Motor abnormal sound intelligent detection system and method based on deep learning

Abstract

The application relates to a deep learning-based motor abnormal sound intelligent detection system and method, wherein the method comprises the following steps: an acquisition module acquires original motor sound based on LabVIEW host computer interface configuration parameters; an adaptive data enhancement module performs adaptive data enhancement processing on discrete acoustic signals to obtain enhanced discrete acoustic signals; a full-link automatic modeling module digitizes and splits the discrete acoustic signals and aligns the discrete sample points; after the normalized discrete acoustic signals are adjusted through A weighting, time-frequency domain conversion is performed, and the preset neural network is inputted for training and optimization to generate a final model file; an inference engine composed of a pre-noise suppression convolution layer and a deep learning model network is constructed through the final model file; and after the inference engine determines the abnormal sound result and confidence, the result is fed back LabVIEW and UI display, result saving and data association operations are performed. The application realizes seamless closed loop by directly calling the encapsulated AI inference unit through the LabVIEW backend.

Description

Technical Field

[0001] This invention belongs to the field of intelligent motor detection technology, and specifically relates to an intelligent detection system and method for motor noise based on deep learning. Background Technology

[0002] In industrial production, motors are key power components, and their operating status directly affects equipment safety and production efficiency. Existing motor condition detection methods mostly rely on manual listening or experience-based judgment. These methods are highly subjective, dependent on operators, and prone to fatigue under long-term continuous operation, leading to unstable detection results and a high misjudgment rate, making it difficult to meet the high reliability and consistency requirements of industrial sites.

[0003] Although existing research has attempted to apply machine learning methods to detect motor noise in order to reduce the subjective influence of human listening, it still faces many challenges in practical applications in industrial settings.

[0004] On the one hand, motor noise signals usually need to be continuously collected at a high sampling rate to retain transient and high-frequency features. On the other hand, intelligent detection models have high computational complexity in feature extraction and inference. The difference in the rhythm of data generation and processing between the two can easily lead to timing mismatch, which in turn causes problems such as data cache backlog, loss or delayed detection results, making it difficult to meet real-time requirements.

[0005] On the other hand, industrial motors operate under normal conditions for a large proportion of their time, making it difficult to obtain abnormal or unusual noise samples. This results in limited and unevenly distributed training data, which limits the robustness and generalization ability of existing models in complex noise environments. Furthermore, control and testing platforms commonly used in industrial settings are mostly based on closed software environments such as LabVIEW, while mainstream artificial intelligence algorithms typically rely on frameworks like Python and PyTorch. The differences in their operating environments and data interaction mechanisms further increase the difficulty of system integration and deployment, thus hindering the engineering application of intelligent noise detection technology in industrial scenarios. Summary of the Invention

[0006] This invention provides a deep learning-based intelligent detection system and method for motor noise, aiming to solve at least one of the technical problems existing in the prior art.

[0007] The technical solution of this invention relates to a deep learning-based intelligent detection method for motor noise. This method is applied to a deep learning-based intelligent detection system for motor noise. The deep learning-based intelligent detection system for motor noise includes a data acquisition module, an adaptive data augmentation module, a full-link automated modeling module, and a real-time inference application module connected sequentially. The deep learning-based intelligent detection method for motor noise includes the following steps: S100, the data acquisition module configures the sampling rate based on a LabVIEW host computer interface. Recording duration By combining microphone sensitivity parameters, the raw motor sound is collected to obtain discrete acoustic signals; S200, the adaptive data enhancement module performs adaptive data enhancement processing on the discrete acoustic signal to obtain the enhanced discrete acoustic signal; The S300 full-link automated modeling module uses a sliding window strategy to digitally segment the enhanced discrete acoustic signal, align the discrete samples of the enhanced discrete acoustic signal, and normalize the enhanced discrete acoustic signal to eliminate amplitude differences under different sampling environments. The S400 full-link automated modeling module performs time-frequency domain conversion on the normalized discrete acoustic signal after A-weighting adjustment and inputs it into a preset neural network for training and optimization, generating the final model file. The S500, a fully automated modeling module, constructs an inference engine consisting of pre-noise-suppressed convolutional layers and a deep learning model network through the final model file. The S600 real-time inference application module determines the abnormal sound result and confidence level through the inference engine and then sends it back. LabVIEW It also performs UI display, result saving, and data association operations.

[0008] Furthermore, step S200 includes: S210, Real-time calculation of the power of discrete acoustic signals ; S220, Based on at least one preset signal-to-noise ratio threshold Dynamically generate matching power At least one Gaussian white noise sequence ; S230, Transform the Gaussian white noise sequence at the sample level. The enhanced discrete acoustic signal is superimposed onto the original discrete acoustic signal.

[0009] Furthermore, in step S300, The sliding window strategy uses the sampling rate Convert the start and end points of the time axis into discrete sampling point indices. The sliding step size To be based on preset time step With sampling rate The integer step size determined by the product is: The sampling point interval for each sliding window segment The calculation logic is as follows: , , in, This is the initial sampling point index. The duration of the sliding window, when When the integer is not an integer, use the floor function. Determine the starting point index to ensure that the slice index meets the discretization requirements of audio processing.

[0010] Furthermore, step S400 includes: S410. The normalized discrete acoustic signal is weighted using A-weighting. S420. Perform a short-time Fourier transform (STFT) on the weighted discrete acoustic signal to obtain a time-frequency characteristic map. S430. Input the time-frequency feature map into the preset neural network for training and optimization, and generate the final model file.

[0011] Furthermore, in step S410, the formula for calculating weight A is: , , in, The frequency of an acoustic signal is expressed in Hertz (Hz). Indicates frequency The weighted value of A at point A, in dB(A); This represents the amplitude response function of A-weighting; This represents a logarithmic operation with base 10; the constants 20.6, 107.7, 737.9, and 12200 are all characteristic frequency parameters.

[0012] Furthermore, step S500 includes: S510. Randomly shuffle the preprocessed discrete acoustic signal and divide it into a training set and a test set according to a preset ratio. S520. Label the samples, with normal fan sounds labeled as category 1 and abnormal / noisy fan sounds labeled as category 0; S530, training deep learning model networks based on the PyTorch deep learning framework; S540, Update the parameters of the deep learning model network by using the Adam optimizer Adaptive Moment Estimation.

