distance are selected to filter out the fault category, and the inverse distance normalization method is used to output the corresponding confidence score.

[0131] The present invention encompasses any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention. To provide a thorough understanding of the present invention, specific details are described in detail below in connection with the preferred embodiments of the present invention, but those skilled in the art will be able to fully understand the present invention without these detailed descriptions. Furthermore, to avoid unnecessary confusion regarding the essence of the present invention, well-known methods, processes, procedures, components, and circuits have not been described in detail.

[0132] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. The motor testing and diagnosis system based on multi-data acquisition is characterized by: It includes a multimodal sensor array module, a heterogeneous data fusion module, a coupling feature mining module, and a migration diagnosis decision module; among which: Multimodal sensing array module: Arranges three-dimensional orthogonal vibration sensors, distributed infrared thermal imagers, and electromagnetic radiation probes to collect the motor's mechanical vibration time domain signal, temperature field spatial distribution data, and near-field electromagnetic radiation spectrum; Heterogeneous data fusion module: used to perform multi-layer time-frequency decomposition on the mechanical vibration time domain signal to generate a time-frequency matrix, perform thermal gradient spatial convolution on the temperature field spatial distribution data to generate a thermal field feature matrix, and extract the fundamental frequency harmonic energy distribution characteristics in the near-field electromagnetic radiation spectrum; Coupling feature mining module: used to calculate the nonlinear correlation coefficient between the time-frequency matrix and the thermal field feature matrix, and fuse the harmonic energy distribution characteristics with the nonlinear correlation results to generate a comprehensive diagnostic factor; Migration diagnosis decision module: used to build a meta-learning-based feature mapping network, map comprehensive diagnostic factors to a shared feature space, and perform cross-model migration diagnosis in combination with a priori fault case library, outputting fault types and confidence scores.

2. The motor testing and diagnosis system based on multi-data acquisition according to claim 1, characterized in that: The multimodal sensing array module includes a vibration sensing unit, a temperature field reconstruction unit, and an electromagnetic spectrum acquisition unit; wherein: Vibration sensing unit: Consists of three piezoelectric accelerometers. The first sensor is installed along the motor's axial direction, while the second and third sensors are symmetrically fixed at 45-degree angles on both sides of the motor's radial direction. The sampling frequency of the first sensor is set to above 10kHz to capture rotor dynamic characteristics, while the sampling frequencies of the second and third sensors are set to below 1kHz to collect bearing vibration signals. Temperature field reconstruction unit: It includes six infrared thermal imagers, each with a 60-degree field of view and 15-degree overlap between adjacent fields of view. Each thermal imager has a built-in uncooled microbolometer, which generates three-dimensional temperature field distribution data through multi-view data fusion. Electromagnetic spectrum acquisition unit: It consists of an axial loop antenna, a radial loop antenna and a stator gap probe. The axial antenna is installed parallel to the end face of the motor to detect the main frequency component of electromagnetic radiation. The radial antenna vertically surrounds the motor casing to collect leakage magnetic harmonics. The stator gap probe is embedded in the motor slot to directionally receive partial discharge signals. The output signals of the three acquisition channels are processed by frequency band selective fusion to form a near-field electromagnetic radiation spectrum.

3. The motor testing and diagnosis system based on multiple data acquisition according to claim 1, characterized in that: The electromagnetic spectrum acquisition unit includes: The filter preprocessing subunit is used to perform bandpass filtering on the three signals. Specifically, the axial loop antenna signal passes through a 1MHz to 30MHz passband filter to extract the main frequency radiation component; the radial loop antenna signal passes through a 100kHz to 5MHz passband filter to retain the harmonic components generated by the motor operation; and the stator gap probe signal passes through a 500kHz to 20MHz passband filter to extract the high-frequency pulse components in the partial discharge signal. Normalization and compression subunit: used to perform amplitude normalization and logarithmic quantization compression processing on each filtered signal to obtain a compressed signal; Spectrum fusion construction subunit: used to convert the normalized three-way compressed signal into the frequency domain and construct the composite near-field electromagnetic radiation spectrum F based on the weighted superposition strategy c (f), its expression is: F c (f)=w1·F1(f)+w2·F2(f)+w3·F3(f), where F1(f), F2(f), and F3(f) represent the spectra of the three signals after fast Fourier transform, respectively; w1, w2, and w3 are spectrum fusion weight coefficients that satisfy the constraint condition of w1+w2+w3=1; and f is a frequency variable.

