A Smart Method for Determining Loud Noise in Speakers Based on Acoustic Response Residual Feature Enhancement and Sample Backflow Training

By generating test excitation signals with known frequency variation patterns, the normal main response components of the loudspeaker are suppressed and residual abnormal components are enhanced using time-frequency mapping relationships. Combined with an abnormal noise classification model, the stability and accuracy issues of loudspeaker abnormal noise detection are solved, achieving stable determination across batches and models. Furthermore, the model is continuously optimized through a data feedback mechanism.

CN123075346APending Publication Date: 2026-11-10FANGBO TECH (SHENZHEN) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611221804.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-12
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing technologies are insufficient to effectively detect abnormal noises such as minor scratches, loose local structures, and narrowband noise in loudspeakers. Traditional electroacoustic indicators are not a reliable substitute for human listening, and the lack of an effective feedback mechanism in production line testing models makes continuous improvement difficult.

Method used

By generating test excitation signals with known frequency variation patterns, the normal main response components are suppressed and the residual abnormal components are enhanced using the time-frequency mapping relationship. The abnormal sound classification model is then used for judgment, and samples with low confidence or inconsistent manual verification are returned to the training library to form a data closed-loop optimization.

Benefits of technology

It improves the stability and accuracy of speaker noise detection, reduces dependence on test volume and ambient noise, achieves stable determination across batches and models, and continuously optimizes the model through a data feedback mechanism.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123075346A_ABST
    Figure CN123075346A_ABST
Patent Text Reader

Abstract

This invention discloses an intelligent method for identifying loudspeaker abnormalities based on acoustic response residual feature enhancement and sample backflow training. The method includes: generating test excitations with known frequency variation patterns and acquiring complete acoustic responses; determining the main excitation frequency based on the excitation time-frequency mapping, and suppressing the normal main response to obtain residual abnormal components; performing time-frequency transformation and local background subtraction on the residual abnormal components to generate a residual time-frequency image; inputting the residual time-frequency image into an abnormality classification model to output predicted labels and confidence scores; and when the confidence score is below a threshold or there is inconsistency between human and machine judgments, binding the samples and backflowing them to the training library. This invention improves the learning stability of the abnormality classification model for defect features by using prior constraints of test excitations to suppress the main response and enhance residual features; and achieves continuous iterative optimization of the production line model through low-confidence sample backflow closure.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of audio signal processing, and in particular to a method for intelligently determining abnormal sounds from loudspeakers based on acoustic response residual feature enhancement and sample backflow training. Background Technology

[0002] During the manufacturing, assembly, transportation, and use of loudspeakers, abnormal noises may occur due to factors such as voice coil rubbing, vibration system eccentricity, local abnormalities in the cone or spider, glue detachment, structural loosening, abnormal magnetic circuit clearance, uneven assembly stress, or component collisions. These abnormal noises often do not manifest as obvious distortion in traditional frequency response curves, but may appear as localized short-term friction sounds, narrowband abnormal pure tones, asynchronous noise, residual ringing, or intermittent impacts. In production line testing scenarios, manual listening suffers from strong subjectivity, significant fatigue effects, obvious reliance on experience, and untraceable results. Traditional electroacoustic indicators such as frequency response, total harmonic distortion, sound pressure level, or global RMS energy, while reflecting some aspects of loudspeaker performance, lack sufficient directivity for minor rubbing, localized structural loosening, narrowband noise, and intermittent abnormal noises, making them unreliable substitutes for manual listening.

[0003] For logarithmic sweep frequency testing, the excitation frequency changes continuously with time, and the main excitation response forms a clear main trajectory in the time-frequency plane. For step-fast sweep frequency testing, each frequency point has a relatively clear hold interval and transition interval. The test excitations themselves contain usable prior constraint information. If this prior constraint is ignored at the detection front end, and only ordinary spectrograms or global energy features are extracted from the complete response, it is impossible to fully distinguish between the normal main response and the abnormal residual components.

