Micro motor defect voiceprint detection method based on deep learning
Through the voiceprint detection method based on deep learning and quantum computing, combined with the causal graph separation algorithm, the problems of insufficient feature expression, cumbersome model optimization and difficulty in multi-dimensional voiceprint processing in micro motor detection are solved, and defect detection with high precision and high interpretability are achieved.
Patent Information
- Application Number
- CN202510568189.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-30
- Publication Date
- 2025-07-11
AI Technical Summary
The voiceprint detection method of micro motors in the prior art has problems such as insufficient feature expression, cumbersome model optimization, difficulty in multi-dimensional voiceprint processing and lack of causal modeling, resulting in insufficient detection accuracy and interpretability.
Using a deep learning-based method, combining quantum feature extraction and causal graph separation algorithms, voiceprint data is obtained through photonic crystal microphones, feature expression is enhanced by quantum computing, NAS optimizes deep learning models, and combining measurement weight data of multiple photonic crystal microphones to perform data statistics and causal modeling to realize the detection of micromotor defects.
It improves the accuracy and interpretability of micro motor defect detection, simplifies the model optimization process, and enhances the comprehensiveness of multi-dimensional soundprint processing and the accuracy of detection.
Smart Images
Figure CN120299478A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of electrode voiceprint detection, and more specifically, to a method for detecting micro-motor defect voiceprints based on deep learning. Background Art
[0002] Micro-motors are widely used in fields such as consumer electronics, industrial automation, and intelligent devices. Their performance and reliability are crucial. Voiceprint detection, as an effective means of motor defect detection, can achieve non-contact and real-time monitoring. Traditional voiceprint detection methods, such as those based on artificial feature extraction and simple machine learning models, have problems such as limited feature expression ability, insufficient model optimization, and weak ability to distinguish complex defect types. With the development of quantum computing and deep learning, combining quantum feature extraction with deep learning can enhance feature expression and improve detection accuracy. At the same time, the causal graph separation algorithm can be used to model the causal relationship between voiceprint features and defect types, improving detection accuracy and interpretability. Against this background, a method for detecting micro-motor defect voiceprints based on deep learning is proposed to improve detection efficiency and reliability and meet the requirements of precise detection in industrial production.
[0003] The prior art has problems such as insufficient feature expression, cumbersome model optimization, difficulty in processing multi-dimensional voiceprints, and lack of causal relationship modeling. Summary of the Invention
[0004] In order to overcome the problems in the prior art such as insufficient feature expression, cumbersome model optimization, difficulty in processing multi-dimensional voiceprints, and lack of causal relationship modeling, the present invention discloses a method for detecting micro-motor defect voiceprints based on deep learning, which can effectively solve the above technical problems.
[0005] To solve the above technical problems, the technical solution of the present invention is as follows:
[0006] A method for detecting micro-motor defect voiceprints based on deep learning includes the following steps:
[0007] Obtain first voiceprint data of the micro-motor within a first time period; the first voiceprint data is obtained based on a photonic crystal microphone arranged around the micro-motor; the first voiceprint data includes acoustic feature data in multiple dimensions;
[0008] Perform quantum feature extraction processing on the acoustic feature data in multiple dimensions in the first voiceprint data to obtain voiceprint reference data of the micro-motor within the first time period; the quantum feature extraction processing combines the principle of quantum computing to enhance the expression ability of features;
[0009] Obtain second voiceprint data of the micro-motor at the current detection moment; the current detection moment is a detection moment after the first time period, and the second voiceprint data includes acoustic feature data in multiple dimensions;
[0010] Perform data statistical processing on the acoustic feature data of the first voiceprint data and the second voiceprint data in each dimension to obtain the data statistical value corresponding to each dimension;
[0011] Based on the data statistical value corresponding to each dimension and the second voiceprint data, use a deep learning model optimized by NAS to determine the voiceprint feature data of the micro-motor;
[0012] Based on the voiceprint feature data and the voiceprint reference data, combine the causal graph separation algorithm to determine the defect state of the micro-motor at the current detection moment.
