A machine learning-based audio system failure analysis method and system

Abstract

The application relates to the technical field of fault detection, and discloses a machine learning sound system fault analysis method and system, which comprises the following steps: performing abnormal feature mining on audio features to obtain abnormal feature clusters, and marking the abnormal feature clusters as to-be-verified fault samples; performing fault feature decoupling on the to-be-verified fault samples to obtain an enhanced fault feature set; performing knowledge distillation training on a preset first deep diagnosis network, and migrating fault judgment knowledge learned by the first deep diagnosis network to a preset second diagnosis architecture to obtain an embedded diagnosis engine; performing feature matching on running parameter flow and performing deviation degree evaluation to obtain multi-dimensional abnormal evidence; performing root cause tracing on the multi-dimensional abnormal evidence to obtain fault positioning information and cause speculation data; and performing structured integration on the fault positioning information and the cause speculation data to obtain an intelligent analysis report; and the application can improve the efficiency of sound system fault analysis.

Description

Technical Field

[0001] This invention relates to the field of fault detection technology, and in particular to a machine learning-based method and system for fault analysis of audio systems. Background Technology

[0002] Current audio system fault analysis relies heavily on manual inspection or simple feature comparison. These methods are difficult to accurately capture hidden abnormal features in audio signals, and are prone to missed faults or misjudgments. They cannot identify and warn of potential system faults in advance, and often can only be repaired after the fault actually occurs, which greatly increases the time and manpower costs of system operation and maintenance.

[0003] Existing technologies do not thoroughly decouple fault characteristics, making it difficult to effectively distinguish the characteristic differences corresponding to different fault modes. At the same time, fault diagnosis-related knowledge is difficult to effectively transfer between diagnostic models of different architectures and is difficult to adapt to diverse audio system equipment. This results in low overall efficiency of fault diagnosis, and the generated fault analysis results lack structured integration, failing to provide comprehensive and sufficient support for tracing the root causes of faults. Summary of the Invention

[0004] This invention provides a machine learning-based method and system for fault analysis of audio systems to solve the problems mentioned in the background section.

[0005] To achieve the above objectives, the present invention provides a machine learning-based audio system fault analysis method, comprising: S1. Anomaly feature mining is performed on the audio features of the target audio system using unsupervised clustering to obtain anomaly feature clusters of the target audio system, and the anomaly feature clusters are marked as fault samples to be verified. S2. Decouple the fault features of the fault sample to be verified, and reconstruct the fault mode based on the decoupled features to obtain the enhanced fault feature set of the target audio system. S3. Based on the enhanced fault feature set, a knowledge distillation training is performed on the preset first deep diagnostic network, and the fault judgment knowledge learned in the first deep diagnostic network is transferred to the preset second diagnostic architecture to obtain the embedded diagnostic engine of the target audio system. S4. Based on the embedded diagnostic engine, feature matching is performed on the operating parameter stream of the target audio system, and the deviation between the matching result and the enhanced fault feature set is evaluated to obtain multidimensional abnormal evidence of the target audio system. S5. Perform root cause analysis on the multidimensional abnormal evidence to obtain the fault location information and cause prediction data of the target audio system. S6. The fault location information and the cause prediction data are structured and integrated to obtain an intelligent analysis report of the target audio system.

[0006] In a preferred embodiment, the step of mining abnormal features of the audio features of the target audio system through unsupervised clustering to obtain abnormal feature clusters of the target audio system, and marking the abnormal feature clusters as fault samples to be verified, includes: Acquire multi-channel raw audio signals of the target audio system while it is in operation; The audio features of the multi-channel raw audio signal are subjected to noise filtering to obtain the cleaned feature set of the target audio system; Region identification is performed on the data density distribution in the cleaning feature set, and cluster centers are extracted based on a preset neighborhood radius parameter to obtain candidate cluster centers for the target audio system. Based on the candidate cluster centers, the cleaning feature set is divided into feature spaces to obtain the abnormal feature clusters of the target audio system; The feature data in the abnormal feature cluster are labeled with attributes to obtain the fault sample to be verified of the target audio system.

[0007] In a preferred embodiment, the step of decoupling the fault features of the fault sample to be verified and reconstructing the fault mode based on the decoupled features to obtain the enhanced fault feature set of the target audio system includes: Multivariate covariance analysis was performed on the fault samples to be verified to obtain the feature subspace of the target audio system; Physical mechanism mapping is performed on the feature subspace to obtain the physical semantic feature set of the target audio system; Based on the physical semantic feature set, the conditional generative adversarial framework of the target sound system is trained adversarially to obtain the synthetic feature generator of the target sound system. The synthetic feature generator is used to generate samples to obtain the synthetic feature sample set of the target audio system; The physical semantic feature set and the synthetic feature sample set are mixed and optimized to obtain the enhanced fault feature set of the target audio system.