[0004] Therefore, it is necessary to propose a method for intelligently determining loudspeaker abnormalities by using the test excitation time-frequency relationship to enhance the residual features of the complete acoustic response. This method can highlight defect-related components before they enter the AI ​​model and bind the original response, residual signal, residual image, manual label, prediction results and test configuration to achieve stable training, online determination, low-confidence sample return and production line traceability. Summary of the Invention

[0005] To address the aforementioned problems in existing technologies, the present invention aims to provide a method for intelligently determining loudspeaker abnormalities based on acoustic response residual feature enhancement and sample reflow training, comprising the following steps: Step S1: Generate a preset test excitation signal with a known frequency variation pattern, and read the test configuration parameters.

[0006] Step S2: Acquire the complete acoustic response signal generated by the loudspeaker under test under the action of the test excitation signal.

[0007] Step S3: Preprocess the complete acoustic response signal, including DC removal, amplitude adjustment, time delay estimation, effective segment truncation, and calibration compensation.

[0008] Step S4: Establish a time-frequency mapping relationship based on the type of the test excitation signal.

[0009] Step S5: Determine the main excitation frequency corresponding to each analysis frame based on the time-frequency mapping relationship, and construct a suppression parameter that varies with frequency using the main excitation frequency as a reference to suppress the normal main response component in the complete acoustic response signal to obtain the residual abnormal component.

[0010] Step S6: Perform energy normalization on the residual anomalous components to map the residual responses under different test conditions to a consistent amplitude reference system.

[0011] Step S7: Perform time-frequency transformation on the residual anomalous components after energy normalization, and perform local background subtraction and image mapping on the obtained residual energy matrix to generate residual time-frequency image or residual feature matrix.

[0012] Step S8: Input the residual time-frequency image or residual feature matrix into the trained heterophony classification model and output the predicted label and confidence score.

[0013] Step S9: Map the predicted label to a qualified or unqualified class according to the preset label mapping rules, and output the final judgment result in combination with the confidence threshold.

[0014] Step S10: When the prediction confidence is lower than the preset threshold or the manual review result is inconsistent with the model output, the complete response signal, residual signal, residual time-frequency image, prediction result and manual review label of the current sample are bound and then fed back to the training sample library.

[0015] Furthermore, step S5 includes the following sub-steps: Step S501: Perform frame-by-frame processing on the complete acoustic response signal, and determine the main excitation frequency corresponding to each analysis frame according to the time-frequency mapping relationship.

[0016] Step S502: Using the main excitation frequency as a reference, construct a tracking high-pass cutoff frequency that varies with frequency, and determine the tracking high-pass cutoff frequency according to the following formula: , In the formula, For the first Tracking the high-pass cutoff frequency of each analysis frame. For the first The main excitation frequency corresponding to each analysis frame. For configurable tracking factors, The value range is from 0.5 to 1.5; Step S503: Use the tracking high-pass cutoff frequency to perform high-pass filtering on the current analysis frame to filter out normal main response components below the main excitation frequency.

[0017] Step S504: Overlap and splice the high-pass filtered frame signals to obtain residual abnormal components.

[0018] Furthermore, step S7 includes the following sub-steps: Step S701: Perform a short-time Fourier transform on the residual anomalous components after energy normalization to obtain the residual energy matrix.

[0019] Step S702: For each time-frequency point in the residual energy matrix, the quantile statistics within the neighborhood of that time-frequency point are used as the local background energy estimate, which is determined by the following formula: , In the formula, For local background energy estimation at time-frequency point (m,k), The result is the short-time Fourier transform of the residual anomalous components. The neighborhood window is centered at the time frequency point (m,k). (·) indicates taking the p-th percentile of the energy value in the neighborhood, where p ranges from 10 to 50.

[0020] Step S703: Subtract the corresponding local background energy estimate from the original energy values ​​at each time frequency point, and perform zero truncation on the negative values ​​to obtain the residual energy matrix after subtracting the local background.

