Motor internal defect detection method based on multi-source data fusion

Abstract

The invention discloses a motor internal defect detection method based on multi-source data fusion, and particularly relates to the field of motor detection.The method comprises the steps that stator current, stator voltage and temperature signals during motor operation are collected, and a multi-source original data set is generated; the data set is preprocessed, a wavelet threshold algorithm is adopted to denoise current and voltage signals, a moving average algorithm is adopted to smooth temperature signals, linear normalization is adopted to unify amplitude, time interpolation is adopted to synchronize data of different sampling frequencies, and a preprocessed signal set is obtained; establishing a multi-source data fusion model containing a feature extraction layer, a correlation layer and a fusion reconstruction layer, and generating a fusion feature vector; and constructing a deep neural network recognition model based on the fused feature vector, and outputting a defect type recognition result. According to the method, a single signal information gap is complemented through multi-source data fusion, the data quality is improved through accurate preprocessing, the defect recognition accuracy and reliability are improved through a deep recognition model, and the method is suitable for detecting various internal defects of the motor.

Description

Technical Field

[0001] This invention relates to the field of motor testing technology, and more specifically, to a method for detecting internal defects in motors based on multi-source data fusion. Background Technology

[0002] As a core power source in industrial production, energy supply, and transportation equipment, the reliability of electric motors directly determines the stability and safety of the entire system. During long-term operation, electric motors are prone to typical defects such as winding short circuits, rotor eccentricity, bearing wear, and partial discharge due to factors like insulation aging, mechanical wear, and electromagnetic forces. According to industrial equipment failure statistics, these internal defects account for over 60% of motor failures. This not only causes direct economic losses due to motor downtime for repairs but can also lead to safety accidents such as fires and power grid fluctuations in severe cases. Therefore, early and accurate detection of internal motor defects has become one of the core requirements for industrial equipment operation and maintenance.

[0003] However, existing technologies have significant shortcomings in practical applications: First, the defect information contained in a single signal is incomplete. For example, the stator current signal is difficult to distinguish between the characteristics of rotor eccentricity and slight bearing wear, and the temperature signal has a delayed response to early electrical defects such as partial discharge, which can easily lead to misjudgment or missed detection of defects. Second, the data preprocessing stage is simple and does not adopt differentiated denoising methods for different signal noise characteristics. Furthermore, it does not solve the problem of asynchronous sampling frequencies of multi-source signals (if used in small quantities), and noise interference and time deviation further reduce data quality. Third, the defect identification models mostly use traditional machine learning algorithms (such as support vector machines and decision trees), which have limited ability to fit complex multi-source features and are difficult to accurately identify various types of internal defects in motors. In particular, the recognition rate for early slight defects is low, which cannot meet the actual needs of high-precision motor detection.

[0004] Furthermore, existing technologies suffer from poor adaptability to various operating conditions. Most detection methods are only validated under stable laboratory conditions (such as rated load and standard power grid environment), while in real-world industrial scenarios, motors often experience complex conditions such as load fluctuations (e.g., switching between 50% and 120% of rated load) and power grid harmonic interference (e.g., harmonic content exceeding 5%). This leads to a significant decrease in the detection performance of existing methods, making them unsuitable for direct application in actual operation and maintenance scenarios. In summary, current motor defect detection technologies still fall short of meeting the practical needs of "early warning, accurate identification, and adaptation to complex operating conditions." A novel detection method is urgently needed that can integrate the advantages of multi-source signals, optimize data processing workflows, and enhance model recognition capabilities. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method for detecting internal defects in motors based on multi-source data fusion. The following scheme addresses the problems of incomplete single signal information, coarse data preprocessing, and low accuracy of recognition models in existing motor defect detection technologies mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for detecting internal defects in a motor based on multi-source data fusion, comprising: S1: collecting multi-source raw data during motor operation to generate a multi-source raw dataset; the multi-source raw data includes stator current signal, stator voltage signal, and temperature signal;

[0007] S2: Perform data preprocessing on the multi-source raw dataset, use wavelet thresholding algorithm to denoise the current signal and voltage signal, use moving average algorithm to perform steady-state smoothing on the temperature signal, use linear normalization algorithm to unify the amplitude of each signal, and use time interpolation-based synchronization algorithm to time-align data with different sampling frequencies to obtain a preprocessed signal set;

[0008] S3: Based on the preprocessed signal set, a multi-source data fusion model is established. The fusion model includes a feature extraction layer, a feature association layer, and a fusion reconstruction layer. The feature extraction layer extracts the time-domain and frequency-domain features of current signals, voltage signals, and temperature signals through convolution calculation. The feature association layer establishes the mapping relationship between signals based on the feature correlation calculation method. The fusion reconstruction layer uses a feature space reconstruction algorithm to nonlinearly combine different signal features to generate a fused feature vector.

[0009] S4: Construct a deep neural network recognition model based on the fused feature vector, train and optimize the parameters of the defect samples through a supervised learning algorithm, and output the defect type recognition result.

[0010] The technical effects and advantages of this invention are as follows:

[0011] 1. This invention collects three types of signals: stator current, voltage, and temperature, covering the electrical and thermal characteristics of the motor. It makes up for the information gap of a single signal and can simultaneously capture the characteristics of multiple defects such as winding short circuit (current imbalance), rotor eccentricity (current periodic distortion), bearing wear (temperature rise), and partial discharge (high-frequency voltage and current pulses), avoiding misjudgment and missed judgment caused by a single signal.

