Gear pair abnormal sound detection method and system based on double microphones and deep learning
By combining dual microphones with deep learning, the problems of strong subjectivity, low efficiency, and weak anti-interference ability in gear pair noise detection have been solved, realizing efficient and reliable automatic detection of gear noise and meeting the needs of industrial production.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-17
- Publication Date
- 2026-04-10
AI Technical Summary
Existing technologies for detecting abnormal noise in gear pairs suffer from problems such as high subjectivity, low efficiency, poor reliability, weak resistance to environmental interference, and reliance on expert experience, making it difficult to meet the requirements of high efficiency, reliability, and consistency in industrial production.
By combining dual microphones with deep learning, signals are collected through the main and reference microphones, preprocessed and analyzed in time and frequency, and noise suppression and feature extraction are performed using a deep learning model to achieve automatic detection of abnormal gear noise.
It achieves objectivity, accuracy, and efficiency in gear noise detection, reduces reliance on manpower and experience, improves detection consistency and production line adaptability, and has self-iterative optimization capabilities.
Smart Images
Figure CN121838809A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial product quality inspection technology, specifically to a method and system for detecting abnormal noises in gear pairs based on dual microphones and deep learning. Background Technology
[0002] In the manufacturing process of automotive transmission gears, the machining accuracy and fitting accuracy of the gear pairs are key indicators determining the performance, lifespan, and NVH (noise, vibration, and harshness) characteristics of the transmission assembly. Among these, whether or not abnormal noise is generated during gear meshing is one of the core criteria for evaluating its quality. Currently, the detection of abnormal noise in the gear factory inspection mainly relies on the following methods:
[0003] Traditional manual auditory testing method: This is the most common testing method currently used in the industry. The specific process is as follows: The gear pair is installed on a special test bench and run under specific speed and load. Experienced testing workers listen to the sound produced by the gear meshing and subjectively judge whether there are any abnormal noises, thereby determining whether the product is qualified.
[0004] This method has the following main drawbacks:
[0005] Highly subjective and inconsistent standards: Test results are highly dependent on workers' personal experience, hearing acuity, and even work mood. The lack of unified, quantifiable judgment standards among different inspectors leads to poor consistency in product inspection quality. Low efficiency: Manual inspection has limited speed, and the longer the workday, the slower the inspection speed due to physical and auditory fatigue, failing to meet the demands of high-speed factory production and becoming a bottleneck for improving production efficiency. Easily fatigued and unreliable: Workers experience auditory fatigue and decreased attention after long hours, leading to a significant increase in missed and false detection rates, making it difficult to guarantee the stability of product quality control. Difficult to pass on and quantify testing experience and methods: The experience of testing experts is "tacit knowledge," difficult to effectively record, standardize, and pass on to new employees. Lack of archived testing data: The lack of objective data records for the testing process and results makes it impossible to trace and analyze in case of quality disputes, hindering digital management and quality improvement in the production process.
[0006] Automated detection methods based on traditional signal processing: To overcome the shortcomings of manual detection, some improved detection schemes have been proposed in existing technologies. These schemes typically use a single microphone to collect the sound of gear operation, simply recording the decibel data as the detection data, or going further, utilizing traditional digital signal processing techniques (such as spectrum analysis, wavelet transform, envelope analysis, etc.) to extract the time-domain or frequency-domain features of the sound signal, and then using preset thresholds for anomaly judgment. However, such methods also have significant limitations in practical industrial applications:
[0007] Poor environmental interference resistance: Industrial environments are noisy, with various strong interferences such as drive motor noise, background mechanical noise, and personnel activity noise. Single-microphone systems struggle to effectively separate the target sound source (gear meshing sound) from background noise, leading to inaccurate feature extraction and a high false alarm rate. Reliance on expert feature engineering: The effectiveness of the method heavily depends on the developers' prior knowledge of acoustic signals and gear fault modes, requiring careful manual design and selection of features (such as energy values and harmonic components in specific frequency bands). This process is cumbersome, time-consuming, and lacks generalization ability for new and complex abnormal noise patterns. Difficult threshold setting: Preset fixed thresholds are difficult to adapt to natural signal changes caused by tool wear, minor fluctuations in operating conditions, etc., resulting in poor flexibility and a high risk of misjudgment. Summary of the Invention
[0008] The purpose of this invention is to provide a precipitation prediction system and method based on cloud parameter monitoring to solve the problems mentioned in the background art.
