Micro motor winding defect online detection method based on pulsed eddy current and deep learning

Abstract

The application discloses a micro motor winding defect online detection method based on pulse eddy current and deep learning, relates to the field of nondestructive testing technology, and comprises the following steps: collecting original pulse eddy current detection data containing time domain signals and frequency domain features through a multi-channel eddy current sensor array; performing data denoising and feature enhancement on the data by using a Maxwell equation constraint network; generating a defect feature candidate set in combination with a multi-scale defect generative adversarial network to obtain target eddy current data; applying a chaotic frequency modulation pulse sequence to the target eddy current data as an excitation signal; collecting a response signal by using a phase synchronization lock-in technology to obtain enhanced eddy current data; extracting corresponding electromagnetic feature vectors containing physical features and deep learning features; comparing the electromagnetic feature vectors with a preset defect feature library; determining a defect detection result in combination with a dynamic excitation response difference; and determining sensor detection weights for data fusion, constructing a dynamically updated defect feature library, and jointly fine-tuning a network.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of nondestructive testing, more particularly, to a micro motor winding defect online detection method based on pulsed eddy current and deep learning. BACKGROUND

[0002] Micro motors are widely used in many fields due to their small size and low power, and the detection of their winding defects is crucial. However, existing detection methods have many drawbacks and cannot meet the actual needs. With the development of science and technology, people have begun to explore more advanced detection methods. Pulsed eddy current detection technology brings new means for motor winding defect detection due to its unique principle. Deep learning technology has shown great ability in image recognition and signal processing. Combining pulsed eddy current technology with deep learning for micro motor winding defect online detection has become a current research hotspot. However, this technology still faces many challenges in practical application. Raw pulsed eddy current detection data is greatly affected by noise, and how to effectively denoise and enhance useful features is a major problem. The extraction and recognition of defect features are not accurate enough to fully and accurately reflect the true situation of winding defects. In addition, the adaptability of the detection model is poor, and the detection effect will decrease significantly when facing different motor models and operating conditions. Moreover, the reliability of the detection results needs to be improved, and false positives and false negatives may occur.

[0003] Existing micro motor winding defect detection methods have problems such as inaccurate detection, strong noise interference, and difficult feature extraction. SUMMARY

[0004] To overcome the problems of inaccurate detection, strong noise interference, and difficult feature extraction in existing micro motor winding defect detection methods, the present application discloses a micro motor winding defect online detection method based on pulsed eddy current and deep learning, which can effectively solve the above technical problems.

[0005] To solve the above technical problems, the technical solution of the present application is as follows:

[0006] The micro motor winding defect online detection method based on pulsed eddy current and deep learning includes the following steps:

[0007] Obtain raw pulsed eddy current detection data of the micro motor during operation. The raw pulsed eddy current detection data is collected by a multi-channel eddy current sensor array and contains multi-dimensional data including time domain signals and frequency domain features.

[0008] Based on the pre-set Maxwell equation constraint network, denoise and feature enhancement are performed on the raw pulsed eddy current detection data. A multi-scale defect generative adversarial network is used to generate a defect feature candidate set, and target eddy current data is obtained.

[0009] The chaotic frequency modulation pulse sequence is applied to the target eddy current data as an excitation signal, a phase synchronization frequency locking technology is used to collect a response signal, and enhanced eddy current data containing dynamic characteristics are obtained;

[0010] An electromagnetic feature vector corresponding to the enhanced eddy current data is extracted, and the electromagnetic feature vector includes physical features based on Maxwell equation constraints and multi-dimensional features extracted by deep learning;

[0011] The electromagnetic feature vector is compared with a preset defect feature library, and a dynamic excitation response difference is combined to determine a defect detection result of the motor winding.