[0013] Furthermore, in step S530, the Cross-Entropy Loss function is used as the optimization objective to measure the difference between the model's predicted probability and the true label. Its expression is as follows: in, The total number of categories; For the sample to belong to the first True one-hot encoded labels for each category; The model predicts that the sample belongs to the first... The probability values ​​of each category.

[0014] In step S540, the Adam optimizer, Adaptive Moment Estimation, dynamically adjusts the adaptive learning rate of different weight parameters in the deep learning model network by calculating the first and second moment estimates of the gradient. During the training phase of the deep learning model network, the Adam optimizer, Adaptive Moment Estimation, sets standard hyperparameters, including the initial learning rate. First-order momentum decay coefficient Second-order momentum decay coefficient and constants used to maintain numerical stability ; The training strategy for the deep learning model network is to use a stochastic gradient descent strategy with a batch size of 32 and a training epochs of 20. Combined with the adaptive characteristics of the Adam optimizer's Adaptive Moment Estimation, it can effectively overcome the local minima of the loss function and significantly reduce oscillations during training.

[0015] Furthermore, step S600 includes: The S610 is controlled by LabVIEW backend, which asynchronously wakes up the encapsulated inference engine via CLI command line commands, and achieves efficient scheduling of the detection process through non-intrusive asynchronous calls; The S620 inference engine loads the final model file for intelligent analysis and sends the judgment results and confidence scores back to LabVIEW via an intermediate text file. S630 enables full lifecycle detection and traceability by automatically associating product serial numbers and classifying and storing abnormal waveforms.

[0016] Furthermore, this invention proposes a deep learning-based intelligent detection system for motor noise, used to implement the deep learning-based intelligent detection method for motor noise. The deep learning-based intelligent detection system for motor noise comprises sequentially connected... The acquisition module is used to acquire raw discrete acoustic signals; The adaptive data augmentation module is used to perform adaptive data augmentation processing on discrete acoustic signals. The end-to-end automated modeling module performs numerical processing, A-weighting, time-frequency transformation, and neural network optimization on the enhanced discrete acoustic signal to build an inference engine. The real-time inference application module is used to detect and determine abnormal noises from motors and to... Interacting with LabVIEW .

[0017] Furthermore, the present invention also proposes a computer-readable storage medium storing program instructions thereon, which, when executed by a processor, implement the aforementioned deep learning-based intelligent detection method for motor noise.

[0018] Compared with existing technologies, the present invention has the following characteristics: The aforementioned deep learning-based intelligent detection system and method for motor noise encapsulates data preprocessing, multi-scale trimming modules, and deep learning models into a single, independently running executable unit, asynchronously invoked from the LabVIEW backend. This allows for seamless closed-loop operation by directly calling the encapsulated AI inference unit from the LabVIEW backend. Attached Figure Description

[0019] Figure 1 This is a flowchart of a deep learning-based intelligent detection method for motor noise.

[0020] Figure 2 This is a flowchart illustrating a deep learning-based intelligent detection method for motor noise.

[0021] Figure 3 This is a schematic diagram of the training accuracy of Example 3 of the intelligent detection method for motor noise based on deep learning.

[0022] Figure 4 This is a data acquisition diagram for Example 1 of the intelligent detection method for motor noise based on deep learning.

[0023] Figure 5 This is a schematic diagram of the host computer page for Example 1 of the intelligent detection method for motor noise based on deep learning.

[0024] Figure 6 This is a test set confusion matrix diagram for Example 3 of the intelligent detection method for motor noise based on deep learning.

[0025] Figure 7 This is a data slice diagram of Example 2 of the intelligent detection method for motor noise based on deep learning.

[0026] Figure 8 The diagram shows the CNN+Resnet18 model for Example 2 of the intelligent detection method for motor noise based on deep learning. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] The following will provide a clear and complete description of the concept, specific structure, and technical effects of the present invention in conjunction with the embodiments and accompanying drawings, so as to fully understand the purpose, solution, and effects of the present invention.

[0029] It should be noted that, unless otherwise specified, when a feature is referred to as "fixed" or "connected" to another feature, it can be directly fixed or connected to the other feature, or indirectly fixed or connected to the other feature. The singular forms "a," "described," and "the" used herein are also intended to include the plural forms, unless the context clearly indicates otherwise. Furthermore, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing particular embodiments only and not for limiting the invention. The term "and / or" as used herein includes any combination of one or more of the associated listed items.

[0030] It should be understood that although the terms first, second, third, etc., may be used in this disclosure to describe various elements, these elements should not be limited to these terms. These terms are only used to distinguish elements of the same type from each other. For example, a first element may also be referred to as a second element without departing from the scope of this disclosure, and similarly, a second element may also be referred to as a first element. Any and all instances or exemplary language (“e.g.,” “such as,” etc.) used herein are intended only to better illustrate embodiments of the invention and, unless otherwise required, do not impose a limitation on the scope of the invention. Furthermore, the industry term “pose” as used herein refers to the position and orientation of an element relative to a spatial coordinate system.

[0031] Reference Figures 1 to 8This invention provides a deep learning-based intelligent detection system and method for motor noise. The deep learning-based intelligent detection method for motor noise is applied to the deep learning-based intelligent detection system for motor noise. The deep learning-based intelligent detection system for motor noise includes a data acquisition module, an adaptive data augmentation module, a full-link automated modeling module, and a real-time inference application module connected in sequence. (Refer to...) Figure 1 The deep learning-based intelligent detection method for motor noise includes the following steps: S100, the data acquisition module configures the sampling rate based on a LabVIEW host computer interface. Recording duration By combining microphone sensitivity parameters, the raw motor sound is collected to obtain discrete acoustic signals; S200, the adaptive data enhancement module performs adaptive data enhancement processing on the discrete acoustic signal to obtain the enhanced discrete acoustic signal; The S300 full-link automated modeling module uses a sliding window strategy to digitally segment the enhanced discrete acoustic signal, aligning the discrete samples of the enhanced discrete acoustic signal. To eliminate amplitude differences under different sampling environments, the enhanced discrete acoustic signal is normalized, and the calculation formula is as follows: , in, This represents the normalized signal sample points. This represents the original discrete acoustic signal sample points; The S400 full-link automated modeling module performs time-frequency domain conversion on the normalized discrete acoustic signal after A-weighting adjustment and inputs it into a preset neural network for training and optimization, generating the final model file. The S500, a fully automated modeling module, constructs an inference engine consisting of pre-noise-suppressed convolutional layers and a deep learning model network through the final model file. The S600 real-time inference application module determines the abnormal sound result and confidence level through the inference engine and then sends it back. LabVIEW It also performs UI display, result saving, and data association operations.