[0021] Step S704: Perform frequency range truncation, energy range unification, and image size normalization on the residual energy matrix after removing the local background to generate a residual time-frequency image.

[0022] Furthermore, a transition segment processing step is included between step S4 and step S5: For the step-fast frequency sweep type, at the beginning of each frequency point, transitional samples related to frequency switching are discarded, and the discard length is determined by the larger of a preset multiple of the current frequency's corresponding period or a preset fixed number of samples; for the logarithmic frequency sweep type, frame boundary artifacts and main response leakage are reduced through overlapping framing and smoothing window processing.

[0023] Furthermore, in step S9: The preset label mapping rules are configured by the user to define which predicted labels belong to the qualified category and which predicted labels belong to the unqualified category. The predicted labels include multiple categories such as normal, voice coil rubbing, cone abnormality, structural looseness, noise, minor abnormality and severe abnormality.

[0024] The final determination result is further combined with conventional electroacoustic indicators and engineering threshold lines for comprehensive judgment.

[0025] The output of the heterophony classification model can be configured to serve only as an auxiliary prompt, or to participate in the final determination when the confidence level meets the threshold condition.

[0026] Furthermore, in step S10: The returned sample package contains the complete response signal, residual signal, residual time-frequency image, sample number, prediction label, confidence level, manual verification label, test configuration parameters, and model version information.

[0027] The conditions for triggering reflux also include the operator actively choosing to add the sample to the training library, abnormal drift in consecutive batches, insufficient number of specified label samples, or failure of model version verification.

[0028] Furthermore, step S4 includes the following sub-steps: Step S401: Determine whether the type of the test excitation signal is logarithmic sweep frequency type or step fast sweep frequency type.

[0029] Step S402: When the test excitation signal is a logarithmic frequency sweep type, establish a continuous time frequency mapping function based on the logarithmic relationship between instantaneous frequency and time, according to the starting frequency, ending frequency and total sweep duration.

[0030] Step S403: When the test excitation signal is a step fast sweep frequency type, establish a segmented time frequency mapping table based on the start time, holding time and frequency value of each frequency point, and mark the frequency switching interval between adjacent frequency points with transition marks.

[0031] Furthermore, in step S6, the energy normalization process employs at least one of local root-mean-square energy normalization, total response root-mean-square energy normalization, or frequency band energy normalization to map the residual responses under different sound pressure levels, different sensitivities, and different tooling coupling conditions to a preset amplitude reference system.

[0032] Furthermore, the following steps are included before step S8: Step S11: Establish a correspondence between multiple sets of residual time-frequency images or residual feature matrices and manually confirmed abnormal sound tags to form a training sample library. The abnormal sound tags are divided into multiple categories according to the defect type. Step S12: Train the heterophony classification model using the training sample library. The heterophony classification model is any one of a convolutional neural network, an image classification model based on transfer learning, or a traditional machine learning classifier. The model output includes the predicted label, the probability value of each label, and the highest confidence level.

[0033] Furthermore, the step S10 is followed by the following steps: Step S13: The training module generates an updated abnormal sound classification model, and the deployment module applies the updated model to the production line prediction module; Step S14: Record the model version, deployment time, and sample prediction information corresponding to each test, so that the same product can be traced back to the corresponding complete response signal, residual time-frequency image, prediction model version, and manual review record based on the sample number.

[0034] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention systematically suppresses the normal main response by utilizing the prior time-frequency constraints of the test stimulus before the artificial intelligence judgment, thereby enhancing the time-frequency contrast of residual abnormal components such as rubbing, noise, looseness and abnormal pure tone that were originally masked by the main response energy in the complete acoustic response signal. This reduces the learning dependence of the abnormal sound classification model on irrelevant factors such as test volume, product sensitivity, tooling coupling and environmental background noise, enabling the model to focus on abnormal residual features directly related to defects rather than normal sound features, thus maintaining a stable abnormal sound judgment capability under cross-batch and cross-model conditions.

