An industrial product abnormal sound detection method based on deep learning

Abstract

The application discloses an industrial product abnormal sound detection method based on deep learning, which comprises a model establishment stage and an online detection stage. The model establishment stage comprises the following steps: collecting sound signals of product samples under different running states, and marking good products and defective products by a sound listener; processing the sound signals to generate abnormal sound feature training images, and training the feature images by using a CNN-LSTM hybrid model to obtain an abnormal sound classification model. The online detection stage comprises the following steps: collecting sound signals and vibration signals of products to be detected in real time, processing the sound signals in the same way to generate abnormal sound feature detection images, inputting the feature detection images into the model, and outputting defective product probabilities and confidence scores; meanwhile, extracting feature indexes of the vibration signals as independent determination parameters; and comprehensively determining the defective product probabilities output by the model and vibration abnormality scores. The application can replace artificial sound detection, improve detection consistency and efficiency, and significantly reduce the missed detection rate.

Description

Technical Field

[0001] This invention relates to the field of industrial acoustic testing technology, specifically to a deep learning-based method for detecting abnormal noise in industrial products, applicable to the detection of abnormal noise quality in industrial products such as motors, electric motors, and fans. Background Technology

[0002] In the industrial manufacturing sector, the sound quality of many rotating industrial products (such as motors, fans, etc.) is an important indicator of their quality. Due to reasons such as bearing wear, rotor imbalance, assembly errors, or foreign matter contamination, these products may produce abnormal sounds in addition to their normal operating sounds, which are known as unusual noises.

[0003] Traditionally, abnormal sound detection relies primarily on manual listening. Companies typically recruit or train personnel with exceptional hearing (commonly known as "golden ears") to assess different operating modes of products on the production line. This method has the following drawbacks:

[0004] 1. Low efficiency: A skilled listener can only test a limited number of samples per hour, which cannot meet the pace of large-scale production.

[0005] 2. Highly subjective, the judgment standards of different listeners and the same listener under different fatigue states are inconsistent, resulting in a high rate of missed detection and unstable batch quality;

[0006] 3. Prolonged, high-intensity exposure to sound can cause irreversible damage to the operator's hearing;

[0007] 4. As the requirements for quietness in products continue to increase, the human ear's ability to distinguish sounds is gradually approaching its limit.

[0008] Manual listening is insufficient to meet the demands of large-scale, high-standard abnormal sound detection. To address these issues, existing technologies have developed automated detection methods based on acoustic signal processing, such as frequency analysis, time-domain filtering, frequency response curve analysis, and traditional machine learning classification methods.

[0009] For example, see Chinese invention patent CN119125874B, which discloses a noise-based stepper motor fault detection method and system. The method includes acquiring the working state of the stepper motor, acquiring audio information of the stepper motor under different working states based on noise detection points, creating an operating audio model as audio features; statistically analyzing the working states and audio features, creating a mapping relationship between audio features and working states; in the practical application stage, acquiring the working period, acquiring actual audio information based on noise detection points, determining audio features based on actual audio information and working period, and acquiring working parameters based on the mapping relationship as prediction results.

[0010] The aforementioned methods have significant limitations: the analysis data is singular, typically relying solely on sound or vibration signals. These methods usually require signal processing experts to manually design features and set fixed thresholds. Since industrial noises often manifest as weak, transient, and highly coupled nonlinear signals with the dominant frequency, existing methods struggle to capture them effectively. Furthermore, the acoustic characteristics vary greatly between different product models, requiring extensive manual adjustments and lacking generalization ability.

[0011] Based on the above, the inventors propose the following technical solutions. Summary of the Invention

[0012] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a method for detecting abnormal noise in industrial products based on deep learning, so as to achieve automation, objectivity and high accuracy of abnormal noise detection.