[0013] Preferably, obtaining the voiceprint reference data of the micro-motor in the first time period includes:
[0014] Use qubits to represent the acoustic feature data of each dimension, and evolve the quantum state through quantum gate operations to achieve quantum entanglement and superposition of features;
[0015] Measure the quantum state, map the quantum features to the classical space, and obtain the data feature value corresponding to each dimension;
[0016] Based on the data feature value corresponding to each dimension, calculate the voiceprint data weight of the corresponding dimension;
[0017] Determine the voiceprint data weight of each dimension as the voiceprint reference data of the micro-motor in the first time period.
[0018] Preferably, the data statistical value includes the data average value and data variance of each dimension; determining the voiceprint feature data of the micro-motor includes:
[0019] Based on the acoustic feature data of the second voiceprint data in each dimension and the data average value of the corresponding dimension, determine the voiceprint data difference corresponding to each dimension;
[0020] Use the voiceprint data difference of each dimension and the data variance of the corresponding dimension as inputs, and process them through a deep learning model optimized by NAS to obtain the voiceprint feature data corresponding to each dimension of the micro-motor.
[0021] Preferably, determining the defect state of the micro-motor at the current detection moment includes:
[0022] Calculate the product between the voiceprint feature data of each dimension and the voiceprint reference data of the corresponding dimension to obtain the weighted feature value corresponding to each dimension of the micro-motor;
[0023] Sum the weighted eigenvalue corresponding to each dimension to obtain the total voiceprint feature value of the micro-motor at the current detection moment;
[0024] Based on the total voiceprint feature value, combine with the causal graph separation algorithm to determine the defect state of the micro-motor at the current detection moment; the causal graph separation algorithm is based on the causal relationship graph to model and reason about the causal relationship between the voiceprint feature and the defect type;
[0025] Specifically: when the total voiceprint feature value meets the preset defect feature threshold condition, use the causal graph separation algorithm to analyze the voiceprint feature to determine the specific defect type;
[0026] When the total voiceprint feature value does not meet the preset defect feature threshold condition, it is determined that the micro-motor has no obvious defect at the current detection moment.
[0027] Preferably, the number of the photonic crystal microphones is multiple; the first voiceprint data includes multiple groups of data corresponding to each photonic crystal microphone; the second voiceprint data includes multiple groups of data corresponding to each photonic crystal microphone; the method includes:
[0028] Determine the measurement weight data of each photonic crystal microphone; the measurement weight data is determined based on the measurement position of the photonic crystal microphone around the micro-motor;
[0029] Calculate the total voiceprint feature value corresponding to each photonic crystal microphone;
[0030] Based on the measurement weight data of each photonic crystal microphone and the total voiceprint feature value corresponding to the photonic crystal microphone, determine the weighted total voiceprint feature value of each photonic crystal microphone;
[0031] Based on the weighted total voiceprint feature value of each photonic crystal microphone, combine with the causal graph separation algorithm to determine the defect state of the micro-motor.
[0032] Preferably, the first time period includes multiple detection time nodes; after obtaining the second voiceprint data of the micro-motor at the current detection moment, the method includes:
[0033] Based on the second voiceprint data corresponding to the current detection moment and the voiceprint data corresponding to multiple adjacent detection time nodes before the current detection moment, update the first voiceprint data to obtain the updated first voiceprint data; the number of detection time nodes corresponding to the updated first voiceprint data is the same as the number of detection time nodes included in the first time period.
[0034] Preferably, after obtaining the second voiceprint data of the micro motor at the current detection moment, the method includes:
[0035] Calculating the minimum distance corresponding to each dimension between the first voiceprint data and the second voiceprint data;
[0036] In the case where the minimum distance of at least one dimension is greater than a preset distance threshold, correcting the first voiceprint data.
