A motor operation noise source identification method based on deep learning

Abstract

The present application relates to the field of data processing, and more particularly to a motor running noise source identification method based on deep learning. The method comprises: collecting the motor external acoustic signal and the q-axis current instruction sequence for time-frequency transformation, and extracting the characteristic tensor by using the double-flow network; combining the signal correlation and the harmonic bias factor to calculate the electromagnetic noise gating coefficient, and constructing an adaptive soft filter; accurately stripping the electromagnetic noise and the environmental noise from the acoustic signal, extracting the mechanical fault characteristic energy, and then identifying the noise category and generating the maintenance instruction. The present application is based on the sound-electricity causal correlation mechanism, realizes the effective decoupling and classification of the mixed noise source, overcomes the misjudgment problem caused by frequency aliasing, and significantly improves the accuracy and operation and maintenance efficiency of the motor fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of data processing, and more specifically to a method for identifying motor operating noise sources based on deep learning. Background Technology

[0002] Motors are widely used in precision manufacturing equipment where extremely low noise and smooth operation are required. During motor operation, abnormal noise is often an early sign of malfunction, and its source environment is extremely complex. This type of noise mainly includes mechanical noise caused by bearing wear, guide rail scraping, or interference between the moving and stator; electromagnetic noise caused by improper PWM switching frequency, current harmonics, or parameter tuning; and aerodynamic noise generated by high-speed motion. Accurately identifying the noise source is crucial for equipment maintenance and parameter adjustment.

[0003] However, existing noise detection methods are mostly based on single-mode sound signal analysis, such as spectral analysis using Fast Fourier Transform. In actual operating conditions, electromagnetic noise and mechanical noise often highly overlap in the frequency domain. For example, the high-frequency howling caused by current harmonics of a specific frequency is difficult to distinguish from the high-frequency vibration sound generated by early pitting of a bearing in a spectrum. Based solely on sound signals, it is difficult for technicians to determine whether it is necessary to adjust the electrical parameters of the servo drive or replace mechanical components.

[0004] This spectral aliasing leads to high trial-and-error costs and low maintenance efficiency. Technicians often need to rely on experience for repeated attempts or use replacement methods to troubleshoot hardware one by one, making it impossible to achieve rapid and accurate fault location. Therefore, there is an urgent need for a method that can effectively decouple electromagnetic noise from mechanical noise and achieve accurate noise source identification to improve the intelligent operation and maintenance level of precision manufacturing equipment. Summary of the Invention

[0005] To address the problem that existing technologies rely solely on single-mode sound signals, making it difficult to distinguish between electromagnetic noise and mechanical noise with overlapping frequency domains, resulting in low fault diagnosis accuracy and low maintenance efficiency, this invention provides a deep learning-based method for identifying motor operating noise sources.

[0006] This invention provides a deep learning-based method for identifying motor operating noise sources, employing the following technical solution:

[0007] A deep learning-based method for identifying motor operating noise sources includes: acquiring external acoustic signals and q-axis current command sequences from a servo driver during motor operation testing; performing time synchronization calibration and short-time Fourier transform on the external acoustic signals and q-axis current command sequences to obtain acoustic time-frequency maps and current time-frequency maps; constructing a dual-stream feature extraction network containing two parallel convolutional neural network branches to extract features from the acoustic and current time-frequency maps and perform normalization processing to obtain acoustic feature tensors and current feature tensors; calculating cross-modal electromagnetic noise gating coefficients based on the dot product of the acoustic and current feature tensors and a preset harmonic bias factor, wherein the electromagnetic noise gating coefficients characterize the explanatory power of current fluctuations on sound generation; constructing an adaptive soft filter based on the electromagnetic noise gating coefficients to filter out components determined to be electromagnetic noise from the spectral amplitude of the external acoustic signals and subtract an environmental noise threshold; calculating mechanical fault feature energy; and identifying the type of mechanical fault in the motor based on the mechanical fault feature energy.

[0008] This invention introduces the q-axis current command sequence within a servo driver as a second mode, utilizing the physical causality of electromagnetic noise generated by electromagnetic force to construct a cross-modal signal analysis framework. By learning the temporal mapping relationship between current fluctuations and sound fluctuations through a deep learning model, the portion of the sound signal related to current fluctuations can be accurately isolated as electromagnetic noise, while the remaining residual portion, which cannot be explained by current and whose intensity exceeds the noise floor, is mechanical noise. This achieves precise decoupling and classification of noise sources at the source level.