[0012] 2. Differentiated preprocessing algorithms are adopted for different signal characteristics. Wavelet threshold denoising improves the signal-to-noise ratio of current and voltage signals by ≥10dB, moving average smoothing reduces the standard deviation of temperature signals by ≥30%, linear normalization eliminates dimensional differences, and time interpolation synchronization ensures that the time deviation of multi-source signals is ≤0.5ms, providing a high-quality data foundation for subsequent feature extraction and recognition.

[0013] 3. The multi-source data fusion model extracts multi-dimensional features through convolution, establishes signal association through correlation analysis, and strengthens key features through attention mechanism. Combined with fully connected deep neural networks and supervised learning, it achieves a test set accuracy of ≥92% and an F1 score of ≥0.90 for each defect category. In particular, its ability to identify early minor defects is significantly better than traditional methods. Moreover, the model has strong generalization ability and is suitable for motor inspection of different models and working conditions. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the overall method flow of the present invention;

[0015] Figure 2 This is a schematic diagram of the preprocessing flow of the present invention;

[0016] Figure 3 This is a flowchart illustrating the construction process of the multi-source data fusion model of the present invention.

[0017] Figure 4 This is a flowchart illustrating the process of obtaining the defect type identification results according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] As attached Figure 1 The method for detecting internal defects in motors based on multi-source data fusion, as shown, includes:

[0020] S1: Collect multi-source raw data during motor operation to generate a multi-source raw dataset; the multi-source raw data includes stator current signal, stator voltage signal and temperature signal;

[0021] This embodiment requires a detailed explanation of the multi-source raw data acquisition process as follows:

[0022] Before collecting multi-source raw data, it is necessary to complete the recording of target motor parameters and the setting of operating conditions. First, record the basic parameters of the motor to be tested, including motor model, rated power, rated voltage, rated current, rated speed and number of phases. These parameters are used for subsequent matching of sensor range and determining whether the data is within the normal operating range. Second, set the motor operating conditions, which need to cover two core scenarios: one is the normal operating state, used to collect defect-free baseline data; the other is the preset defect state, including winding short circuit (such as artificially creating 1-turn or 2-turn short circuits), rotor eccentricity (such as adjusting rotor concentricity to 5% or 10% eccentricity), bearing wear (such as selecting bearings with wear of 0.1mm or 0.2mm), and partial discharge (such as simulating partial discharge scenarios through high-voltage equipment). Each operating condition needs to be collected under different load conditions (such as 50% rated load, 100% rated load, and 120% rated load) to ensure coverage of the typical load range in actual motor operation.

[0023] It is important to further explain that the acquisition method for stator current signals requires careful attention to sensor selection, installation location, and signal output processing. For sensor selection, a closed-loop Hall effect current sensor is used. This sensor features non-contact measurement characteristics, a response bandwidth ≥10kHz (capturing high-frequency defect characteristics in the current), a range adapted to 1.5 times the motor's rated current, strong anti-interference capabilities, and avoidance of the influence of power grid harmonics on the measurement. For installation location, three Hall effect current sensors are connected in series in the A, B, and C phase power supply lines of the motor stator windings. The installation points must be close to the motor terminals (≤10cm from the terminals) to reduce measurement errors caused by line losses. During installation, ensure good insulation between the sensors and the lines, using insulated terminals for fixation, and align the sensor probe with the centerline of the conductor to avoid decreased measurement accuracy due to angular deviations. For signal output processing, the sensors convert the acquired current signal into a standard 0-5V analog voltage signal, which is then connected to the independent analog input channel of the data acquisition card via a shielded cable (with the shield grounded). Each current sensor corresponds to one channel to prevent crosstalk between channels.

[0024] The acquisition method for stator voltage signals must match the motor voltage level while ensuring installation safety. For sensor selection, use voltage divider sensors or Hall effect sensors with a withstand voltage ≥ twice the motor's rated voltage, a response bandwidth ≥ 10kHz, and a measurement accuracy ≤ 0.5%, ensuring accurate capture of subtle changes in the voltage waveform. For installation, connect three voltage sensors in parallel to the A, B, and C phase input terminals of the motor stator windings, respectively, and fix them in the motor control cabinet using insulating brackets, maintaining a safe distance (≥ 5cm) from high-voltage lines. If voltage divider sensors are used, ensure the power rating of the divider resistors is matched to avoid overheating and damage. For signal output processing, the sensors step down the high-voltage signal to a 0-5V analog voltage signal, which is then connected to the dedicated analog input channel of the data acquisition card (separate from the current signal channel) via an independent shielded cable. The channel input impedance is set to 1MΩ to reduce signal attenuation.