[0013] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: a method for detecting abnormal noise in industrial products based on deep learning, the method comprising: a model building stage and an online detection stage; the model building stage includes the following steps: Step S1, Sample collection: acquiring sound signals of multiple product samples under different operating states, and having a listener manually determine whether each product sample is a good or defective product by listening to it; marking the sound signal segments corresponding to the product samples determined to be defective as abnormal samples, and marking the sound signal segments corresponding to the product samples determined to be good as normal samples, thereby obtaining an original sound signal sample set with the good / defective product labels; Step S2, The training feature extraction steps are as follows: Time-frequency transformation and image preprocessing are performed on each sound signal in the labeled original sound signal sample set to generate an abnormal sound feature training image; Step S3, model training: Using the abnormal sound feature training image as input and the corresponding good / defective product label as a supervision signal, supervised learning training is performed on the deep learning network model to obtain an abnormal sound classification model; The online detection stage includes: Step S4, data acquisition: Sound signals are acquired through acoustic sensors under different operating states of the product under test; vibration signals are acquired through vibration sensors; Step S5, detection feature extraction: The sound signal of the product under test is processed in the same way as in Step S2 to generate an abnormal sound feature detection image; Step S6, online recognition: The abnormal sound feature detection image is input into the abnormal sound classification model, and the defective product probability and confidence score are output; Step S7, comprehensive judgment: Based on the defective product probability and the confidence score, combined with the vibration anomaly score of the vibration signal, the product under test is determined to be a good or defective product.

[0014] Furthermore, in the above technical solution, in steps S1 and S4, different operating states include the operating states of the product under different speeds, different loads, or different working modes; the sampling frequency of the sound signal is not less than 48 kHz, and the sampling frequency of the vibration signal is not less than 10 kHz; a non-contact microphone is used to collect the sound signal, and a contact vibration sensor is used to collect the vibration signal.

[0015] Furthermore, in the above technical solution, the time-frequency transformation and image preprocessing in steps S2 and S5 include: first, background noise suppression and high-pass filtering of the sound signal; then, short-time Fourier transform of the filtered sound signal to obtain a time-frequency spectrum; next, converting the time-frequency spectrum into a Mel spectrum using a Mel filter bank; and finally, performing logarithmic compression and grayscale mapping on the Mel spectrum to generate the abnormal sound feature image.

[0016] Furthermore, in the above technical solution, the Mel spectrum has a frequency band range of 200 Hz to 15 kHz, and 128 to 512 Mel filter banks are used for frequency mapping, resulting in an image size of 224×224 pixels.

[0017] Furthermore, in the above technical solution, in step S3, the deep learning network model CNN-LSTM hybrid architecture is used, wherein the CNN part uses ResNet or MobileNet structure to extract spatial features of the image, the LSTM part is used to capture temporal change patterns under different time windows, and the model output layer is a binary classification Softmax layer, which outputs the probability of good products and the probability of defective products.

[0018] Furthermore, in the above technical solution, the comprehensive judgment method in step S7 is as follows: when the probability of defective products output by the sound classification model is greater than or equal to the preset confidence threshold, it is directly judged as a defective product; when the probability of defective products is less than the confidence threshold but the vibration abnormality score exceeds the preset vibration threshold, it is judged as a defective product; in other cases, it is judged as a good product.

[0019] Furthermore, in the above technical solution, in step S7, the vibration anomaly score is obtained by performing time-domain and frequency-domain analysis on the vibration signal, extracting at least one of the indicators among effective value, peak value, kurtosis and specific frequency band energy, and obtaining a vibration anomaly score between 0 and 1 after normalization processing. The larger the value, the more severe the vibration anomaly. The vibration threshold is preset according to the product testing requirements.

[0020] Furthermore, in the above technical solution, in step S7, when the difference between the defective product probability and the confidence threshold is within a preset range, the abnormal sound feature detection image of the product to be tested is pushed to the quality inspector's terminal for manual review; the judgment result of the manual review is periodically used to incrementally fine-tune the abnormal sound classification model.

[0021] Furthermore, in the above technical solution, in step S3, when the number of product samples determined to be defective in step S1 is less than a preset threshold, a diffusion generation model is used to perform variant synthesis on the sound signals of existing defective product samples to generate synthetic abnormal data to expand the training set, so that the ratio of good samples to abnormal samples in the training set reaches a preset equilibrium range.

[0022] Furthermore, in the above technical solution, in step S1, the listeners are at least two professional listeners who have passed hearing test certification; for samples where the two listeners' judgments are inconsistent, a third professional listener will make an arbitration judgment; the consistency of the manual listening judgments is not less than 95%.