[0035] This invention automatically feeds samples with insufficient confidence or discrepancies between manual verification and model output during production line testing back to the training sample library by binding complete response signals, residual signals, residual time-frequency images, prediction results, and manual verification labels. Combined with model training, deployment, and version tracking mechanisms, this forms a complete data closed loop from production line testing to model training and deployment. This allows the abnormal sound classification model to continuously accumulate boundary samples and misjudged samples for iterative optimization as the production line runs. This solves the problem that traditional production line detection models are difficult to continuously improve after deployment due to the lack of an effective feedback mechanism. At the same time, the complete data package bound to each test provides verifiable evidence for subsequent quality traceability and dispute reproduction. Attached Figure Description

[0036] Figure 1 This is an exemplary flowchart of the determination method of the present invention.

[0037] Figure 2 This is an exemplary flowchart of the steps for obtaining abnormal components according to the present invention.

[0038] Figure 3 This is an exemplary flowchart of the steps for generating residual time-frequency images according to the present invention.

[0039] Figure 4 A flowchart illustrating the exemplary steps for establishing mapping relationships in this invention.

[0040] Figure 5 A flowchart illustrating the exemplary steps of constructing a classification model according to the present invention.

[0041] Figure 6 This is an exemplary flowchart of the test result recording steps of the present invention. Detailed Implementation

[0042] The present invention will be further described below with reference to specific embodiments.

[0043] like Figure 1 The above illustrates a method for intelligently determining loudspeaker noise based on acoustic response residual feature enhancement and sample reflow training, as provided in this embodiment. The method includes the following steps: Step S1: Generate a preset test excitation signal with a known frequency variation pattern, and read the test configuration parameters.

[0044] Step S2: Acquire the complete acoustic response signal generated by the loudspeaker under test under the action of the test excitation signal.

[0045] Step S3: Preprocess the complete acoustic response signal. The preprocessing includes DC removal, amplitude adjustment, time delay estimation, effective segment truncation, and calibration compensation.

[0046] Step S4: Establish a time-frequency mapping relationship based on the type of test excitation signal.

[0047] Step S5: Determine the main excitation frequency corresponding to each analysis frame based on the time-frequency mapping relationship, and construct a suppression parameter that varies with frequency using the main excitation frequency as a reference to suppress the normal main response component in the complete acoustic response signal and obtain the residual abnormal component. Step S6: Perform energy normalization on the residual anomalous components to map the residual responses under different test conditions to a consistent amplitude reference system; Step S7: Perform time-frequency transformation on the residual anomalous components after energy normalization, and perform local background subtraction and image mapping on the obtained residual energy matrix to generate residual time-frequency image or residual feature matrix; Step S8: Input the residual time-frequency image or residual feature matrix into the trained heterophony classification model, and output the predicted label and confidence score; Step S9: Map the predicted labels to qualified or unqualified classes according to the preset label mapping rules, and output the final judgment result in combination with the confidence threshold; Step S10: When the prediction confidence is lower than the preset threshold or the manual review result is inconsistent with the model output, the complete response signal, residual signal, residual time-frequency image, prediction result and manual review label of the current sample are bound and then fed back to the training sample library.

[0048] like Figure 2 As shown, step S5 includes the following sub-steps: Step S501: Perform frame-by-frame processing on the complete acoustic response signal and determine the main excitation frequency corresponding to each analysis frame according to the time-frequency mapping relationship; Step S502: Using the main excitation frequency as a reference, construct a tracking high-pass cutoff frequency that varies with frequency, and determine the tracking high-pass cutoff frequency according to the following formula: , In the formula, For the first Tracking the high-pass cutoff frequency of each analysis frame. For the first The main excitation frequency corresponding to each analysis frame. For configurable tracking factors, The value range is from 0.5 to 1.5; Step S503: Use the tracking high-pass cutoff frequency to perform high-pass filtering on the current analysis frame to filter out normal main response components below the main excitation frequency; Step S504: Overlap and splice the high-pass filtered frame signals to obtain residual abnormal components.