[0037] Preferably, the correcting the first voiceprint data includes:
[0038] Calculating the difference between the acoustic feature data of the second voiceprint data in each dimension and the acoustic feature data of the corresponding dimension at the previous detection time node at the current detection moment, to obtain the data offset value corresponding to each dimension;
[0039] Based on the acoustic feature data of the first voiceprint data in each dimension and the data offset value corresponding to each dimension, correcting the first voiceprint data.
[0040] An electronic device, comprising: a memory and at least one processor, wherein instructions are stored in the memory, and at least one of the processors calls the instructions in the memory, so that the device executes each step of the detection method as described above.
[0041] A computer-readable storage medium, on which instructions are stored, and when the instructions are executed by a processor, each step of the detection method as described above is implemented.
[0042] Compared with the prior art, the beneficial effects of the present invention are as follows: Through quantum feature extraction and processing, the present invention uses quantum bits to represent acoustic feature data, realizes the entanglement and superposition of features through quantum gate operations, and then maps the quantum features to the classical space through quantum state measurement to obtain data eigenvalues. Combining with the weight of voiceprint data, voiceprint reference data is formed, enhancing the expression ability of features and solving the problem of insufficient feature extraction in the prior art; adopting a deep learning model optimized by NAS can automatically and quickly find the optimal model structure and parameters in the network structure, improve the performance of the detection model, simplify the optimization process, and solve the problem of cumbersome model optimization in the prior art; performing data statistical processing on the first voiceprint data and the second voiceprint data in each dimension to obtain data statistical values, and then determining the voiceprint feature data. At the same time, integrating the measurement weight data and the total voiceprint feature value of multiple photonic crystal microphones realizes the comprehensive processing of multi-dimensional voiceprint data, improves the accuracy and comprehensiveness of detection, and solves the problem of difficult multi-dimensional voiceprint processing in the prior art; based on the causal graph separation algorithm, modeling and reasoning the causal relationship between voiceprint features and defect types. When the total voiceprint feature value meets the preset conditions, analyzing and determining the specific defect type provides a basis for the detection result and solves the problem of missing causal relationship modeling in the prior art. Description of the Drawings
[0043] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only exemplary, and for those of ordinary skill in the art, without creative efforts, other implementation drawings can be obtained according to the provided drawings.
[0044] Figure 1 It is a flowchart of the method for detecting the defects of a micro motor based on deep learning. Detailed Embodiments
[0045] The drawings are only for illustrative purposes and cannot be construed as limitations on this patent;
[0046] To better illustrate this embodiment, some components in the drawings will be omitted, enlarged or reduced, which do not represent the actual size of the product;
[0047] For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.
[0048] The following will further illustrate the technical solutions of the present invention with reference to the drawings and embodiments.
[0049] Embodiment
[0050] A method for detecting the defective sound pattern of a micro-motor based on deep learning, comprising the following steps:
[0051] Obtain the first sound pattern data of the micro-motor within the first time period; the first sound pattern data is obtained based on a photonic crystal microphone arranged around the micro-motor; the first sound pattern data includes acoustic feature data in multiple dimensions;
[0052] Perform quantum feature extraction processing on the acoustic feature data in multiple dimensions in the first sound pattern data to obtain the sound pattern reference data of the micro-motor within the first time period; the quantum feature extraction processing combines the principles of quantum computing to enhance the expression ability of features;
[0053] Obtain the second sound pattern data of the micro-motor at the current detection moment; the current detection moment is the detection moment after the first time period, and the second sound pattern data includes acoustic feature data in multiple dimensions;
[0054] Based on the acoustic feature data of the first sound pattern data and the second sound pattern data in each dimension, perform data statistical processing to obtain the data statistical value corresponding to each dimension;
[0055] Based on the data statistical value corresponding to each dimension and the second sound pattern data, use a deep learning model optimized by NAS to determine the sound pattern feature data of the micro-motor;
[0056] Based on the sound pattern feature data and the sound pattern reference data, combine the causal graph separation algorithm to determine the defective state of the micro-motor at the current detection moment.