[0023] This invention addresses the technical problems of existing technologies for detecting abnormal noises in industrial products, such as high subjectivity, inconsistent standards, high false negative rates, and limited recognition capabilities of traditional automated detection methods due to reliance on manual listening. It proposes a deep learning-based method for detecting abnormal noises in industrial products, which offers the following advantages compared to existing technologies:

[0024] 1. This invention first uses professional listeners (golden ears) for offline annotation, and then performs deep learning feature extraction, which eliminates the subjective inconsistency and fatigue problems of manual online listening inspection, resulting in high detection consistency.

[0025] 2. This invention uses the conversion of acoustic signals into images, thereby effectively capturing weak transient abnormal sound features.

[0026] 3. This invention also fuses vibration signals as an independent judgment parameter, and collects sound signals and vibration signals simultaneously for comprehensive judgment. The sound classification model is the main criterion, and the vibration anomaly score is the auxiliary criterion, which effectively avoids the misjudgment that may occur under specific working conditions by a single sensor and can further improve the reliability of detection.

[0027] 4. After offline annotation and deep learning, the present invention has a fast online detection speed, which can meet the needs of high-speed production lines. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with relevant embodiments. Obviously, the described embodiments are only a part of the embodiments of this invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0029] This invention provides a method for automatically detecting abnormal sounds in industrial products based on deep learning. The entire process can be divided into two stages: a model building stage and an online detection stage.

[0030] I. Model Building Phase.

[0031] Step S1, Sample Collection

[0032] First, it is necessary to collect operating sound and vibration data from a certain number of product samples. These product samples include both good and defective products. Sound signals from multiple product samples under different operating conditions are acquired. These different operating conditions include the product's operation under different speeds, loads, or operating modes. Sound signals are collected using a non-contact microphone, with a sampling frequency of no less than 48kHz. During acquisition, sound is recorded using a microphone, and vibration is measured using a vibration sensor.

[0033] Then, after the data collection is completed, each product sample is manually judged by listening to determine whether it is a good or defective product. To improve the accuracy of labeling, preferably, at least two professional listeners who have passed hearing tests and are certified (i.e., "golden ears") are arranged to make independent judgments; for samples whose judgments by the two listeners are inconsistent, a third professional listener will make an arbitration judgment, and the consistency requirement for labeling is not less than 95%. The sound signal segments corresponding to the product samples judged as defective are marked as abnormal samples, and the sound signal segments corresponding to the product samples judged as good are marked as normal samples, thereby obtaining the original sound signal sample set with good / defective product labels.

[0034] Of course, collecting vibration data in this step is not mandatory. The purpose of collecting vibration data from product samples is to determine a reasonable vibration threshold during the next stage of online testing.

[0035] Step S2, Training Feature Extraction

[0036] This step involves processing the acquired sound signal, converting it into an image. This conversion is necessary because deep learning models excel at image processing. The conversion principle is as follows: first, background noise is removed; then, faint unusual sounds are amplified; and finally, mathematical transformations are used to convert the sound signal into a time-spectrum graph, where the horizontal axis represents time, the vertical axis represents frequency, and the sound intensity is represented by color depth or different colors.

[0037] By performing time-frequency transformation and image preprocessing on each sound signal in the labeled original sound signal sample set obtained in step S1, training images representing the abnormal sound features are generated. Specifically, this includes the following sub-steps:

[0038] S21: Background noise suppression processing. A steady-state segment without obvious abnormal noise is extracted from the sound signal as a background noise template. The background noise template is subtracted from the original sound signal using spectral subtraction or adaptive filtering to suppress the main periodic operating noise during normal product operation and to highlight the faint abnormal noise characteristics.

[0039] S22: High-pass filter. A high-pass filter is applied to the audio signal after background noise suppression. The cutoff frequency is preferably 200Hz to filter out low-frequency environmental noise and further highlight abnormal sound characteristics.

[0040] S23: Short-Time Fourier Transform (STFT). The filtered audio signal undergoes a short-time Fourier transform to obtain a time-frequency spectrum, converting the one-dimensional time-domain signal into a two-dimensional time-frequency domain signal.

[0041] S24: Mel frequency mapping. The time-frequency spectrum is converted into a Mel frequency spectrum using a Mel filter bank. The number of Mel filter banks is preferably 128 to 512, and the frequency range is preferably 200Hz to 15kHz, which can cover the main sound-emitting frequency bands of the mechanical structures of rotating products such as motors.

[0042] S25: Image Generation. The Mel spectrogram is logarithmically compressed and grayscale mapped to generate a fixed-size (e.g., 224×224 pixels) image of the abnormal sound features. This image contains not only frequency domain features but also temporal pattern information of the sound signal changing over time.

