Automatic detection method and system for noise defect of electroacoustic device

Abstract

The invention relates to the technical field of acoustic performance testing of electronic products, in particular to an AI model-based automatic detection method and system for noise defects of electroacoustic devices such as a loudspeaker and a receiver. The system comprises a hardware acquisition module, a data processing module and a core AI analysis engine, and adopts a two-stage architecture of unsupervised anomaly detection and supervised defect classification: in the first stage, only a non-defective sample is needed to train a generative model, and high-sensitivity anomaly detection is realized by reconstructing a difference spectrogram; in the second stage, accumulated defective product data is utilized to train a classification model, and automatic identification of defect types such as air leakage and particle loosening is completed. According to the invention, cold start deployment, low omission ratio, interpretable diagnosis and model sustainable evolution are realized, and the method is suitable for automatic quality inspection of large-scale electroacoustic devices.

Description

Technical Field

[0001] This invention belongs to the field of audio signal processing and artificial intelligence technology, and in particular relates to an automated detection method and system for noise defects in electroacoustic devices. Background Technology

[0002] In the field of electronic product manufacturing, especially in the production of electroacoustic devices such as loudspeakers (SPK) and receivers (RCV), noise defect detection is a key quality control link to ensure the acoustic performance of products and user experience.

[0003] Currently, the industry commonly uses manual listening or setting fixed thresholds based on simple acoustic indicators (such as total harmonic distortion (THD)) for noise assessment. However, these traditional methods have significant limitations. Manual listening relies heavily on the operator's subjective experience, which is not only inefficient and susceptible to fatigue, but also lacks standardized criteria, resulting in poor consistency of test results. While threshold-based automated detection methods have a certain degree of objectivity, they struggle to effectively distinguish between abnormal noises from the product itself (such as air leakage, particle loosening, mechanical friction, and distortion) and environmental noise or the background noise of the testing system, leading to high rates of missed and false detections.

[0004] Furthermore, existing solutions typically only output a binary "good / defective" result, failing to identify and classify the specific defect types of defective products. This severely restricts the production line's ability to quickly trace quality issues and optimize processes. More critically, if traditional supervised machine learning methods are used to build defect classification models, a large amount of labeled and class-balanced defective sample data is required as a training basis. However, in the early stages of actual production, the number of defective products is small and the defect type coverage is incomplete, making the construction cycle of high-quality models long and costly, which is difficult to meet the needs of rapid deployment and iteration on the production line.

[0005] Therefore, there is an urgent need for an intelligent audio noise detection technology that can achieve high-precision anomaly detection without the need for a large number of defective samples, and has the ability to continuously learn and diagnose defects, so as to break through the bottleneck problem in the current automated testing of electroacoustic devices. Summary of the Invention

[0006] This invention addresses the technical problems of existing electroacoustic device noise detection technologies, which rely on manual listening or acoustic index judgment based on fixed thresholds, including high subjectivity, high false negative rate, high false positive rate, unidentifiable defect type, and difficulty in model cold start. It provides an automated method and system for detecting noise defects in electroacoustic devices. The method constructs a two-stage intelligent analysis architecture of "anomaly detection + defect classification," achieving high-precision anomaly identification with only good product samples. It also utilizes defective product data accumulated during production to automatically determine defect types, thus forming an interpretable, evolvable, and fully automated acoustic quality inspection solution.

[0007] This invention provides an automatic noise defect detection system for electroacoustic devices, comprising three main parts: a hardware acquisition module, a data processing module, and a core AI analysis engine. The hardware acquisition module acquires multi-channel audio response signals of the device under test under standard excitation in an acoustically isolated environment. The data processing module preprocesses the raw audio signals and converts them into a two-dimensional time-spectrum graph. The core AI analysis engine employs a decoupled two-stage deep learning model structure: the first stage performs unsupervised anomaly detection, and the second stage performs supervised defect classification; the two stages work together to complete the entire process from anomaly detection to root cause diagnosis.