[0012] Preferably, the Maxwell equation constraint network based on the preset is used to denoise and enhance the original pulse eddy current detection data, and a multi-scale defect generative adversarial network is used to generate a defect feature candidate set, and the target eddy current data specifically includes:

[0013] The original pulse eddy current detection data is input into the Maxwell equation constraint network, the network output is optimized through the loss function embedded with electromagnetic equations, and data denoising and physical feature extraction are realized;

[0014] The denoised data is input into the multi-scale defect generative adversarial network to generate a defect feature candidate set under different scales;

[0015] Based on a preset feature screening rule, the target eddy current data is screened from the candidate set.

[0016] Preferably, the chaotic frequency modulation pulse sequence is applied to the target eddy current data as an excitation signal, a phase synchronization frequency locking technology is used to collect a response signal, and enhanced eddy current data containing dynamic characteristics are obtained specifically includes:

[0017] A chaotic frequency modulation pulse sequence with ergodicity and non-periodicity is generated, and the frequency, amplitude and pulse width dynamically change with the chaotic mapping;

[0018] The chaotic frequency modulation pulse sequence is loaded to the excitation end of the eddy current sensor, and the phase relationship between the excitation and response signals is locked by using the phase synchronization frequency locking technology;

[0019] The eddy current response signal after excitation is collected and recorded, and enhanced eddy current data containing dynamic characteristics are constructed.

[0020] Preferably, the electromagnetic feature vector corresponding to the enhanced eddy current data is extracted, and the electromagnetic feature vector includes physical features based on Maxwell equation constraints and multi-dimensional features extracted by deep learning specifically includes:

[0021] The magnetic field strength and eddy current density physical features of the enhanced eddy current data are calculated through the Maxwell equation constraint network;

[0022] The enhanced eddy current data is input into a convolutional neural network to extract multi-dimensional deep learning features of time domain waveforms and frequency domain spectral lines;

[0023] The physical features and deep learning features are fused to form an electromagnetic feature vector.

[0024] Preferably, the comparison of the electromagnetic feature vector with the preset defect feature library, combined with the dynamic excitation response difference, determines the defect detection result of the micro motor winding, specifically comprising:

[0025] The cosine similarity of the electromagnetic feature vector and the standard feature in the defect feature library is calculated to obtain a static matching score;

[0026] The response signal difference under chaotic dynamic excitation is analyzed, and a dynamic feature deviation degree is calculated;

[0027] Based on the weighted sum of the static matching score and the dynamic feature deviation degree, compared with a preset threshold, it is determined whether the micro motor winding has defects and the defect type.

[0028] Preferably, the multi-channel eddy current sensor array is arranged in the axial and circumferential directions of the micro motor winding, and the detection areas of each sensor partially overlap, and the method further comprises:

[0029] The detection weight of each eddy current sensor is determined, and the detection weight is calculated based on the electromagnetic coupling degree of the sensor position and the winding structure;

[0030] The data collected by each sensor is weighted and fused to generate global eddy current detection data.

[0031] Preferably, the method further comprises:

[0032] Based on the historical detection data and the defect type, a dynamically updated defect feature library is constructed, and the defect feature library is optimized in real time through an incremental learning algorithm;

[0033] After completing N detections, the Maxwell equation constraint network and the multi-scale defect generative adversarial network are jointly fine-tuned using new detection data.

[0034] Preferably, an electronic device comprises a memory and at least one processor, the memory stores instructions, and at least one processor invokes the instructions in the memory to make the device execute each step of the detection method as described above.

[0035] Preferably, a computer readable storage medium stores instructions, and the instructions are executed by a processor to implement each step of the detection method as described above.

[0036] Preferably, a computer program product comprising instructions which, when executed, cause the steps of the detection method as described above to be performed.