[0009] To achieve the above objectives, the present invention provides the following technical solution: an automatic gear noise detection method based on dual microphones and deep learning, comprising the following steps:
[0010] S1. Signal Acquisition: On the gear test bench, audio signals are acquired synchronously using a main microphone and a reference microphone; the main microphone is pointed at the gear pair meshing point to acquire a mixed signal containing gear sound, drive device noise and environmental noise; the reference microphone is pointed at the gear pair test device and away from the gear meshing point to acquire a reference noise signal containing drive device noise and environmental noise.
[0011] S2. Signal preprocessing: The acquired mixed signal and reference noise signal are preprocessed by filtering, windowing and other methods to enhance the effective spectral components and reduce spectral leakage.
[0012] S3. Time-frequency analysis: Perform short-time Fourier transform on the preprocessed signal to generate the mixed spectrum of the mixed signal and the noise reference spectrum of the reference noise signal, respectively.
[0013] S4. Deep Learning Model Inference: The mixed spectrogram and the noise reference spectrogram are used as inputs to a pre-trained deep learning model, which outputs the classification probability of the gear noise. The deep learning model is a neural network structure that can suppress noise features and enhance the acoustic features of the target.
[0014] S5. Result Judgment and Output: Based on the classification probability output by the model, determine whether the gear pair is qualified and output the detection result.
[0015] Furthermore, the deep learning model in S4 includes a feature extraction branch, a noise suppression module, and a classification module;
[0016] The feature extraction branch includes a first convolutional neural network and a second convolutional neural network, which are used to extract high-level features from the mixed spectrogram and the noisy reference spectrogram, respectively;
[0017] The noise suppression module, connected to the output of the feature extraction branch, is used to adaptively weight or filter the mixed features based on the noise reference features, suppressing the noise-related feature components and enhancing the feature components related to gear noise; this module is preferably a neural network layer based on an attention mechanism.
[0018] The classification module, connected to the output of the noise suppression module, is used to classify the purified features and output the final classification result.
[0019] Furthermore, S2 also includes audio signal pre-emphasis processing, which pre-emphasizes the original audio signal. High-frequency enhancement is performed to compensate for the attenuation of high-frequency components:
[0020] in, It is the signal after pre-emphasis. is the pre-emphasis coefficient, and n is the sampling point index; it can compensate for the high-frequency attenuation of the audio signal, improve the signal-to-noise ratio (SNR), and facilitate feature extraction.
[0021] Furthermore, the windowing process in S2 involves frequency division and windowing, and its formula is as follows:
[0022]
[0023]
[0024] in, It is the signal of the m-th frame; H is a window function of length N; H is the frame shift; and N is the frame length.
[0025] Furthermore, in the feature extraction branch, high-level features are extracted from the spectrogram, and the feature generation uses short-time Fourier transform, as shown in the following formula:
[0026] in, It is the k-th frequency component (complex number) of the m-th frame, and its amplitude Used to draw sound spectrograms.
[0027] Furthermore, the attention mechanism implementation process of the noise suppression module includes:
[0028] S401. Perform global average pooling on the noisy reference feature map Ref extracted by the second convolutional neural network to obtain the channel statistical vector. The calculation formula is:
[0029]
[0030] Where: B is the batch size, C is the number of feature channels, F is the frequency dimension, and T is the time dimension;
[0031] S402, Channel statistical vector The input consists of a subnetwork composed of two fully connected layers. The first fully connected layer compresses the number of channels to C->C / r, where r is the reduction ratio, 1<r≤16. After passing through the ReLU activation function, the second fully connected layer restores the number of channels to C.
[0032] S403, via the Sigmoid function:
[0033]
[0034] The sub-network output is mapped to the 0~1 interval to generate an attention weight matrix M with the same size as the hybrid feature map;
[0035] S404, using element-wise multiplication The mixed feature map U extracted from the main branch is weighted and filtered to obtain the purified target feature map. .
[0036] Furthermore, the classification module employs a Softmax classifier, whose decision formula is:
[0037]
[0038] in: It is the probability that the sample belongs to category i (K=2, i.e. qualified / unqualified); It is the output (logits) of the last linear layer of the model. =[z qualified, z unqualified].
[0039] Furthermore, the value range of the pre-weighting coefficient α is 0.9≤α≤0.98, preferably α=0.95.