[0043] Step S3, Model Training

[0044] Using the abnormal sound feature training image generated in step S2 as input and the corresponding good / defective product labels as supervision signals, supervised learning training is performed on the deep learning network model to obtain the abnormal sound classification model.

[0045] The deep learning network model preferably adopts a CNN-LSTM hybrid architecture. The CNN part uses a convolutional neural network structure such as ResNet or MobileNet to extract spatial texture features from the anomaly image; the LSTM part is used to capture temporal variation patterns under different time windows. The model output layer is a binary classification softmax layer, outputting the probability of a good product and the probability of a defective product.

[0046] In this step, a deep learning model is trained using the abnormal sound feature training images generated in step S2 and the corresponding non-defective / defective products. A convolutional neural network (CNN) is responsible for identifying the shape and texture features of abnormal sounds from the images; a recurrent neural network (LSTM) is responsible for analyzing the variation pattern of these features over time. After training is completed, the model learns to judge whether a product is qualified from the sound images.

[0047] During the training process, the cross-entropy loss function and the Adam optimizer are used for parameter optimization. In view of the actual situation in industrial scenarios where the number of defective product samples is often far less than that of non-defective product samples, which leads to the problem of unbalanced collected samples, the following method can be adopted to solve it: if the number of product samples determined as defective in step S1 is less than the preset threshold, a diffusion generation model is used to perform variant synthesis on the sound signals of the existing defective product samples, generate synthetic abnormal data to expand the training set, so that the quantity ratio of non-defective samples to abnormal samples in the training set reaches the preset equilibrium range, thereby improving the training effect and generalization ability of the model.

[0048] 2. Online Detection Stage

[0049] Briefly, the online detection stage is to collect the sound signal and vibration signal of each product on the actual production line, similarly convert the sound signal into an image, and input the image into the trained model. The model will automatically output an abnormality probability and calculate a vibration score at the same time. Then, judgment is made based on the set sound threshold and vibration threshold. The judgment rule is: as long as the abnormality probability is greater than or equal to the sound threshold, or the vibration score is greater than or equal to the vibration threshold, the product is judged as a defective product; otherwise, it is judged as a non-defective product. These two thresholds can be adjusted at any time according to product requirements. If the judgment of the model is uncertain and the abnormality probability is near the threshold, the corresponding image can be pushed to the operation screen for the quality inspector to make a decision, and the decisions of the quality inspector are collected and used periodically to fine-tune the model, so that the model becomes more and more accurate with use. Specifically, the online detection stage includes the following steps.

[0050] Step S4: Data Collection.

[0051] Under different operating states of the product to be tested, the sound signal is collected by a non-contact microphone, with the same sampling frequency not lower than 48kHz. The vibration signal is collected by a contact vibration sensor, with the sampling frequency not lower than 10kHz. The contact vibration sensor is used to collect low-frequency signals, and the non-contact microphone is used to collect high-frequency signals, and the two complement each other.

[0052] Step S5: Detection Feature Extraction.

[0053] The sound signal of the product under test is processed in exactly the same way as in step S2—that is, background noise suppression, high-pass filtering, short-time Fourier transform, Mel frequency mapping, logarithmic compression and grayscale mapping are performed in sequence to generate an abnormal sound feature detection image.

[0054] Step S6: Online identification.

[0055] The abnormal sound feature detection image generated in step S5 is input into the abnormal sound classification model trained in step S3, and the model outputs the probability of defective products and the confidence score.

[0056] The confidence score is the absolute value of the difference between the probability of a defective product and the probability of a good product output by the binary classification Softmax layer. The larger the value, the more certain the model's judgment. When the probability of a defective product is greater than the probability of a good product, the model tends to classify it as a defective product; otherwise, it tends to classify it as a good product.

[0057] Step S7, comprehensive judgment.

[0058] Based on the defect probability and confidence score output by the sound model, combined with the vibration anomaly score of the vibration signal, the product under test is comprehensively determined to be either good or defective. The specific determination method is as follows:

[0059] First, preset the sound confidence threshold Ta and the vibration threshold Tv.

[0060] If the defective product probability Pbad output by the sound classification model is greater than or equal to Ta, then the product under test is directly determined to be defective.