[0037] Compared with the prior art, the present application has the beneficial effects that: the prior method is difficult to accurately detect the winding defects of the micro motor, while the present application combines the pulse eddy current technology with deep learning to more accurately identify defects, the multi-channel eddy current sensor array collects multi-dimensional data containing time domain and frequency domain information, after denoising and feature enhancement by the Maxwell equation constraint network, the noise interference is effectively removed and the useful features are highlighted; the multi-scale defect generative adversarial network generates different scale defect feature candidate sets, filters out the target eddy current data, and makes the defect features more comprehensive and accurate, thereby improving the detection accuracy; the original pulse eddy current detection data is easily affected by noise, the present application optimizes the network output by embedding the loss function of electromagnetism equation, realizes data denoising and physical feature extraction, can effectively suppress noise interference, ensure data quality and feature extraction accuracy, and ensure the stability and reliability of the subsequent detection process; on the one hand, the Maxwell equation constraint network is used to calculate the physical features such as magnetic field strength and eddy current density, based on the electromagnetism theory, which can accurately reflect the physical nature of the winding defects; on the other hand, the data is input into the convolutional neural network to extract multi-dimensional deep learning features of time domain waveform and frequency domain spectrum, mine the potential law of the data, and fuse the electromagnetic feature vectors formed by the physical features and deep learning features to more comprehensively represent the winding defect characteristics, thereby providing a basis for accurately detecting defects; the present application considers the detection weights of different sensors, determines the weights according to the electromagnetic coupling degree of the sensor position and the winding structure, and performs data fusion to generate global eddy current detection data, so that the detection result is more representative, in addition, a dynamically updated defect feature library is constructed, which is optimized in real time through an incremental learning algorithm, adapts to different motor models and operating conditions, and maintains good detection performance; when comparing the electromagnetic feature vector with the defect feature library, not only the static matching score is calculated, but also the dynamic feature deviation is analyzed, and the defects are judged by combining the two weighted sums, the static features and dynamic response differences are comprehensively considered, the misjudgment and omission rates are reduced, and the reliability of the detection result is ensured; the phase synchronous lock-in technology is used to collect the response signal, the signal collection efficiency and accuracy are improved, the detection process is ensured to be fast, and the entire detection process has high automation degree, which is realized through electronic equipment, computer readable storage medium and computer program product, improves the detection efficiency, and meets the rapid detection demand in actual application. BRIEF DESCRIPTION OF DRAWINGS

[0038] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings needed to be used in the following embodiment or prior art description will be briefly introduced. Obviously, the drawings in the following description are only exemplary, and those skilled in the art can also obtain other drawings from the provided drawings without creating any inventive labor.

[0039] Figure 1 The step diagram of the micro motor winding defect online detection method based on pulsed eddy current and deep learning. DETAILED DESCRIPTION

[0040] The accompanying drawings are only used for illustrative description and cannot be understood as a limitation of the patent;

[0041] In order to better illustrate the embodiment, some components in the drawings may be omitted, enlarged or reduced, and do not represent the actual size of the product;

[0042] It can be understood by those skilled in the art that some well-known structures and their descriptions in the drawings may be omitted.

[0043] The technical solutions of the present application will be further described below in combination with the drawings and embodiments.

[0044] EMBODIMENT

[0045] The micro motor winding defect online detection method based on pulsed eddy current and deep learning comprises the following steps:

[0046] Obtain the original pulsed eddy current detection data of the micro motor in operation, the original pulsed eddy current detection data is collected by a multi-channel eddy current sensor array, and the multi-dimensional data containing time domain signals and frequency domain features;

[0047] Based on the preset Maxwell equation constraint network, the original pulsed eddy current detection data is denoised and feature enhanced, a defect feature candidate set is generated by combining a multi-scale defect generative adversarial network, and target eddy current data is obtained;

[0048] A chaotic frequency modulation pulse sequence is applied to the target eddy current data as an excitation signal, a phase synchronization lock-in technology is used to collect a response signal, and enhanced eddy current data containing dynamic features are obtained;

[0049] An electromagnetic feature vector corresponding to the enhanced eddy current data is extracted, the electromagnetic feature vector includes physical features based on the Maxwell equation constraint and multi-dimensional features extracted by deep learning;

[0050] The electromagnetic feature vector is compared with a preset defect feature library, and a dynamic excitation response difference is combined to determine the defect detection result of the micro motor winding.