[0057] The obtaining of the sound pattern reference data of the micro-motor within the first time period includes:
[0058] Use qubits to represent the acoustic feature data in each dimension, and evolve the quantum state through quantum gate operations to achieve quantum entanglement and superposition of features;
[0059] Measure the quantum state, map the quantum features to the classical space, and obtain the data feature values corresponding to each dimension;
[0060] Based on the data feature values corresponding to each dimension, calculate the sound pattern data weights corresponding to the dimensions;
[0061] Determine the sound pattern data weights of each dimension as the sound pattern reference data of the micro-motor within the first time period.
[0062] The data statistical values include the data average value and data variance of each dimension; the determination of the sound pattern feature data of the micro-motor includes:
[0063] Determine the voiceprint data difference corresponding to each dimension based on the acoustic feature data of the second voiceprint data in each dimension and the data average value of the corresponding dimension;
[0064] Use the voiceprint data difference of each dimension and the data variance of the corresponding dimension as inputs, and process them through a deep learning model optimized by NAS to obtain the voiceprint feature data corresponding to the micro-motor in each dimension.
[0065] The determination of the defect state of the micro-motor at the current detection moment includes:
[0066] Calculate the product between the voiceprint feature data of each dimension and the voiceprint reference data of the corresponding dimension to obtain the weighted eigenvalue corresponding to the micro-motor in each dimension;
[0067] Sum the weighted eigenvalues corresponding to each dimension to obtain the total voiceprint feature value of the micro-motor at the current detection moment;
[0068] Based on the total voiceprint feature value, combined with the causal graph separation algorithm, determine the defect state of the micro-motor at the current detection moment; the causal graph separation algorithm is based on a causal relationship graph to model and reason about the causal relationship between voiceprint features and defect types;
[0069] Specifically: when the total voiceprint feature value meets the preset defect feature threshold condition, use the causal graph separation algorithm to analyze the voiceprint features to determine the specific defect type;
[0070] When the total voiceprint feature value does not meet the preset defect feature threshold condition, determine that there is no obvious defect in the micro-motor at the current detection moment.
[0071] The number of the photonic crystal microphones is multiple; the first voiceprint data includes multiple groups of data corresponding to each photonic crystal microphone; the second voiceprint data includes multiple groups of data corresponding to each photonic crystal microphone; the method includes:
[0072] Determine the measurement weight data of each photonic crystal microphone; the measurement weight data is determined based on the measurement position of the photonic crystal microphone around the micro-motor;
[0073] Calculate the total voiceprint feature value corresponding to each photonic crystal microphone;
[0074] Based on the measurement weight data of each photonic crystal microphone and the total voiceprint feature value corresponding to the photonic crystal microphone, determine the weighted total voiceprint feature value of each photonic crystal microphone;
[0075] Based on the total weighted voiceprint feature value of each photonic crystal microphone, combined with the causal graph separation algorithm, determine the defect state of the micro-motor.
[0076] The first time period includes multiple detection time nodes; after obtaining the second voiceprint data of the micro-motor at the current detection moment, the method includes:
[0077] Based on the second voiceprint data corresponding to the current detection moment and the voiceprint data corresponding to multiple adjacent detection time nodes before the current detection moment, update the first voiceprint data to obtain updated first voiceprint data; the number of detection time nodes corresponding to the updated first voiceprint data is the same as the number of detection time nodes included in the first time period.
[0078] After obtaining the second voiceprint data of the micro-motor at the current detection moment, the method includes:
[0079] Calculate the minimum distance corresponding to each dimension between the first voiceprint data and the second voiceprint data.
[0080] In the case where the minimum distance of at least one dimension is greater than a preset distance threshold, correct the first voiceprint data.