[0040] Furthermore, the training process of the pre-trained deep learning model includes:
[0041] Audio data of qualified gear pairs and gear pairs with abnormal noise under different working conditions were collected. After processing through steps S1 to S3, a training dataset containing a mixed spectrogram, a noise reference spectrogram, and corresponding labels (qualified / unqualified) was constructed.
[0042] Using the cross-entropy loss function as the optimization objective, the Adam optimizer is used to iteratively train the deep learning model;
[0043] After training, the model accuracy is verified using a test dataset. When the accuracy is ≥99%, the model training is considered complete and it can be used for subsequent inference.
[0044] A gear pair abnormal noise detection system based on dual microphones and deep learning, employing the aforementioned gear pair abnormal noise detection method based on dual microphones and deep learning, includes:
[0045] A dual-microphone array is set on the gear test bench, including a main microphone and a reference microphone. The main microphone is pointed at the meshing point of the gear pair to collect a mixed signal containing gear sound, drive device noise and environmental noise. The reference microphone is pointed at the gear pair test device and away from the gear meshing point to collect a reference noise signal containing drive device noise and environmental noise. The two microphones collect audio signals synchronously.
[0046] The signal processing module, connected to the dual microphone array, is used to preprocess and perform short-time Fourier transform on the mixed signal and the reference noise signal to generate the corresponding mixed spectrogram and noise reference spectrogram.
[0047] The deep learning processing module is connected to the signal processing module and has a pre-trained deep learning model built in. It receives the two spectrograms and outputs the classification probability of the gear noise.
[0048] The result output module is connected to the deep learning processing module, which determines whether the gear pair is qualified based on the classification probability and outputs the detection result.
[0049] Compared with the prior art, the beneficial effects of the present invention are:
[0050] By using deep learning models to replace human hearing, the detection results are completely objective, eliminating the influence of factors such as human differences and fatigue, and ensuring the stability and traceability of product quality.
[0051] With its unique dual-microphone setup and matching dual-branch deep learning model, it can actively learn and separate noise features, thus accurately capturing faint gear noises even under strong background interference, solving the fatal weakness of poor anti-interference in traditional single-microphone solutions.
[0052] It avoids the complex feature engineering design process that relies on expert experience in traditional methods; the deep learning model can automatically learn the most effective discriminative features directly from the original spectrogram and has a stronger generalization ability for unknown fault types and compound faults.
[0053] The entire inspection process can be completed in seconds or even milliseconds, which greatly improves inspection efficiency and can keep up with the production rhythm of automated production lines, providing a technical foundation for achieving 100% online full inspection.
[0054] The detection data collected by the system can be continuously used to optimize and retrain the deep learning model, so that the model's judgment ability can be continuously improved with the accumulation of production data, and it has the ability to self-iterate and optimize.
[0055] This invention represents a qualitative leap in the dimensions of detection consistency, accuracy, efficiency, adaptability, and quality management. All major performance indicators can be greatly improved, while significantly reducing reliance on manpower and experience, thus solving a long-standing technical pain point in this field. Attached Figure Description
[0056] Figure 1 This is a schematic diagram of a sound acquisition, processing, and computing system.
[0057] Figure 2 This is a schematic diagram of an automatic gear noise detection process based on dual microphones and deep learning.
[0058] Figure 3 A schematic diagram of a deep learning feature extraction network model;
[0059] Figure 4 Weight matrix A schematic diagram illustrating the generation process. Detailed Implementation
[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0061] Please see Figure 1 —4. This invention provides a technical solution: an automatic gear noise detection method based on dual microphones and deep learning, comprising the following steps:
[0062] S1. Signal Acquisition: On the gear test bench, audio signals are acquired synchronously using a main microphone and a reference microphone; the main microphone is pointed at the gear pair meshing point to acquire a mixed signal containing gear sound, drive device noise and environmental noise; the reference microphone is pointed at the gear pair test device and away from the gear meshing point to acquire a reference noise signal containing drive device noise and environmental noise.
[0063] S2. Signal preprocessing: The acquired mixed signal and reference noise signal are preprocessed by filtering, windowing and other methods to enhance the effective spectral components and reduce spectral leakage.
[0064] S3. Time-frequency analysis: Perform short-time Fourier transform on the preprocessed signal to generate the mixed spectrum of the mixed signal and the noise reference spectrum of the reference noise signal, respectively.