[0025] Temperature signal acquisition requires differentiated selection and installation for different heat-generating parts of the motor, covering three key locations: stator windings, bearings, and the motor housing. For sensor selection, embedded platinum resistance thermometers are used for stator winding temperature measurement, with a range of -50~200℃ and an accuracy of ±0.1℃, and can be directly embedded in the winding end gaps. K-type thermocouples are used for bearing temperature measurement, with a response time ≤1s and a range of 0~300℃, suitable for the high-temperature environment of the bearing area. Non-contact infrared temperature sensors are used for the motor housing surface temperature measurement, with a measurement distance of 5~10cm and an accuracy of ±0.5℃, avoiding interference from mechanical contact with the housing. Regarding installation location and method, the stator winding ends should be cleaned before installing the platinum resistance thermometers. To remove dust and oil, the sensor is adhered to the surface of the winding copper wire using high-temperature resistant thermally conductive adhesive, and then secured with fiberglass ribbon to ensure a tight fit between the sensor and the winding. The K-type thermocouple is adhered to the outer ring surface of the bearing using high-temperature adhesive with a layer thickness ≤0.5mm to reduce thermal resistance. The infrared sensor is fixed to the middle of the motor housing using a bracket, with the lens facing an area on the housing surface without heat dissipation holes to avoid airflow affecting the measurement. For signal output processing, the resistance signals or mV-level voltage signals output by the platinum resistance thermometer and K-type thermocouple need to be converted into 0-5V analog signals by a signal conditioning module before being connected to the data acquisition card. The infrared sensor directly outputs digital signals, which are connected to the data acquisition card via an I2C communication interface.

[0026] It should be further explained that the sampling parameters for multi-source raw data need to be configured differently according to the changing characteristics of each signal to ensure a balance between data validity and economy. Regarding sampling frequency, the sampling frequency for stator current and stator voltage signals is set to 10kHz~50kHz, with the specific value determined based on the motor's rated frequency, which can capture high-frequency defect characteristics in the current and voltage. The sampling frequency for temperature signals is set to 1Hz~10Hz. Because temperature changes slowly, a lower sampling frequency can reduce redundant data and lower storage pressure. Regarding sampling duration, continuous sampling should be conducted for 30 minutes to 1 hour under each operating condition, covering three stages: motor startup (startup process approximately 10~30 seconds), steady-state operation (accounting for more than 80% of the sampling time), and load switching (if load switching is included), ensuring that the data includes the characteristics of different operating stages of the motor. Finally, the collected multi-source raw data is integrated into a multi-source raw dataset.

[0027] S2: Perform data preprocessing on the multi-source raw dataset, use wavelet thresholding algorithm to denoise the current signal and voltage signal, use moving average algorithm to perform steady-state smoothing on the temperature signal, use linear normalization algorithm to unify the amplitude of each signal, and use time interpolation-based synchronization algorithm to time-align data with different sampling frequencies to obtain a preprocessed signal set;

[0028] The data preprocessing process described in this embodiment is as follows:

[0029] The specific method for denoising current and voltage signals using wavelet thresholding algorithms needs to be designed with parameters tailored to the signal characteristics; the specific steps include:

[0030] The first step is wavelet decomposition: Select a suitable wavelet basis for non-stationary electrical signals, and perform multi-level decomposition on current and voltage signals. The number of decomposition levels is determined according to the signal sampling frequency. For example, a 20kHz sampled signal is decomposed into 5 levels to ensure that high-frequency noise is concentrated in levels 1-3 and defect features are concentrated in levels 4-5, thereby obtaining the low-frequency approximation coefficients and high-frequency detail coefficients of each level.

[0031] The second step, threshold selection and processing: A soft thresholding function is used to process high-frequency detail coefficients. The threshold calculation formula adopts an adaptive threshold based on Stein's unbiased risk estimation, as follows: Where N is the signal length and σ is the noise standard deviation, estimated by the median absolute deviation of the first-layer high-frequency coefficients, i.e. c 1为 Wavelet decomposition of the signal to be denoised: first-level high-frequency coefficients; set coefficients with absolute values ​​less than the threshold to zero, and subtract the threshold from coefficients with absolute values ​​greater than the threshold;

[0032] Step 3, Wavelet Reconstruction: Perform wavelet inverse transform on the processed high-frequency detail coefficients and the unprocessed low-frequency approximation coefficients to reconstruct the denoised current signal and voltage signal. Verify the effect by comparing the signal-to-noise ratio (SNR) of the signal before and after denoising, and require the SNR to increase by ≥ 10 dB.

[0033] It should be further noted that the specific method of using the moving average algorithm to perform steady-state smoothing on the temperature signal needs to balance the smoothing effect and feature retention. The specific methods include:

[0034] Determination of Window Size: Set the sliding window length according to the temperature change period, and use an adaptive window. Under normal operating conditions, the window length is 30 sampling points, corresponding to 3 seconds to 30 seconds, because the temperature sampling frequency is 1 - 10 Hz; when the detected temperature change rate > 0.5 °C / min, the window length is halved to 15 sampling points to avoid over-smoothing from masking the abnormal upward trend of temperature.