[0009] Furthermore, time synchronization calibration is performed on the external acoustic signal and the q-axis current command sequence, including: calculating the cross-correlation function of the external acoustic signal and the q-axis current command sequence; determining the time lag corresponding to the peak value of the cross-correlation function; and shifting the external acoustic signal according to the time lag to achieve time alignment between the external acoustic signal and the q-axis current command sequence.

[0010] Furthermore, the process of constructing the harmonic bias factor includes: reading the number of pole pairs and real-time speed of the motor, calculating the fundamental frequency of the motor, calculating the integer multiples of the fundamental frequency, and constructing a harmonic sequence; reading the PWM switching frequency of the servo driver, calculating the integer multiples of the PWM switching frequency, and constructing a switching frequency sequence; constructing a sensitive frequency set, which includes all frequency points in the harmonic sequence and the switching frequency sequence; constructing a harmonic bias factor with an initial value of zero, and setting the value of the harmonic bias factor to a preset bias strength constant at the frequency position corresponding to the sensitive frequency set.

[0011] This invention introduces a harmonic bias factor, incorporating prior physical knowledge of motors. When feature extraction is unclear, it acts as a fallback, acting as an expert rule to prevent missed detections and further improving the robustness of the identification, ensuring that electromagnetic noise at specific frequency points can be preferentially identified.

[0012] Furthermore, the electromagnetic noise gating coefficient satisfies the following relationship:

[0013]

[0014] in, For time and frequency The electromagnetic noise gating factor at the location, Let the acoustic feature tensor be... Let the current characteristic tensor be... Harmonic bias factor This represents the dot product operation. It is an exponential function with the natural constant as its base.

[0015] This invention evaluates the directional consistency of feature tensors through dot product operations and combines it with the Sigmoid activation form to transform the interpretability of current to sound into probability values ​​between 0 and 1, thereby achieving a refined evaluation of electromagnetic noise components and providing a reliable weighting basis for subsequent signal separation.

[0016] Furthermore, the characteristic energy of the mechanical fault satisfies the following relationship:

[0017]

[0018] in, For time Mechanical fault characteristic energy, For time and frequency The spectral amplitude of the external acoustic signal at that location, The environmental noise threshold. The electromagnetic noise gating coefficient is... This is the function for finding the maximum value.

[0019] This invention dynamically adjusts the weights by using a gating coefficient and introduces an environmental noise threshold, effectively eliminating logical loopholes caused by background noise and signal strength differences, ensuring that only genuine mechanical fault signals are retained and amplified, and significantly improving the signal-to-noise ratio.

[0020] Furthermore, the bias intensity constant is calibrated based on the saturation characteristics of the activation function in the calculation formula of the electromagnetic noise gating coefficient, as well as experimental results.

[0021] Furthermore, the step of identifying the mechanical fault type of the motor based on the mechanical fault characteristic energy includes: extracting the time series of mechanical fault characteristic energy; calculating the root mean square value and kurtosis value of the time series, comparing them with preset impact thresholds and energy thresholds to determine the fault type; and outputting fault diagnosis results, which include normal, bearing wear, and rotor friction.

[0022] Furthermore, the deep learning-based motor noise source identification method further includes: calculating the mean value of the electromagnetic noise gating coefficient; determining whether electromagnetic whistling and mechanical faults exist based on the mean value; generating electrical parameter optimization instructions if electromagnetic whistling exists; and generating hardware repair instructions if mechanical faults exist.

[0023] Furthermore, the environmental noise threshold is obtained by: collecting ambient sound for a preset duration through a microphone array while the motor is stationary; and calculating the average value of the spectral amplitude of the ambient sound as the environmental noise threshold.

[0024] Furthermore, the generated electrical parameter optimization instructions specifically include adjusting the current loop gain parameters and adjusting the filter parameters; the generated hardware repair instructions specifically include replacing the bearings and lubricating the guide rails.