[0065] S4. Deep Learning Model Inference: The mixed spectrogram and the noise reference spectrogram are used as inputs to a pre-trained deep learning model, which outputs the classification probability of the gear noise. The deep learning model is a neural network structure that can suppress noise features and enhance the acoustic features of the target.
[0066] S5. Result Judgment and Output: Based on the classification probability output by the model, determine whether the gear pair is qualified and output the detection result.
[0067] The deep learning model in S4 includes a feature extraction branch, a noise suppression module, and a classification module;
[0068] The feature extraction branch includes a first convolutional neural network and a second convolutional neural network, which are used to extract high-level features from the mixed spectrogram and the noisy reference spectrogram, respectively;
[0069] The noise suppression module, connected to the output of the feature extraction branch, is used to adaptively weight or filter the mixed features based on the noise reference features, suppressing the noise-related feature components and enhancing the feature components related to gear noise; this module is preferably a neural network layer based on the attention mechanism.
[0070] The classification module, connected to the output of the noise suppression module, is used to classify the purified features and output the final classification result.
[0071] S2 also includes audio signal pre-emphasis processing. Due to air absorption and the frequency response limitations of recording equipment, the energy attenuation of high-frequency components in sound signals is much faster than that of low-frequency components when propagating through the air and passing through the recording equipment. This results in an unbalanced spectrum in the original, unprocessed audio signal, with strong low-frequency energy and weak high-frequency energy. In this test, the normal meshing sound of gear pairs usually contains rich low-frequency components, while many abnormal noises caused by defects often manifest as brief, sharp transient signals, whose energy is mainly concentrated in the high-frequency range. Without pre-emphasis, these already weak high-frequency signals representing defects will be submerged in strong low-frequency background noise, making it difficult for subsequent spectrum analysis to effectively extract these key features. To compensate for the high-frequency components, audio signal pre-emphasis processing is required first in signal processing.
[0072] Audio signal pre-emphasis processing on the original audio signal High-frequency enhancement is performed to compensate for the attenuation of high-frequency components:
[0073] in, It is the signal after pre-emphasis. is the pre-emphasis coefficient, and n is the sampling point index; it can compensate for the high-frequency attenuation of the audio signal, improve the signal-to-noise ratio (SNR), and facilitate feature extraction.
[0074] The characteristics of sound signals change over time. For example, a gear noise might only occur during a brief moment of engagement. The statistical characteristics of a 2-second audio signal vary. Such signals are called non-stationary signals. The traditional Fourier Transform (FT) processes the global signal, assuming that all its statistical characteristics are constant. If you perform the FT on the entire non-stationary signal, you only get a global average spectrum, which tells you which frequency components are in the sound, but completely loses the temporal information of when these frequency components appear. To obtain the time-varying spectrum of the signal, i.e., the spectrogram, thus preserving both frequency and temporal information, we need to perform frequency division processing on the acquired audio signal.
[0075] However, framing is mathematically equivalent to multiplying an infinitely long continuous signal by a rectangular window function (with a value of 1 within the frame length and 0 outside the frame length). This abrupt truncation causes huge signal abrupt changes at the beginning and end of the frame. According to the properties of the Fourier transform, abrupt changes in the time domain introduce a large number of unwanted high-frequency components into the frequency domain, causing the calculated spectrum to become blurred, and energy to "leak" across the entire frequency band. This obscures the true frequency characteristics, making them impossible to accurately identify. Windowing replaces the default rectangular window by multiplying each frame signal with a non-rectangular window function that smoothly transitions at both ends. The advantage of this is that it greatly mitigates the abrupt changes at the beginning and end of the frame, making the start and end of the signal smoother, thus significantly reducing spectral leakage and allowing the calculated spectrum to more accurately reflect the frequency components of the signal within the frame. Framing and windowing are essential preprocessing steps for generating high-quality spectrograms.
[0076] The windowing process in S2 involves frequency division and windowing, and its formula is as follows:
[0077]
[0078]
[0079] in, It is the signal of the m-th frame; H is a window function of length N; H is the frame shift; and N is the frame length.
[0080] In order to provide deep learning models with effective input data that they can "understand" and "learn" from, the collected and preprocessed sound data needs to be "translated" into a spectrogram that can clearly show the essential characteristics of the sound. The abnormal noise of gears will show a pattern that is completely different from the normal sound on the spectrogram. These patterns are the "features" that we need to let the model learn, so that the algorithm in the field of computer vision (CNN) can play its role and achieve a perfect match between the raw audio data and the deep learning model architecture.