[0035] Calculation Method: For the temperature signal sequence (T(m) (m = 1, 2,..., n)), the calculation formula for the signal T'(m) after moving average is: , where T(m) is the value of the original temperature signal at time m; m is the time index (m = 1 corresponds to the first sampling point, m = n corresponds to the last sampling point); n is the total number of sampling points of the temperature signal; T'(m) is the value of the smoothed temperature signal at time m; W is the sliding window length (i.e., the number of sampling points participating in the average); i is the index within the window (from 0 to W - 1, corresponding to the first W sampling points within the window); when m < W (i.e., the starting segment of the signal, the window cannot fully cover the first W points), use forward averaging: , where i = 1 to m are all sampling points from the first sampling point to the current time m, ensuring the effectiveness of the starting segment of the signal. After smoothing, verify the reduction rate of the standard deviation of the temperature signal (required ≥ 30%), and at the same time ensure that the deviation of the temperature peak time ≤ 2%, to avoid losing key temperature features.

[0036] Specifically, the unification process in this embodiment is as follows:

[0037] Use the linear normalization algorithm to perform unification processing on the amplitudes of each signal, eliminate the dimensional differences of current, voltage, and temperature, and avoid the model over-focusing on high-amplitude signals; in the original signal, the amplitude of the current signal is usually 0 - dozens of A, the voltage signal is 0 - hundreds of V, and the temperature signal is 0 - dozens of °C, with significant differences in the numerical ranges. Direct fusion will cause the model to over-focus on high-amplitude signals; linear normalization eliminates the dimensional influence by mapping each signal to the same numerical interval.

[0038] Specifically, the time alignment process in this embodiment is as follows:

[0039] The specific method for time alignment of data with different sampling frequencies using a time interpolation-based synchronization algorithm requires establishing a unified time axis. Time alignment maps low-sampling-rate signals to a high-sampling-rate time axis using interpolation methods, including:

[0040] Time axis unification: The time axis is based on the timestamp of the denoised current signal (accurate to the millisecond level), denoted as t. k (k=1,2,...,M), where M is the amount of current signal data), where t k This represents the absolute time of the k-th current sampling point;

[0041] Interpolation processing: For temperature signals, interpolation is performed based on their original timestamp t' j (j=1,2,...,N), where N is the amount of temperature signal data, t' j (where t is the absolute time of the j-th temperature sampling point), and linear interpolation is used to calculate each reference time t. k The corresponding temperature value T(t) k The calculation formula is: , where T(t) k (t) is the reference time. k The corresponding interpolated temperature value; Indicates t k Located between the j-th and j+1-th original timestamps of the temperature signal; T(t' j ) represents the temperature signal at the original timestamp t' j The value of T(t' j+1 ) represents the temperature signal at the original timestamp t' j+1 The value of t; k -t' j For t k The time difference with the previous temperature timestamp; t' j+1 -t' j The interval between two adjacent temperature timestamps.

[0042] It should be further explained that for voltage signals, since they have the same sampling frequency as current signals (both are high-frequency signals), only timestamp calibration is required (a deviation ≤ 1ms is considered synchronization, and a deviation greater than 1ms is used to take the signal value of the closer timestamp); then alignment verification is performed: 100 time points are randomly selected, and the time deviation between the aligned voltage, temperature and current signals is calculated. The deviation is required to be ≤ 0.5ms to ensure the consistency of multi-source signals in the time dimension.

[0043] After the above wavelet threshold denoising, moving average smoothing, linear normalization and time alignment processing, the final preprocessed signal set specifically includes: the denoised and normalized stator current signal, the denoised and normalized stator voltage signal, and the smoothed and normalized temperature signal.

[0044] S3: A multi-source data fusion model is established based on the preprocessed signal set. The fusion model includes a feature extraction layer, a feature association layer, and a fusion reconstruction layer. The feature extraction layer extracts the time-domain and frequency-domain features of current, voltage, and temperature signals through convolution calculations, and supplements time-series segmentation processing to ensure signal correlation. The feature association layer establishes the mapping relationship between signals based on the feature correlation calculation method and quantifies the feature dependence degree under different fault conditions. The fusion reconstruction layer combines different signal features nonlinearly based on the feature space reconstruction algorithm, dynamically allocates feature weights according to the fault type, and generates a fusion feature vector for defect identification.

[0045] This embodiment requires a detailed explanation of the multi-source data fusion model establishment process as follows:

[0046] First, a feature extraction layer is established. This layer extracts the time-domain and frequency-domain features of current, voltage, and temperature signals through convolution calculations. Differentiated convolution structures need to be designed for the characteristics of different signals. Simultaneously, time-series segmentation processing is added to preserve the temporal correlation of the signals. The preprocessed signal set is a time-series signal (based on a unified time axis, with a current / voltage sampling frequency of 30kHz and a temperature sampling frequency of 5Hz). All signals are first segmented at 1s / segment, with each segment containing 30,000 current / voltage sampling points and 5 temperature sampling points. Time interpolation ensures the time alignment of multiple source signals within each segment. Then, a sliding window-style local feature extraction is performed through convolution calculations, effectively capturing the temporal variation patterns of the signals. Frequency domain distribution characteristics: For the preprocessed current, voltage, and temperature signals, a one-dimensional convolutional neural network (1D-CNN) is used to extract local temporal features: For the current and voltage signals, the focus is on abrupt changes and periodic fluctuations. Two convolutional sub-layers are set up. The first sub-layer uses 16 convolutional kernels (size 10, stride 2) with the ReLU activation function to extract high-frequency local features (such as the rate of change of current over 10 consecutive time steps). The second sub-layer uses 32 convolutional kernels (size 20, stride 4) to extract low-frequency trend features (such as the fluctuation of the average voltage over 20 consecutive time steps). For the temperature signal, the focus is on the slow upward trend of temperature. One convolutional sub-layer is set up with 8 convolutional kernels. A kernel (size 50, stride 10) with ReLU activation function is used to extract long-term temperature change features (e.g., cumulative temperature increment over 50 consecutive time steps). For frequency domain feature extraction, a Fast Fourier Transform (FFT) is first performed on the current and voltage signals to obtain their spectra (frequency resolution 1Hz). Temperature signals, due to their slow change, are directly analyzed based on time-domain spectrum (extracting low-frequency features from 0~0.1Hz using a Short-Time Fourier Transform). For harmonic components in the current and voltage signals' spectra (e.g., the 2nd and 4th harmonics during rotor eccentricity), a one-dimensional convolution is used to process the spectral sequence. Two convolutional sub-layers are set up. The first sub-layer uses 16 convolutional kernels (size 5, stride 1) to extract specific frequency features. For the energy concentration characteristics, the second sub-layer uses 32 convolutional kernels (size 10, stride 2) to extract the spectral distribution characteristics within a wide bandwidth. For the temperature signal, targeting the low-frequency component of 0~0.1Hz in the spectrum, one convolutional sub-layer (8 convolutional kernels, size 3, stride 1) is used to extract the energy change characteristics of the low-frequency band. Finally, the time-domain and frequency-domain convolutional features are flattened to form a single-source feature set of current feature vector (dimension 128), voltage feature vector (dimension 128), and temperature feature vector (dimension 32, integrating the time-domain / frequency-domain features of three measurement points: stator winding, bearing, and housing). This ensures that each feature corresponds to the motor operating state in the same time period and avoids correlation distortion caused by the mixing of features across time periods.

[0047] The feature association layer establishes the mapping relationship between signals based on the feature correlation calculation method. It requires quantifying the dependence between different source features and selecting physically meaningful associated feature pairs based on the physical laws of motor faults. A method combining Pearson correlation coefficient (measuring linear correlation) and mutual information (measuring nonlinear correlation) is adopted: the formula for Pearson correlation coefficient is... Where X and Y are features from different sources, cov(X,Y) is the covariance, and σ X σ Y The standard deviation is ; the mutual information formula is . Where p(x,y) is the joint probability distribution, and p(x) and p(y) are the marginal probability distributions; a correlation threshold is set (Pearson correlation coefficient absolute value ≥ 0.6, mutual information ≥ 0.3), and feature pairs are screened in combination with the physical characteristics of the fault: such as "current second harmonic component - voltage second harmonic component" when the rotor is eccentric (Pearson correlation coefficient ≥ 0.7, ≤ 0.3 under normal operating conditions), "current high-frequency pulse - winding temperature" when the winding is short-circuited (mutual information ≥ 0.5, ≤ 0.2 under normal operating conditions), "vibration characteristics (indirectly reflected by current signal) - bearing temperature" when the bearing is worn (mutual information ≥ 0.4, ≤ 0.15 under normal operating conditions), etc., and an association matrix C is constructed (dimension is (number of current features + number of voltage features) × number of temperature features = 256 × 32), and the matrix elements are the correlation indicators (Pearson correlation coefficient or mutual information value) of the corresponding feature pairs, which not only retains the key correlation information of multi-source features, but also eliminates false correlations without physical meaning (such as the weak correlation between isolated current features and temperature features caused by power grid interference).

[0048] The fusion reconstruction layer uses a feature space reconstruction algorithm to nonlinearly combine different signal features and dynamically assigns feature weights based on fault type to generate a fusion feature vector for defect identification. It employs an algorithm combining kernel principal component analysis (KPCA) with an attention mechanism. The KPCA input is the associated feature matrix (dimension N). sample ×(128+128+32), where N sample Given the number of samples (each sample corresponds to one segment of signal correlation features), a radial basis function (RBF) kernel is selected. (γ=0.01) determined through 5-fold cross-validation), after mapping the features to a high-dimensional space, the principal components corresponding to the top 100 largest eigenvalues ​​are selected to form an intermediate feature vector (dimensional 100). This process can retain more than 95% of the fault feature information and effectively reduce feature dimensionality and redundancy. In the dynamic attention mechanism, the mutual information MI(g) between each dimension of the intermediate feature and the defect label is first calculated according to the fault type classification. k(e.g., under the winding fault label, the MI(g25,D) of the 25th dimension (corresponding to voltage distortion characteristics) is 0.6; under the bearing fault label, the MI(g78,D) of the 78th dimension (corresponding to temperature peak characteristics) is 0.55), and then the attention weight is obtained through normalization. This enables dynamic allocation of feature weights under different fault types: for example, when identifying winding short circuits, the weight α of the voltage distortion-related dimension is adjusted. 25 =0.3; When identifying bearing wear, the weight α of the temperature peak correlation dimension. 78 =0.25; the final fused feature vector is F=[α1g1,α2g2,...,α 100 g 100 With a dimension of 100, the vector, verified by t-SNE visualization, can improve the feature differentiation between the composite fault of "bearing wear + inter-turn short circuit" and the single fault by 40%, and integrates single-source details and multi-source correlation information, which can be directly used for defect identification.