[0051] The preset Maxwell equation constraint network is used to denoise and enhance the features of the original pulsed eddy current detection data, a defect feature candidate set is generated by combining a multi-scale defect generative adversarial network, and target eddy current data is obtained, which specifically comprises:

[0052] The original pulsed eddy current detection data is input into the Maxwell equation constraint network, and the network output is optimized by embedding a loss function of the electromagnetic equation to achieve data denoising and physical feature extraction.

[0053] The denoised data is input into a multi-scale defect generative adversarial network to generate a candidate set of defect features at different scales.

[0054] Based on preset feature selection rules, target eddy current data are selected from the candidate set.

[0055] The step of applying a chaotic frequency-modulated pulse sequence as an excitation signal to the target eddy current data and acquiring the response signal using phase synchronization frequency locking technology to obtain enhanced eddy current data containing dynamic features specifically includes:

[0056] Generate a chaotic frequency-modulated pulse sequence with ergodicity and aperiodicity, whose frequency, amplitude and pulse width change dynamically with the chaotic mapping;

[0057] A chaotic frequency-modulated pulse sequence is applied to the excitation end of the eddy current sensor, while the phase relationship between the excitation and response signals is locked using phase synchronization frequency locking technology.

[0058] Collect and record the eddy current response signal after excitation to construct enhanced eddy current data containing dynamic features.

[0059] The extraction of the electromagnetic feature vector corresponding to the enhanced eddy current data, wherein the electromagnetic feature vector includes physical features based on Maxwell's equation constraints and multidimensional features extracted by deep learning, specifically includes:

[0060] Physical characteristics of magnetic field strength and eddy current density in enhanced eddy current data were calculated using Maxwell equation-constrained networks.

[0061] The enhanced eddy current data is input into a convolutional neural network to extract multi-dimensional deep learning features of time-domain waveforms and frequency-domain spectral lines.

[0062] By integrating physical features with deep learning features, an electromagnetic feature vector is formed.

[0063] The step of comparing the electromagnetic feature vector with a preset defect feature library and determining the defect detection result of the micro-motor winding by combining the dynamic excitation response difference specifically includes:

[0064] Calculate the cosine similarity between the electromagnetic feature vector and the standard features in the defect feature library to obtain the static matching score;

[0065] Analyze the differences in response signals under chaotic dynamic excitation and calculate the dynamic characteristic deviation.

[0066] Based on the weighted sum of static matching score and dynamic feature deviation, and compared with a preset threshold, it is determined whether there are defects in the winding of the micro-motor and the type of defects.

[0067] The multi-channel eddy current sensor array is arranged along the axial and circumferential directions of the micro-motor winding, and the detection areas of each sensor partially overlap. The method further includes:

[0068] The detection weight of each eddy current sensor is determined, and the detection weight is calculated based on the electromagnetic coupling degree between the sensor position and the winding structure;

[0069] The data collected by each sensor are weighted and fused to generate global eddy current detection data.

[0070] The method further includes:

[0071] A dynamically updated defect feature library is constructed based on historical detection data and defect types, and the defect feature library is optimized in real time through an incremental learning algorithm.

[0072] After each N tests, the Maxwell equation constraint network and the multi-scale defect generative adversarial network are jointly fine-tuned using the new test data.

[0073] An electronic device includes: a memory and at least one processor, the memory storing instructions, wherein at least one processor invokes the instructions in the memory to cause the device to perform the steps of the detection method as described above.

[0074] A computer-readable storage medium storing instructions that, when executed by a processor, implement the steps of the detection method described above.

[0075] A computer program product comprising instructions that, when executed, cause the steps of the detection method described above to be performed.

