A method, medium and system for detecting and classifying abnormal noise of an automobile seat driver

Abstract

The application discloses an automobile seat driver abnormal sound detection and classification method, medium and system, relates to the technical field of automobile part quality inspection, and comprises the following steps: collecting audio signals in the forward and reverse rotation process of a seat driver and marking; frame extraction is performed on the samples to form a frame-level feature matrix; an abnormal sound detection model is constructed, the attention score of the frame-level feature matrix is calculated and normalized as a weight, the frame-level features are weighted and aggregated, then mapped to an abnormal probability to realize normal / abnormal judgment; the SHAP analysis is used to calculate the feature contribution vector of the abnormal sample, and a classification template is generated by weighted average according to the abnormal sound subdivision type; after the to-be-tested audio is identified as abnormal by the model, the contribution vector is calculated and compared with each template in terms of similarity to complete the abnormal sound subdivision and classification. The medium and system are used to implement the above method. The application retains abnormal sound time dimension dynamic evolution information, is compatible with audio length fluctuation, has high recall rate and interpretability, and realizes full-process automatic quality inspection.

Description

Technical Field

[0001] This application relates to the field of automotive parts quality inspection technology, and in particular to a method, medium and system for detecting and classifying abnormal noises in automotive seat drive systems. Background Technology

[0002] As a crucial component of vehicle comfort and safety, the operational stability of automotive seat actuators directly impacts user experience and corporate reputation. However, during actual production, due to mechanical wear, assembly errors, or poor lubrication, seat actuators may generate various types of abnormal noises, such as friction sounds, clicking sounds, and whistling sounds. Therefore, abnormal noise detection is an essential step before automotive seat actuators leave the factory and has long been a challenge for many manufacturers.

[0003] Currently, most manufacturers rely on manual listening in soundproof rooms for noise detection. This method has the following drawbacks: different quality inspectors have varying levels of skill, hearing fatigue, and standards for judging abnormal sounds, making it easy to miss or misjudge, and making it difficult to guarantee consistent test results. This results in high customer complaint and return rates for seat drives after they leave the factory. Manual inspection is slow and cannot meet the rapid testing needs of large-scale production lines, affecting production cycle and delivery speed. Maintaining soundproof room facilities and professional noise detection personnel long-term limits business development. Manual inspection cannot trace and analyze the acoustic data of seat drives, nor can it support subsequent root cause analysis and process improvement.

[0004] In recent years, artificial intelligence technology has driven the upgrading of industrial quality inspection towards "intelligent and refined" processes, and abnormal noise detection methods based on machine learning have become a research hotspot. However, existing methods still have the following shortcomings: Most existing methods use global statistical features of the entire audio segment, such as the mean and standard deviation, losing information about the dynamic evolution of abnormal noise over time and making it difficult to capture instantaneous abnormal noise patterns; existing deep learning black-box models struggle to explain to process engineers "why an abnormality is identified," limiting their credibility and widespread application in industrial settings; existing methods mostly focus on the binary classification problem of "presence or absence of abnormal noise," lacking refined identification and interpretable attribution of abnormal noise types, thus limiting their application value in process optimization and fault tracing, and failing to meet the full-process requirements of "detection-diagnosis-improvement" in intelligent industrial manufacturing. Summary of the Invention

[0005] One objective of this application is to provide a method for detecting and classifying abnormal noises from automotive seat drives that can address at least one of the deficiencies in the aforementioned background art.

[0006] Another objective of this application is to provide a medium for implementing the above-described method for detecting and classifying abnormal noises in automotive seat drives.

[0007] Another object of this application is to provide a system for implementing the above-described method for detecting and classifying abnormal noises in automotive seat drives.

[0008] To achieve at least one of the above objectives, one aspect of this application provides a method for detecting and classifying abnormal noises from a car seat drive system, comprising the following steps: acquiring audio signals of the forward and reverse rotation processes of the seat drive system in a standard quiet room environment; independently labeling the forward and reverse rotation processes to indicate whether there is abnormal noise and the sub-type of the abnormal noise, forming an audio dataset with multi-label annotations; performing frame segmentation on the samples in the audio dataset and extracting multi-dimensional features from each frame to form a frame-level feature matrix; constructing and training an abnormal noise detection model based on attention pooling, using the frame-level feature matrix corresponding to the sample as input, calculating the attention score for each frame and normalizing it into attention weights, and using the attention weights... The multi-dimensional features corresponding to the frame-level feature matrix are weighted to obtain aggregated features, which are then mapped to anomaly probabilities for determining whether the audio is normal or abnormal. For samples in the audio dataset that are determined to be abnormal, the SHAP value analysis method is used to calculate the multi-dimensional feature contribution vector of each sample. For each sub-category of abnormal audio, the multi-dimensional feature contribution vectors of all samples belonging to that sub-category are weighted and averaged to generate a classification template for that sub-category. For the audio signal to be tested that is identified as abnormal by the abnormal audio detection model, the multi-dimensional feature contribution vector is calculated and compared with the similarity of each classification template to perform sub-category classification of the audio signal to be tested.

[0009] Preferably, the abnormal sound detection model includes an attention scoring network, a weight normalization layer, a weighted summation layer, and a linear mapping layer. The attention scoring network is used to calculate the attention score for each frame of the frame-level feature matrix. The weight normalization layer normalizes the attention score into attention weights using the Softmax function. The weighted summation layer uses the obtained attention weights to perform weighted summation on the multi-dimensional features corresponding to all frames of the frame-level feature matrix to obtain aggregated features. The aggregated features are concatenated with the directional features and then input into the linear mapping layer, which maps them to anomaly probabilities using the Sigmoid function.

[0010] Preferably, the expression for calculating the anomaly probability P is as follows: ; ; ; In the formula, This represents the Sigmoid function. This represents feature concatenation, where N represents the number of frames corresponding to the frame-level feature matrix. This represents the attention weights corresponding to the t-th frame in the frame-level feature matrix. This represents the eigenvector corresponding to the t-th frame in the frame-level feature matrix. Indicates directional characteristics, Indicates the output layer weights. Both b and b represent bias terms. and These represent the attention scores corresponding to the t-th and j-th frames of the frame-level feature matrix, respectively. This represents the attention output weight vector. This represents the activation function. This represents the frame feature mapping matrix.

[0011] Preferably, the classification templates corresponding to each sub-type of abnormal sound are deployed on the production line. After each batch of test samples is classified into different sub-types of abnormal sound on the production line, a stratified confidence weighting strategy is used to screen the effective sample pool. The batch-weighted SHAP mean vector is calculated based on the multi-dimensional feature contribution vectors corresponding to each sample in the effective sample pool. Based on the obtained batch-weighted SHAP mean vector, the classification templates used for the classification of abnormal sound sub-types of the next batch of test samples are updated using an adaptive smoothing factor.

[0012] Preferably, the screening of the effective sample pool includes the following process: obtaining the anomalous probability corresponding to the audio anomalous sample in the current batch based on the anomalous sound detection model; calculating the cosine similarity between the audio anomalous sample in the current batch and the classification template corresponding to each anomalous sound sub-type; if the cosine similarity corresponding to the audio anomalous sample exceeds a preset first similarity threshold, and the corresponding anomalous probability exceeds a preset first probability threshold, the audio anomalous sample is classified as a high-confidence sample; if the cosine similarity corresponding to the audio anomalous sample is between a preset first similarity threshold and a preset second similarity threshold, and the corresponding anomalous probability is between a preset first probability threshold and a preset second probability threshold, the audio anomalous sample is classified as a medium-confidence sample; if the cosine similarity corresponding to the audio anomalous sample is lower than a preset second similarity threshold, or the corresponding anomalous probability is lower than a preset second probability threshold, the audio anomalous sample is classified as a low-confidence sample; and summarizing the high-confidence samples and medium-confidence samples to construct an effective sample pool; wherein, the first similarity threshold is greater than the second similarity threshold, and the first probability threshold is greater than the second probability threshold.