[0049] It should be further explained that traditional multi-source data fusion methods often use simple linear combinations (such as feature splicing and weighted summation) or ordinary principal component analysis (PCA) for dimensionality reduction. The former does not consider the nonlinear correlation between features and the adaptability of fault types, while the latter can only handle linear feature relationships. This leads to serious overlap between the features of composite faults and single faults in the low-dimensional space. Visualizing the feature distribution of existing technologies through t-SNE reveals that the overlap rate between the composite fault feature of "bearing wear + inter-turn short circuit" and the single fault feature of "bearing wear" exceeds 60%, and the overlap rate with the single fault feature of "inter-turn short circuit" exceeds 50%. The feature discrimination can only reach 35%-45%, which is one of the core reasons why the accuracy of existing technologies in identifying composite defects is generally lower than 75%.

[0050] The synergistic effect of KPCA and dynamic attention mechanism demonstrates a breakthrough advantage in t-SNE visualization verification: On the one hand, KPCA maps nonlinear multi-source correlation features to a high-dimensional space through radial basis function (RBF) kernels, breaking the linear limitations of ordinary PCA and enabling the initial separation of originally overlapping features in the high-dimensional space; on the other hand, the dynamic attention mechanism strengthens the weights of key features for different fault types—in the composite fault of "bearing wear + inter-turn short circuit", it enhances the bearing temperature peak feature (78th dimension, weight α). 78 The contribution of α = 0.25 is retained, while the voltage distortion characteristics corresponding to the inter-turn short circuit (25th dimension, weight α) are preserved. 25=0.18), avoiding the loss of composite information caused by the dominance of a single feature; after reducing the high-dimensional fusion features to 2-dimensional space for visualization through t-SNE, the Euclidean distance between the composite fault feature cluster of "bearing wear + inter-turn short circuit" and the single "bearing wear" feature cluster is increased from 1.2±0.3 in the prior art to 2.5±0.4, and the Euclidean distance between the composite fault feature cluster of "bearing wear + inter-turn short circuit" and the single "inter-turn short circuit" feature cluster is increased from 1.0±0.2 to 2.3±0.3. The feature discrimination is increased from about 40% in the prior art to more than 80%, and the feature overlap rate is reduced to less than 15%, which completely solves the problem of confusion between composite fault and single fault features in the prior art.

[0051] Meanwhile, the fusion feature vector's ability to integrate single-source details and multi-source correlation information is superior to existing technologies: existing fusion features often suffer from "loss of single-source details" or "breakage of multi-source correlations," resulting in insufficient ability to capture early composite faults; while the fusion feature vector of this invention includes single-source details such as a 15kHz high-frequency pulse of stator current and a gradual increase of bearing temperature of 0.8℃ / min, and establishes a cross-source correlation of "high-frequency pulse of current - voltage distortion - gradual increase of temperature" through mutual information analysis of the feature correlation layer, giving the fusion feature vector the characteristics of "traceable details and verifiable correlations"—in t-SNE visualization, even an early composite fault with a wear of 0.1mm + 1 turn of short circuit can have its feature cluster clearly separated from the feature clusters of normal operating conditions and single early faults, effectively filling the gap in existing technologies for early composite defect detection.

[0052] S4: Construct a deep neural network recognition model based on the fused feature vector, train and optimize the parameters of the defect samples through supervised learning algorithm, and output the recognition results and confidence scores of different defect types by combining the robustness verification of industrial working conditions.

[0053] This embodiment requires specific explanation of the process of constructing a deep neural network recognition model as follows:

[0054] The structural design of the deep neural network recognition model needs to match the dimension of the fused feature vector and the defect classification requirements (including early defects and composite defects). A fully connected deep structure of "input layer - hidden layer - dropout layer - output layer" is adopted. The input layer dimension is consistent with the fused feature vector dimension (the fused feature vector dimension is 100, so the number of neurons in the input layer is set to 100), responsible for receiving the fused feature vector and passing it to the hidden layer. The hidden layer has three fully connected layers, extracting high-order defect features through layer-by-layer nonlinear transformation: the first hidden layer contains 256 neurons (enhancing the ability to capture early features), and the activation function is LeakyReLU (formula: f(a) = max(0.01a, a), where a is the neuron input, and the negative slope of 0.01 is determined through cross-validation), specifically preserving early defects. The first hidden layer contains weak features (such as the voltage distortion feature of 0.5% in the early stage of winding insulation aging); the second hidden layer contains 128 neurons with the activation function ReLU (formula f(a)=max(0,a)) to integrate multi-source associated features (such as the coupling mode of "current spike-temperature slow rise"); the third hidden layer contains 64 neurons with the activation function still ReLU, further compressing the feature dimension and strengthening the composite fault mode (such as the mixed feature of "bearing wear + inter-turn short circuit"); to prevent overfitting, a Dropout layer (dropoutrate=0.3, randomly deactivating 30% of neurons) is added after the third hidden layer; the output layer contains 6 neurons (corresponding to 4 single defects + 1 composite defect + 1 normal state, a total of 6 classes), and the activation function is softmax (formula f(a)=max(0,a)). , where z i (The input to the i-th neuron in the output layer) outputs the probability value of each category (the sum of the probabilities is 1), which makes it easier to intuitively determine the defect type and distinguish between single and compound faults.