[0008] The hardware acquisition module includes a standard excitation system, a microphone array, and a synchronous data acquisition card. The hardware acquisition module is housed in a soundproof enclosure to isolate external environmental noise from interfering with the test results. The device under test (DUT) is placed at the center of the enclosure; the DUT is a speaker or receiver. The standard excitation system inputs a standardized test signal to the DUT; the test signal is pink noise or a logarithmic sweep signal with a frequency range of 20Hz to 20kHz. The microphone array includes at least three omnidirectional, low-self-noise, acoustically calibrated microphones, arranged in a non-collinear spatial distribution, preferably in a triangular layout, to achieve multi-angle, high-fidelity pickup of the DUT's acoustic response.

[0009] Furthermore, the synchronous data acquisition card is connected to the output of each microphone, with a sampling rate greater than or equal to 48kHz and a quantization accuracy greater than or equal to 24bit, ensuring strict synchronization of multiple audio signals in the time domain and phase.

[0010] Furthermore, the data processing module includes a signal preprocessing unit and a time-frequency transformation unit; the signal preprocessing unit performs the following operations in sequence: first, it uses spectral subtraction or adaptive filtering algorithms to suppress fixed background noise in the original audio signal; then, it compensates for the signal amplitude according to the sensitivity calibration parameters of each microphone; next, it normalizes the compensated signal according to the sound pressure level to make the test results of different batches comparable; finally, it standardizes the signal amplitude, mapping it to the [0,1] interval or converting it into a distribution form with zero mean and unit variance; The time-frequency transformation unit performs a short-time Fourier transform (STFT) on the preprocessed time-domain signal to generate a two-dimensional time-frequency spectrum X, where the horizontal axis represents time (unit: seconds), the vertical axis represents frequency (unit: Hz), and the pixel grayscale or color represents the energy intensity (unit: dB) of the corresponding time-frequency point. The STFT parameters are configured as a window length of 1024 points, a Hanning window function, and an inter-frame overlap rate of 75% to balance time resolution and frequency resolution.

[0011] Specifically, the core AI analysis engine includes a first-stage general anomaly detection model and a second-stage refined defect classification model, which together form a decoupled two-stage architecture.

[0012] The first-stage general anomaly detection model is a generative neural network, specifically the generator part of a denoising autoencoder or generative adversarial network. This model is trained using only a large number of good-quality time-frequency spectrograms X as input, aiming to learn the distribution patterns of normal audio signals in the time-frequency domain and possess the ability to reconstruct the input spectrograms with high fidelity. The first-stage general anomaly detection model includes the following steps in the inference phase: Input the spectrum X to be measured into the model to obtain the reconstructed spectrum Y, and calculate the difference spectrum D=|XY|; Perform quantitative analysis on the difference spectrum D: Integrate the energy difference along the frequency axis for each time frame to obtain the "noise-time curve", which is used to locate the specific time when the noise occurs. Integrating the energy difference at each frequency point along the time axis yields the "noise-frequency curve," which is used to identify the frequency bands where noise energy is concentrated. Further calculate the overall anomaly score S, which is the L2 norm or total energy value of all pixel values ​​in the difference spectrum D; if S is greater than a preset threshold, the sample is determined to be a defective product and the second stage model is triggered; at the same time, the original spectrum X, the difference spectrum D, the anomaly score S and the determination result are automatically stored in the defective product database for subsequent model training.

[0013] Furthermore, the second-stage refined defect classification model is an image classification neural network, specifically a convolutional neural network (CNN) or a visual Transformer (ViT). The training data for this model comes from the defect spectrum X accumulated during the operation of the first-stage model, and the corresponding defect category labels are marked by humans or expert systems. The defect categories include, but are not limited to, "air leakage", "particle loosening", "scratching", and "harmonic distortion". During the inference stage, the original spectrum X is only input into the second-stage model after the first-stage model determines that the sample is abnormal. The model outputs the probability distribution and corresponding confidence scores of various defects, completing the fine-grained classification of defective products. As the number of samples and categories in the defective product database continue to expand, the model updates the network parameters through incremental learning mechanisms or periodic full retraining, gradually improving the classification accuracy and generalization ability.