[0081] The correcting the first voiceprint data includes:
[0082] Calculate the difference between the acoustic feature data of the second voiceprint data in each dimension and the acoustic feature data of the corresponding dimension at the previous detection time node of the current detection moment to obtain the data offset value corresponding to each dimension.
[0083] Based on the acoustic feature data of the first voiceprint data in each dimension and the data offset value corresponding to each dimension, correct the first voiceprint data.
[0084] An electronic device includes: a memory and at least one processor, wherein instructions are stored in the memory, and at least one of the processors invokes the instructions in the memory so that the device executes each step of the detection method as described above.
[0085] A computer-readable storage medium stores instructions thereon, and when the instructions are executed by a processor, each step of the detection method as described above is implemented.
[0086] In a specific implementation, please refer to Figure 1, a micro-motor was selected as the detection object and installed on a stable experimental bench. Five photonic crystal microphones were evenly arranged around the micro-motor. These microphones have high sensitivity and broadband response characteristics and can accurately capture the weak and complex acoustic fingerprint signals generated during the operation of the motor. The distance between each microphone and the surface of the motor housing was maintained at about 10 cm to ensure the effective pickup of the signals.
[0087] The photonic crystal microphones were connected to a high-precision data acquisition card through coaxial cables. The data acquisition card has multi-channel input functions. The sampling frequency was set to 48000 Hz and the resolution was 24 bits, which can digitize the acoustic fingerprint signals with high enough precision. At the same time, a high-performance computer was used as the data processing terminal, and deep learning frameworks such as TensorFlow or PyTorch and related signal processing software libraries were installed to provide the hardware and software environment for quantum feature extraction, deep learning model training, and the operation of the causal graph separation algorithm.
[0088] The micro-motor was started to run stably under rated load and speed. The first time period was set as the first 10 minutes after the motor started running. During this period, the acoustic fingerprint signals were collected in real time through the five photonic crystal microphones. Each microphone collected 48000 data points per second. After being converted by the data acquisition card, the first acoustic fingerprint data was obtained. These data contain acoustic feature information in multiple dimensions, such as frequency, amplitude, phase, etc. These original acoustic fingerprint data were stored in the database of the computer.
[0089] The acoustic feature data in each dimension, such as the frequency dimension and the amplitude dimension, were read from the first acoustic fingerprint data in the database. The data elements in each dimension were represented as qubits (quantum bits). Quantum gates such as Hadamard gates, Pauli-X gates, and Pauli-Z gates were used to operate on the qubits to achieve the evolution of the quantum state, generating the quantum entanglement and superposition effects of the features, so that the original classical acoustic feature data was extended and enhanced in the quantum space, thereby improving the expression ability of the features.
[0090] The quantum state after the operation of the quantum gates was measured to map the quantum features back to the classical space, and the corresponding data feature values in each dimension were obtained. For example, for the acoustic feature data in the frequency dimension, a value reflecting its quantum features was obtained after quantum processing. This value contains deeper feature information in the frequency dimension. Based on these data feature values, through a specific weight calculation algorithm, such as based on the magnitude and distribution of the feature values, the weights corresponding to the acoustic fingerprint data in each dimension were calculated. These weights were combined to form the acoustic fingerprint reference data of the micro-motor in the first time period.
[0091] During the continuous operation of the motor, every 5 minutes is taken as a detection period, and the current detection moment is the starting moment of each detection period. At the current detection moment, 5 photonic crystal microphones are used again to collect 10 seconds of voiceprint data as the second voiceprint data, which also contains acoustic feature data in multiple dimensions. After being converted by the data acquisition card, it is transmitted to the computer for processing.
[0092] Correspond the first voiceprint data and the second voiceprint data according to the dimensions. For each dimension, such as the frequency dimension, calculate the data average value and data variance in the first voiceprint data and the second voiceprint data to obtain the corresponding data statistical values for each dimension. For example, for the frequency dimension, calculate the average value μ1 and variance σ12 of the frequency values in the first voiceprint data, and the average value μ2 and variance σ22 of the frequency values in the second voiceprint data.