[0032] Compared with existing technologies, the present invention has the following characteristics: The aforementioned deep learning-based intelligent detection system and method for motor noise encapsulates data preprocessing, multi-scale trimming modules, and deep learning models into a single, independently running executable unit, asynchronously invoked from the LabVIEW backend. This allows for seamless closed-loop operation by directly calling the encapsulated AI inference unit from the LabVIEW backend.

[0033] This invention solves the following technical problems: 1. Limitations of manual inspection: It solves the problems of fatigue misjudgment, strong subjectivity and unstable test results caused by the reliance on manual listening in traditional motor quality inspection.

[0034] 2. Engineering integration and compatibility challenges: Through decoupled encapsulation, the differences in operating environment and interaction mechanism between industrial environment (LabVIEW) and artificial intelligence framework (Python / PyTorch) are resolved, greatly reducing the deployment threshold.

[0035] 3. Timing and synchronization of data processing: This solves the problem of feature breakage and cumulative error caused by mismatched sliding window step size in high sampling rate audio processing.

[0036] 4. Generalization bottleneck of small sample modeling: Gaussian white noise injection algorithm with adaptive signal-to-noise ratio (SNR) This effectively solves the pain points of difficulty in obtaining abnormal noise samples from motors and the tendency of models to overfit.

[0037] The effects achieved by this invention are: 1. Extremely high detection accuracy: The system achieves an overall accuracy of no less than 99% in imbalanced datasets and maintains "zero false positives" for normal samples.

[0038] 2. Fast online cycle time: The complete inspection time for a single product (including recording, reasoning, and saving) is no more than 20 seconds, which fully meets the needs of real-time full inspection in industrial production lines.

[0039] 3. Full lifecycle traceability: The system automatically associates product serial numbers and categorizes and stores abnormal waveforms, providing closed-loop data support for subsequent process optimization and algorithm iteration.

[0040] 4. High robustness and environmental adaptability: The decision-level fusion mechanism significantly enhances the system's resistance to environmental background noise and weak abnormal sounds.

[0041] The present invention has the following technical improvements: 1. Preprocessing algorithm standardization: The microphone sensitivity calibration process is introduced, and the performance differences between different batches of sensors are eliminated through standardized gain processing.

[0042] 2. Adaptive noise injection algorithm: capable of adjusting noise injection based on input signal power. Dynamically generating noise sequences with matched power enables automated data augmentation that more closely resembles real-world operating conditions.

[0043] 3. Discrete sample-driven preprocessing module: ensuring the sliding window step size Using integers not only eliminates errors but also enables rapid locking of signals in the stationary phase, improving inference efficiency.

[0044] Specifically, for step S100, the discrete sample point precise alignment and multi-scale pruning logic: unlike conventional random slicing, this invention introduces a sampling rate-based... Integer step size constraint mechanism. Based on the preset step size. Calculate the number of discrete samples Furthermore, it mandates that the sample point offset be an integer, thereby eliminating sample point offset at the index level. A pruning module is integrated within the inference engine, which automatically slices a single recording across multiple time intervals by setting the starting sample point and length, significantly expanding the breadth of the feature distribution without increasing recording costs.

[0045] For step S100, the following technical features are present: (1) High-frequency lossless acquisition: ensure that the original signal covers high-frequency and transient abnormal sound characteristics, so as to retain the most authentic technical manuscript for subsequent analysis.

[0046] (2) Standardized gain control: The original acquisition level is standardized through the microphone sensitivity calibration process to ensure that the original data acquired by different batches of hardware have a consistent benchmark.

[0047] Furthermore, step S200 includes: S210, Real-time calculation of the power of discrete acoustic signals ; S220, Based on at least one preset signal-to-noise ratio threshold Dynamically generate matching power At least one Gaussian white noise sequence ; S230, Transform the Gaussian white noise sequence at the sample level. The enhanced discrete acoustic signal is superimposed onto the original discrete acoustic signal.

[0048] The system calls a backend algorithm to process the unenhanced raw dataset recorded in S100. First, the raw signal power is calculated. Then, based on the preset signal-to-noise ratio threshold Dynamically generate Gaussian white noise sequences with matched power Finally passed The sample-level overlay operation generates the enhanced dataset.

[0049] For step S200, the following technical features are present: (1) Dynamic environment simulation: Controlled noise is superimposed on the original signal to effectively simulate the acoustic background interference under complex industrial production lines, which significantly improves the generalization ability of the model in small sample scenarios.

[0050] (3) Sample diversity expansion: By superimposing noise of different intensities on the same original audio segment, the feature distribution breadth can be multiplied without increasing the recording cost.

[0051] Specifically, refer to Figure 2 This invention addresses the problems of scarce abnormal sound samples and sampling point offset by performing sample-level processing on the original signal: (1) Adaptive noise injection: The system calculates the original signal power in real time. A Gaussian white noise sequence with matched power is dynamically generated based on a preset signal-to-noise ratio (SNR). ,pass The superposition operation generates enhanced samples.

[0052] (2) Precise alignment of discrete sample points: The system forces a sliding step size during the slicing process. This is an integer to ensure accurate correspondence between the start and end indices of the sliding window, eliminating accumulated errors in high-frequency audio processing. The original audio sampling rate is... The duration of the sliding window is The sliding step size is Unit: s.

[0053] (3) Logical connection between modules 1 (corresponding to Arrow 1): Data-driven and input mapping logic. That is, the high-quality feature dataset generated by the above enhancement methods serves as the training benchmark and input source for the subsequent automated modeling module.