4. The motor testing and diagnosis system based on multiple data acquisition according to claim 1, characterized in that: The heterogeneous data fusion module includes a time-frequency decomposition unit, a thermal field extraction unit, an electromagnetic spectrum analysis unit and a feature fusion unit; wherein: Time-frequency decomposition unit: used to perform multi-layer time-frequency decomposition processing on the mechanical vibration time domain signal from the vibration sensing unit. It uses a step-by-step deconstruction method to reconstruct and decompose the signal at different time scales and frequency scales, and extract a two-dimensional time-frequency matrix containing time locality and frequency distribution; Thermal field extraction unit: used to perform thermal gradient convolution calculation on the three-dimensional temperature field spatial distribution data generated by the temperature field reconstruction unit. By constructing a multi-directional thermal gradient template in the spatial domain and performing directional convolution operations along each dimension of the temperature field tensor, it extracts local thermal slope change information and overall thermal diffusion trend, and finally generates a thermal field feature matrix. Electromagnetic spectrum analysis unit: used to extract the fundamental frequency harmonic energy distribution from the near-field electromagnetic radiation spectrum from the electromagnetic spectrum acquisition unit. By locating the main frequency position and its integer multiple frequency point area in the frequency domain, the spectrum energy of each harmonic bandwidth interval is integrated and calculated, and the characteristic vector containing multiple harmonic component energy indicators is output; Feature fusion unit: used to map the time-frequency matrix, thermal field feature matrix and harmonic energy vector to a unified dimensional structure and construct a third-order feature tensor.

5. The motor testing and diagnosis system based on multiple data acquisition according to claim 4, characterized in that: The thermal field extraction unit includes: Gradient template construction subunit: used to construct spatial gradient templates in three directions, defining the first-order difference kernel along the X, Y and Z spatial dimensions of the temperature field tensor respectively; Spatial convolution calculation subunit: used to perform gradient convolution on the input three-dimensional temperature field tensor in the X, Y and Z directions respectively to obtain the gradient distribution tensor in each direction; Thermal field feature generation subunit: used to integrate the gradient distribution tensors in each direction into the gradient mode field; then perform projection averaging along a main observation surface to finally obtain a two-dimensional thermal field feature matrix.

6. The motor testing and diagnosis system based on multiple data acquisition according to claim 5, characterized in that: The electromagnetic spectrum analysis unit comprises: Main frequency positioning subunit: used to scan the composite near-field electromagnetic radiation spectrum, identify and locate the frequency position of the global energy peak, and use it as the reference main frequency point; Harmonic interval division sub-unit: Based on the determined reference main frequency point, a series of frequency band areas centered on integer multiples of the main frequency are generated to provide clear interval boundaries for energy calculation; Energy product operator unit: used to accumulate the spectrum power in each harmonic interval. By summing the power values of each frequency band, the overall energy value of the corresponding order harmonics is obtained, forming an energy index that can directly reflect the intensity of electromagnetic disturbance; Feature vector output subunit: used to arrange and combine the harmonic energy indicators of each order in the order of frequency order from low to high, and output the harmonic energy feature vector.