[0076] For specific implementation details, please refer to [link / reference]. Figure 1 On the experimental platform for the operation of the micro motor, various testing equipment and devices are set up to ensure that the micro motor is in normal operating condition and its speed is stable at the set value. A multi-channel eddy current sensor array is arranged in the axial and circumferential directions of the micro motor winding. The detection areas of each sensor partially overlap to ensure comprehensive detection of the winding. For example, an 8-channel eddy current sensor array is selected, with each sensor evenly distributed at different positions in the axial and circumferential directions of the winding.

[0077] The aforementioned multi-channel eddy current sensor array is used to collect raw pulse eddy current detection data during the operation of the micro-motor in real time. This data includes multi-dimensional data of time-domain signals and frequency-domain characteristics. The sampling rate of the time-domain signal is 1MHz, and the frequency range of the frequency-domain characteristics is 1kHz to 100kHz. The collected raw data is initially stored, and the data storage format can adopt common binary format or ASCII code format, etc.

[0078] The raw pulsed eddy current detection data is input into a constrained network based on the preset Maxwell equations. The network can be structured using a CNN (convolutional neural network) architecture, embedding loss functions of electromagnetic equations, such as the mean square error loss function combined with Maxwell equation terms.

[0079] The network is trained and optimization algorithms, such as the Adam optimizer, are used to continuously adjust the network parameters so that the network output can achieve data denoising and physical feature extraction, resulting in denoised data. For example, the number of training iterations is set to 1000 and the learning rate is 0.001.

[0080] The denoised data is input into a multi-scale defect generative adversarial network (GAN), which includes a generator and a discriminator. The generator is used to generate a candidate set of defect features at different scales, and the discriminator is used to distinguish between real defect features and generated defect features.

[0081] By training the GAN, the candidate set of defect features generated by the generator can be made to simulate winding defect features of different scales as realistically as possible, such as open circuits, short circuits, and inter-turn short circuits, as well as defect features of different sizes, shapes and locations. During the training process, the network structure of both the generator and the discriminator can be set to a multilayer perceptron (MLP) structure, and the number of training iterations is about 2000.

[0082] According to the preset feature selection rules, target eddy current data is selected from the generated defect feature candidate set. The feature selection rules can be based on indicators such as the significance, stability, and correlation of defect features. For example, the significance threshold can be set to 0.8, the stability threshold to 0.9, and the correlation threshold to 0.7. The defect feature candidate set that meets these threshold conditions is used as the target eddy current data.

[0083] A chaotic frequency-modulated pulse sequence with ergodicity and aperiodicity is generated. Its frequency, amplitude and pulse width change dynamically with the chaotic mapping. For example, the chaotic sequence is generated by using the Logistic chaotic mapping and mapped to the frequency, amplitude and pulse width parameters of the pulse sequence. The frequency range is 5kHz to 50kHz, the amplitude range is 1V to 5V and the pulse width range is 50ns to 500ns.

[0084] A chaotic frequency-modulated pulse sequence is applied to the excitation end of the eddy current sensor. Simultaneously, phase synchronization frequency locking technology is used to lock the phase relationship between the excitation and response signals. Through phase synchronization frequency locking technology, it can be ensured that the acquired response signal and the excitation signal have the same phase, thereby improving the accuracy and stability of signal acquisition. The eddy current response signal after excitation is acquired and recorded to construct enhanced eddy current data containing dynamic characteristics. The acquisition time can be set from 10ms to 100ms, and the sampling rate of the acquired data is the same as that of the original data.

[0085] The physical characteristics of magnetic field strength and eddy current density in enhanced eddy current data are calculated using a constrained network based on Maxwell's equations. By utilizing Ampere's circuital law and Faraday's law of electromagnetic induction from Maxwell's equations, the enhanced eddy current data are analyzed and calculated to obtain the physical characteristic values ​​of magnetic field strength and eddy current density.