[0054] In some embodiments, for step S220, different signal-to-noise ratio thresholds are added by superimposing Gaussian white noise of different intensities onto the same segment of the original discrete acoustic signal. Gaussian white noise sequence This allows for a significant expansion of the feature distribution breadth without increasing recording costs.

[0055] Furthermore, in step S300, The sliding window strategy uses the sampling rate Convert the start and end points of the time axis into discrete sampling point indices. The sliding step size To be based on preset time step With sampling rate The integer step size determined by the product is: The sampling point interval for each sliding window segment The calculation logic is as follows: , , in, This is the initial sampling point index. The duration of the sliding window, when When the integer is not an integer, use the floor function. Determine the starting point index to ensure that the slice index meets the discretization requirements of audio processing.

[0056] Specifically, a sliding window strategy is used to digitally segment the enhanced audio. The system mandates a specific sliding step size. The integer is used to determine the exact discrete sampling point index for each slice window. and .

[0057] For step S300, the following technical features are present: (1) Sample-level index alignment: Eliminate accumulated errors from the underlying index level and solve the phase shift and feature breakage problems caused by step size rounding under high sampling rate.

[0058] (2) Local feature refinement: By capturing transient abnormal noise features during motor operation through multi-scale slicing, the sensitivity of the model to detailed anomalies is improved.

[0059] Furthermore, step S400 includes: S410. The normalized discrete acoustic signal is weighted using A-weighting. S420. Perform a short-time Fourier transform (STFT) on the weighted discrete acoustic signal to obtain a time-frequency characteristic map. S430. Input the time-frequency feature map into the preset neural network for training and optimization, and generate the final model file.

[0060] Specifically, the sliced ​​audio is processed using A-weighting to reflect the characteristics of human ear response, and then converted into a time-frequency map using Short Time Fourier Transform (STFT). The feature map is then input into a neural network (such as CNN+ResNet) for training and optimization, generating the final model file.

[0061] For step S400, the following technical features are present: (1) Auditory perception weighting: Combine the A-weighting formula to calculate the amplitude response of each frequency, highlight the different frequency bands that the human ear is sensitive to, and filter out irrelevant high and low frequency noise.

[0062] (2) Customized residual architecture: Preliminary noise suppression is performed using the pre-convolutional layer to improve the accuracy of the residual network in recognizing weak acoustic features in industrial environments.

[0063] Furthermore, in step S410, the formula for calculating weight A is: , , in, The frequency of an acoustic signal is expressed in Hertz (Hz). Indicates frequency The weighted value of A at point A, in dB(A); This represents the amplitude response function of A-weighting; This represents a logarithmic operation with base 10; the constants 20.6, 107.7, 737.9, and 12200 are all characteristic frequency parameters.

[0064] Specifically, the end-to-end automated modeling module transforms the processed data stream into a model file with diagnostic capabilities. For the enhanced signal, to better align with the subjective perception characteristics of the human ear for different frequencies, it employs A-weighting to weight the acoustic signal. Essentially, this constructs a psychoacoustic bandpass filter that simulates the auditory perception characteristics of the human ear. By executing the A-weighting algorithm, this system selectively suppresses low-frequency (typically below 500Hz) and high-frequency (typically above 10kHz) components, while significantly highlighting the mid-frequency band (1kHz - 4kHz) to which the human ear is most sensitive. This allows for the precise capture of subtle abnormal noise characteristics generated by the cooling fan during operation, such as bearing friction and fan blade rubbing.

[0065] For step S420, the Short-Time Fourier Transform (STFT) is a method for analyzing non-stationary signals. It transforms the signal from the time domain to the time-frequency domain, simultaneously acquiring the frequency components of the signal at different time points, thus displaying the change of the signal's spectrum over time. The Short-Time Fourier Transform (STFT) is represented as follows: In the formula, Indicates time; Represents the original signal; Describes the window function, where It is the center time point of the window function, indicating that the signal is subjected to spectral analysis at that time position.

[0066] Furthermore, step S500 includes: S510. Randomly shuffle the preprocessed discrete acoustic signal and divide it into a training set and a test set according to a preset ratio. S520. Label the samples, with normal fan sounds labeled as category 1 and abnormal / noisy fan sounds labeled as category 0; S530, training deep learning model networks based on the PyTorch deep learning framework; S540, Update the parameters of the deep learning model network by using the Adam optimizer Adaptive Moment Estimation.

[0067] Specifically, an inference engine is constructed, consisting of a pre-noise-suppressed convolutional layer and a deep learning model network, wherein the inference engine is a deep residual architecture. Preprocessed samples are randomly shuffled and divided into training and testing sets at a ratio of 90% to 10%. This division ensures complete data independence between the training and evaluation phases, avoiding overestimation of model performance due to data leakage. Sample labels are assigned based on manual listening results, with normal fans defined as category 1 and abnormal (noisy) fans defined as category 0.

[0068] A pre-noise suppression convolutional layer consisting of convolutional layers, batch normalization (BN), and ReLU activation function is added before the deep learning feature extraction network to construct a customized noise suppression pre-module.

[0069] The aforementioned preprocessing algorithm and model file are packaged into a standalone executable unit (.exe). After the inference engine runs, it sends back the decision results and confidence scores, which are then processed by... LabVIEW Read and display the data in the UI, save the results, and associate the data.

[0070] For step S500, the following technical features are present: (1) Non-intrusive integrated architecture: It realizes the decoupled deployment of the A algorithm and the LabVIEW control environment, and solves the problem of building a complex Python-dependent environment in industrial sites.

[0071] (2) High-efficiency asynchronous scheduling: Through backend collaborative calls, a rapid detection cycle of less than 20 seconds for a single product is achieved, meeting the real-time full inspection requirements of the production line.

[0072] Furthermore, in step S530, The Cross-Entropy Loss function is used as the optimization objective to measure the difference between the model's predicted probabilities and the true labels. Its expression is as follows: in, The total number of categories; For the sample to belong to the first True one-hot encoded labels for each category; The model predicts that the sample belongs to the first... The probability values ​​of each category.