[0014] The present invention provides an automated method for detecting noise defects in electroacoustic devices, which specifically includes the following steps: S1. Start the standard excitation system to drive the device under test to emit sound, and at the same time control the microphone array to synchronously acquire multi-channel audio signals; S2. Input the acquired raw audio signal into the data processing module, and sequentially perform background noise suppression, microphone sensitivity compensation, sound pressure level normalization and amplitude standardization processing; S3. Perform a short-time Fourier transform on the preprocessed audio signal to generate a two-dimensional time-spectrum graph X; S4. Input the time-frequency spectrum X into the first-stage general anomaly detection model to obtain the reconstructed spectrum Y, and calculate the difference spectrum D=|XY|; S5. Based on the difference spectrum D, generate the difference volume-time curve by integrating along the frequency axis, generate the difference volume-frequency curve by integrating along the time axis, and calculate the overall anomaly score S. S6. Determine if the abnormal score S exceeds the preset threshold: If S≤threshold, it is determined to be a good product and the process ends; if S>threshold, it is determined to be a defective product and proceed to S7. S7. Input the original spectrum X into the second-stage refined defect classification model and output the defect category and its confidence level; S8. The system automatically generates a structured test report, which includes the judgment result (good / defective product), defect type, confidence level, volume-time curve, volume-frequency curve, and key time-frequency region slices.

[0015] Furthermore, it also includes: S9. The original spectrum X, difference spectrum D, judgment result and defect label corresponding to the defective product are automatically archived to the defective product database, forming a closed-loop feedback mechanism of "detection → labeling → training → optimization".

[0016] The preset threshold is determined by statistically analyzing the distribution of abnormal scores S on the good product sample set. Specifically, 99.5% of the abnormal scores S of the good product samples is taken as the initial threshold, and it is dynamically adjusted according to newly collected good product data during system operation to adapt to fluctuations in production line processes. The incremental learning mechanism adopts a knowledge distillation or elastic weight solidification strategy to prevent the model from forgetting existing category features during the training of new categories. The structured test report is output in JSON or XML format, supporting integration with enterprise MES or QMS systems to achieve automatic flow and traceability of quality data.

[0017] The technical advantages of this invention are as follows: By employing an unsupervised anomaly detection model that relies solely on good product samples, it solves the problem that traditional supervised learning methods cannot be deployed in the early stages of production lines due to the scarcity of defective product samples, thus enabling rapid cold start of the detection system; by using a generative model to quantify the reconstruction residuals of the time-spectrum diagram, weak, non-steady-state, and non-periodic noise defects (such as transient particle sounds and intermittent air leakage sounds) can be effectively captured, significantly reducing the false negative rate; by introducing a second-stage supervised classification model, the traditional "good / bad" binary classification is expanded to multi-class defect diagnosis, providing clear direction for production process improvement; through the visualization output of the difference spectrum D, the volume-time curve, and the volume-frequency curve, the AI ​​decision-making process becomes interpretable, facilitating engineer review and model debugging; by integrating detection, classification, data archiving, and model updates into a unified system architecture, the entire process is automated, and the model performance is iteratively improved based on continuously accumulated defective product data, forming an intelligent quality inspection system with self-evolution capabilities.