[0013] Preferred batch-weighted SHAP mean vector The calculation expression is as follows: ; In the formula, M represents the total number of samples in the effective sample pool. This represents the sample confidence level of the i-th sample in the valid sample pool. The sample confidence level of a high-confidence sample is 1, and the sample confidence level of a medium-confidence sample is 0.5. This represents the anomaly probability of the i-th sample in the valid sample pool. This represents the multi-dimensional feature contribution vector of the i-th sample in the effective sample pool.

[0014] The preferred expression for updating the classification template is: ; ; In the formula, This represents a classification template used for further subdividing and categorizing abnormal sounds in the next batch of samples to be tested. This represents the classification template used for further subdividing and categorizing heterophonic sounds in the current k batches of test samples. This represents the adaptive smoothing factor corresponding to the current k batches of samples to be tested. This represents the batch-weighted SHAP mean vector corresponding to the current k batches of samples to be tested. This represents the preset base smoothing factor. This represents the average cosine similarity of the previous batch of samples to be tested.

[0015] Preferably, when updating the classification template, to avoid the bias of the classification template being excessively dominated by samples of a single subtype of abnormal sounds, an adaptive anomaly ratio constraint mechanism is set up. The specific content is as follows: Real-time monitoring of the sample proportion of each subtype of abnormal sounds in the effective sample pool during the update process, and recording the number of effective samples of each subtype of abnormal sounds in each batch of test samples; When the sample proportion of an abnormal sound of a certain subtype of abnormal sounds in the effective sample pool shifts for multiple consecutive batches and the shift exceeds a preset threshold, it is determined that the classification template corresponding to the subtype of abnormal sounds has a risk of bias, triggering the reconstruction of the classification template; When reconstructing the classification template, the classification template is recalculated from the current effective sample pool according to the sample confidence and anomaly probability weights, and the historical cumulative bias is reset.

[0016] Another aspect of this application provides a system for implementing the above-described method for detecting and classifying abnormal noises from an automotive seat drive unit. This system is deployed on a production line equipped with an abnormal noise detection station and an abnormal noise classification station. The system includes an audio acquisition module, an abnormal noise detection module, and an abnormal noise classification module. The audio acquisition module is deployed at the abnormal noise detection station and is used to acquire audio signals generated during the operation of the seat drive unit. The abnormal noise detection module is deployed at the abnormal noise detection station and is used to receive the audio signal to be tested and extract a frame-level feature matrix. Based on the extracted frame-level feature matrix, the system determines whether the audio signal is normal or abnormal. If the abnormal sound is detected... If the audio signal is normal, the current seat driver is sent to the qualified area; otherwise, the current seat driver is sent to the abnormal sound classification station, and the characteristic sequence and abnormal probability of the corresponding audio signal are sent to the abnormal sound classification module. The abnormal sound classification module is deployed at the abnormal sound classification station on the production line and has built-in classification templates for each sub-type of abnormal sound. The abnormal sound classification module calculates the multi-dimensional feature contribution vector of the corresponding audio signal based on the received feature sequence and abnormal probability, compares the similarity with each classification template, outputs the specific sub-type of abnormal sound, and sends the seat driver to the corresponding classification area according to the classification result.

[0017] Another aspect of this application provides a computer-readable storage medium storing a computer program; when the computer program is executed by a processor, it implements the above-described method for detecting and classifying abnormal noises from a car seat driver.

[0018] Compared with the prior art, the beneficial effects of this application are as follows: (1) This application employs a frame-segmentation processing strategy to finely segment the audio signal and extract multi-dimensional features within each frame to form a frame-level feature matrix that retains the time-axis order. This approach overcomes the shortcomings of traditional methods that use global statistical features of the entire audio segment and lose time-varying information about abnormal sounds. It can completely record the fluctuation of feature indicators with the time axis. Especially for abnormal sound types such as stuttering sounds, which have intermittent and sudden time-domain features, frame-level features can effectively capture their transient change patterns, providing a richer and more accurate feature basis for subsequent abnormal sound detection and classification.

[0019] (2) This application introduces a trainable attention pooling layer into the abnormal sound detection model. By calculating the attention score for each frame and obtaining the attention weight after normalization, the attention weight is obtained by weighted summation of all frame-level feature vectors, and the frame-level feature matrix with variable time dimension is adaptively compressed into aggregated features with fixed dimension. This mechanism does not require fixed-length pruning or zero-padding of the audio, and can be compatible with the slight changes in audio duration caused by trigger fluctuations, individual differences in seat drivers, etc. in actual production line acquisition, which significantly enhances the robustness and generalization ability of the model under complex production line conditions.

[0020] (3) After determining the audio anomalies of the samples, this application uses the SHAP value analysis method to calculate the multi-dimensional feature contribution vector of each sample, and constructs classification templates for various sub-types of abnormal sounds based on this. This can clearly characterize the contribution direction and degree of each dimension feature in the model judgment, providing process engineers with traceable and understandable basis for abnormal sound attribution. At the same time, by comparing the similarity between the SHAP contribution vector of the test sample and each classification template, the abnormal sound sub-classification is achieved, making the classification results interpretable, which helps technicians quickly locate the physical root cause of abnormal sounds and effectively supports subsequent process improvement and fault tracing. Attached Figure Description

[0021] Figure 1 This is a schematic diagram illustrating the working steps of the method for detecting and classifying abnormal noises from automotive seat actuators in this application.

[0022] Figure 2 This is a schematic diagram of the overall workflow of the method for detecting and classifying abnormal noises from automotive seat drive systems in this application.

[0023] Figure 3 This is a schematic diagram of the system for implementing the method for detecting and classifying abnormal noises from automotive seat actuators in this application. Detailed Implementation

[0024] The present application will now be further described in conjunction with specific embodiments. It should be noted that, in the description of this specification, the use of terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicates that the specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms should not be construed as necessarily referring to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0025] In the description of this application, it should be noted that the terms "center", "lateral", "longitudinal", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise", etc., which indicate the orientation and positional relationship based on the orientation or positional relationship shown in the accompanying drawings, are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and should not be construed as limiting the specific protection scope of this application.

[0026] It should be noted that the terms "first," "second," etc., in the specification and claims of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0027] In this application, unless otherwise expressly specified and limited, the terms "installation," "connection," "joining," and "fixing," etc., should be interpreted broadly. For example, they can refer to a connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0028] In this application, unless otherwise expressly specified and limited, "above" or "below" the second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of the second feature includes the first feature being directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature includes the first feature being directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0029] The terms “comprising” and “having”, and any variations thereof, in the specification and claims of this application are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or device.

[0030] One aspect of this application provides a method for classifying and detecting abnormal noises from automotive seat actuators, such as... Figure 1 As shown, one preferred embodiment includes the following steps: S100: Audio signals of the seat drive's forward and reverse rotation processes are collected in a standard quiet room environment. The forward and reverse rotation processes are independently labeled to indicate whether there are abnormal noises and the sub-types of abnormal noises, forming an audio dataset with multi-label annotations.