[0073] In step S540, the Adam optimizer, Adaptive Moment Estimation, dynamically adjusts the adaptive learning rate of different weight parameters in the deep learning model network by calculating the first and second moment estimates of the gradient. During the training phase of the deep learning model network, the Adam optimizer, Adaptive Moment Estimation, sets standard hyperparameters, including the initial learning rate. First-order momentum decay coefficient Second-order momentum decay coefficient and constants used to maintain numerical stability ; The training strategy for the deep learning model network is to use a stochastic gradient descent strategy with a batch size of 32 and a training epochs of 20. Combined with the adaptive characteristics of the Adam optimizer's Adaptive Moment Estimation, it can effectively overcome the local minima of the loss function and significantly reduce oscillations during training.

[0074] Reference Figure 2 The second logical connection between modules (corresponding to Arrow 2) is the connection between model integration and engineering deployment. That is, the independent model file (.exe) generated after offline training and optimization is integrated into the online detection system as the core inference kernel.

[0075] Furthermore, step S600 includes: The S610 is controlled by LabVIEW backend, which asynchronously wakes up the encapsulated inference engine via CLI command line commands, and achieves efficient scheduling of the detection process through non-intrusive asynchronous calls; The S620 inference engine loads the final model file for intelligent analysis and sends the judgment results and confidence scores back to LabVIEW via an intermediate text file. S630 enables full lifecycle detection and traceability by automatically associating product serial numbers and classifying and storing abnormal waveforms.

[0076] Specifically, the deep learning-based intelligent detection method for motor noise uses non-intrusive communication logic. LabVIEW controls the front-end to record and generate the original audio. After the inference engine runs, it sends back the judgment result and confidence level, which is then read by LabVIEW and displayed in the UI.

[0077] The real-time online detection and closed-loop application module achieves production line-level intelligent discrimination through cross-platform collaboration: Non-intrusive asynchronous calls: The LabVIEW-controlled backend asynchronously wakes up the encapsulated inference engine via command line (CLI) instructions, enabling efficient scheduling of the detection process.

[0078] Decoupled communication mechanism: The inference engine loads the model file for intelligent analysis and sends the judgment results and confidence scores back to LabVIEW via communication methods such as intermediate text files (.txt).

[0079] Closed-loop quality control: The system automatically associates product serial numbers and categorizes and stores abnormal waveforms to achieve full lifecycle detection and traceability.

[0080] For step S600, automated judgment and closed-loop quality traceability are performed. LabVIEW reads the judgment conclusion and confidence level of the inference engine in real time. The system automatically associates the current product's serial number (SN) and stores the detection metadata and original waveform files according to the results.

[0081] For step S600, the following technical features are present: (2) Closed-loop management: Real-time storage and traceability of detection data are realized, providing reliable data support for subsequent production process optimization.

[0082] (2) Fault sample retention: Automatically retain the original data of NG (abnormal) samples to provide a real negative sample library for the continuous iteration of AI models.

[0083] Furthermore, this invention proposes a deep learning-based intelligent detection system for motor noise, used to implement the deep learning-based intelligent detection method for motor noise. The deep learning-based intelligent detection system for motor noise comprises sequentially connected... The acquisition module is used to acquire raw discrete acoustic signals; The adaptive data augmentation module is used to perform adaptive data augmentation processing on discrete acoustic signals. The end-to-end automated modeling module performs numerical processing, A-weighting, time-frequency transformation, and neural network optimization on the enhanced discrete acoustic signal to build an inference engine. The real-time inference application module is used to detect and determine abnormal noises from motors and to... Interacting with LabVIEW .

[0084] Furthermore, the present invention also proposes a computer-readable storage medium storing program instructions thereon, which, when executed by a processor, implement the aforementioned deep learning-based intelligent detection method for motor noise.

[0085] In a specific embodiment, such as Embodiment 1: Online Automatic Detection System for Abnormal Noise from Laptop Cooling Fans and Cross-Platform Integration This embodiment discloses an intelligent quality inspection system for laptop graphics card cooling fans, which focuses on the construction of the hardware acquisition environment and the software and hardware collaboration logic based on a decoupled architecture.

[0086] 1. Hardware acquisition environment setup (refer to...) Figure 4 ): The system first constructs a high signal-to-noise ratio physical detection environment to obtain high-fidelity raw acoustic data of motor operation: Controlled sound field environment: An industrial soundproof box with sound insulation performance of not less than the preset sound insulation threshold (such as 30-60dB) across the entire frequency band is used. When the ambient noise of the external factory environment is 80dBA, the background noise inside the box can be stably controlled within 50dB.

[0087] High-precision sensor matrix: High-precision acoustic sensors (such as B&K 4955, but not limited to) are selected and installed in the vertical and horizontal directions relative to the central axis of the fan under test, with the detection distance set within a preset range of 10mm to 150mm. This two-dimensional sensor arrangement ensures that both axial mechanical friction noise and radial hydrodynamic noise of the motor can be captured simultaneously.

[0088] 2. Graphical host computer interface logic: UI functional layout: such as Figure 5 As shown, the LabVIEW interface integrates parameter configuration (such as recording duration of 5s and target rotation speed of 3600 RPM), real-time waveform display area, FFT spectrum, and octave band analysis curve.

[0089] Reference Figure 5 The host computer interface mainly integrates parameter configuration, signal monitoring, and intelligent diagnostic functions. Users can set the recording duration, speed (RPM), and host computer parameters through the top area, and click "Start Record" to start recording motor sound. The middle and lower parts of the interface display the time-domain waveform, frequency-domain curve (FFT curve), and analysis feature map of the sound signal in real time, realizing visualized monitoring of the acquired signal. The right area is used to call the pre-trained model for A-model analysis and judgment, display the detection results (OK / NG) in real time, and automatically save the data.

[0090] Asynchronous scheduling and communication: The LabVIEW backend asynchronously wakes up the encapsulated inference engine (.exe) via command-line instructions.

[0091] Real-time result feedback: The system reads the .txt log generated by the inference engine and displays the judgment conclusion in real time in the large status bar on the right side of the interface. As shown in the figure, when the system reads Result:1, the interface is prominently marked as "PASS".