[0055] It should be further explained that the implementation of the supervised learning algorithm needs to be based on labeled fused feature samples, focusing on strengthening the learning of early defect samples. The prediction bias is calculated by improving the loss function and backpropagation is used to update the network parameters. First, the training data is prepared: the fused feature vector is divided into training set, validation set and test set in a ratio of 7:1:2. The sample labels are one-hot encoded (6 types of labels are: normal state [1,0,0,0,0,0], winding short circuit [0,1,0,0,0,0], rotor eccentricity [0,0,1,0,0,0], bearing wear [0,0,0,1,0,0], partial discharge [0,0,0,0,1,0], and composite defect [0,0,0,0,0,1]). For early defect samples (such as bearing wear ≤0.1mm, winding insulation aging degree ≤10%), a data augmentation strategy of "time flip + Gaussian noise addition (signal-to-noise ratio 30dB)" is adopted to augment the original 500 early defect samples. The sample size was expanded to 1500 sets, addressing the scarcity of early defect samples in industrial scenarios. This data expansion strategy ultimately increased the original 500 sets of early samples to 1500 sets, an expansion of 200%, and more than 98% of the newly added samples retained the weak features of early defects (verified through frequency domain analysis, the amplitude deviation of the defect feature frequency components (such as the second harmonic component corresponding to bearing wear) of the expanded samples was ≤8%). Compared with existing technologies, this not only solves the problem of "scarcity of samples" but also avoids the drawback of "distortion of expanded sample features"—existing technologies often have a cross-validation accuracy of ≤75% for early defects due to insufficient samples or feature distortion. However, this invention, through this expansion strategy, can improve the cross-validation accuracy of the model on early defect samples to over 88%, providing key sample support for subsequent deep neural network recognition models to achieve an early defect recognition rate of ≥90%, truly meeting the actual needs of difficult collection and limited types of early defect samples in industrial scenarios.

[0056] The loss function uses cross-entropy loss with an early defect weight factor (formula is...). ), where b is the batch size (set to 32 to balance training efficiency and parameter update stability), y k,i This represents the true label (0 or 1) of the k-th sample. Let w be the probability predicted by the model that the k-th sample belongs to the i-th class. i Class weights (early defect and compound defect samples w) i =2.0, normal / critical defect sample w i =1.0), by weighting the model to focus on key defects; the optimizer is Adam (adaptive moment estimation), the initial learning rate is set to 0.001, and the learning rate is dynamically adjusted (the learning rate is halved when the validation set loss does not decrease for 5 consecutive rounds), and the momentum parameters β1=0.9 and β2=0.999 to ensure fast convergence and stability.

[0057] The model training and parameter optimization process requires multiple iterations to improve network performance. Regularization strategies are introduced to prevent overfitting and enhance generalization ability under varying conditions. The training iteration count is set to 200 rounds (maximum iteration limit). Each iteration performs batch gradient descent on the training set: the training set is divided into several batches of 32, and the loss value is calculated after each batch is input into the network. The weights and biases (e.g., the weights w of the l-th layer) are updated using the backpropagation algorithm. l The update formula is Where η is the current learning rate, For the loss function on w l The gradient); three strategies are used in parameter optimization: L2 regularization (formula is...). , λ=0.0001), l represents the hidden layer index (l=1,2,3 corresponds to 3 hidden layers). This represents the connection weight between the i-th neuron in layer l and the j-th neuron in the previous layer. This regularization term penalizes excessively large weight values, constraining network complexity to avoid overfitting. Early stopping (training stops when the validation set accuracy fails to improve for 10 consecutive rounds and the loss increases) preserves the optimal parameters. 5% of interference samples (such as non-defect signals caused by power grid harmonics and load fluctuations) are mixed into the training set to enhance the model's anti-interference ability. Key indicators are monitored in real time during training: training set accuracy ≥ 95%, validation set accuracy ≥ 92% (early defect validation set accuracy ≥ 90%), ensuring that the model learns fully and has generalization ability.

[0058] After model training, performance is verified through multi-dimensional testing, and recognition results are output. The recognition results must include defect type and confidence level. The fused feature vectors of the test set are input into the optimized model to verify the reliability of early defect recognition and recognition under varying operating conditions. The test set includes four categories: basic samples (normal + severe defects), early defect samples (expanded), composite defect samples, and varying operating condition samples (30%-120% rated load, 5%-10% grid harmonics). This is significantly different from existing technologies that only use basic samples (normal + severe defects) under stable laboratory conditions for testing. The test set of this invention has a wider coverage that better matches actual industrial scenarios, effectively verifying the model's performance in areas where existing technologies are weak, such as "early defect concealment identification," "distinguishing between cross-features of complex defects," and "resistance to varying operating conditions." This avoids misjudgments of model generalization ability due to a single test sample. The optimized model outputs probability values ​​for six categories, and the category with the highest probability is taken as the recognition result (e.g., if a sample outputs probabilities of [0.02, 0.89, 0.03, 0.05, 0.01], the recognition result is "early rotor eccentricity"). Simultaneously, the probability value of that category is also output as... The confidence level is 87% (for this case). Model performance is evaluated using five dimensions: overall accuracy ≥ 92%; early defect identification rate ≥ 90%; composite defect accuracy ≥ 88%; accuracy at 120% load ≥ 92%; false positive rate ≤ 2% under 5% harmonic interference. This is compared to the typical performance of traditional machine learning models (such as support vector machines and decision trees)—overall accuracy 78%-85%, early defect identification rate 60%-70%, composite defect accuracy 70%-75%, accuracy at 120% load ≤ 75%, and false positive rate 15% under 5% harmonic interference. The invention achieves significant breakthroughs in all indicators, with a 20% reduction in early defect recognition rate, an increase of over 17 percentage points in accuracy at 120% load, and a 13-18 percentage point reduction in false positive rate under 5% harmonic interference. This effectively solves the problems of insufficient early weak defect capture and poor adaptability to complex working conditions in existing technologies. The final output recognition results are presented in a structured form (e.g., "Defect type: early bearing wear, confidence level: 94%; working condition adaptability: applicable to 75% rated load"), which can be directly used for the detection and diagnosis of internal defects in motors.