[0093] Under the deep learning framework, using the existing voiceprint data sample sets of the micro motor in normal and faulty states, the optimal deep learning model architecture is automatically searched through the Neural Architecture Search (NAS) algorithm. The NAS algorithm will try different combinations of the number of network layers, the number of neurons, activation functions, etc. to find the model with the best performance in voiceprint feature extraction and defect detection tasks. During the training process, the first voiceprint data and the corresponding voiceprint reference data are used as training samples. After multiple iterative trainings, the model can learn the mapping relationship from the data statistical values and the second voiceprint data to the voiceprint feature data.
[0094] Input the acoustic feature data of the second voiceprint data in each dimension and the corresponding data average value of each dimension into the trained NAS-optimized deep learning model. The model will output the voiceprint feature data corresponding to each dimension of the micro motor, and these feature data comprehensively reflect the abnormal conditions of the motor voiceprint in each dimension at the current detection moment.
[0095] Multiply the voiceprint feature data (obtained from the deep learning model) of each dimension by the voiceprint reference data (obtained from quantum feature extraction) of the corresponding dimension to obtain the weighted feature value corresponding to each dimension of the micro motor. For example, in the frequency dimension, the voiceprint feature data is F and the voiceprint reference data is f, then the weighted feature value is F×f. Sum up the weighted feature values of all dimensions to obtain the total voiceprint feature value of the micro motor at the current detection moment. This total value is an index that comprehensively reflects the overall abnormal degree of the motor voiceprint.
[0096] Based on a large number of defect sample data of micro motors and the corresponding voiceprint feature knowledge, a causal relationship graph between voiceprint features and defect types is constructed. In this graph, the nodes represent voiceprint features, such as energy changes in specific frequency bands, amplitude fluctuations, etc. and defect types, such as bearing wear, stator winding short circuit, rotor imbalance, etc. The edges represent the causal relationship between voiceprint features and defect types. For example, bearing wear will cause a significant increase in energy in certain specific frequency bands.
[0097] Compare the total value of the calculated voiceprint features with the preset defect feature threshold. If the total value of the voiceprint features exceeds this threshold, it indicates that the motor may have defects. At this time, use the causal graph separation algorithm to conduct reasoning and analysis on the causal relationship graph. According to the contribution of the weighted feature values of each dimension in the total value of the voiceprint features, find the possible defect types corresponding to these voiceprint feature values in the causal relationship graph. Combine reasoning rules and probability calculations to determine the specific defect type and output and display the results. For example, display information such as detecting bearing wear defects on the computer interface and suggesting timely maintenance. If the total value of the voiceprint features does not reach the threshold, it is determined that the micro motor has no obvious defects at the current detection moment, and a status prompt indicating normal motor operation is output.
[0098] According to the installation positions of 5 photonic crystal microphones around the micro motor, analyze their sensitivities to the voiceprint signals of different parts of the motor. For example, the microphone close to the motor bearing responds more strongly to the voiceprint signals generated by the bearing, while the microphone close to the stator winding is more sensitive to the voiceprint signals related to the winding. Based on this difference in position sensitivity, determine the measurement weight data of each photonic crystal microphone. The weight values can be obtained through experimental calibration or calculated based on a physical model. The weight data reflects the relative importance of different microphones in the overall defect detection.
[0099] After obtaining the total value of the voiceprint features corresponding to each photonic crystal microphone, multiply the total value by the corresponding measurement weight data to obtain the weighted total value of the voiceprint features of each microphone. Sum and fuse the weighted total values of the voiceprint features of all microphones to obtain a comprehensive total value of the voiceprint features. Use this fused comprehensive value to combine the causal graph separation algorithm again to make a final determination of the defect status of the micro motor. Compared with the determination result using only the data of a single microphone, the determination after fusing the data of multiple microphones is more reliable and accurate, and can more comprehensively reflect the overall operating state of the motor.