[0086] Enhanced eddy current data is input into a convolutional neural network to extract multi-dimensional deep learning features of time-domain waveforms and frequency-domain spectra. For example, the structure of a convolutional neural network may include an input layer, several convolutional layers, pooling layers, fully connected layers, and an output layer. The trained network model extracts features from the enhanced eddy current data to obtain time-domain waveform features, such as waveform amplitude, width, rise time, and fall time, and frequency-domain spectral features, such as frequency components, amplitude spectrum, and phase spectrum.

[0087] By integrating physical features and deep learning features, an electromagnetic feature vector is formed. The physical features such as magnetic field strength and eddy current density obtained from the above calculations, as well as the deep learning features such as time-domain waveforms and frequency-domain spectra, are combined to form a complete electromagnetic feature vector, which is used for subsequent defect detection and comparison.

[0088] The extracted electromagnetic feature vectors are compared with a preset defect feature library, which can be constructed based on historical detection data and defect types. The library contains electromagnetic feature vector samples of different types and degrees of winding defects.

[0089] The cosine similarity between the electromagnetic feature vector and the standard feature in the defect feature library is calculated to obtain the static matching score. The formula for calculating the cosine similarity is: cosθ=(A·B) / (|A|·|B|), where A and B are the electromagnetic feature vector and the standard feature vector, respectively, and θ is the angle between them. The static matching score ranges from -1 to 1, and the larger the value, the higher the similarity.

[0090] The differences in response signals under chaotic dynamic excitation are analyzed, and the dynamic characteristic deviation is calculated. By comparing the differences in dynamic characteristics, such as frequency response characteristics, amplitude change characteristics, and phase change characteristics, between the current response signal and the response signal under normal conditions, the dynamic characteristic deviation is calculated. This deviation can reflect the degree of change in the dynamic characteristics of winding defects.

[0091] The presence and type of defects in the micro-motor winding are determined by comparing the weighted sum of the static matching score and the dynamic feature deviation with a preset threshold. For example, if the weight of the static matching score is set to 0.6, the weight of the dynamic feature deviation to 0.4, and the preset threshold to 0.7, a defect is determined if the weighted sum is greater than or equal to 0.7. The defect type is then determined based on the comparison results with different defect type features in the defect feature library. If the weighted sum is less than 0.7, the winding is determined to be normal.

[0092] The test results (whether defects exist and the type of defects) are displayed on the human-machine interface of the electronic device so that operators can understand the operating status of the micro-motor windings in a timely manner. At the same time, the test results are stored in the memory, and the storage format can be consistent with the original data storage format.

[0093] The above detection method can be implemented by a specially designed electronic device, which includes a memory and at least one processor. The memory stores instructions, and the processor calls the instructions in the memory to execute the various steps of the above detection method.

[0094] Alternatively, the detection method can be programmed into a computer program and stored on a computer-readable storage medium, such as a hard disk, optical disk, or USB flash drive. When the program is run, it causes the computer to execute the steps of the aforementioned detection method.

[0095] The same or similar labels correspond to the same or similar parts;

[0096] The terms used to describe positional relationships in the accompanying drawings are for illustrative purposes only and should not be construed as limiting this patent.

[0097] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. For those skilled in the art, other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all implementation methods here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. An online detection method for winding defects in micromotors based on pulsed eddy currents and deep learning, characterized in that, Includes the following steps: The raw pulsed eddy current detection data of the micro motor during operation is acquired. The raw pulsed eddy current detection data is collected by a multi-channel eddy current sensor array and contains multi-dimensional data including time domain signals and frequency domain characteristics. The original pulsed eddy current detection data is denoised and feature-enhanced based on a pre-defined Maxwell equation constraint network. A multi-scale defect generative adversarial network is then used to generate a defect feature candidate set to obtain the target eddy current data. A chaotic frequency-modulated pulse sequence is applied to the target eddy current data as an excitation signal, and the response signal is acquired using phase synchronization frequency locking technology to obtain enhanced eddy current data containing dynamic characteristics. Extract the electromagnetic feature vector corresponding to the enhanced eddy current data. The electromagnetic feature vector includes physical features based on Maxwell equation constraints and multidimensional features extracted by deep learning. The electromagnetic feature vector is compared with a preset defect feature library, and the defect detection result of the micro-motor winding is determined by combining the difference in dynamic excitation response.