[0081] In the feature extraction branch, high-level features are extracted from the spectrogram. The feature generation uses Short Time Fourier Transform (STFT), as shown in the following formula:
[0082] in, It is the k-th frequency component (complex number) of the m-th frame, and its amplitude Used to draw sound spectrograms.
[0083] The implementation process of the attention mechanism in the noise suppression module includes:
[0084] S401. Perform global average pooling on the noisy reference feature map Ref extracted by the second convolutional neural network to obtain the channel statistical vector. The calculation formula is:
[0085]
[0086] Where: B is the batch size, C is the number of feature channels, F is the frequency dimension, and T is the time dimension;
[0087] S402, Channel statistical vector The input consists of a subnetwork composed of two fully connected layers. The first fully connected layer compresses the number of channels to C->C / r, where r is the reduction ratio, 1<r≤16. After passing through the ReLU activation function, the second fully connected layer restores the number of channels to C.
[0088] S403, via the Sigmoid function:
[0089]
[0090] The sub-network output is mapped to the 0~1 interval to generate an attention weight matrix M with the same size as the hybrid feature map;
[0091] S404, using element-wise multiplication The mixed feature map U extracted from the main branch is weighted and filtered to obtain the purified target feature map. .
[0092] This is the feature map extracted from the main branch. It is a three-dimensional tensor, which can be understood as a batch of "special photos" extracted from the spectrogram of the main microphone, containing both the gear signal we are interested in and anomalous noise.
[0093] This is the attention weight matrix. This is related to... A matrix of identical size, where each value is between 0 and 1. It can be viewed as a smart mask or filter.
[0094] A value of 1 means: "The feature here is very important; it is most likely a gear sound. Please retain it completely!"
[0095] A value of 0 means: "The features here are entirely background noise or drive noise; please filter them out completely!"
[0096] A value between 0 and 1 means: "Signal and noise are mixed here, please suppress them partially."
[0097] Element-wise multiplication. This is not matrix multiplication, but rather the direct multiplication of corresponding values in two matrices.
[0098] The purified output feature map. The enhanced target signal features after noise suppression.
[0099] It is not pre-set, but dynamically generated by the model based on the reference noise signal. This generation process is key to achieving intelligent noise reduction, and the attention weight matrix... The generation process is as follows:
[0100] Step 1: Input: Feature map Ref extracted by the reference branch network from the spectrogram of the reference microphone.
[0101] Step 2: Global Average Pooling: Compress the spatial dimensions (frequency F and time T) of Ref to obtain a channel statistics vector that aggregates global information. The calculation formula is as follows:
[0102] Step 3: Nonlinear Transformation: Vector This is then passed through a small subnetwork (typically consisting of two fully connected layers with a ReLU activation function in between). The first fully connected layer compresses the number of channels (C->C / r, where r is the reduction ratio), and the second layer restores the original number of channels (C / r -> C). This process is used to learn the complex correlations between channels.
[0103] Step 4: Sigmoid Activation: The output of the subnetwork is finally passed through a Sigmoid function, which maps each value to the interval between 0 and 1, thus obtaining the final attention weights. The Sigmoid function is defined as follows: .
[0104] This formula and process demonstrate the essential difference between this invention and simple applications of deep learning:
[0105] Targeted inhibition: Formula It achieves precise filtering at the feature level. The model does not treat all features equally, but rather adaptively weights each unit of the feature map based on the learned knowledge.
[0106] Dynamic Adaptation: Weight Matrix It is dynamically generated. For different input samples, the model generates different "filters" to achieve adaptive noise suppression, rather than using a fixed filter. This allows it to cope with ever-changing industrial noise environments.
[0107] The physical meaning is clear: the entire process perfectly simulates an intelligent sound engineer.
[0108] Input: What does the background noise sound like? (Refer to branch features)
[0109] Decision: Generate a tuning scheme (calculate the weight matrix) — "Which frequency bands should be reduced, and which should be retained?"
[0110] Execution: Process the main audio track according to the plan (element-wise multiplication), and finally output a clean audio spectrogram (purified features). ).
[0111] The innovation lies in the strong coupling between the algorithm and the dual-microphone hardware design. The effectiveness of the algorithm depends on the dual-channel signals provided by the hardware, while the value of the hardware design is realized through the powerful performance of the algorithm. Furthermore, this method can be applied to various monitoring scenarios using different training datasets, demonstrating strong generalization capabilities.