[0100] After obtaining the second voiceprint data at the current detection moment, update the first voiceprint data within the first time period by combining it with the voiceprint data corresponding to multiple adjacent detection time nodes before the current detection moment, such as the previous 3 detection cycles, a total of 15 minutes. Using a sliding window method, discard the earliest collected part of the data and add the latest second voiceprint data, so that the updated first voiceprint data always contains the latest period of time, for example, still the voiceprint feature information within 10 minutes, and the number of corresponding detection time nodes is the same as the number within the initially set first time period, ensuring the timeliness and accuracy of subsequent quantum feature extraction and voiceprint reference data calculation.
[0101] Calculate the minimum distance corresponding to each dimension between the first voiceprint data and the second voiceprint data. Distance calculation can use methods such as Euclidean distance and Manhattan distance. For example, in the frequency dimension, calculate the distance between the frequency feature vectors in the first voiceprint data and the second voiceprint data. If the minimum distance of at least one dimension is greater than the preset distance threshold (this threshold can be determined by statistically analyzing the fluctuation range of normal motor voiceprint data), then trigger the correction process of the first voiceprint data to ensure that the first voiceprint data can adapt to the slow performance changes that may occur in the motor or the voiceprint feature drift caused by external environmental factors.
[0102] For the situation that needs to be corrected, first calculate the difference between the acoustic feature data of the second voiceprint data in each dimension and the acoustic feature data of the corresponding dimension at the previous detection time node of the current detection moment to obtain the data offset value corresponding to each dimension. Then add the data offset value corresponding to each dimension to the acoustic feature data of the first voiceprint data to correct the first voiceprint data, so that the corrected first voiceprint data can better reflect the actual operating state characteristics of the motor at present.
[0103] The same or similar reference numerals correspond to the same or similar components;
[0104] The terms describing the positional relationship in the drawings are only for illustrative purposes and should not be construed as a limitation of this patent;
[0105] Obviously, the above embodiments of the present invention are only examples for clearly illustrating the present invention, and are not limitations on the implementation manners of the present invention. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to enumerate all the implementation manners here. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention should be included in the protection scope of the claims of the present invention.
Claims
1. A method for detecting defects in micro motors based on deep learning, characterized in that, Including the following steps: Obtain the first voiceprint data of the micro-motor within the first time period; the first voiceprint data is obtained based on a photonic crystal microphone arranged around the micro-motor; the first voiceprint data includes acoustic feature data in multiple dimensions; Perform quantum feature extraction processing on the acoustic feature data in multiple dimensions in the first voiceprint data to obtain the voiceprint reference data of the micro-motor within the first time period; the quantum feature extraction processing combines the principle of quantum computing to enhance the expression ability of features; Obtain the second voiceprint data of the micro-motor at the current detection moment; the current detection moment is the detection moment after the first time period, and the second voiceprint data includes acoustic feature data in multiple dimensions; Perform data statistical processing on the acoustic feature data of the first voiceprint data and the second voiceprint data in each dimension to obtain the data statistical value corresponding to each dimension; Based on the data statistical value corresponding to each dimension and the second voiceprint data, use a deep learning model optimized by NAS to determine the voiceprint feature data of the micro-motor; Based on the voiceprint feature data and the voiceprint reference data, combine the causal graph separation algorithm to determine the defect state of the micro-motor at the current detection moment.
2. The detection method according to claim 1, characterized in that, The obtaining of the voiceprint reference data of the micro-motor within the first time period includes: Use qubits to represent the acoustic feature data of each dimension, and evolve the quantum state through quantum gate operations to achieve quantum entanglement and superposition of features; Measure the quantum state, map the quantum features to the classical space, and obtain the data feature value corresponding to each dimension; Based on the data feature value corresponding to each dimension, calculate the voiceprint data weight corresponding to the dimension; Determine the voiceprint data weight of each dimension as the voiceprint reference data of the micro-motor within the first time period.