[0025] This invention clarifies specific maintenance actions for different fault types, enabling the intelligent operation and maintenance system to directly output executable operation suggestions, such as adjusting the current loop gain to eliminate electromagnetic whistling, or replacing bearings to solve mechanical wear, thus forming a complete technical closed loop from diagnosis to execution, further reducing the technical threshold and trial-and-error costs for operation and maintenance personnel.

[0026] The present invention has the following technical effects:

[0027] This invention overcomes the limitations of traditional single-mode auscultation by utilizing the physical causality between current and sound to achieve mathematical decoupling of electromagnetic noise and mechanical noise in the feature space. Even if the frequencies of the two overlap, they can be distinguished based on the essential characteristic of whether they fluctuate with current, greatly reducing the false alarm rate.

[0028] By introducing an environmental noise threshold and normalizing the feature tensor, logical loopholes caused by background noise and signal strength differences are eliminated, ensuring that only genuine mechanical fault signals are preserved and amplified. Furthermore, the introduction of a harmonic bias factor cleverly integrates prior physical knowledge of the motor, acting as a fallback for expert rules when feature extraction is unclear, further enhancing the robustness of the recognition.

[0029] This invention eliminates the need for additional vibration sensors, utilizing only existing microphones and internal current data from the driver. Advanced diagnostics are achieved through algorithmic soft fusion, reducing hardware costs and deployment complexity. The identification results directly guide subsequent actions; distinguishing electromagnetic noise directly guides electrical parameter optimization, and distinguishing mechanical noise guides hardware repair, forming a complete intelligent operation and maintenance logic loop. This significantly reduces trial-and-error costs and improves maintenance efficiency. Attached Figure Description

[0030] Figure 1 This is a flowchart of a deep learning-based method for identifying motor operating noise sources, provided in an embodiment of the present invention.

[0031] Figure 2 This is a multimodal signal feature analysis diagram according to an embodiment of the present invention;

[0032] Figure 3 This is a diagram showing the noise source decoupling and diagnosis results of an embodiment of the present invention. Detailed Implementation

[0033] This invention provides a deep learning-based method for identifying motor operating noise sources, referring to... Figure 1 This includes steps S1-S4.

[0034] S1: Acquire the external acoustic signal and the q-axis current command sequence inside the servo driver during motor operation test, perform time synchronization calibration and short-time Fourier transform on the external acoustic signal and the q-axis current command sequence to obtain the acoustic time-frequency diagram and the current time-frequency diagram.

[0035] Specifically, during motor operation testing, two signals are collected simultaneously.

[0036] (1) The sound of the motor running is collected by an industrial-grade high-sensitivity microphone array deployed near the motor and defined as the external acoustic signal. In this embodiment, the sampling rate is 44.1kHz, covering the high-frequency howling band.

[0037] (2) Acquire the q-axis current command sequence inside the servo driver and define it as the internal current signal. In this embodiment, the q-axis current command of the motor control is read in real time through the high-speed data bus of the servo driver, such as EtherCAT. This signal directly reflects the fluctuation of the electromagnetic torque.

[0038] Because sound travels through the air with a physical delay, the system first calculates the cross-correlation function of the two signals and finds the time lag corresponding to the cross-correlation peak. For example, if the calculated time lag is 3ms, then the system will... A 3ms forward shift achieves microsecond-level synchronization of the acoustic and electrical signals. Subsequently, short-time Fourier transforms are performed on the aligned signals to obtain the acoustic time-frequency diagrams. Current time-frequency diagram .

[0039] Figure 2 This is a multimodal signal feature analysis diagram according to an embodiment of the present invention. Figure 2 (a) is the spectrum of the collected motor running sound. There is obvious aliasing noise, as well as sudden pulses and continuous high-frequency energy. When relying solely on the sound spectrum, it is difficult to distinguish these noises from different sources in terms of frequency. Figure 2 (b) is the spectrum of the q-axis current command sequence acquired synchronously. It can be seen that the spectrum is clean and clear, and the electromagnetic noise corresponds highly to the current harmonics, while the mechanical noise has no corresponding fluctuation in the current.

[0040] In this way, by calculating and compensating for the time lag through the cross-correlation function, the signal asynchrony problem caused by the sound propagation delay is eliminated, ensuring the accuracy of subsequent time-series correlation analysis and laying a precise data foundation for cross-modal feature fusion.