[0018] In summary, the technical solution provided by this invention has achieved a significant transformation in the field of noise detection for electroacoustic devices, from experience-based to data-driven, from qualitative judgment to quantitative diagnosis, and from static models to dynamic evolution. Its technical details are specific, its logic is rigorous, and its feasibility is strong. It is suitable for automated acoustic quality control in large-scale manufacturing scenarios such as consumer electronics, car audio, and communication terminals. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the structure of the tested component in the soundproof box in an embodiment of the present invention; Figure 2 This is a schematic diagram of the automatic detection system for noise defects in electroacoustic devices according to an embodiment of the present invention. Figure 3 This is a flowchart illustrating the automatic detection method for noise defects in electroacoustic devices in an embodiment of the present invention. Figure 4 This is a schematic diagram of the data processing flow of the automatic detection method for noise defects in electroacoustic devices in an embodiment of the present invention. Detailed Implementation

[0020] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] like Figures 1-4 As shown, the overall architecture of the automatic noise defect detection system for electroacoustic devices described in this invention comprises three main parts: a hardware acquisition module, a data processing module, and a core AI analysis engine. The hardware acquisition module, housed in a closed, soundproof enclosure 1, is used to acquire the multi-channel audio response signals of the device under test (DUT) 2 under standard excitation in an acoustically isolated environment. The data processing module preprocesses the raw audio signals and converts them into a two-dimensional time-spectrum graph. The core AI analysis engine employs a decoupled two-stage deep learning model structure. The first stage performs unsupervised anomaly detection, and the second stage performs supervised defect classification. These two stages work together to complete the entire process from anomaly detection to root cause diagnosis.

[0022] In practical deployment, the device under test (DUT) 2 is a loudspeaker (SPK) or receiver (RCV), placed in the center of the soundproof enclosure 1 to ensure that its sound emission direction matches the spatial layout of the microphone array 3. A standardized test signal is input to the DUT 2 via a standard excitation system. This standardized test signal is either pink noise or a logarithmic sweep signal, covering a frequency range of 20Hz to 20kHz, with a duration of 5 seconds. The microphone array 3 consists of three omnidirectional, low-self-noise, acoustically calibrated microphones arranged in a non-collinear equilateral triangle. Each microphone is positioned at a preset distance from the DUT 2, such as 30cm, to achieve multi-angle, high-fidelity pickup of the DUT 2's sound response. A high-precision multi-channel synchronous data acquisition card is connected to the outputs of each microphone, with a sampling rate set to 96kHz and a quantization precision of 24bit, ensuring strict synchronization of the three audio signals in the time domain and phase, and avoiding spectral distortion caused by phase mismatch between channels.

[0023] After acquisition, the raw audio signal is sent to the data processing module. The signal preprocessing unit first uses spectral subtraction based on a statistical model to suppress fixed background noise: by acquiring 10 seconds of background noise samples under no-excitation conditions, calculating its average power spectral density, and subtracting this background spectrum frame by frame from the formal test signal, retaining the effective acoustic response components. Subsequently, according to the sensitivity calibration parameters (unit: mV / Pa) provided by each microphone at the factory, the amplitude of the three signals is compensated to eliminate the gain deviation introduced by individual microphone differences. Next, the compensated signal is normalized according to the reference sound pressure level (94dBSPL@1kHz) to make the test results of different batches and different production lines comparable. Finally, the signal amplitude is standardized and mapped to the [0,1] interval. Specifically, the global mean is subtracted from each frame of the signal, divided by the global standard deviation, and then compressed to the target interval through linear transformation to adapt to the input requirements of the subsequent neural network.

[0024] After preprocessing, the time-frequency transformation unit performs short-time Fourier transform (STFT) on the three audio signals. The STFT parameters are configured with a window length of 1024 points (corresponding to a time window of approximately 10.7ms), a Hanning window function, and a frame shift of 256 points (i.e., an inter-frame overlap rate of 75%), thus achieving a balance between time and frequency resolution. The transform generates three two-dimensional time-frequency spectra. In each spectra, the horizontal axis represents time (0–5 seconds, 188 frames in total), and the vertical axis represents frequency (0–48kHz, 513 frequency points in total). The pixel value represents the energy intensity (in dB, logarithmically compressed) at the corresponding time and frequency point. To improve model robustness, the system stacks the three spectra along the channel dimension to form a three-dimensional tensor of size 513×188×3, which is then input into the core AI analysis engine as the original time-frequency spectra X.