[0031] It should be noted that a standard soundproof room environment generally refers to an acoustic testing environment with a background noise level ≤20dB. This is used to eliminate interference from environmental noise on the acquired signals, ensuring the accuracy of subsequent feature extraction and model judgment. Because the mechanical transmission path, load characteristics, and gear meshing state of a car seat drive differ in the forward and reverse directions, its acoustic characteristics also differ significantly. Therefore, it is necessary to label the forward and reverse processes separately. This step provides high-quality benchmark data for subsequent supervised learning by establishing a multi-label, finely annotated audio dataset. This audio dataset not only distinguishes between normal and abnormal audio but also accurately labels the specific types of abnormal sounds, enabling the subsequent model to perform fine-grained abnormal sound classification.

[0032] S200: Perform frame segmentation on the samples in the audio dataset and extract multi-dimensional features from each frame to form a frame-level feature matrix.

[0033] It should be noted that existing abnormal sound detection methods mostly use global statistical features of the entire audio segment, such as mean and standard deviation, which loses the dynamic evolution information of abnormal sounds over time and makes it difficult to capture instantaneous abnormal sound patterns. This step transforms the raw audio signal into a structured frame-level feature matrix, thus fully preserving the dynamic evolution information of each feature index over time.

[0034] S300: Construct and train an abnormal sound detection model based on attention pooling. Take the frame-level feature matrix corresponding to the sample as input, calculate the attention score of each frame and normalize it into attention weights. Use the attention weights to weight the multi-dimensional features corresponding to the frame-level feature matrix to obtain aggregated features and map them into abnormal probabilities, which are used to determine whether the audio of the sample is normal or abnormal.

[0035] It's important to understand that attention pooling is a learnable sequence-to-vector compression method. Its core idea is to allow the model to automatically learn the importance of each frame's features to the final judgment, then perform a weighted sum of all frame features, thus compressing the variable-length frame-level feature sequence into a fixed-length vector. The attention score is a scalar value calculated by the attention scoring network for each frame. A higher score indicates greater importance of the frame in the overall audio, meaning the more crucial the abnormal sound information contained in that frame. The attention weights are probability distributions obtained by normalizing the attention scores of each frame using the Softmax function, ensuring that the sum of the weights of all frames is 1; a higher weight indicates a greater contribution of that frame to the weighted summation. The anomaly probability is a value between 0 and 1, obtained by mapping the model's output layer using the Sigmoid function, used to quantify the likelihood of a sample being judged as abnormal; the closer the anomaly probability is to 1, the more likely the sample is abnormal, and the closer it is to 0, the more likely it is normal.

[0036] S400: For samples in the audio dataset that are determined to be audio abnormal, the SHAP value analysis method is used to calculate the multi-dimensional feature contribution vector of each sample. For each sub-category of abnormal sound, the multi-dimensional feature contribution vectors of all samples belonging to that sub-category of abnormal sound are weighted and averaged to generate the classification template of that sub-category of abnormal sound.

[0037] It's important to understand that SHAP value analysis is a model interpretation method based on Shapley values ​​in game theory. It calculates a contribution value for each feature of each sample, representing the magnitude and direction of that feature's influence in classifying the sample into a particular category. The core advantage of SHAP values ​​is that the sum of the feature contribution values ​​equals the difference between the model's predicted output and the baseline value; and when model changes increase the influence of a feature, the SHAP value of that feature does not decrease. A multi-dimensional feature contribution vector is a vector formed by arranging the SHAP contribution values ​​of each feature dimension in a fixed order after applying SHAP value analysis to an anomaly detection model. Each dimension of this vector corresponds to a feature, and the sign of the value indicates the direction in which the feature prompts the model to classify it as an anomaly (positive value indicates a positive contribution, negative value indicates a negative contribution). The absolute value represents the degree of contribution.

[0038] S500: For audio signals identified as audio abnormalities by the abnormal sound detection model, calculate multi-dimensional feature contribution vectors and compare their similarity with each classification template to perform abnormal sound sub-classification of the audio signals under test.

[0039] Understandably, this application employs a frame-segmentation strategy to finely segment the audio signal and extract multi-dimensional features within each frame, forming a frame-level feature matrix that preserves the time-axis order. This approach overcomes the shortcomings of traditional methods that use global statistical features of the entire audio segment, resulting in the loss of time-varying information about abnormal sounds. It can completely record the fluctuation of feature indicators over time. Especially for abnormal sound types such as stuttering sounds, which have intermittent and sudden temporal characteristics, frame-level features can effectively capture their transient change patterns, providing a richer and more accurate feature foundation for subsequent abnormal sound detection and classification.

[0040] This application introduces a trainable attention pooling layer into the abnormal sound detection model. By calculating the attention score for each frame and normalizing it to obtain the attention weight, a weighted sum of all frame-level feature vectors is performed, adaptively compressing the frame-level feature matrix with a variable time dimension into a fixed-dimensional aggregated feature. This mechanism eliminates the need for fixed-length pruning or zero-padding of the audio, and can accommodate minor variations in audio duration caused by trigger fluctuations, individual differences in seat drivers, etc., in actual production line acquisition. This significantly enhances the model's robustness and generalization ability under complex production line conditions.

[0041] After determining audio anomalies in the samples, this application employs the SHAP value analysis method to calculate the multi-dimensional feature contribution vector of each sample, and constructs classification templates for various subcategories of abnormal sounds based on this vector. This clearly characterizes the contribution direction and degree of each feature in the model's determination, providing process engineers with a traceable and understandable basis for attributing abnormal sounds. Simultaneously, by comparing the similarity between the SHAP contribution vector of the test sample and each classification template, the abnormal sound subcategorization is achieved, making the classification results interpretable. This helps technicians quickly locate the physical root cause of abnormal sounds, effectively supporting subsequent process improvements and fault tracing.

[0042] In a specific embodiment, when performing step S100, the audio signal is acquired using an industrial-grade high-sensitivity microphone, with a preferred sampling rate of 44.1kHz, a preferred bit depth of 16bit, and a preferred WAV lossless format. During audio signal acquisition, a 0.5-second silence interval is typically provided between the forward and reverse rotation processes in each audio signal, which can be accurately segmented using the energy threshold method. When labeling each audio signal, experienced listening personnel determine whether there are any abnormal sounds in the forward and reverse rotation segments; if abnormal sounds are found, the type of abnormal sound is further labeled, including but not limited to stuttering, friction, and howling, thus forming an audio dataset with multiple labels. The acquired samples for each model of seat actuator must cover two main categories: qualified and unqualified parts. The judgment rule is: if there is an abnormal sound in either the forward or reverse rotation segment, the seat actuator is judged as an unqualified part, i.e., an NG part; only when there is no abnormal sound in both the forward and reverse rotation processes is the seat actuator judged as a qualified part, i.e., an OK part. Among them, qualified parts correspond to normal audio signals, while unqualified parts correspond to abnormal audio signals.

[0043] In a specific embodiment, when performing step S200, as follows: Figure 2 As shown, each audio signal acquired in step S100 needs to be divided into two independent audio segments according to forward and reverse rotation. Then, each audio segment is processed by frame segmentation, and multi-dimensional features of each frame are extracted to form a frame-level feature matrix.