3. The detection method according to claim 1, characterized in that, The data statistical value includes the data average value and data variance of each dimension; the determining of the voiceprint feature data of the micro-motor includes: Based on the acoustic feature data of the second voiceprint data in each dimension and the data average value corresponding to the dimension, determine the voiceprint data difference corresponding to each dimension; Use the voiceprint data difference of each dimension and the data variance corresponding to the dimension as inputs, and process them through a deep learning model optimized by NAS to obtain the voiceprint feature data corresponding to each dimension of the micro-motor.
4. The detection method according to claim 3, wherein The determining of the defect state of the micro-motor at the current detection moment includes: Calculate the product between the voiceprint feature data of each dimension and the voiceprint reference data corresponding to the dimension to obtain the weighted feature value corresponding to each dimension of the micro-motor; Sum up the weighted feature values corresponding to each dimension to obtain the total voiceprint feature value of the micro-motor at the current detection moment; Based on the total voiceprint feature value, combine the causal graph separation algorithm to determine the defect state of the micro-motor at the current detection moment; the causal graph separation algorithm is based on a causal relationship graph to model and reason about the causal relationship between voiceprint features and defect types. Specifically: when the total voiceprint feature value meets the preset defect feature threshold condition, the causal graph separation algorithm is used to analyze the voiceprint features to determine the specific defect type; When the total voiceprint feature value does not meet the preset defect feature threshold condition, it is determined that the micro-motor has no obvious defect at the current detection moment.
5. The detection method according to claim 4, characterized in that, The number of the photonic crystal microphones is multiple; the first voiceprint data includes multiple groups of data corresponding to each photonic crystal microphone; the second voiceprint data includes multiple groups of data corresponding to each photonic crystal microphone; the method includes: Determine the measurement weight data of each photonic crystal microphone; the measurement weight data is determined based on the measurement position of the photonic crystal microphone around the micro-motor. Calculate the total voiceprint feature value corresponding to each photonic crystal microphone. Based on the measurement weight data of each photonic crystal microphone and the total voiceprint feature value corresponding to the photonic crystal microphone, determine the weighted total voiceprint feature value of each photonic crystal microphone. Based on the weighted total voiceprint feature value of each photonic crystal microphone, combined with the causal graph separation algorithm, determine the defect state of the micro-motor.
6. The detection method according to claim 1, wherein The first time period includes multiple detection time nodes; after obtaining the second voiceprint data of the micro-motor at the current detection moment, the method includes: Based on the second voiceprint data corresponding to the current detection moment and the voiceprint data corresponding to multiple adjacent detection time nodes before the current detection moment, update the first voiceprint data to obtain the updated first voiceprint data; the number of detection time nodes corresponding to the updated first voiceprint data is the same as the number of detection time nodes included in the first time period.
7. The detection method according to claim 1, characterized in that After obtaining the second voiceprint data of the micro-motor at the current detection moment, the method includes: Calculate the minimum distance corresponding to each dimension between the first voiceprint data and the second voiceprint data. In the case where the minimum distance of at least one dimension is greater than the preset distance threshold, correct the first voiceprint data.
8. The detection method according to claim 7, wherein The correction of the first voiceprint data includes: Calculate the difference between the acoustic feature data of the second voiceprint data in each dimension and the acoustic feature data of the corresponding dimension at the previous detection time node of the current detection moment to obtain the data offset value corresponding to each dimension. Based on the acoustic feature data of the first voiceprint data in each dimension and the data offset value corresponding to each dimension, correct the first voiceprint data.
9. An electronic device, characterized in that, Includes: A memory and at least one processor, wherein instructions are stored in the memory, and at least one of the processors calls the instructions in the memory to enable the device to execute each step of the detection method according to any one of claims 1-8.
10. A computer-readable storage medium having instructions stored thereon, characterized in that, When the instructions are executed by the processor, each step of the detection method according to any one of claims 1-8 is implemented.