[0025] The core AI analysis engine first calls the first-stage general anomaly detection model. This model employs a denoising autoencoder (DAE) structure, with the encoder consisting of four convolutional layers. Each convolutional layer has a kernel size of 3×3, and the number of channels is 32, 64, 128, and 256 respectively, using ReLU activation function followed by a max pooling layer. The decoder has a symmetrical design, using transposed convolution for upsampling, and finally outputs a reconstructed time-spectrum map Y with the same size as the original time-spectrum map X. During the model training phase, only 100,000 historically accumulated good-quality spectrograms (from mass-production batches of the same model product) are used. The loss function is L1 reconstruction loss, the optimizer is Adam, and the learning rate is set to 1e-4. Training is performed for several rounds until convergence. During the inference phase, the original time-spectrum map X of the current test sample is input into the first-stage general anomaly detection model to obtain the reconstructed time-spectrum map Y. The difference spectrum map calculation module calculates the difference spectrum map D=|XY|, which highlights the abnormal regions that the model cannot reconstruct.

[0026] Subsequently, the anomaly score calculation unit performs quantitative analysis on the difference spectrum D. First, the energy difference is integrated along the frequency axis (vertical axis) for each time frame, i.e., summed over each column of the difference spectrum D, resulting in a one-dimensional vector of length 188, which serves as the "anomaly volume-time curve." This curve clearly reflects the start and end times and duration of the noise. Second, the energy difference is integrated along the time axis (horizontal axis) for each frequency point, i.e., summed over each row of the difference spectrum D, resulting in a one-dimensional vector of length 513, which serves as the "anomaly volume-frequency curve." This curve reveals the concentrated region of anomalous energy in the frequency domain. For example, particle loosening usually manifests as broadband noise in the 2–5 kHz range, while air leakage is concentrated in the low-frequency range (<500 Hz).

[0027] The difference spectrum D is an M×N matrix (M=513 frequency points, N=188 time frames); the matrix element D(i,j) represents the difference energy value at the i-th frequency point and the j-th time frame. Based on this, the overall anomaly score S is calculated, defined as the L2 norm of all pixel values ​​in the difference spectrum D, i.e.: .

[0028] The threshold judgment module compares the anomaly score S with a preset threshold. This threshold is determined by statistically analyzing the distribution of the anomaly scores S across several independent good product samples, taking 99.5% as the initial threshold (e.g., S_th = 125.3). It is dynamically updated every 1000 new good product samples added during system operation to adapt to minor drifts in the production line process. If S ≤ S_th, the product is judged as good, and the testing process ends; if S > S_th, the product is judged as defective, and the second-stage refined defect classification model is triggered.

[0029] The second-stage refined defect classification model employs the VisionTransformer (ViT) architecture. The input is the original temporal spectrum image X (513×188×3), which is first divided into 16×16 image patches. Each patch is flattened and embedded into a 768-dimensional vector space via linear projection, then concatenated with learnable positional codes. This is input to a 12-layer Transformer encoder, and finally, the classification head outputs the probability distribution of four defect categories. The training data for this model comes from several defect spectrum images accumulated during the historical operation of the first-stage general anomaly detection model. These images were manually labeled by acoustic engineers, and the defect categories include "air leakage," "particle loosening," "ringing," and "harmonic distortion." Training uses cross-entropy loss and employs an Elastic Weight Consolidation (EWC) strategy to prevent catastrophic forgetting. When more than 100 new defect samples of a certain type are added, the system automatically starts an incremental training process, fine-tuning only the parameters of the decoder's final layer and the classification head, while retaining the core feature extraction capability.