[0044] It's important to understand that framing an audio segment involves dividing it into several equal-length, partially overlapping frames along the timeline. Each frame typically lasts 20ms to 50ms, with an overlap of 50% to 75% between adjacent frames. The purpose of framing is to ensure short-term stationarity of the audio signal within each frame, allowing for independent signal analysis of each frame. Multi-dimensional feature extraction for each frame refers to extracting quantitative indicators that characterize the physical and perceptual properties of the sound signal from multiple dimensions, including the time domain, frequency domain, time-frequency domain, and psychoacoustics.

[0045] It should be understood that time-domain characteristics include peak value, peak-to-peak value, RMS, kurtosis, skewness, crest factor, impulse factor, margin factor, and waveform factor. Among these, the peak value is the maximum amplitude of the signal within a frame, reflecting the instantaneous maximum impact intensity; the peak-to-peak value is the difference between the maximum and minimum amplitudes of the signal within a frame, reflecting the dynamic range of the signal; RMS is the root mean square value of the signal within a frame, reflecting the average energy level of the signal; kurtosis reflects the steepness of the signal distribution curve and is sensitive to impact anomalies; skewness reflects the asymmetry of the signal distribution and can be used to distinguish between positive and negative impacts; the crest factor is the ratio of peak value to RMS, reflecting the peak characteristics of the signal; the impulse factor is the ratio of peak value to absolute average value and is sensitive to impulse anomalies; the margin factor is the ratio of peak value to root mean square amplitude and is sensitive to wear-related faults; and the waveform factor is the ratio of RMS to absolute average value, reflecting the overall shape characteristics of the signal waveform.

[0046] Frequency domain characteristics include the spectral centroid, spectral bandwidth, and band energy ratio. The spectral centroid is the first moment of the power spectrum, reflecting the concentrated frequency location of spectral energy, and can be used to determine the frequency of abnormal sounds. The spectral bandwidth is the second moment of the power spectrum, reflecting the degree of dispersion of spectral energy. The band energy ratio is calculated by dividing the spectrum into multiple frequency bands, such as 0~2kHz, 2~5kHz, and 5~20kHz, and calculating the proportion of energy in each band to the total energy; this can be used to locate the frequency range of abnormal sound sources.

[0047] The time-frequency domain features include Mel-frequency cepstral coefficients (MFCC), first-order MFCC difference, second-order MFCC difference, and wavelet packet energy. Among these, the Mel-frequency cepstral coefficients are cepstral coefficients extracted after nonlinear mapping of the power spectrum using a Mel-scale filter bank, simulating the characteristics of human auditory perception and providing a good representation of the subjective auditory perception of seat drive noise. The first-order and second-order MFCC differences reflect the rate of change and acceleration of the Mel-frequency cepstral coefficients over time, respectively, and are used to capture the dynamic evolution characteristics of the noise. The wavelet packet energy, through multi-level wavelet packet decomposition of the signal, calculates the time-series sequence of the energy of each terminal node, enabling the capture of the local time-frequency characteristics of the noise at different time-frequency resolutions.

[0048] Psychoacoustic features include loudness, sharpness, and roughness. Loudness, based on the ISO 532-1 Zwicker method, is calculated from the power spectrum via Bark critical band mapping and a specific loudness integral, reflecting the perceived sound intensity. Sharpness, based on the DIN 45692 standard, is calculated using a specific loudness distribution and high-frequency weighted moments, reflecting the proportion of high-frequency components in the sound, used to characterize high-frequency abnormal sounds such as howling. Roughness, based on a simplified Zwicker model, is calculated from the Bark critical band envelope modulation depth and dominant modulation frequency estimation, reflecting the intensity of temporal fluctuations in the sound, used to characterize abnormal sounds with fluctuating characteristics such as friction sounds. These three psychoacoustic features share the Bark critical band energy hub, allowing for three-way derivation with only a single Bark critical band mapping calculation, avoiding repeated Fast Fourier Transform calculations.

[0049] By concatenating all the features extracted from each frame in a fixed order and arranging them in chronological order, the required frame-level feature matrix can be formed. Where R represents the set of real numbers, N represents the number of frames corresponding to the frame-level feature matrix, and C represents the total dimension of the frame-level feature matrix. This matrix fully records the fluctuation of each feature index over time. For example, when the frame shift is 256, the sampling rate is 44.1kHz, and the audio clip length is 2.5s, the number of frames N is approximately 431 frames, and the total dimension C is 64 dimensions, including 9 dimensions in the time domain, 5 dimensions in the frequency domain, 47 dimensions in the time-frequency domain, and 3 dimensions in the psychoacoustic domain; the final frame-level feature matrix is ​​obtained. .

[0050] In a specific embodiment, when performing step S300, as follows: Figure 2 As shown, the abnormal sound detection model includes an attention scoring network, a weight normalization layer, a weighted summation layer, and a linear mapping layer.

[0051] For the attention scoring network, it is configured to compute the attention score for each frame of the frame-level feature matrix. Specifically, for the frame-level feature matrix F = [f1, f2, ..., f...], ... t , ..., f N ] T ; where f t Let t represent the feature vector corresponding to the t-th frame in the frame-level feature matrix. The attention scoring network calculates a scalar score for the feature vector corresponding to each frame. Essentially, the attention scoring network is a single-hidden-layer fully connected neural network. Its input is the feature vector of each frame, and its output is the attention score for that frame. A higher attention score indicates a higher importance of that frame in the overall audio, meaning the more crucial the anomaly information contained in that frame. The parameters of the attention scoring network are automatically optimized through backpropagation during model training, without requiring manual setting of frame importance rules. For ease of understanding, the attention score will be represented by a specific expression below.

[0052] Specifically, the attention score corresponding to the t-th frame of the frame-level feature matrix. The calculation expression is as follows: .

[0053] In the formula, This represents the attention output weight vector, which consists of the trainable parameters of the model. Let W represent the activation function, W represent the frame feature mapping matrix, and represent the trainable parameters of the model. Let represent the eigenvector corresponding to the t-th frame of the frame-level feature matrix, b represent the bias term, and represent the trainable parameters of the model.

[0054] The weight normalization layer is configured to normalize the attention scores into attention weights using the Softmax function. Specifically, for the attention scores of N frames, the weight normalization layer maps them to a probability distribution, i.e., the attention weight distribution. Since the Softmax function is dynamically calculated for the N input elements, this layer can adaptively output a corresponding number of weight distributions regardless of the value of the number of frames N. This is key to the ability of attention pooling to handle variable-length sequences. For ease of understanding, the attention weights will be represented by specific expressions below.

[0055] Specifically, for the attention weights corresponding to the t-th frame in the frame-level feature matrix... The calculation expression is as follows: .

[0056] In the formula, represents the attention score corresponding to the j-th frame of the frame-level feature matrix, and N represents the number of frames corresponding to the frame-level feature matrix.

[0057] For the weighted summation layer, the configuration involves using the obtained attention weights to perform a weighted summation of the multi-dimensional features corresponding to all frames in the frame-level feature matrix, resulting in aggregated features; these aggregated features are then combined with the directional features. (Forward rotation equals +1, reverse rotation equals -1) These are concatenated to obtain the final aggregated global feature vector. Specifically, the weighted summation layer sums the feature vectors of each frame according to their attention weights, resulting in a fixed-dimensional vector that aggregates information from all frames. This aggregated feature is essentially a compressed extraction of effective information from the entire audio sequence. Frames with higher attention weights account for a larger proportion of the aggregated feature, allowing the model to focus on the most critical abnormal sound segments in the audio. The introduction of directional features is due to the significant differences in acoustic characteristics between the forward and reverse rotation of the seat drive. Using directional features as additional input helps the model establish different judgment criteria for audio from different operating directions, improving detection accuracy. For ease of understanding, the global feature vector will be represented by a specific expression below.