[0049] like Figure 3 As shown, step S7 includes the following sub-steps: Step S701: Perform a short-time Fourier transform on the energy-normalized residual anomalous components to obtain the residual energy matrix; Step S702: For each time-frequency point in the residual energy matrix, the quantile statistics within the neighborhood of that time-frequency point are used as the local background energy estimate. The local background energy estimate is determined by the following formula: , In the formula, For local background energy estimation at time-frequency point (m,k), The result is the short-time Fourier transform of the residual anomalous components. The neighborhood window is centered at the time frequency point (m,k). (·) indicates taking the p-th percentile of the energy value in the neighborhood, where p ranges from 10 to 50; Step S703: Subtract the corresponding local background energy estimate from the original energy value at each time frequency point, and perform zero truncation on the negative values ​​to obtain the residual energy matrix after subtracting the local background. Step S704: Perform frequency range truncation, energy range unification, and image size normalization on the residual energy matrix after removing the local background to generate a residual time-frequency image.

[0050] The transition process between steps S4 and S5 also includes a step: For the step-fast frequency sweep type, at the beginning of each frequency point, transitional samples related to frequency switching are discarded, and the discard length is determined by the larger of a preset multiple of the current frequency's corresponding period or a preset fixed number of samples; for the logarithmic frequency sweep type, frame boundary artifacts and main response leakage are reduced through overlapping framing and smoothing window processing.

[0051] In step S9: The preset label mapping rules are configured by the user to define which predicted labels belong to the qualified category and which predicted labels belong to the unqualified category. The predicted labels include multiple categories such as normal, voice coil rubbing, cone abnormality, structural looseness, noise, minor abnormality and severe abnormality. The final judgment is further based on a comprehensive assessment of conventional electroacoustic indicators and engineering thresholds. The output of the heterophony classification model can be configured to serve only as an auxiliary prompt, or to participate in the final judgment when the confidence level meets the threshold condition.

[0052] In step S10: The returned sample package contains the complete response signal, residual signal, residual time-frequency image, sample number, prediction label, confidence level, manual verification label, test configuration parameters and model version information; The conditions for triggering reflux also include the operator actively choosing to add the sample to the training library, abnormal drift in consecutive batches, insufficient number of specified label samples, or failure of model version verification.

[0053] like Figure 4 Step S4 shown includes the following sub-steps: Step S401: Determine whether the test excitation signal is a logarithmic sweep frequency type or a step fast sweep frequency type; Step S402: When the test excitation signal is a logarithmic frequency sweep type, establish a continuous time frequency mapping function based on the logarithmic relationship between instantaneous frequency and time, according to the starting frequency, ending frequency and total sweep duration; Step S403: When the test excitation signal is a step fast sweep frequency type, establish a segmented time frequency mapping table based on the start time, hold time and frequency value of each frequency point, and mark the frequency switching interval between adjacent frequency points with transition marks.

[0054] In step S6, the energy normalization process employs at least one of local root mean square energy normalization, total response root mean square energy normalization, or frequency band energy normalization to map the residual responses under different sound pressure levels, different sensitivities, and different tooling coupling conditions to a preset amplitude reference system.

[0055] like Figure 5 As shown, the following steps are included before step S8: Step S11: Establish a correspondence between multiple sets of residual time-frequency images or residual feature matrices and manually confirmed abnormal sound labels to form a training sample library. The abnormal sound labels are divided into multiple categories according to the defect type. Step S12: Train the heterophony classification model using the training sample library. The heterophony classification model can be any of the following: convolutional neural network, image classification model based on transfer learning, or traditional machine learning classifier. The model output includes the predicted label, the probability value of each label, and the highest confidence level.