[0030] After a product is identified as defective, the system inputs the original time-domain spectrum X into the second-stage refined defect classification model, outputting the confidence scores for various defects, such as "particle loosening: 82%" and "air leakage: 12%". Simultaneously, the system automatically generates a structured test report, including the judgment result (good / defective), defect type, confidence score, anomaly volume-time curve, anomaly volume-frequency curve, a visualization of the difference spectrum D, and key time-frequency region slices (such as the time-frequency window where the abnormal energy peak is located). This report is output in JSON format, with fields including "sample_id", "result", "defect_type", "confidence", "time_curve", "freq_curve", and "spectrogram_url", and can be directly integrated with the enterprise's MES system to achieve automatic flow and traceability of quality data.

[0031] In addition, all original time-domain spectrograms X, difference spectrograms D, anomaly scores S, judgment results, and final defect labels after manual review for products judged as defective are automatically archived into the defective product database. The database adopts a distributed storage architecture, supports multi-dimensional retrieval by time, product model, defect type, etc., and periodically triggers model retraining tasks: when the number of defect samples of a certain type increases by more than 20%, the system schedules a GPU cluster to fully retrain the second-stage refined defect classification model, and after verifying the performance improvement of the new model through A / B testing, it automatically switches the online service version, forming a closed-loop feedback mechanism of "detection → labeling → training → optimization".

[0032] It should be noted that the order of the above steps is "logical" or "exemplary," and is not the only possible order of execution.

[0033] In a practical application on a smartphone speaker production line, this system was deployed at the final inspection station, processing approximately 20,000 components daily. Initially, only 30,000 good product samples were used to train the first-stage general anomaly detection model. Within two weeks, it achieved a 98.7% anomaly detection rate (Recall) and a 99.2% good product pass rate (Specificity), far superior to the traditional THD + manual relistening solution (Recall=86.4%, Specificity=92.1%). After three months of operation, the defect database contained 4,800 labeled samples. The second-stage refined defect classification model achieved an average classification accuracy of 91.5% for the four main types of defects, with a "particle loosening" identification accuracy of 94.2%. Engineers analyzed the abnormal sound-frequency curve 13 and discovered that a batch of air leakage defects was concentrated in the 300–400Hz range. Tracing back to the deviation in the dispensing process parameters, timely adjustments reduced the defect rate by 62%. The entire system achieves unmanned operation throughout the process, with single-piece testing time controlled within 8 seconds, meeting the cycle time requirements of high-speed production lines.

[0034] In summary, this invention achieves its goals through the close collaboration of a hardware acquisition module, a data processing module, and a core AI analysis engine, combined with the attached... Figure 1 The complete signal and data flow path shown realizes full-link automation from raw acoustic signal acquisition, time-frequency feature extraction, unsupervised anomaly detection, supervised defect classification to data closed-loop optimization. It solves the core problems in noise detection of electroacoustic devices, such as sample scarcity, high false negative rate, lack of interpretation, and lack of evolution. It has high engineering feasibility and industrial promotion value.

[0035] The above descriptions are merely some embodiments of the present invention. Those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. An automatic detection system for noise defects in electroacoustic devices, characterized in that, It includes a hardware acquisition module, a data processing module, and a core AI analysis engine. The hardware acquisition module is set in a soundproof box and is used to acquire the multi-channel audio response signal of the device under test under standard excitation in an acoustically isolated environment. The data processing module is used to preprocess the raw audio signal and convert it into a two-dimensional time-spectrum graph. The core AI analysis engine adopts a decoupled two-stage deep learning model structure. The first stage performs unsupervised anomaly detection, and the second stage performs supervised defect classification.

2. The automatic detection system for noise defects in electroacoustic devices as described in claim 1, characterized in that, The hardware acquisition module includes a standard excitation system, a microphone array, and a synchronous data acquisition card; the standard excitation system inputs pink noise or a logarithmic sweep signal with a frequency coverage range of 20Hz to 20kHz to the device under test; the microphone array includes at least three microphones, which are spatially distributed in a non-collinear manner; the at least three microphones are arranged in an equilateral triangle, and each microphone is at a preset distance from the device under test.