[0041] S2: Construct a two-stream feature extraction network containing two parallel convolutional neural network branches to extract features from the acoustic time-frequency map and the current time-frequency map respectively, and perform normalization processing to obtain the acoustic feature tensor and the current feature tensor.

[0042] Specifically, a neural network branch consisting of two parallel convolutions is constructed, defined as the acoustic flow branch and the current flow branch, which serve as feature extractors to independently learn the features of their respective modalities. The acoustic flow branch input... Current flow branch input These two branches have the same number of layers, kernel size, and stride. After multiple convolutions, batch normalization, and dilated convolutions, and with convolution padding enabled, the acoustic feature tensor is extracted. and current characteristic tensor The details are as follows:

[0043] 1. Design of a dual-stream feature extraction network architecture

[0044] Employing a digital twin architecture, it includes two parallel branches that are structurally symmetrical but parameter-independent: an acoustic flow branch and a current flow branch, which map signals of different modes into the same semantic feature space.

[0045] (1) Input data: Acoustic time-frequency diagram Input acoustic flow branch, and output current time-frequency diagram Input current flow branch.

[0046] (2) Feature extraction core module

[0047] Local feature extraction layer: Ordinary convolutional layer, using multiple two-dimensional convolutional kernels, combined with the same padding strategy and ReLU activation function, performs sliding scan on the acoustic time-frequency plot and the current time-frequency plot to extract local features of the time-frequency plot, such as the fundamental frequency line in the current signal and the transient impulse component in the sound signal;

[0048] Data distribution and standardization layer: Batch normalization is used to standardize and correct the distribution of feature data, which is used to accelerate model convergence, prevent internal covariate shift, and ensure the stability of the training process.

[0049] Wide-area temporal perception layer: dilated convolutional layer. By introducing a dilated convolution mechanism and setting different dilation rates, combined with the ReLU activation function, it can capture a wider range of information without increasing model parameters or reducing image resolution. This is crucial for recognizing electromagnetic howls with long temporal correlations.

[0050] (3) Feature space mapping and normalization

[0051] The network's final outputs are acoustic feature tensors. and current characteristic tensor Before output, L2 normalization is performed on the two feature tensors to make the magnitude always 1, ensuring that the dot product of the two feature tensors only reflects the consistency of direction in the feature space.

[0052] 2. Supervised training strategies

[0053] Before actual deployment, the network needs to be pre-trained using a data-driven approach to enable it to distinguish between noise caused by current and noise unrelated to current.

[0054] (1) Sample dataset construction: Construct a training dataset containing triples.

[0055] Benchmark sample: A real servo driver q-axis current command sequence;

[0056] Positive samples: Sound segments that have a strong physical causal relationship with the above q-axis current command sequence, such as recorded electromagnetic howls or synthesized electromagnetic sounds generated by physical models;

[0057] Negative samples: Sound segments that are completely unrelated to the above q-axis current command sequence, such as pure mechanical impact sounds, friction sounds, or the sound of another unrelated motor running.

[0058] (2) Optimization objective and loss function

[0059] The training process uses a triplet loss function, employs the Adam optimizer for iteration, and adjusts the convolution kernel weights and bias terms through the backpropagation algorithm.

[0060] Homogeneous signal aggregation: When the input acoustic signal and current signal have a strong physical causal relationship, i.e., positive samples, such as electromagnetic howling, the optimization algorithm adjusts the network weights so that the dot product result of the feature tensors output by the two, i.e., the cosine similarity, approaches 1.

[0061] Heterogeneous signal separation: When the input acoustic signal is unrelated to the current signal, i.e., a negative sample, such as a mechanical impact, the optimization algorithm adjusts the network weights so that the dot product of the feature tensors output by the two signals, i.e., the cosine similarity, approaches 0.

[0062] Training execution: Input the constructed sample dataset, calculate the loss value using the loss function, and update the back gradient using gradient descent based on the loss value until the loss value converges to complete the network training.

[0063] In this way, the trained network will have accurate cross-modal analysis capabilities, that is, it can accurately map whether current fluctuations cause sound fluctuations into the directional consistency of the feature space.