[0092] Recording execution (without inference): The system first obtains the serial number (SN) of the product under test by scanning a code. Then, under controlled speed conditions, the system drives the acoustic sensor through a multi-channel data acquisition module to collect the sound signal of the motor operation. The acquired data is stored in WAV format, and the file name includes key information such as the product SN, acquisition date, operating speed, and recording duration for subsequent data management and analysis.

[0093] 3. Cross-platform decoupled integration and asynchronous inference logic To address the incompatibility issue between LabVIEW and AI algorithm environments, this embodiment employs a "black box" encapsulation scheme for algorithm deployment: Independent Inference Unit: The Python script containing data preprocessing, pruning modules, and the trained CNN+ResNet18 model is packaged into a single-file executable program Python.exe using a packaging tool.

[0094] Command-line driven: After recording ends, LabVIEW uses the "Execute System Command" function to issue an asynchronous call command: Python.exe "{audio path or file name}.wav". At this time, the inference engine is awakened and begins to independently perform data loading, feature extraction, and intelligent discrimination tasks without consuming LabVIEW's main thread resources.

[0095] 4. Non-intrusive data exchange and result feedback The system achieves cross-platform result retrieval by sharing text logs, ensuring the reliability of the data chain. Log Generation: After the inference engine finishes processing, it will update or generate a specific text log file in real time. Taking a single actual detection as an example, the log records the complete processing chain information: Log Started 20xx-xx-xx 00:11:22 SN: ABCD0123456789 Data Preprocessing Completed (Time: xs) AI Model Loaded | Majority Class: 0 (xx.xx%) Final Result: NG | Model Inference Time: xs (Result:0) Result loop: The LabVIEW backend continuously monitors the log file, obtains the final judgment by reading the specific identification code at the end of the file (e.g., Result:0 represents NG, Result:1 represents OK), and associates it with the current product's SN code for classification and storage.

[0096] Cycle time verification: According to actual tests, the total inspection time of a single item, including preprocessing and model inference, is controlled within 20 seconds (the recording duration during the verification phase can be changed), which fully meets the cycle time requirements of online full inspection on the production line.

[0097] Example 2: Pre-processing and Adaptive Data Augmentation Method for Precise Alignment of Discrete Sample Points This embodiment details how the system backend (within the inference engine) performs fine-grained digitization processing on the acquired unenhanced raw dataset to address the issues of data synchronization and sample scarcity under high sampling rates. The corresponding flowchart illustrates the logical flow from data source to model construction (e.g., Figure 2 Arrow 1 path).

[0098] 1. Precise alignment preprocessing and adaptive data augmentation method for discrete sample points This embodiment details how the system backend processes high sampling rate audio, corresponding to the logical flow from data source to model construction in the flowchart ( Figure 2 In the middle (arrow1 path), to avoid cumulative phase error in high-sampling-rate continuous signal processing, the system executes the following alignment logic: Parameter settings: Set the sliding time step To address transient impact noise, this embodiment sets... Between 0.05s and 2s ( Figure 7 yes Cutting example); Reference Figure 7 OK: Qualified product / normal operating status; NG (Not Good): Unqualified product / abnormal operating status (such as abnormal fan noise); True Label: True label (the actual category of the sample); Predicted Label: Predicted label (the category determined by the model).

[0099] Step size integer constraint: according to the formula (When the sampling rate is 48000) Calculate the step size of the discrete sample points. hour, ;when hour, .

[0100] Index determination: Each slice window Starting sampling point index Strictly follow (in (This is the initial offset).

[0101] Technical effect: Due to Since the value is an integer, the cropping module does not need to perform rounding or resampling when slicing, thus eliminating the problems of sampling point offset and feature breakage in high-frequency audio processing from the index level.

[0102] 2.A Weighting Feature Enhancement Processing Since abnormal noises from cooling fans (such as slight scraping sounds or irregular bearing vibrations) are usually distributed in a specific frequency band, the system introduces an A-weighting algorithm to weight and filter the signal: Frequency response correction: using the A-weighted amplitude response function Gain adjustment is performed on each frequency component: Human ear perception fitting: through formula Convert the signal into a form that conforms to the subjective loudness characteristics of the human ear. Distribution. This step effectively suppressed irrelevant extremely high or low frequency interference, highlighting the core frequency band of the abnormal sound.

[0103] 3. Adaptive noise injection (data augmentation) based on SNR To address the pain point of scarce abnormal sound samples, the system performs automated enhancements on the original audio master obtained by the S100: Energy sensing: Calculating the original signal sampling rate and power .

[0104] Adaptive noise synthesis: based on a preset signal-to-noise ratio (e.g., ... to (Range) Dynamically calculate target noise power Generate Gaussian white noise sequence .

[0105] Overlay processing: Perform sample-level overlay. .

[0106] Technical results: By injecting noise into 10 NG (abnormal) original samples under different SNR, the robustness of the model to complex production line background noise was multiplied without increasing the recording cost.

[0107] 4. Time-frequency domain feature map transformation The enhanced and weighted discrete slice signal is then converted into a time-frequency graph using a short-time Fourier transform (STFT): Time-frequency mapping: applying window functions Perform local spectral analysis on the signal to obtain fine features of frequency changes over time.

[0108] Size standardization: unify the size of the generated time-frequency graphs ( The image is used as the standard input for the subsequent CNN+ResNet network architecture (model diagram as shown). Figure 8 ).

[0109] Reference Figure 8 Input layer: A spectral image with dimensions of 224×224×3. Convolutional layer (Conv): Used to extract local features of the image, such as texture and frequency distribution. Residual block (BasicBlock): Solves the vanishing gradient problem in deep networks through residual connections, consisting of four stages from Layer 1 to Layer 4, with the number of channels increasing progressively from 64 to 512. Pooling layer (Global Avg Pool): Global average pooling is used to reduce feature dimensionality while preserving key information. Fully connected layer (FC): Maps the extracted features to the classification space. Classification output: The final output is a "OK" or "NG" classification result.