3. The automatic detection system for noise defects in electroacoustic devices as described in claim 2, characterized in that, The synchronous data acquisition card has a sampling rate greater than or equal to 48kHz and a quantization accuracy greater than or equal to 24bit.

4. The automatic detection system for noise defects in electroacoustic devices as described in claim 1, characterized in that, The data processing module includes a signal preprocessing unit and a time-frequency transformation unit. The signal preprocessing unit sequentially performs background noise suppression, microphone sensitivity compensation, sound pressure level normalization, and amplitude normalization. The time-frequency transformation unit performs a short-time Fourier transform on the preprocessed audio signal to generate a two-dimensional time-frequency spectrum.

5. The automatic detection system for noise defects in electroacoustic devices as described in claim 1, characterized in that, The core AI analysis engine includes a first-stage general anomaly detection model and a second-stage refined defect classification model. The first-stage general anomaly detection model is a denoising autoencoder or the generator part of a generative adversarial network, trained using only the two-dimensional time-spectrum graphs of good products. The second-stage refined defect classification model is a convolutional neural network or a visual Transformer, trained using the two-dimensional time-spectrum graphs of defective products labeled with defect categories.

6. The automatic detection system for noise defects in electroacoustic devices as described in claim 5, characterized in that, The first-stage general anomaly detection model outputs a reconstructed time-frequency spectrum. The system calculates a difference spectrum based on the original two-dimensional time-frequency spectrum and the reconstructed time-frequency spectrum, and generates anomaly volume-time curves and anomaly volume-frequency curves based on the difference spectrum, while simultaneously calculating the overall anomaly score S.

7. The automatic detection system for noise defects in electroacoustic devices as described in claim 6, characterized in that, The anomaly score S is the L2 norm of all pixel values ​​in the difference spectrum. The system compares the anomaly score S with a preset threshold through the threshold judgment module. If S is greater than the preset threshold, it is determined to be a defective product and the second stage of refined defect classification model is triggered.

8. The automatic detection system for noise defects in electroacoustic devices as described in claim 6, characterized in that, The second-stage refined defect classification model outputs defect categories and confidence levels. The defect categories include air leakage, particle loosening, rubbing rings, and harmonic distortion.

9. The automatic detection system for noise defects in electroacoustic devices as described in claim 1, characterized in that, It also includes a defective product database, which stores the original time-spectrum diagrams, difference spectra, judgment results and defect labels of products judged as defective, and performs incremental learning or periodic full retraining on the second-stage refined defect classification model based on the defective product database.

10. An automated method for detecting noise defects in electroacoustic devices, characterized in that, The automatic detection system for noise defects in electroacoustic devices based on claim 1 includes the following steps: S1. Start the standard excitation system to drive the device under test to emit sound, and at the same time control the microphone array to synchronously acquire multi-channel audio signals; S2. Input the acquired raw audio signal into the data processing module and perform background noise suppression, microphone sensitivity compensation, sound pressure level normalization and amplitude standardization in sequence; S3. Perform a short-time Fourier transform on the preprocessed audio signal to generate a two-dimensional time-spectrum graph X; S4. Input the time-frequency spectrum X into the first-stage general anomaly detection model to obtain the reconstructed spectrum Y, and calculate the difference spectrum D=|XY|; S5. Based on the difference spectrum D, generate the difference volume-time curve by integrating along the frequency axis, generate the difference volume-frequency curve by integrating along the time axis, and calculate the overall anomaly score S. S6. Determine if the abnormal score S exceeds the preset threshold: If S≤threshold, it is determined to be a good product and the process ends; if S>threshold, it is determined to be a defective product and proceed to S7. S7. Input the original spectrum X into the second-stage refined defect classification model and output the defect category and its confidence level; S8. The system automatically generates a structured test report, which includes the judgment result, defect type, confidence level, volume-time curve, volume-frequency curve, and key time-frequency region slices.

Images