[0056] like Figure 6 As shown, the following steps are included after step S10: Step S13: The training module generates an updated abnormal sound classification model, and the deployment module applies the updated model to the production line prediction module; Step S14: Record the model version, deployment time, and sample prediction information corresponding to each test, so that the same product can be traced back to the corresponding complete response signal, residual time-frequency image, prediction model version, and manual review record based on the sample number.

Claims

1. A method for intelligently determining loudspeaker abnormal sounds based on acoustic response residual feature enhancement and sample reflow training, characterized in that, Includes the following steps: Step S1: Generate a preset test excitation signal with a known frequency variation pattern, and read the test configuration parameters; Step S2: Acquire the complete acoustic response signal generated by the loudspeaker under test under the action of the test excitation signal; Step S3: Preprocess the complete acoustic response signal, including DC removal, amplitude adjustment, time delay estimation, effective segment truncation, and calibration compensation; Step S4: Establish a time-frequency mapping relationship based on the type of the test excitation signal; Step S5: Determine the main excitation frequency corresponding to each analysis frame based on the time-frequency mapping relationship, and construct a suppression parameter that varies with frequency using the main excitation frequency as a reference to suppress the normal main response component in the complete acoustic response signal to obtain the residual abnormal component; Step S6: Perform energy normalization on the residual anomalous components to map the residual responses under different test conditions to a consistent amplitude reference system; Step S7: Perform time-frequency transformation on the residual anomalous components after energy normalization, and perform local background subtraction and image mapping on the obtained residual energy matrix to generate residual time-frequency image or residual feature matrix; Step S8: Input the residual time-frequency image or residual feature matrix into the trained heterophony classification model, and output the predicted label and confidence score; Step S9: Map the predicted label to a qualified or unqualified class according to the preset label mapping rules, and output the final judgment result in combination with the confidence threshold; Step S10: When the prediction confidence is lower than the preset threshold or the manual review result is inconsistent with the model output, the complete response signal, residual signal, residual time-frequency image, prediction result and manual review label of the current sample are bound and then fed back to the training sample library.

2. The intelligent method for detecting abnormal noise in loudspeakers based on acoustic response residual feature enhancement and sample backflow training as described in claim 1, characterized in that, Step S5 includes the following sub-steps: Step S501: Perform frame segmentation processing on the complete acoustic response signal, and determine the main excitation frequency corresponding to each analysis frame according to the time-frequency mapping relationship; Step S502: Using the main excitation frequency as a reference, construct a tracking high-pass cutoff frequency that varies with frequency, and determine the tracking high-pass cutoff frequency according to the following formula: , In the formula, For the first Tracking the high-pass cutoff frequency of each analysis frame. For the first The main excitation frequency corresponding to each analysis frame. Configurable tracking factor; Step S503: Use the tracking high-pass cutoff frequency to perform high-pass filtering on the current analysis frame to filter out normal main response components below the main excitation frequency; Step S504: Overlap and splice the high-pass filtered frame signals to obtain residual abnormal components.

3. The intelligent method for detecting abnormal sounds in loudspeakers based on acoustic response residual feature enhancement and sample backflow training as described in claim 1, characterized in that, Step S7 includes the following sub-steps: Step S701: Perform a short-time Fourier transform on the energy-normalized residual anomalous components to obtain the residual energy matrix; Step S702: For each time-frequency point in the residual energy matrix, the quantile statistics within the neighborhood of that time-frequency point are used as the local background energy estimate, which is determined by the following formula: , In the formula, For local background energy estimation at time-frequency point (m,k), The result is the short-time Fourier transform of the residual anomalous components. The neighborhood window is centered at the time frequency point (m,k). (·) indicates taking the p-th percentile of the energy value within the neighborhood; Step S703: Subtract the corresponding local background energy estimate from the original energy value at each time frequency point, and perform zero truncation on the negative values ​​to obtain the residual energy matrix after subtracting the local background. Step S704: Perform frequency range truncation, energy range unification, and image size normalization on the residual energy matrix after removing the local background to generate a residual time-frequency image.