[0059] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0060] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting internal defects in motors based on multi-source data fusion, characterized in that, include: S1: Collect multi-source raw data during motor operation to generate a multi-source raw dataset; the multi-source raw data includes stator current signal, stator voltage signal and temperature signal; S2: Perform data preprocessing on the multi-source raw dataset, use wavelet thresholding algorithm to denoise the current signal and voltage signal, use moving average algorithm to perform steady-state smoothing on the temperature signal, use linear normalization algorithm to unify the amplitude of each signal, and use time interpolation-based synchronization algorithm to time-align data with different sampling frequencies to obtain a preprocessed signal set; S3: Based on the preprocessed signal set, a multi-source data fusion model is established. The fusion model includes a feature extraction layer, a feature association layer, and a fusion reconstruction layer. The feature extraction layer extracts the time-domain and frequency-domain features of the current signal, voltage signal, and temperature signal through convolution calculation. The feature association layer establishes a mapping relationship between signals based on the feature correlation calculation method; the fusion reconstruction layer uses the feature space reconstruction algorithm to nonlinearly combine the features of different signals to generate a fused feature vector. S4: Construct a deep neural network recognition model based on the fused feature vector, train and optimize the parameters of the defect samples through a supervised learning algorithm, and output the defect type recognition result.

2. The method for detecting internal defects in motors based on multi-source data fusion according to claim 1, characterized in that: The stator current signal is used to reflect the electromagnetic operating state of the motor, the stator voltage signal is used to reflect power supply fluctuations and coil excitation characteristics, and the temperature signal is used to reflect thermal imbalance and insulation aging trends. The acquired signals are synchronously calibrated in the form of timestamps to generate a multi-source raw dataset.

3. The method for detecting internal defects in motors based on multi-source data fusion according to claim 1, characterized in that: The wavelet thresholding algorithm includes performing multi-level wavelet decomposition on the current signal and voltage signal to separate high-frequency detail coefficients and low-frequency approximation coefficients; during the thresholding process, a threshold range is set according to the frequency band energy distribution of the signal, and coefficients of high-frequency detail coefficients are filtered out and amplitudes are corrected. Then, wavelet reconstruction is performed to obtain the denoised current signal and voltage signal.

4. The method for detecting internal defects in motors based on multi-source data fusion according to claim 1, characterized in that: The moving average algorithm includes determining the length of the sliding window based on the sampling frequency of the temperature signal, performing mean calculation on the temperature sampling points within the window, sliding the window sequentially and updating the sequence values ​​to obtain a smoothed temperature signal.

5. The method for detecting internal defects in motors based on multi-source data fusion according to claim 1, characterized in that: The feature extraction layer includes a temporal convolution structure and a frequency convolution structure. The temporal convolution structure performs stepwise convolution calculations on current signals, voltage signals, and temperature signals to extract local variation features and periodic patterns. The frequency convolution structure performs multi-layer convolution operations on the amplitude sequence of the signal after spectral transformation to extract feature information of frequency distribution and energy concentration intervals, forming a feature set containing multi-source features.

6. The method for detecting internal defects in motors based on multi-source data fusion according to claim 1, characterized in that: The feature association layer includes feature correlation calculation and matrix construction steps. By performing linear and nonlinear correlation analysis on current features, voltage features and temperature features, the degree of dependence between features is calculated. Feature pairs are filtered and combined to construct a signal feature correlation matrix.

7. The method for detecting internal defects in motors based on multi-source data fusion according to claim 1, characterized in that: In the fusion and reconstruction layer, the nonlinear mapping algorithm maps features to a high-dimensional space through an adapted kernel function, selects principal components and retains fault features, and dynamically assigns weights based on feature and defect labels, ultimately generating a fusion feature vector.

8. The method for detecting internal defects in motors based on multi-source data fusion according to claim 1, characterized in that: In the supervised learning of the deep neural network recognition model, the sample labels adopt an encoding method adapted to multi-class recognition. The performance indicators of the training set and the validation set are monitored in real time during the training process. The learning rate is adjusted according to the performance of the validation set. Training stops when the performance of the validation set meets the preset requirements. The model output results include the defect type and the corresponding confidence level.

Images