[0110] Example 3: Model Training and Performance Validation on Imbalanced Datasets This embodiment verifies the discrimination accuracy and system operating speed of the present invention when processing extremely imbalanced data by conducting closed-loop tests on a large number of samples on an actual production line.

[0111] 1. Dataset Construction and Distribution Description Training set distribution: To improve the model's ability to capture anomalous features, the sample was expanded using adaptive data augmentation techniques (as described in S200). The training set ultimately contained 4655 samples of the OK category and 310 samples of the NG category.

[0112] Test set distribution: Blind testing was conducted using unenhanced, raw real-world data, with a total of 2289 samples. Among them, there were 2185 samples in the OK category (normal fans) and only 104 samples in the NG category (noisy fans).

[0113] Challenge Analysis: The ratio of normal to abnormal samples in the test set is close to 21:1. Under such an extremely imbalanced distribution, abnormal sound features are often masked by the sound of high-speed rotating airflow, placing extremely high demands on the feature extraction depth of the model.

[0114] 2. Model training stability: Designing a CNN+ResNet model: The deep learning model used consists of a pre-convolutional CNN (CNN) layer and a ResNet18 backbone network. First, a 3×3 convolutional layer with 64 output channels is added to the input, along with batch normalization and ReLU activation functions, for preliminary feature extraction and noise suppression of the time-frequency map. The feature maps are then fed into a standard ResNet18 structure, including a Conv1 layer (7×7 convolution, 64 channels, stride 2), a max-pooling layer, and four residual module groups composed of BasicBlocks with 64, 128, 256, and 512 channels, respectively. The final class decision is made at the end of the network through global average pooling and fully connected layers.

[0115] The training performance of the dataset above was achieved by iterative training using the PyTorch framework (e.g., 20 epochs). Convergence speed: refer to Figure 3 The training set accuracy (Train Acc) and validation set accuracy (Val Acc) quickly stabilized and converged to around 1.00 after the 10th round.

[0116] Robustness proof: The smooth training curves demonstrate that the customized CNN+ResNet18 architecture has excellent learning stability in small sample and imbalanced data environments, without obvious overfitting.

[0117] 3. Core Detection Accuracy Indicators By introducing precise alignment of discrete sample points and an A-weighted feature enhancement algorithm, the system demonstrates excellent stability in actual tests. Overall accuracy: The model achieved an overall accuracy of over 99% in the blind test with no fewer than 2,000 samples.

[0118] Extremely low false positives and zero false negatives: based on the confusion matrix (e.g. Figure 6 )show: NG category: All 104 abnormal samples were correctly identified, achieving "zero missed detection" and effectively solving the engineering pain point of weak friction sound identification.

[0119] OK Category: Only 1 out of 2185 normal samples was mistakenly identified as abnormal, which greatly ensured the first-pass yield of the production line and reduced the cost of manual re-inspection.

[0120] System runtime and efficiency verification: This embodiment records the entire lifecycle time from triggering data acquisition to outputting results, verifying the efficiency of the asynchronous collaborative architecture. Total inspection cycle time: The complete inspection cycle for a single product is consistently within 20 seconds.

[0121] Backend processing efficiency: Based on the .txt log records generated by the inference engine, the core time consumption distribution is as follows: Data preprocessing time (including discrete slicing, A-weighting, and graph transformation): within 2 seconds; Deep learning model inference time: less than 1 second; Total time for result aggregation and I / O: comparable to inference time.

[0122] Conclusion: This cycle time perfectly matches the full inspection speed requirements of current high-performance laptop production lines, proving that cross-platform asynchronous call mechanisms can effectively eliminate the risk of data backlog.

[0123] It should be understood that the method steps in the embodiments of the present invention can be implemented or carried out by computer hardware, a combination of hardware and software, or by computer instructions stored in a non-transitory computer-readable storage medium. The method can use standard programming techniques. Each program can be implemented in a high-level procedural or object-oriented programming language to communicate with the computer system. However, if necessary, the program can be implemented in assembly or machine language. In any case, the language can be a compiled or interpreted language. Furthermore, for this purpose, the program can run on a programmed application-specific integrated circuit (ASIC).

[0124] Furthermore, the procedures described herein may be performed in any suitable order unless otherwise indicated herein or otherwise clearly contradicted by the context. The procedures described herein (or variations and / or combinations thereof) may be executed under the control of one or more computer systems configured with executable instructions, and may be implemented by hardware or a combination thereof as code (e.g., executable instructions, one or more computer programs, or one or more applications) that commonly executes on one or more processors. The computer program comprises a plurality of instructions executable by one or more processors.

[0125] Furthermore, the method can be implemented operatively connected to any suitable type of computing platform, including but not limited to personal computers, minicomputers, mainframes, workstations, networked or distributed computing environments, standalone or integrated computer platforms, or communicating with charged particle tools or other imaging systems, etc. Aspects of the invention can be implemented as machine-readable code stored on a non-transitory storage medium or device, whether removable or integrated into a computing platform, such as a hard disk, optical read and / or write storage medium, RAM, ROM, etc., such that it is readable by a programmable computer, and when the storage medium or device is read by the computer, it can be used to configure and operate the computer to perform the processes described herein. Furthermore, the machine-readable code, or portions thereof, can be transmitted via wired or wireless networks. The invention described herein includes these and other different types of non-transitory computer-readable storage media when such media comprises instructions or programs that implement the steps described above in conjunction with a microprocessor or other data processor. When programmed according to the methods and techniques described in the invention, the invention may also include the computer itself.

[0126] A computer program can be applied to input data to perform the functions described herein, thereby transforming the input data to generate output data stored in non-volatile memory. The output information can also be applied to one or more output devices, such as a display. In a preferred embodiment of the invention, the transformed data represents physical and tangible objects, including specific visual depictions of physical and tangible objects generated on the display.

[0127] The above description is merely a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention, as long as they achieve the technical effects of the present invention by the same means, should be included within the scope of protection of the present invention. Within the scope of protection of the present invention, the technical solutions and / or implementation methods can have various modifications and variations.