[0064] S3: Based on the dot product of the acoustic feature tensor and the current feature tensor, and combined with the harmonic bias factor constructed based on the motor physical parameters, the electromagnetic noise gating coefficient across modes is calculated. The electromagnetic noise gating coefficient characterizes the explanatory power of current fluctuations on sound generation.

[0065] Specifically, in order to evaluate the contribution of current fluctuations to sound generation, this invention designs a cross-modal signal correlation gating model. Its core is: if the sound at a certain time and frequency is caused by current, i.e. electromagnetic noise, then the acoustic features and current features at that point should have a high degree of directional consistency, i.e., strong correlation; conversely, if the sound originates from mechanical friction, the acoustic features and current features at that point should be uncorrelated.

[0066] Based on this physical logic, the electromagnetic noise gating coefficient The relation is:

[0067]

[0068] in, The term represents the dot product operation between the acoustic feature tensor and the current feature tensor, which is equivalent to the cosine similarity. The larger the value, the more consistent the directions of the two feature tensors are, meaning that the changing trends of sound fluctuations and current fluctuations are highly synchronized. It is the harmonic bias factor.

[0069] This represents the explanatory power of the current signal for the sound signal at time t and frequency f, ranging from (0,1). The closer the value is to 1, the greater the probability that the noise at that location is electromagnetic noise; the closer the value is to 0, the greater the probability that the noise at that location is mechanical noise.

[0070] exist In the relation, Adding the result to the dot product, when the frequency f is in the sensitive frequency band, the system will allow... Equal to the bias strength constant At this time If the value approaches 1, the frequency is directly identified as electromagnetic noise; if the frequency f is not in the sensitive frequency band, Keep it at 0. It depends on the similarity between acoustic and electrical characteristics.

[0071] Assuming that at a certain sensitive frequency, acoustic and current characteristics are highly correlated, i.e. The value is 0.9, and this frequency is at a harmonic point, i.e. Then ultimately If the value approaches 1, it is determined to be electromagnetic noise.

[0072] Assuming that at a certain frequency, acoustic and current characteristics are uncorrelated, The term is 0.1, and this frequency is not at a harmonic point, i.e. ,but for If the directions are opposite, If the term is -0.5, then It will be smaller and is judged to be non-electromagnetic noise.

[0073] Thus, by combining the dot product operation of the feature tensor with the harmonic bias of the physical prior, a gating mechanism that integrates data-driven and physical rules is constructed, which can accurately assess the degree to which the sound signal contains current components and effectively prevent missed judgments and misjudgments.

[0074] Harmonic bias factor It is a tensor with sparse properties in the frequency dimension, which mathematically participates directly as a bias term in the activation function calculation of the electromagnetic noise gating coefficient relationship. The core is to calculate the sensitive frequencies of electromagnetic noise using the deterministic physical parameters of the motor, and preset a value at these sensitive frequency positions. This allows for the capture of weak correlations between current and sound even when the signal is weak or environmental interference is severe, avoiding missed detections. The specific construction method is as follows:

[0075] (1) The system reads the number of motor pole pairs P and the real-time speed n in real time through the communication interface of the servo driver, and calculates the fundamental frequency. ;

[0076] (2) Multiply the calculated fundamental frequency by an integer multiple to construct the sequence. In this example, a multiple of 5 is used, i.e.: It is named a harmonic sequence;

[0077] (3) Read the switching frequency of the driver PWM In this example, the fundamental frequency and its second harmonic are selected to construct the sequence, namely: ;

[0078] (4) Finally, take and The union of these values ​​is labeled as the sensitive frequency. Initialization. It is zero at the sensitive frequency. The value is set as the bias intensity constant. .

[0079] It should be noted that, The calibration is based on a comprehensive analysis of the saturation characteristics of the Sigmoid function and experimental results.

[0080] (1) First, the analysis is based on the saturation characteristics of the Sigmoid function: when the acoustic characteristics at a certain time and frequency are not correlated with the current characteristics, that is... The term approaches 0; and, since a 95% confidence level is generally considered a highly reliable threshold in statistical and engineering applications, therefore, combined with If Approaching 0.95, we can obtain Choose 3, therefore choose 3. As theoretical initial values;

[0081] (2) Construct a search interval centered on the theoretical initial value. In this embodiment, the interval is [1.0, 5.0], and the scan is performed with a step size of 0.5. Collect a mixed sample set of known electromagnetic noises such as current harmonic howling at a specific frequency, and known mechanical noises such as bearing wear noise, and perform experimental calibration. Calculate different The false negative rate is selected from the values ​​below the specified values, with the lowest false negative rate being chosen. .