[0058] Specifically, the expression for the global feature vector z is as follows: .

[0059] In the formula, This represents feature concatenation, where N represents the number of frames corresponding to the frame-level feature matrix. This represents the attention weights corresponding to the t-th frame in the frame-level feature matrix. This represents the eigenvector corresponding to the t-th frame in the frame-level feature matrix. Indicates directional characteristics.

[0060] For the linear mapping layer, the configuration involves inputting the global feature vector into the linear mapping layer and mapping it to anomaly probabilities using the Sigmoid function. Specifically, after concatenating the aggregated features and directional features, the linear mapping layer uses a linear transformation to map the concatenated features to a scalar, which is then compressed to the (0, 1) interval by the Sigmoid function to obtain the final anomaly probability. For ease of understanding, the anomaly probability will be represented by a specific expression below.

[0061] Specifically, the expression for calculating the anomaly probability P is as follows: .

[0062] In the formula, Let z represent the sigmoid function, and z represent the global eigenvector. This represents the output layer weights, which are trainable parameters of the model. This represents the bias term, which is a trainable parameter of the model.

[0063] It should be noted that all the trainable parameters of the above models are automatically optimized during the model training process through the backpropagation algorithm and optimizer, without the need for manual setting. For ease of understanding, the training process of the abnormal sound detection model will be described in detail below.

[0064] In one specific embodiment, training the abnormal sound detection model includes the following process: S310: Divide the audio dataset with multi-label annotations into training set, validation set and test set.

[0065] It should be understood that the training set is used for learning the model's parameters, that is, continuously updating the trainable parameters of the network through forward and backward propagation; the validation set is used for model hyperparameter tuning, including the selection of hyperparameters and the triggering of early stopping strategies; the test set is used for the final performance evaluation after the model training is completed. The test set data does not participate in any form of parameter update or hyperparameter selection during the training process to ensure the objectivity of the evaluation results.

[0066] When dividing the training, validation, and test sets, the corresponding sample ratios are different. Generally, the training set has the largest number of samples, followed by the validation set, and the test set has the smallest. For example, the sample ratios for the training, validation, and test sets can be 7:2:1. To avoid data skew causing a decrease in the accuracy of subsequent SHAP classification, it is necessary to ensure that the proportions of the three types of abnormal sounds in the outlier samples of the training, validation, and test sets are consistent. For example, in the training, validation, and test sets, the proportions of the three types of abnormal sounds—stuttering sounds, friction sounds, and howling sounds—are all 1:1:1.

[0067] S320: Perform feature standardization on the frame-level feature matrix corresponding to each sample in the training set.

[0068] It should be noted that because the frame-level feature matrix F contains features with different dimensions such as time domain, frequency domain, and psychoacoustic, its numerical range is extremely large. Directly inputting it into the attention network would lead to large numerical features dominating the scoring calculation. Furthermore, due to the fluctuation in the duration of production line audio, the number of frames N is not fixed, making it impossible to directly apply the traditional StandardScaler standardization, which requires a fixed sample-feature matrix dimension. Therefore, it is necessary to perform global standardization of all frame-level feature matrices in the training set according to their dimensions before attention pooling.

[0069] For ease of understanding, the following example uses the j-th feature in the t-th frame of the frame-level feature matrix to illustrate the specific normalization process: .

[0070] In the formula, This represents the j-th feature in the t-th frame of the standardized frame-level feature matrix. This represents the j-th feature in the t-th frame of the unnormalized frame-level feature matrix. and Let represent the mean and standard deviation of all frames in the j-th dimension of the frame-level feature matrix of all samples in the training set, respectively.

[0071] S330: The frame-level feature matrix of the training set samples is used as the input of the abnormal sound detection model. The attention scoring network calculates the attention score for each frame of the frame-level feature matrix. The weight normalization layer normalizes the attention scores of all frames into attention weights using the Softmax function. The weighted summation layer uses the attention weights to perform weighted summation on the multi-dimensional features corresponding to N frames of the frame-level feature matrix to obtain aggregated features. These aggregated features are then concatenated with the directional features to form a global feature vector. The linear mapping layer maps the global feature vector to anomaly probabilities through linear transformation and the Sigmoid function.

[0072] S340: During training, cross-entropy loss with class weights is used, and the weight of abnormal sound samples is greater than that of normal samples. After training, the false negative rate of abnormal sound samples in the test set is ≤0.1%, and the false positive rate is ≤5%.

[0073] It should be understood that cross-entropy loss with class weights is based on standard cross-entropy loss, but assigns different weights to different classes to address the class imbalance problem. Since the number of defective parts in industrial scenarios is far less than the number of qualified parts, using standard cross-entropy loss would cause the model to favor predicting qualified parts and severely miss defective parts. Therefore, in this embodiment, it is preferable to set the weight of abnormal noise samples to 3-5 times that of normal samples. This forces the model to pay more attention to minority class abnormal noise samples during training, thereby effectively improving the recall rate of abnormal noise samples.

[0074] It should also be understood that the false negative rate refers to the proportion of defective parts that are mistakenly identified as acceptable parts, while the false positive rate refers to the proportion of acceptable parts that are mistakenly identified as defective parts. In industrial quality inspection scenarios, missing a single defective product may lead to customer complaints or even recall risks, while falsely identifying a acceptable product only incurs a small amount of rework costs. Therefore, when training the model, to ensure the strict principle of "better to falsely identify than to falsely identify" in industrial quality inspection scenarios, the abnormal sound sample recall rate is used as the core evaluation indicator. The abnormal sound sample recall rate is the proportion of correctly detected abnormal sound samples out of all actual abnormal sound samples, and its sum with the false negative rate is 1.

[0075] It's important to understand that while model parameters are updated based on the loss function on the training set during each training cycle, a decrease in loss on the training set does not necessarily mean improved performance on unseen data. For example, overfitting can cause the model to become overly adapted to noise and specific patterns in the training set, leading to decreased generalization ability. Therefore, in this embodiment, the final model parameters cannot be selected solely based on performance on the training set. Thus, a validation set is specifically set up for selecting the final model parameters. Specifically, after each training cycle, the current model is evaluated on the validation set. Based on the false negative rate of abnormal sound samples in the validation set, the final model parameters are selected from multiple candidate models generated in different training cycles for validation on the test set.

[0076] It's also important to know that during model training, a total number of training epochs is typically set to terminate the training process. However, during training, the model may converge before the total number of training epochs is reached. Therefore, this embodiment includes an early stopping mechanism for model training. Specifically, after each training epoch, the current model is evaluated on a validation set, and the false negative rate of abnormal sound samples in the validation set is monitored as the training epochs change. If the false negative rate does not increase within a preset number of consecutive training epochs, training is terminated early.

[0077] In a specific embodiment, when performing step S400, as follows: Figure 2 As shown, the construction of the classification template includes the following process: The multi-dimensional features corresponding to the frame-level feature matrix are arranged in a set order to form a global feature naming list. Therefore, when calculating the multi-dimensional feature contribution vector using the SHAP value analysis method, the generation strictly follows the order of the global feature naming list, ensuring that the classification template is completely consistent with the number of dimensions and feature semantics of the audio signal data under test, avoiding classification errors due to disordered order. After obtaining the multi-dimensional feature contribution vectors corresponding to the samples in the audio dataset that are determined to be audio anomalies, the initial classification template for each sub-type of abnormal audio is calculated using an anomaly probability-weighted average method. The specific calculation expression is as follows: .