Claims

1. A deep learning-based motor abnormal sound intelligent detection method, characterized in that, The deep learning-based intelligent detection method for motor noise is applied to a deep learning-based intelligent detection system for motor noise. This system includes a data acquisition module, an adaptive data augmentation module, a full-link automated modeling module, and a real-time inference application module connected sequentially. The deep learning-based intelligent detection method for motor noise includes the following steps: S100, the data acquisition module configures the sampling rate based on a LabVIEW host computer interface. Recording duration By combining microphone sensitivity parameters, the raw motor sound is collected to obtain discrete acoustic signals; S200, the adaptive data enhancement module performs adaptive data enhancement processing on the discrete acoustic signal to obtain the enhanced discrete acoustic signal; The S300 full-link automated modeling module uses a sliding window strategy to digitally segment the enhanced discrete acoustic signal, align the discrete samples of the enhanced discrete acoustic signal, and normalize the enhanced discrete acoustic signal to eliminate amplitude differences under different sampling environments. The S400 full-link automated modeling module performs time-frequency domain conversion on the normalized discrete acoustic signal after A-weighting adjustment and inputs it into a preset neural network for training and optimization, generating the final model file. The S500, a fully automated modeling module, constructs an inference engine consisting of pre-noise-suppressed convolutional layers and a deep learning model network through the final model file. The S600 real-time inference application module uses the inference engine to determine the abnormal sound results and confidence levels, then sends them back to LabVIEW for UI display, result saving, and data association operations.

2. The intelligent detection method for motor noise based on deep learning according to claim 1, characterized in that, Step S200 includes: S210, Real-time calculation of the power of discrete acoustic signals ; S220, Based on at least one preset signal-to-noise ratio threshold Dynamically generate matching power At least one Gaussian white noise sequence ; S230, Transform the Gaussian white noise sequence at the sample level. The enhanced discrete acoustic signal is superimposed onto the original discrete acoustic signal.

3. The intelligent detection method for motor noise based on deep learning according to claim 1, characterized in that, In step S300, The sliding window strategy uses the sampling rate Convert the start and end points of the time axis into discrete sampling point indices. The sliding step size To be based on preset time step With sampling rate The integer step size determined by the product is: The sampling point interval for each sliding window segment The calculation logic is as follows: , , in, This is the initial sampling point index. The duration of the sliding window, when When the integer is not an integer, use the floor function. Determine the starting point index to ensure that the slice index meets the discretization requirements of audio processing.

4. The intelligent detection method for motor noise based on deep learning according to claim 1, characterized in that, Step S400 includes: S410. The normalized discrete acoustic signal is weighted using A-weighting. S420. Perform a short-time Fourier transform (STFT) on the weighted discrete acoustic signal to obtain a time-frequency characteristic map. S430. Input the time-frequency feature map into the preset neural network for training and optimization, and generate the final model file.

5. The intelligent detection method for motor noise based on deep learning according to claim 4, characterized in that, In step S410, the formula for calculating weight A is: ， ， in, The frequency of an acoustic signal is expressed in Hertz (Hz). Indicates frequency The weighted value of A at the location is expressed in dB(A). This represents the amplitude response function of A-weighting; This represents a logarithmic operation with base 10; the constants 20.6, 107.7, 737.9, and 12200 are all characteristic frequency parameters.

6. The intelligent detection method for motor noise based on deep learning according to claim 1, characterized in that, Step S500 includes: S510. Randomly shuffle the preprocessed discrete acoustic signal and divide it into a training set and a test set according to a preset ratio. S520. Label the samples, with normal fan sounds labeled as category 1 and abnormal / noisy fan sounds labeled as category 0; S530, training deep learning model networks based on the PyTorch deep learning framework; S540, Update the parameters of the deep learning model network by using the Adam optimizer Adaptive Moment Estimation.

7. The intelligent detection method for motor noise based on deep learning according to claim 6, characterized in that, In step S530, the Cross-Entropy Loss function is used as the optimization objective to measure the difference between the model's predicted probability and the true label. Its expression is as follows: in, The total number of categories; For the sample to belong to the first Each category has a true unique hot-coded label; The model predicts that the sample belongs to the first... The probability values ​​of each category. In step S540, the Adam optimizer, Adaptive Moment Estimation, dynamically adjusts the adaptive learning rate of different weight parameters in the deep learning model network by calculating the first and second moment estimates of the gradient. During the training phase of the deep learning model network, the Adam optimizer, Adaptive Moment Estimation, sets standard hyperparameters, including the initial learning rate. First-order momentum decay coefficient Second-order momentum decay coefficient and constants used to maintain numerical stability. ; The training strategy for the deep learning model network is to use a stochastic gradient descent strategy with a batch size of 32 and a training epochs of 20. Combined with the adaptive characteristics of the Adam optimizer's Adaptive Moment Estimation, it can effectively overcome the local minima of the loss function and significantly reduce oscillations during training.

8. The intelligent detection method for motor noise based on deep learning according to claim 1, characterized in that, Step S600 includes: The S610 is controlled by LabVIEW backend, which asynchronously wakes up the encapsulated inference engine via CLI command line commands, and achieves efficient scheduling of the detection process through non-intrusive asynchronous calls; The S620 inference engine loads the final model file for intelligent analysis and sends the judgment results and confidence scores back to LabVIEW via an intermediate text file. S630 enables full lifecycle detection and traceability by automatically associating product serial numbers and classifying and storing abnormal waveforms.

9. A deep learning-based intelligent detection system for motor noise, used to implement the deep learning-based intelligent detection method for motor noise as described in any one of claims 1 to 8, characterized in that, The aforementioned deep learning-based intelligent detection system for motor noise includes sequentially connected... The acquisition module is used to acquire raw discrete acoustic signals; The adaptive data augmentation module is used to perform adaptive data augmentation processing on discrete acoustic signals. The end-to-end automated modeling module performs numerical processing, A-weighting, time-frequency transformation, and neural network optimization on the enhanced discrete acoustic signal to build an inference engine. The real-time inference application module is used to detect and determine abnormal noises from motors and interact with LabVIEW.

10. A computer-readable storage medium, characterized in that, It stores program instructions, which, when executed by a processor, implement the deep learning-based intelligent detection method for motor noise as described in any one of claims 1 to 8.

Images