[0082] S4: Construct an adaptive soft filter based on the electromagnetic noise gating coefficient, filter out components identified as electromagnetic noise from the spectral amplitude of the external acoustic signal and subtract the environmental noise threshold, calculate the mechanical fault characteristic energy, and identify the mechanical fault type of the motor based on the mechanical fault characteristic energy.

[0083] Specifically, using The components identified as electromagnetic noise are filtered out from the sound spectrum energy of the external acoustic signal, and the environmental noise threshold is subtracted. The remaining residual energy is the mechanical fault characteristic energy.

[0084] Mechanical Fault Characteristic Energy The relationship is as follows:

[0085]

[0086] in, Characterizing the total energy of mechanical noise at time t, if If the value shows periodic peaks, it indicates mechanical impact such as bearing erosion or gear tooth breakage. It is time and frequency The spectral amplitude at that location is determined by the external acoustic signal. Obtained by performing a Fourier transform; It is the environmental noise threshold, used to prevent irrelevant noise in the environment from being mistakenly identified as mechanical fault characteristic energy. When the motor is stopped, the background sound in the test environment is collected, and the spectral amplitude of the background sound is calculated, which is the environmental noise threshold.

[0087] It should be noted that, The term represents the weighting of mechanical components; for electromagnetic noise, When the frequency approaches 1, the sound energy at that frequency is suppressed; for non-electromagnetic noise, As the frequency approaches zero, the sound energy at that frequency is preserved. The term is squared to make the system focus on mechanical noises that do not contain current characteristics, thereby improving the signal-to-noise ratio.

[0088] The motor bearing exhibits early pitting corrosion, emitting a periodic, weak noise during operation. This sound is not reflected in the current signal, therefore the corresponding... The value is extremely low; despite some ambient noise, the instantaneous amplitude of the periodic, weak noise is significantly higher than the noise floor. Ultimately... The system exhibited a clear peak at the corresponding time point. The system identified this and accurately diagnosed it as bearing wear.

[0089] Finally, extract Let x be a time series containing N data points, i.e. ; Calculate the root mean square value of x and kurtosis value ,in Let x be the mean of x;

[0090] Mathematically, the kurtosis value of a standard normal distribution is 3. When the motor is running normally, the mechanical noise is close to a normal distribution, and the kurtosis value will fluctuate around 3. Therefore, the impact threshold is set to 4. The measurement and calculation are performed once the motor is newly installed and confirmed to be in good working order. As a benchmark, the energy threshold is set to 2.5 times the benchmark value;

[0091] First, determine If the impact threshold is exceeded, it indicates periodic impact and is determined to be bearing wear; if the impact threshold is not exceeded, but... If the energy value is greater than the energy threshold, it indicates a persistent energy anomaly, which is determined to be motioner friction; if K and If none of the above conditions are met, the condition is considered normal. The final fault diagnosis result is output, such as: Normal, Bearing Wear, Moving Part Friction.

[0092] Figure 3 This is a diagram showing the noise source decoupling and diagnosis results of an embodiment of the present invention. Figure 3 (c) in the figure is the original mixed noise waveform, which is a complex mixed signal containing periodic high-amplitude pulses. Figure 3 In (d) of the diagram, the blue dashed line represents the decoupled electromagnetic noise, corresponding to high... In this section, the solid red line represents the decoupled mechanical residuals, corresponding to high... part. Figure 3 (e) shows the system's probabilistic diagnostic results for the fault category, indicating that the main fault source is mechanical bearing wear.

[0093] Thus, through adaptive soft filters and focusing enhancement mechanisms, electromagnetic noise is effectively removed from the mixed signal and environmental noise is suppressed, extracting pure mechanical fault features and greatly improving the signal-to-noise ratio and accuracy of mechanical fault diagnosis.