7. The motor testing and diagnosis system based on multiple data acquisition according to claim 1, characterized in that: The coupling feature mining module includes a correlation coefficient calculation unit, a dynamic weight allocation unit, and a diagnostic factor fusion unit; wherein: Correlation coefficient calculation unit: used to flatten the time-frequency matrix and the thermal field feature matrix into a data sequence of the same dimension according to the corresponding time-space index, and calculate the nonlinear correlation coefficient ρ between the two sets of data sequences based on the mutual information algorithm MI ; Dynamic weight allocation unit: used to allocate weights according to the nonlinear correlation coefficient ρ MI The relative size of each order harmonic energy index dynamically generates a set of fusion weights {w v , w e }, where w v represents the weight of the vibration-thermal field coupling feature; w e Represents the overall weight of the harmonic energy characteristics; Diagnostic factor fusion unit: used to perform weighted fusion processing on the harmonic energy vector and the nonlinear correlation coefficient according to the dynamic weight to generate the final comprehensive diagnostic factor vector D. The calculation formula is: Among them, D is the comprehensive diagnostic factor of the final output; Indicates the average value of harmonic energy index.

8. The motor testing and diagnosis system based on multiple data acquisition according to claim 1, characterized in that: The migration diagnosis decision module includes a feature mapping unit, a shared space alignment unit, a fault matching judgment unit and a confidence score output unit; wherein: Feature Mapping Unit: This unit is used to construct a lightweight feature mapping network based on a meta-learning framework. It uses a few-shot support set training approach and uses diagnostic factor and label pairs from historical motor samples to quickly initialize model parameters and obtain a generalizable embedded feature extractor. Shared space alignment unit: used to convert the input comprehensive diagnostic factor into a shared embedding space through the mapping network; Fault matching and discrimination unit: used to perform distance measurement and category matching between the target feature vector in the shared space and the labeled feature vector in the prior fault case library; Confidence score output unit: used to calculate the final fault type confidence score based on the similarity distribution between the target feature and the case library samples.

9. The motor testing and diagnosis system based on multiple data acquisition according to claim 8, characterized in that: The fault matching determination unit includes: Feature normalization subunit: used to perform unified normalization processing on the embedding vector obtained after the target comprehensive diagnostic factor is input into the feature mapping network and the embedding vector in the case library; Vector distance calculation subunit: used to calculate the Euclidean distance between the target sample embedding vector and each labeled sample vector in the case library. Let the target vector be z t , the i-th case vector is z i , then the Euclidean distance d between the two i The expression is: i =||z t -z i ||2; Similarity sorting subunit: used to sort all the calculated distance values d i Sort from small to large and select the first K neighboring cases with the smallest distance as the candidate reference sample set; Category determination subunit: It is used to count the fault category with the highest frequency among the first K nearest neighbor samples as the final diagnosis result of the target sample and output the final fault type label.

10. A motor testing and diagnosis method based on multiple data acquisition is implemented by the motor testing and diagnosis system based on multiple data acquisition according to any one of claims 1 to 9, characterized in that: The following steps are involved: S1: Deploy three-dimensional orthogonal vibration sensors, distributed infrared thermal imagers, axial loop antennas, radial loop antennas, and stator gap probes to obtain mechanical vibration time domain signals, temperature field spatial distribution data, and near-field electromagnetic radiation spectra. S2: Perform multi-layer short-time Fourier decomposition on the mechanical vibration time domain signal to generate a two-dimensional time-frequency matrix. Perform three-dimensional thermal gradient convolution on the temperature field spatial distribution data to generate a thermal field feature matrix. At the same time, locate the main frequency and each order harmonic interval in the electromagnetic radiation spectrum, extract the corresponding energy eigenvectors, and construct the three types of features into a multidimensional tensor according to a unified indexing rule. S3: Based on the mutual information algorithm, the nonlinear correlation coefficient between the time-frequency matrix and the thermal field characteristic matrix is calculated, and the comprehensive diagnostic factor vector is formed through the dynamic weight allocation mechanism in combination with the harmonic energy characteristics; S4: Utilizes a feature mapping network pre-trained using a meta-learning strategy and projects the comprehensive diagnostic factor vector into a shared embedding space of fixed dimensionality; S5: In the shared embedding space, the Euclidean distance between the target embedding vector and the labeled embedding vector in the prior fault case library is calculated. Several neighboring samples with the smallest distance are selected to filter out the fault category, and the inverse distance normalization method is used to output the corresponding confidence score.

Images