[0078] In the formula, This represents the initial classification template for the subdivision of Class A unusual sounds, where A∈{1,2,3}; where 1 corresponds to a stuttering sound, 2 corresponds to a friction sound, and 3 corresponds to a whistling sound. This represents the total number of samples in the valid sample pool corresponding to the A-class heterophonic sub-category. This represents the anomaly probability of the i-th sample in the effective sample pool corresponding to the A-type anomaly sub-category. This represents the multi-dimensional feature contribution vector of the i-th sample in the effective sample pool corresponding to the A-type heterophony sub-category.

[0079] It should be understood that the effective sample pool corresponding to the Class A anomaly sub-type can consist of some or all of the samples in the audio dataset constructed in step S100 that are labeled as Class A anomaly sub-type. It is important to note that all samples in the effective sample pool are audio anomaly samples.

[0080] Understandably, in order to ensure the accuracy of the classification template in classifying different sound subcategories, the classification template can be verified for accuracy. For ease of understanding, the verification process will be described simply below.

[0081] Specifically, audio anomaly samples are selected from the audio dataset to construct training and testing sets. Based on the samples in the training set, classification templates corresponding to each sub-type of anomaly are constructed. Multi-dimensional feature contribution vectors of all audio anomaly samples in the testing set are obtained, and the cosine similarity between each multi-dimensional feature contribution vector and the classification template corresponding to each sub-type of anomaly is calculated. The specific calculation expression is as follows: .

[0082] In the formula, This represents the multi-dimensional feature contribution vector of the test sample. This represents the initial classification template corresponding to the subdivision of Class A heterophonic sounds.

[0083] It should be noted that the closer the cosine similarity value is to 1, the more similar the sample is to that sub-category of heterophony. The sub-category of heterophony corresponding to the classification template with the highest cosine similarity is selected as the final classification result of the test sample.

[0084] The true subcategories of abnormal sounds in all samples of the statistical test set are analyzed, and the macro-average F1 score of the audio abnormal samples is calculated. If the calculated macro-average F1 score is greater than 85%, it indicates that the obtained classification template meets the subcategorization accuracy requirements of the audio signal data; otherwise, the frame-level feature matrix is ​​re-extracted or the abnormal sound detection module is retrained to recalculate the classification template until the subcategorization accuracy requirements are met.

[0085] It is worth noting that, due to factors such as changes in product batches, switching of component suppliers, and fluctuations in environmental conditions (temperature, humidity), the acoustic characteristics of the seat drive will slowly shift, leading to a gradual increase in the feature difference between the classification template and newly added samples, and a continuous decrease in classification accuracy. Therefore, this embodiment proposes an adaptive update mechanism that enables the classification template to continuously evolve in accordance with changes in production line conditions; for ease of understanding, a detailed description will be provided below.

[0086] In a specific embodiment, such as Figure 2 As shown, the adaptive update mechanism for the classification template includes the following process: the classification templates corresponding to each sub-type of abnormal sound are deployed on the production line; after each batch of samples to be tested is classified into sub-types of abnormal sound on the production line, a hierarchical confidence weighting strategy is used to screen the effective sample pool; the batch-weighted SHAP mean vector is calculated based on the multi-dimensional feature contribution vectors corresponding to each sample in the effective sample pool; based on the obtained batch-weighted SHAP mean vector, the classification templates used for the sub-type classification of abnormal sound on the next batch of samples to be tested are updated using an adaptive smoothing factor.

[0087] In a specific example, the effective sample pool refers to the set of high-quality samples that, after stratified confidence screening, are deemed suitable for updating the classification template. The screening of the effective sample pool includes the following process: obtaining the anomalous probability corresponding to the audio anomalous samples in the current batch based on the anomalous sound detection model; calculating the cosine similarity between the audio anomalous samples in the current batch and the classification templates corresponding to each anomalous sound sub-type. If the cosine similarity corresponding to an audio anomalous sample exceeds a preset first similarity threshold, and the corresponding anomalous probability exceeds a preset first probability threshold, the audio anomalous sample is classified as a high-confidence sample. If the cosine similarity corresponding to an audio anomalous sample is between a preset first similarity threshold and a preset second similarity threshold, and the corresponding anomalous probability is between a preset first probability threshold and a preset second probability threshold, the audio anomalous sample is classified as a medium-confidence sample. If the cosine similarity corresponding to an audio anomalous sample is lower than a preset second similarity threshold, or the corresponding anomalous probability is lower than a preset second probability threshold, the audio anomalous sample is classified as a low-confidence sample. The high-confidence samples and medium-confidence samples are then combined to construct the effective sample pool; wherein the first similarity threshold is greater than the second similarity threshold, and the first probability threshold is greater than the second probability threshold.

[0088] It should be understood that low-confidence samples may represent novel anomalous sound patterns or noise interference, and therefore are not included in the effective sample pool for updating the classification template to avoid contaminating the classification template. The specific values ​​of the first similarity threshold, second similarity threshold, first probability threshold, and second probability threshold can be set according to the actual needs of those skilled in the art. For example, the first similarity threshold can be preset to 0.85, the second similarity threshold to 0.7, the first probability threshold to 0.8, and the second probability threshold to 0.7. Then, samples with cosine similarity ≥ 0.85 and anomaly probability ≥ 0.8 are high-confidence samples; samples with 0.7 ≤ cosine similarity < 0.85 and 0.7 ≤ anomaly probability < 0.8 are medium-confidence samples; and samples with cosine similarity < 0.7 or anomaly probability < 0.7 are low-confidence samples and are not included in the update.

[0089] In a specific example, the batch-weighted SHAP mean vector refers to the weighted average of the multi-dimensional feature contribution vectors of all samples in the effective sample pool, calculated by weighting the sample confidence level and the anomaly probability. It represents the overall feature contribution pattern of the effective samples in that batch. For ease of understanding, the batch-weighted SHAP mean vector... The calculation expression is as follows: .

[0090] In the formula, M represents the total number of samples in the effective sample pool. This represents the sample confidence level of the i-th sample in the valid sample pool. The sample confidence level of a high-confidence sample is 1, and the sample confidence level of a medium-confidence sample is 0.5. This represents the anomaly probability of the i-th sample in the valid sample pool. This represents the multi-dimensional feature contribution vector of the i-th sample in the effective sample pool.

[0091] In a specific example, the adaptive smoothing factor is a coefficient between 0 and 1, used to control the weighting of old template information and new sample information during the classification template update process. The expression for updating the classification template based on the adaptive smoothing factor is: ; .

[0092] In the formula, This represents a classification template used for further subdividing and categorizing abnormal sounds in the next batch of samples to be tested. This represents the classification template used for further subdividing and categorizing heterophonic sounds in the current k batches of test samples. This represents the adaptive smoothing factor corresponding to the current k batches of samples to be tested. This represents the batch-weighted SHAP mean vector corresponding to the current k batches of samples to be tested. This represents the preset basic smoothing factor, used to control the minimum update magnitude of the classification template. The specific value range can be 0.5 to 0.7. This represents the average cosine similarity of the previous batch of samples to be tested.