4. The intelligent method for detecting abnormal noise in loudspeakers based on acoustic response residual feature enhancement and sample backflow training as described in claim 1, characterized in that, A transition segment processing step is also included between step S4 and step S5: For the step-fast frequency sweep type, at the beginning of each frequency point, transitional samples related to frequency switching are discarded, and the discard length is determined by the larger of a preset multiple of the current frequency's corresponding period or a preset fixed number of samples; for the logarithmic frequency sweep type, frame boundary artifacts and main response leakage are reduced through overlapping framing and smoothing window processing.

5. The intelligent method for detecting abnormal noise in loudspeakers based on acoustic response residual feature enhancement and sample backflow training as described in claim 1, characterized in that, In step S9: The preset label mapping rules are configured by the user to define which predicted labels belong to the qualified category and which predicted labels belong to the unqualified category. The predicted labels include multiple categories such as normal, voice coil rubbing, cone abnormality, structural looseness, noise, minor abnormality and severe abnormality. The final determination result is further combined with conventional electroacoustic indicators and engineering threshold lines for comprehensive judgment. The output of the heterophony classification model can be configured to serve only as an auxiliary prompt, or to participate in the final determination when the confidence level meets the threshold condition.

6. The intelligent method for detecting abnormal sounds in loudspeakers based on acoustic response residual feature enhancement and sample reflow training as described in claim 1, characterized in that, In step S10: The returned sample package contains the complete response signal, residual signal, residual time-frequency image, sample number, prediction label, confidence level, manual verification label, test configuration parameters and model version information; The conditions for triggering reflux also include the operator actively choosing to add the sample to the training library, abnormal drift in consecutive batches, insufficient number of specified label samples, or failure of model version verification.

7. The intelligent method for detecting abnormal noise in loudspeakers based on acoustic response residual feature enhancement and sample backflow training as described in claim 1, characterized in that, Step S4 includes the following sub-steps: Step S401: Determine whether the type of the test excitation signal is logarithmic sweep frequency type or step fast sweep frequency type; Step S402: When the test excitation signal is a logarithmic frequency sweep type, establish a continuous time frequency mapping function based on the logarithmic relationship between instantaneous frequency and time, according to the starting frequency, ending frequency and total sweep duration; Step S403: When the test excitation signal is a step fast sweep frequency type, establish a segmented time frequency mapping table based on the start time, holding time and frequency value of each frequency point, and mark the frequency switching interval between adjacent frequency points with transition marks.

8. The intelligent method for detecting abnormal sounds in loudspeakers based on acoustic response residual feature enhancement and sample reflow training as described in claim 1, characterized in that, In step S6, the energy normalization process employs at least one of local root-mean-square energy normalization, total response root-mean-square energy normalization, or frequency band energy normalization to map the residual responses under different sound pressure levels, different sensitivities, and different tooling coupling conditions to a preset amplitude reference system.

9. The intelligent method for detecting abnormal noise in loudspeakers based on acoustic response residual feature enhancement and sample backflow training as described in claim 1, characterized in that, The following steps are included before step S8: Step S11: Establish a correspondence between multiple sets of residual time-frequency images or residual feature matrices and manually confirmed abnormal sound tags to form a training sample library. The abnormal sound tags are divided into multiple categories according to the defect type. Step S12: Train the heterophony classification model using the training sample library. The heterophony classification model is any one of a convolutional neural network, an image classification model based on transfer learning, or a traditional machine learning classifier. The model output includes the predicted label, the probability value of each label, and the highest confidence level.

10. The intelligent method for detecting abnormal sounds in loudspeakers based on acoustic response residual feature enhancement and sample backflow training as described in claim 1, characterized in that, The following steps are included after step S10: Step S13: The training module generates an updated abnormal sound classification model, and the deployment module applies the updated model to the production line prediction module; Step S14: Record the model version, deployment time, and sample prediction information corresponding to each test, so that the same product can be traced back to the corresponding complete response signal, residual time-frequency image, prediction model version, and manual review record based on the sample number.