2. The detection method according to claim 1, characterized in that, The preset Maxwell equation-constrained network denoises and enhances the features of the original pulsed eddy current detection data, and combines it with a multi-scale defect generative adversarial network to generate a defect feature candidate set, resulting in target eddy current data, specifically including: The original pulsed eddy current detection data is input into the Maxwell equation constraint network, and the network output is optimized by embedding a loss function of the electromagnetic equation to achieve data denoising and physical feature extraction. The denoised data is input into a multi-scale defect generative adversarial network to generate a candidate set of defect features at different scales. Based on preset feature selection rules, target eddy current data are selected from the candidate set.

3. The detection method according to claim 1, characterized in that, The step of applying a chaotic frequency-modulated pulse sequence as an excitation signal to the target eddy current data and acquiring the response signal using phase synchronization frequency locking technology to obtain enhanced eddy current data containing dynamic features specifically includes: Generate a chaotic frequency-modulated pulse sequence with ergodicity and aperiodicity, whose frequency, amplitude and pulse width change dynamically with the chaotic mapping; A chaotic frequency-modulated pulse sequence is applied to the excitation end of the eddy current sensor, while the phase relationship between the excitation and response signals is locked using phase synchronization frequency locking technology. Collect and record the eddy current response signal after excitation to construct enhanced eddy current data containing dynamic features.

4. The detection method according to claim 1, characterized in that, The extraction of the electromagnetic feature vector corresponding to the enhanced eddy current data, wherein the electromagnetic feature vector includes physical features based on Maxwell's equation constraints and multidimensional features extracted by deep learning, specifically includes: Physical characteristics of magnetic field strength and eddy current density in enhanced eddy current data were calculated using Maxwell equation-constrained networks. The enhanced eddy current data is input into a convolutional neural network to extract multi-dimensional deep learning features of time-domain waveforms and frequency-domain spectral lines. By integrating physical features with deep learning features, an electromagnetic feature vector is formed.

5. The detection method according to claim 1, characterized in that, The step of comparing the electromagnetic feature vector with a preset defect feature library and determining the defect detection result of the micro-motor winding by combining the dynamic excitation response difference specifically includes: Calculate the cosine similarity between the electromagnetic feature vector and the standard features in the defect feature library to obtain the static matching score; Analyze the differences in response signals under chaotic dynamic excitation and calculate the dynamic characteristic deviation. Based on the weighted sum of static matching score and dynamic feature deviation, and compared with a preset threshold, it is determined whether there are defects in the winding of the micro-motor and the type of defects.

6. The detection method according to claim 1, characterized in that, The multi-channel eddy current sensor array is arranged along the axial and circumferential directions of the micro-motor winding, and the detection areas of each sensor partially overlap. The method further includes: The detection weight of each eddy current sensor is determined, and the detection weight is calculated based on the electromagnetic coupling degree between the sensor position and the winding structure; The data collected by each sensor are weighted and fused to generate global eddy current detection data.

7. The detection method according to claim 1, characterized in that, The method further includes: A dynamically updated defect feature library is constructed based on historical detection data and defect types, and the defect feature library is optimized in real time through an incremental learning algorithm. After each N tests, the Maxwell equation constraint network and the multi-scale defect generative adversarial network are jointly fine-tuned using the new test data.

8. An electronic device, characterized in that, include: The device includes a memory and at least one processor, the memory storing instructions, and at least one processor invoking the instructions in the memory to cause the device to perform the steps of the detection method as described in any one of claims 1-7.

9. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the various steps of the detection method as described in any one of claims 1-7.

10. A computer program product, characterized in that, The computer program product includes instructions that, when executed, cause the steps of the detection method as described in any one of claims 1-7 to be performed.

Images