[0093] Understandably, when the production line is stable, the average cosine similarity of the previous batch of test samples is close to 1, the adaptive smoothing factor increases to close to 1, the weight of the old classification template in the update formula is greater, the classification template is updated smoothly, and excessive fluctuations are avoided. When the production line fluctuates, the average cosine similarity of the previous batch of test samples decreases, the adaptive smoothing factor decreases synchronously, the weight of the new sample information in the update formula increases, the classification template is updated faster, and it can quickly adapt to new working conditions.

[0094] It is important to note that during the classification template update process, if samples of a certain subtype of unusual speech consistently occupy an excessively high proportion in the effective sample pool, the classification template may shift excessively towards that subtype, causing a decrease in the classification accuracy of other subtypes of unusual speech. Therefore, in this embodiment, as... Figure 2 As shown, to avoid the bias of the classification template being overly dominated by samples of a single anomaly subtype, an adaptive anomaly ratio constraint mechanism is set; for ease of understanding, a detailed description will be provided below.

[0095] In one specific embodiment, the adaptive anomaly ratio constraint mechanism includes the following: real-time monitoring of the sample proportion of each anomaly subtype in the effective sample pool during the update process, and recording the number of effective samples of each anomaly subtype in each batch of test samples. When the sample proportion of anomalies of a certain anomaly subtype in the effective sample pool shifts for multiple consecutive batches and the shift exceeds a preset threshold, it is determined that the classification template corresponding to that anomaly subtype has a bias risk, triggering the reconstruction of the classification template. When reconstructing the classification template, the classification template is recalculated from the current effective sample pool according to the sample confidence and anomaly probability weights, resetting the historical cumulative bias.

[0096] It is important to note the difference between classification template reconstruction and classification template updating: classification template updating involves fine-tuning the existing classification template while retaining most of the historical information; while classification template reconstruction discards the current classification template and recalculates it using all high-quality samples in the current effective sample pool. The purpose of reconstruction is to eliminate the distortion of the classification template caused by long-term accumulated bias, enabling the classification template to quickly restore its balanced representation ability of the three types of anomalies.

[0097] It is understood that the technical solution of this application establishes a complete closed-loop logic of "initial classification template construction, sample classification, and online adaptive update based on hierarchical confidence level"; adopts a three-level confidence level weighted strategy to ensure the reliability of classification template updates; combines confidence level weight and anomaly probability to calculate batch mean, thereby improving the purity of classification templates; and introduces an adaptive smoothing factor to dynamically adjust the update step size according to production line stability, thereby achieving intelligent adaptation of slow updates during stable periods and fast updates during changing periods.

[0098] Another aspect of this application provides a system for implementing the above-described method for detecting and classifying abnormal noises in automotive seat drive systems, deployed on a production line equipped with an abnormal noise detection station and an abnormal noise classification station; such as Figure 3 As shown, one preferred embodiment includes an audio acquisition module, an abnormal sound detection module, and an abnormal sound classification module.

[0099] The noise detection station and the noise classification station operate independently yet collaboratively. The noise detection station focuses on rapid screening to ensure production line cycle time, while the noise classification station focuses on detailed diagnosis and adaptive updating of classification templates to ensure classification accuracy and long-term stability. The two stations communicate via a production line data bus, transmitting intermediate data such as feature sequences and anomaly probabilities.

[0100] The audio acquisition module is deployed at the abnormal noise detection station to collect audio signals during the operation of the seat drive.

[0101] Specifically, the core hardware of the audio acquisition module includes a high-sensitivity measurement microphone, an audio acquisition card, and an industrial control computer. During deployment, the microphone is installed at the testing station in a soundproof room, maintaining a constant distance from the seat driver under test to ensure consistent acquisition conditions. The audio acquisition module supports parameter presets and batch acquisition for multiple seat driver models, and the acquired audio is stored in lossless WAV format. The audio acquisition module sends the acquired raw audio signal to the abnormal sound detection module for further processing.

[0102] The abnormal noise detection module is deployed at the abnormal noise detection station. It receives the audio signal to be tested and extracts a frame-level feature matrix. Based on the extracted frame-level feature matrix, it determines whether the audio signal is normal or abnormal. If the audio signal is determined to be normal, the current seat driver is sent to the qualified area; otherwise, the current seat driver is sent to the abnormal noise classification station, and the corresponding feature sequence and abnormal probability of the audio signal are sent to the abnormal noise classification module.

[0103] Specifically, the abnormal sound detection module integrates frame-level feature extraction, attention pooling, and classification detection functions. It receives real-time audio signals from the audio acquisition module, segments the audio according to forward / reverse rotation, and labels directional features. It employs a frame-segmentation strategy to extract four main categories of physical features to form a frame-level feature matrix. Attention scoring, Softmax normalization, and weighted summation address the issue of inconsistent time dimensions caused by duration fluctuations. After concatenating the aggregated features and directional features, the result and anomaly probability are output through a linear mapping layer activated by Sigmoid. This enables rapid sorting on the production line and sends the feature sequences and anomaly probabilities of abnormal samples to the abnormal sound classification module.

[0104] It should be noted that the abnormal noise detection module immediately switches to the next seat driver to be inspected after sending the feature sequence and anomaly probability, without waiting for the classification station to complete the classification. This ensures that the efficient cycle time of the abnormal noise detection station is not affected by the classification time. Seat drivers identified as having audio abnormalities by the abnormal noise detection module can be sorted to the abnormal noise classification station.

[0105] The abnormal noise classification module is deployed at the abnormal noise classification station on the production line. It has built-in classification templates for each sub-type of abnormal noise. The abnormal noise classification module calculates the multi-dimensional feature contribution vector of the corresponding audio signal based on the received feature sequence and abnormal probability, compares the similarity with each classification template, outputs the specific sub-type of abnormal noise, and sends the seat driver to the corresponding classification area according to the classification result. The classification area is divided into friction sound area, stuttering sound area and howling sound area according to the sub-type of abnormal noise.

[0106] It's important to know that after each batch of noise classification is completed, the noise classification module automatically executes the aforementioned adaptive update mechanism for the classification template, ensuring that the template continuously evolves with the production line conditions. The noise classification module also automatically summarizes noise type data and generates daily / weekly / monthly reports on the percentage of each type of noise, facilitating statistical analysis and process defect investigation by technical personnel.

[0107] Another aspect of this application provides a computer-readable storage medium, in a preferred embodiment of which a computer program is stored on the storage medium; when the computer program is executed by a processor, it implements the above-described method for detecting and classifying abnormal noises from a car seat drive.

[0108] The basic principles, main features, and advantages of this application have been described above. Those skilled in the art should understand that this application is not limited to the above embodiments. The embodiments and descriptions in the specification are merely the principles of this application. Various changes and modifications can be made to this application without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claims. The scope of protection claimed by this application is defined by the appended claims and their equivalents.