[0112] The classification module uses the Softmax classifier, and its decision formula is:
[0113]
[0114] in: It is the probability that the sample belongs to category i (K=2, i.e. qualified / unqualified); It is the output (logits) of the last linear layer of the model. =[z qualified, z unqualified].
[0115] The value range of the pre-weighting coefficient α is 0.9≤α≤0.98, with α=0.95 being preferred.
[0116] The training process of a pre-trained deep learning model includes:
[0117] Audio data of qualified gear pairs and gear pairs with abnormal noise under different working conditions were collected. After processing through steps S1 to S3, a training dataset containing a mixed spectrogram, a noise reference spectrogram, and corresponding labels (qualified / unqualified) was constructed.
[0118] Using the cross-entropy loss function as the optimization objective, the Adam optimizer is used to iteratively train the deep learning model;
[0119] After training, the model accuracy is verified using a test dataset. When the accuracy is ≥99%, the model training is considered complete and it can be used for subsequent inference.
[0120] like Figure 1 As shown: A gear pair abnormal noise detection system based on dual microphones and deep learning, employing the gear pair abnormal noise detection method based on dual microphones and deep learning as described in claim 1, characterized in that it includes:
[0121] A dual-microphone array is set on the gear test bench, including a main microphone and a reference microphone. The main microphone is pointed at the meshing point of the gear pair to collect a mixed signal containing gear sound, drive device noise and environmental noise. The reference microphone is pointed at the gear pair test device and away from the gear meshing point to collect a reference noise signal containing drive device noise and environmental noise. The two microphones collect audio signals synchronously.
[0122] The signal processing module, connected to the dual microphone array, is used to preprocess and perform short-time Fourier transform on the mixed signal and the reference noise signal to generate the corresponding mixed spectrogram and noise reference spectrogram.
[0123] The deep learning processing module is connected to the signal processing module and has a pre-trained deep learning model built in. It receives the two spectrograms and outputs the classification probability of the gear noise.
[0124] The result output module is connected to the deep learning processing module, which determines whether the gear pair is qualified based on the classification probability and outputs the detection result.
[0125] The sound acquisition, processing, and computing system specifically comprises a dual-microphone array consisting of microphone 1 and microphone 2, which serve as the main microphone and reference microphone as described above. The main microphone is placed near the gear meshing point to acquire a mixed signal of gear sound, drive noise, and ambient noise. This is a traditional signal source. The microphone is positioned as close as possible to the drive unit and as far away from the gear meshing point as possible; its core task is to specifically acquire the noise generated by the drive unit and background ambient noise, assuming that the acquired gear target signal is very weak. This provides a near-"pure" noise reference signal.
[0126] The signal preprocessing module, deep learning inference module, and result determination and display module run in the sound acquisition and processing computing system. They are used to perform signal preprocessing, time-frequency analysis steps, and pre-trained deep learning models, respectively, to load the models and perform inference; and finally, to display the detection results.
Claims
1. A gear abnormal sound automatic detection method based on double microphone and deep learning, characterized in that, Comprising the following steps: S1, signal acquisition: on the gear test bench, the audio signals are synchronously collected by the main microphone and the reference microphone; the main microphone is directed to the gear pair meshing point for collecting the mixed signal containing gear sound, driving device noise and environmental noise; the reference microphone is directed to the gear pair test device and away from the gear meshing point for collecting the reference noise signal containing driving device noise and environmental noise; S2, signal preprocessing: the collected mixed signal and reference noise signal are respectively pre-filtered, windowed and pre-processed to enhance the effective frequency spectrum components and reduce the frequency spectrum leakage; S3, time-frequency analysis: the pre-processed signal is subjected to short-time Fourier transform to generate a mixed sound spectrogram corresponding to the mixed signal and a noise reference sound spectrogram corresponding to the reference noise signal; S4, deep learning model inference: the mixed sound spectrogram and the noise reference sound spectrogram are input into the pre-trained deep learning model, and the classification probability of the gear abnormal sound is output by the model; the deep learning model is a neural network structure capable of suppressing noise features and enhancing target acoustic features; S5, result determination and output: according to the classification probability output by the model, it is determined whether the gear pair is qualified, and the detection result is output.