[0061] If Pbad is less than Ta, but the vibration anomaly score Vs is greater than or equal to Tv, it is also judged as a defective product.

[0062] All other cases are judged as good products.

[0063] The vibration anomaly score is obtained by performing time-domain and frequency-domain analysis on the vibration signal, extracting at least one of the following indicators: effective value (RMS, reflecting vibration energy), peak value (reflecting impact), kurtosis (reflecting impulsive anomaly), and specific frequency band energy (such as the frequency band corresponding to the characteristic frequency of bearing failure). After normalization, a vibration anomaly score between 0 and 1 is obtained, with a larger value indicating a more severe vibration anomaly.

[0064] Both the sound confidence threshold and the vibration threshold can be dynamically adjusted according to the product testing requirements. The adjustment does not require retraining the abnormal sound classification model. When the testing requirements are strict, the threshold is lowered to reduce missed detections; when the testing requirements are lenient, the threshold is raised to reduce over-detection.

[0065] When the absolute value of the difference between the probability of defective products and the confidence threshold is less than 0.1, the abnormal sound feature detection image of the product to be tested is pushed to the quality inspector's terminal for manual review; the judgment result of the manual review is used periodically to incrementally fine-tune the abnormal sound classification model to form a closed-loop optimization.

[0066] The following detailed description uses specific examples.

[0067] Taking a certain model of stepper motor as an example, a total of 500 units were sampled, with three test modes set: 300 rpm under no-load, 1200 rpm under no-load, and 900 rpm under rated load of 0.2 N·m. Each motor ran in the three modes sequentially, and after each mode had been running stably for 10 seconds, a 1-3 second signal was collected. The microphone sampling rate was 48 kHz, and the vibration sensor sampling rate was 10 kHz. The vibration sensor was mounted on the motor housing.

[0068] Then, three audiophiles certified by the International Association of Audio Engineers (IAE) were hired to evaluate the sound of each motor in three different modes using professional monitoring headphones in a soundproof room. The evaluation was conducted by two independent auditors, with a third person arbitrating in case of discrepancies. Ultimately, 58 defective products were identified.

[0069] For the acquired 1-3 second signal, the most stable 0.5 second period is first extracted as the background noise template, and spectral subtraction is used for noise reduction, followed by filtering with a fourth-order Butterworth high-pass filter. Then, the Mel spectrum is calculated, and the resulting two-dimensional array is interpolated and scaled to a 224×224 pixel grayscale PNG image. At this point, obvious bright spots or bright stripes will be visible on the image of the defective sample, corresponding to abnormal noise events.

[0070] A CNN-LSTM model was constructed. The CNN part used a pre-trained ResNet-34. The LSTM part consisted of two layers, with the classification layer outputting two classes. The AdamW optimizer was used for training, and the model was trained for 50 epochs. Since there were only 58 defective samples in the training set, a diffusion model was used to generate 100 synthetic images of abnormal sounds, which were then mixed with the real samples. Finally, the model's accuracy was verified.

[0071] Once the model's accuracy meets the target, during production line testing, each motor automatically operates in three modes. Sound signals are collected via non-contact microphones, and a Mel-ray diagram is generated and inferred in real time. Simultaneously, vibration signals are collected via contact vibration sensors. Combined with the vibration anomaly score of the vibration signals, the product under test is comprehensively judged as either good or defective.

[0072] Of course, the above description is only a specific embodiment of the present invention and is not intended to limit the scope of the present invention. All equivalent changes or modifications made to the structure, features and principles described in the claims of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A method for detecting abnormal noises in industrial products based on deep learning, characterized in that, This abnormal sound detection method includes: a model building stage and an online detection stage; The model building phase includes the following steps: Step S1, Sample Collection: Acquire sound signals of multiple product samples under different operating conditions. The listener will manually listen to each product sample to determine whether it is a good or defective product. The sound signal segments corresponding to the product samples determined to be defective will be marked as abnormal samples, and the sound signal segments corresponding to the product samples determined to be good will be marked as normal samples, thus obtaining an original sound signal sample set with the good / defective product labels. Step S2, Training Feature Extraction Step: Perform time-frequency transformation and image preprocessing on each sound signal in the labeled original sound signal sample set to generate a training image of abnormal sound features; Step S3, Model Training: Using the abnormal sound feature training image as input and the corresponding good / defective product label as supervision signal, supervised learning training is performed on the deep learning network model to obtain the abnormal sound classification model; The online detection phase includes: Step S4, Data Acquisition: Under different operating conditions of the product under test, sound signals are acquired through acoustic sensors; vibration signals are acquired through vibration sensors. Step S5, Feature Extraction: The sound signal of the product under test is processed in the same way as in step S2 to generate an abnormal sound feature detection image; Step S6, Online Identification: Input the abnormal sound feature detection image into the abnormal sound classification model, and output the probability of defective products and confidence score; Step S7, Comprehensive Judgment: Based on the defective product probability and the confidence score, combined with the vibration anomaly score of the vibration signal, determine whether the product to be tested is a good product or a defective product.