Claims

1. A method for identifying motor operating noise sources based on deep learning, characterized in that, include: The external acoustic signals and the q-axis current command sequence inside the servo driver are acquired during motor operation testing. Time synchronization calibration and short-time Fourier transform are performed on the external acoustic signals and the q-axis current command sequence to obtain the acoustic time-frequency diagram and the current time-frequency diagram. A two-stream feature extraction network containing two parallel convolutional neural network branches is constructed to extract features from the acoustic time-frequency map and the current time-frequency map respectively and perform normalization processing to obtain the acoustic feature tensor and the current feature tensor. Based on the dot product of acoustic feature tensor and current feature tensor, and combined with the harmonic bias factor constructed based on motor physical parameters, the electromagnetic noise gating coefficient across modes is calculated. The electromagnetic noise gating coefficient characterizes the explanatory power of current fluctuations on sound generation. An adaptive soft filter is constructed based on the electromagnetic noise gating coefficient. The component identified as electromagnetic noise is filtered out from the spectral amplitude of the external acoustic signal and the environmental noise threshold is subtracted. The mechanical fault characteristic energy is calculated, and the mechanical fault type of the motor is identified based on the mechanical fault characteristic energy. The electromagnetic noise gating factor satisfies the following relationship: , For time and frequency The electromagnetic noise gating factor at the location, For acoustic feature tensor, For current characteristic tensor, Harmonic bias factor This represents the dot product operation. It is an exponential function with the natural constant as its base; The characteristic energy of mechanical faults satisfies the following relationship: , For time Mechanical fault characteristic energy, For time and frequency The spectral amplitude of the external acoustic signal at the location, The environmental noise threshold. This is the function for finding the maximum value.

2. The method for identifying motor operating noise sources based on deep learning according to claim 1, characterized in that, Time synchronization calibration of external acoustic signals and q-axis current command sequences, including: Calculate the cross-correlation function between the external acoustic signal and the q-axis current command sequence; Determine the time lag corresponding to the peak value of the cross-correlation function; The external acoustic signal is shifted according to the time lag to achieve timing alignment between the external acoustic signal and the q-axis current command sequence.

3. The method for identifying motor operating noise sources based on deep learning according to claim 1, characterized in that, The process of constructing the harmonic bias factor includes: Read the number of pole pairs and real-time speed of the motor, calculate the fundamental frequency of the motor, calculate the integer multiples of the fundamental frequency, and construct a harmonic sequence; read the PWM switching frequency of the servo driver, calculate the integer multiples of the PWM switching frequency, and construct a switching frequency sequence; construct a sensitive frequency set, which contains all frequency points in the harmonic sequence and the switching frequency sequence. Construct a harmonic bias factor with an initial value of zero, and set the value of the harmonic bias factor to a preset bias intensity constant at the frequency position of the corresponding sensitive frequency set.

4. The method for identifying motor operating noise sources based on deep learning according to claim 3, characterized in that, The bias intensity constant is calibrated based on the saturation characteristics of the activation function in the calculation formula of the electromagnetic noise gating coefficient and experimental results.

5. The method for identifying motor operating noise sources based on deep learning according to claim 1, characterized in that, The method of identifying the type of mechanical fault in a motor based on its characteristic energy includes: Time series for extracting characteristic energy of mechanical faults; Calculate the root mean square value and kurtosis value of the time series, and compare them with the preset impact threshold and energy threshold to determine the fault type; Output fault diagnosis results, including normal, bearing wear, and rotor friction.

6. The method for identifying motor operating noise sources based on deep learning according to claim 1, characterized in that, The method further includes: Calculate the mean value of the electromagnetic noise gating coefficient; Determine whether there is electromagnetic whistling or mechanical failure based on the average value; If electromagnetic whistling is present, an electrical parameter optimization command will be generated; if mechanical faults are present, a hardware repair command will be generated.

7. The method for identifying motor operating noise sources based on deep learning according to claim 1, characterized in that, The environmental noise threshold is obtained as follows: When the motor is stationary, ambient sound is collected for a preset duration using a microphone array; the average value of the ambient sound spectrum amplitude is calculated as the ambient noise threshold.

8. The method for identifying motor operating noise sources based on deep learning according to claim 6, characterized in that, The generated electrical parameter optimization instructions specifically include adjusting the current loop gain parameters and adjusting the filter parameters; the generated hardware repair instructions specifically include replacing bearings and lubricating guide rails.

Images