Claims

1. A method for detecting and classifying abnormal noises from a car seat drive unit, characterized in that, Includes the following steps: Audio signals of the seat drive's forward and reverse rotation processes were collected in a standard quiet room environment. The forward and reverse rotation processes were independently labeled to indicate whether there were any abnormal noises and the sub-types of abnormal noises, forming an audio dataset with multi-label annotations. The samples in the audio dataset are segmented into frames, and multi-dimensional features of each frame are extracted to form a frame-level feature matrix. An attention pooling-based abnormal sound detection model is constructed and trained. The frame-level feature matrix corresponding to the sample is used as input. The attention score of each frame is calculated and normalized to attention weight. The attention weight is used to weight the multi-dimensional features corresponding to the frame-level feature matrix to obtain aggregated features and map them to abnormal probabilities, which are used to determine whether the audio of the sample is normal or abnormal. For samples in the audio dataset that are identified as audio abnormalities, the SHAP value analysis method is used to calculate the multi-dimensional feature contribution vector of each sample. For each sub-category of abnormal sounds, the multi-dimensional feature contribution vectors of all samples belonging to that sub-category of abnormal sounds are weighted and averaged to generate a classification template for that sub-category of abnormal sounds. For audio signals identified as having audio abnormalities by the abnormal sound detection model, a multi-dimensional feature contribution vector is calculated and compared with the similarity of each classification template to perform sub-classification of abnormal sounds in the audio signals under test.

2. The method for detecting and classifying abnormal noises from a car seat drive unit as described in claim 1, characterized in that, The abnormal sound detection model includes an attention scoring network, a weight normalization layer, a weighted summation layer, and a linear mapping layer; The attention scoring network is used to calculate the attention score for each frame of the frame-level feature matrix; The weight normalization layer normalizes the attention score into attention weights using the Softmax function; The weighted summation layer uses the obtained attention weights to perform weighted summation on the multi-dimensional features corresponding to all frames of the frame-level feature matrix to obtain aggregated features; The aggregated features and directional features are concatenated and then input into a linear mapping layer, where they are mapped to anomaly probabilities using the Sigmoid function.

3. The method for detecting and classifying abnormal noises in a car seat drive system as described in claim 2, characterized in that, The expression for calculating the anomaly probability P is as follows: ； ； ； In the formula, This represents the Sigmoid function. This represents feature concatenation, where N represents the number of frames corresponding to the frame-level feature matrix. This represents the attention weights corresponding to the t-th frame in the frame-level feature matrix. This represents the eigenvector corresponding to the t-th frame in the frame-level feature matrix. Indicates directional characteristics, Indicates the output layer weights. Both b and 'b' represent bias terms. and These represent the attention scores corresponding to the t-th and j-th frames of the frame-level feature matrix, respectively. This represents the attention output weight vector. This represents the activation function. This represents the frame feature mapping matrix.

4. The method for detecting and classifying abnormal noises from a car seat drive unit as described in any one of claims 1-3, characterized in that, The classification templates corresponding to each sub-type of abnormal sound are deployed on the production line. After the production line completes the classification of the sub-type of abnormal sound for each batch of samples to be tested, a stratified confidence weighted strategy is used to screen the effective sample pool. Calculate the batch-weighted SHAP mean vector based on the multi-dimensional feature contribution vectors corresponding to each sample in the effective sample pool; Based on the obtained batch-weighted SHAP mean vector, the classification template used for the next batch of test samples to classify heterophonic sub-types is updated using an adaptive smoothing factor.

5. The method for detecting and classifying abnormal noises in a car seat drive system as described in claim 4, characterized in that, The screening of the effective sample pool includes the following process: Based on the abnormal sound detection model, obtain the abnormal probability corresponding to the audio abnormal sample in the current batch; calculate the cosine similarity between the audio abnormal sample in the current batch and the classification template corresponding to each sub-type of abnormal sound; If the cosine similarity of an abnormal audio sample exceeds a preset first similarity threshold, and the corresponding abnormal probability exceeds a preset first probability threshold, the abnormal audio sample is classified as a high-confidence sample. If the cosine similarity of an abnormal audio sample is between a preset first similarity threshold and a preset second similarity threshold, and the corresponding abnormal probability is between a preset first probability threshold and a preset second probability threshold, then the abnormal audio sample is classified as a medium confidence sample. If the cosine similarity of an abnormal audio sample is lower than a preset second similarity threshold, or the corresponding abnormal probability is lower than a preset second probability threshold, the abnormal audio sample is classified as a low-confidence sample. A pool of high-confidence and medium-confidence samples is constructed by combining them. Among them, the first similarity threshold is greater than the second similarity threshold, and the first probability threshold is greater than the second probability threshold.

6. The method for detecting and classifying abnormal noises in a car seat drive system as described in claim 5, characterized in that, Batch-weighted SHAP mean vector The calculation expression is as follows: ； In the formula, M represents the total number of samples in the effective sample pool. This represents the sample confidence level of the i-th sample in the valid sample pool. The sample confidence level of a high-confidence sample is 1, and the sample confidence level of a medium-confidence sample is 0.

5. This represents the anomaly probability of the i-th sample in the valid sample pool. This represents the multi-dimensional feature contribution vector of the i-th sample in the effective sample pool.

7. The method for detecting and classifying abnormal noises in a car seat drive system as described in claim 6, characterized in that, The expression for updating the classification template is: ； ； In the formula, This represents a classification template used for further subdividing and categorizing abnormal sounds in the next batch of samples to be tested. This represents the classification template used for further subdividing and categorizing heterophonic sounds in the current k batches of test samples. This represents the adaptive smoothing factor corresponding to the current k batches of samples to be tested. This represents the batch-weighted SHAP mean vector corresponding to the current k batches of samples to be tested. This represents the preset basic smoothing factor. This represents the average cosine similarity of the previous batch of samples to be tested.

8. The method for detecting and classifying abnormal noises in a car seat drive system as described in claim 7, characterized in that, When updating the classification template, to avoid samples from a single anomaly subtype excessively dominating the classification template's bias, an adaptive anomaly ratio constraint mechanism is set up, the details of which are as follows: Real-time monitoring of the proportion of samples of each sub-type of abnormal sound in the effective sample pool during the update process, and recording the number of effective samples of each sub-type of abnormal sound in each batch of samples to be tested; When the proportion of a certain sub-type of aberration in the effective sample pool shifts for multiple consecutive batches and the shift exceeds a preset threshold, it is determined that the classification template corresponding to the sub-type of aberration has a risk of bias, and the reconstruction of the classification template is triggered. When reconstructing the classification template, the classification template is recalculated from the current valid sample pool according to the sample confidence and anomaly probability, and the historical cumulative bias is reset.

9. A system for implementing the method for detecting and classifying abnormal noises in automotive seat drive systems as described in any one of claims 1-8, deployed on a production line, the production line being equipped with an abnormal noise detection station and an abnormal noise classification station; characterized in that, The system includes: Audio acquisition module; the audio acquisition module is deployed at the abnormal noise detection station and is used to acquire audio signals during the operation of the seat drive. Anomaly detection module; The anomaly detection module is deployed at the anomaly detection station and is used to receive the audio signal to be tested and extract the frame-level feature matrix. Based on the extracted frame-level feature matrix, the audio signal is judged to be normal or abnormal. If the audio signal is judged to be normal, the current seat driver is sent to the qualified area; otherwise, the current seat driver is sent to the anomaly classification station and the feature sequence and anomaly probability of the corresponding audio signal are sent to the anomaly classification module. Anomaly classification module; The anomaly classification module is deployed at the anomaly classification station on the production line and has built-in classification templates for each anomaly sub-type; The anomaly classification module calculates the multi-dimensional feature contribution vector of the corresponding audio signal based on the received feature sequence and anomaly probability, compares the similarity with each classification template, outputs the specific anomaly sub-type, and sends the seat driver to the corresponding classification area according to the classification result.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program; when the computer program is executed by a processor, it implements the method for detecting and classifying abnormal noises in a car seat drive as described in any one of claims 1-8.

Images