2. The dual-microphone and deep learning based automatic detection method of gear abnormal sound according to claim 1, characterized in that, The deep learning model in S4 includes a feature extraction branch, a noise suppression module and a classification module; The feature extraction branch includes a first convolutional neural network and a second convolutional neural network, respectively used for extracting high-level features from the mixed sound spectrogram and the noise reference sound spectrogram; The noise suppression module is connected to the output end of the feature extraction branch and is used for adaptively weighting or filtering the mixed features based on the noise reference features, suppressing the feature components related to noise and enhancing the feature components related to gear abnormal sound; The module is preferably a neural network layer based on attention mechanism; The classification module is connected to the output end of the noise suppression module and is used for classifying the purified features to output the final classification result.
3. The dual-microphone and deep learning based automatic detection method of gear abnormal sound according to claim 1, characterized in that, The S2 also includes an audio signal pre-emphasis process, which pre-emphasizes the original audio signal to compensate for the attenuation of high frequency components: wherein, is the pre-emphasized signal, is a pre-emphasis coefficient, n is a sample point index; the high-frequency attenuation of the audio signal can be compensated, the signal-to-noise ratio (SNR) is improved, and feature extraction is facilitated.
4. The dual-microphone and deep learning based automatic detection method of gear abnormal sound according to claim 1, characterized in that, The windowing process in S2 is divided into frequency and windowing, and the formula is as follows: wherein is the mth frame signal; is a window function of length N; H is the frame shift; N is the frame length.
5. The dual-microphone and deep learning based automatic detection method of gear abnormal sound according to claim 2, characterized in that, The short-time Fourier transform is used for feature generation in the spectrum feature extraction branch, and the formula is as follows: wherein, is the kth frequency component (complex number) of the mth frame, whose magnitude for mapping the sound.
6. The dual-microphone and deep learning based automatic detection method of gear abnormal sound according to claim 2, characterized in that, The attention mechanism implementation process of the noise suppression module includes: In S401, global average pooling is performed on the noise reference feature map Ref extracted by the second convolutional neural network to obtain a channel statistical vector The calculation formula is Wherein: B is the batch size, C is the feature channel number, F is the frequency dimension, and T is the time dimension; S402, Channel statistical vector The input consists of a subnetwork composed of two fully connected layers. The first fully connected layer compresses the number of channels to C->C / r, where r is the reduction ratio, 1<r≤16. After passing through the ReLU activation function, the second fully connected layer restores the number of channels to C. S403, through the Sigmoid function: Map the subnetwork output to the 0~1 interval to generate an attention weight matrix M consistent with the size of the mixed feature map; S404, adopt element level multiplication The mixed feature map U extracted by the main branch is weighted filtered to obtain a purified target feature map .
7. The dual-microphone and deep learning based automatic detection method of gear abnormal sound according to claim 2, characterized in that, The classification module adopts a Softmax classifier, and its decision formula is: where: is the probability that the sample belongs to class i (K = 2, i.e. pass / fail); is the output of the last linear layer of the model (logits); = [zpass, zfail].
8. The dual-microphone and deep learning based automatic detection method of gear abnormal sound according to claim 3, characterized in that, The pre-emphasis coefficient a is in the range of 0.9≤a≤0.98, and preferably a=0.
95.
9. The dual-microphone and deep learning based automatic detection method of gear abnormal sound according to claim 2, characterized in that, The training process of the pre-trained deep learning model includes: Collect audio data of qualified gear pairs and gear pairs with abnormal sound under different working conditions, process them through S1~S3 steps, and construct a training data set containing mixed sound spectrograms, noise reference sound spectrograms and corresponding labels (qualified / unqualified); Taking the cross-entropy loss function as the optimization objective, the deep learning model is iteratively trained by using the Adam optimizer; After training, the model accuracy is verified by the test data set, and when the accuracy is ≥99%, the model training is completed and used for subsequent inference.
10. A dual-microphone and deep learning based gear pair abnormal sound detection system, adopting the dual-microphone and deep learning based gear pair abnormal sound detection method of claim 1, characterized in that, Comprising: The double microphone array is arranged on the gear test platform, and includes a main microphone and a reference microphone. The signal processing module is connected with the double microphone array, and is used for pre-processing and short-time Fourier transform of the mixed signal and the reference noise signal, to generate corresponding mixed spectrogram and noise reference spectrogram. The deep learning processing module is connected with the signal processing module, and is internally provided with a pre-trained deep learning model. The result output module is connected with the deep learning processing module, and is used for judging whether the gear pair is qualified according to the classification probability and outputting a detection result.