2. The method for detecting abnormal noise in industrial products based on deep learning according to claim 1, characterized in that, In steps S1 and S4, different operating states include the operating states of the product under different speeds, different loads, or different working modes; the sampling frequency of the sound signal is not less than 48 kHz, and the sampling frequency of the vibration signal is not less than 10 kHz; a non-contact microphone is used to collect the sound signal, and a contact vibration sensor is used to collect the vibration signal.

3. The method for detecting abnormal noise in industrial products based on deep learning according to claim 1, characterized in that, In steps S2 and S5, the time-frequency transformation and image preprocessing include: First, the audio signal undergoes background noise suppression and high-pass filtering. Then, a short-time Fourier transform is performed on the filtered audio signal to obtain the time-spectrum diagram; Next, the time-frequency spectrum is converted into a Mel spectrum using a Mel filter bank; Finally, the Mel spectrogram is logarithmically compressed and grayscale mapped to generate the heterophonic feature image.

4. The method for detecting abnormal noise in industrial products based on deep learning according to claim 3, characterized in that, The Mel spectrum has a frequency band ranging from 200 Hz to 15 kHz, and uses 128 to 512 Mel filter banks for frequency mapping, resulting in an image size of 224×224 pixels.

5. The method for detecting abnormal noise in industrial products based on deep learning according to claim 1, characterized in that, In step S3, the deep learning network model uses a CNN-LSTM hybrid architecture. The CNN part uses a ResNet or MobileNet structure to extract spatial features of the image, and the LSTM part is used to capture temporal change patterns under different time windows. The model output layer is a binary classification Softmax layer, which outputs the probability of good products and the probability of bad products.

6. The method for detecting abnormal noise in industrial products based on deep learning according to claim 1, characterized in that, In step S7, the comprehensive judgment method is as follows: when the probability of defective products output by the sound classification model is greater than or equal to the preset confidence threshold, it is directly judged as a defective product; when the probability of defective products is less than the confidence threshold but the vibration abnormality score exceeds the preset vibration threshold, it is judged as a defective product; in other cases, it is judged as a good product.

7. A method for detecting abnormal noise in industrial products based on deep learning according to claim 1 or 6, characterized in that, In step S7, the vibration anomaly score is obtained by performing time-domain and frequency-domain analysis on the vibration signal, extracting at least one of the indicators among effective value, peak value, kurtosis and specific frequency band energy, and obtaining a vibration anomaly score between 0 and 1 after normalization. The larger the value, the more severe the vibration anomaly. The vibration threshold is preset according to the product testing requirements.

8. The method for detecting abnormal noise in industrial products based on deep learning according to claim 6, characterized in that, In step S7, when the difference between the defective product probability and the confidence threshold is within a preset range, the abnormal sound feature detection image of the product to be tested is pushed to the quality inspector's terminal for manual review; the judgment result of the manual review is periodically used to incrementally fine-tune the abnormal sound classification model.

9. The method for detecting abnormal noise in industrial products based on deep learning according to claim 1, characterized in that, In step S3, when the number of product samples identified as defective in step S1 is less than a preset threshold, a diffusion generation model is used to synthesize variants of the sound signals of existing defective product samples to generate synthetic abnormal data to expand the training set, so that the ratio of good samples to abnormal samples in the training set reaches a preset equilibrium range.

10. The method for detecting abnormal noise in industrial products based on deep learning according to claim 1, characterized in that, In step S1, the listeners are at least two professional listeners who have passed hearing test certification; for samples where the two listeners' judgments are inconsistent, a third professional listener will make an arbitration judgment; the consistency of the manual listening judgments is not less than 95%.