[0008] In a preferred embodiment, the step of training a preset first deep diagnostic network with knowledge distillation based on the enhanced fault feature set, and transferring the fault judgment knowledge learned in the first deep diagnostic network to a preset second diagnostic architecture to obtain an embedded diagnostic engine for the target audio system, includes: Based on the enhanced fault feature set, teacher training data containing feature samples and corresponding fault labels is constructed; The teacher training data is input into a preset first deep diagnostic network for parameter optimization training to obtain the convergent first deep diagnostic network of the target audio system. The logical relationship between deep feature representations and the final classification layer output is extracted from the first deep diagnostic network that has been trained and converged, thereby obtaining the knowledge abstraction rules of the target audio system; Based on the knowledge abstraction rules, the preset second diagnostic architecture is structurally adapted, and the second diagnostic architecture is fine-tuned and trained using the teacher training data to obtain the embedded diagnostic engine of the target audio system.

[0009] In a preferred embodiment, the step of performing feature matching on the operating parameter stream of the target audio system based on the embedded diagnostic engine, and evaluating the deviation between the matching results and the enhanced fault feature set, yields multidimensional anomaly evidence of the target audio system, including: Based on the feature primitives of the embedded diagnostic engine, online feature distillation is performed on the operating parameter stream of the target audio system to obtain the real-time feature primitive sequence of the target audio system. The consistency between the real-time feature primitive sequence and the enhanced fault feature set is measured to obtain the mismatch tensor of the target audio system; Based on the attention map of the embedded diagnostic engine, the mismatch tensor is weighted and focused to construct the joint deviation index matrix of the target audio system; Based on a preset fault mode knowledge graph, the joint deviation index matrix is ​​decomposed into anomaly contribution to obtain a structured anomaly mode description of the target audio system. By encapsulating the structured anomaly pattern description, the temporal context of the real-time feature primitive sequence, and the attention map into an evidence chain, multidimensional anomaly evidence of the target audio system is obtained.

[0010] In a preferred embodiment, the step of weighted focusing of the mismatch tensor based on the attention map of the embedded diagnostic engine to construct a joint deviation index matrix of the target audio system includes: Based on the feature activation history recorded by the embedded diagnostic engine during the knowledge distillation training process, the discrimination importance of different feature primitives in the enhanced fault feature set under various fault modes is statistically analyzed to obtain the static attention weight distribution of the target audio system. Based on the temporal context information and instantaneous statistical characteristics of the real-time feature primitive sequence, the static attention weight distribution is dynamically modulated to obtain a dynamic attention map of the target audio system. Based on the dynamic attention map, redundancy suppression is performed on the mismatch tensor to obtain the focusing abnormal signal component of the target audio system; The focused abnormal signal components are aggregated along the feature dimension and time dimension at multiple scales, and the weights in the dynamic attention map are used as aggregation coefficients to construct the joint deviation index matrix of the target audio system.

[0011] In a preferred embodiment, the step of tracing the multidimensional abnormal evidence to obtain fault location information and cause prediction data for the target audio system includes: Subgraph isomorphic matching is performed on the abnormal pattern features encapsulated in the multidimensional abnormal evidence to obtain the initial hypothesis space of the target audio system; Based on the temporal priority and statistical dependence of the feature primitive sequences in the multidimensional anomaly evidence, conflict resolution is performed on the initial hypothesis space to obtain the refined hypothesis network of the target audio system. Using the nodes in the refined hypothesis network as node states, the relationships in the refined hypothesis network as edge weights, and the deviation index in the multidimensional anomaly evidence as observation evidence, a dynamic neural causal graph of candidate fault source nodes in the target audio system is constructed to obtain the causal confidence score and explanatory contribution of the target audio system. Based on the causal confidence score and explanatory contribution, the candidate fault source nodes are optimized and ranked to obtain the fault location information of the target audio system. By performing link backtracking on the candidate fault source nodes, the cause prediction data of the target audio system is obtained.

[0012] In a preferred embodiment, the formulas for calculating the causal confidence score and the explanatory contribution are as follows: , ; In the formula, For the candidate fault source node number Causal confidence scores for each node For the candidate fault source node number The node and the first of the observed evidence nodes The strength of causal influence between each piece of evidence. The first of the observed evidence nodes Deviation index of each evidence node, The temporal distance between the candidate fault source node and the observed evidence node in the refined hypothesis network. The preset temporal decay coefficient, For the first The candidate fault source node and the first The prior association degree of each of the historical failure modes in the failure propagation knowledge graph. For the first The significance score of each of the historical failure modes in the current operating cycle. The preset prior fusion coefficients, For the first The total number of observation evidence nodes connected to each of the candidate fault source nodes. For the first The total number of historical failure modes associated with each of the candidate failure source nodes. For the first The observed evidence node for the first The explanatory contribution of each of the candidate fault source nodes.

[0013] In a preferred embodiment, the step of structurally integrating the fault location information and the cause prediction data to obtain an intelligent analysis report of the target audio system includes: Logically associate the component identifiers and fault type codes in the fault location information with the dominant cause paths in the cause prediction data to obtain standardized diagnostic entries for the target audio system. The standardized diagnostic items are supplemented with information from the list of influencing factors in the causal inference data to obtain an extended diagnostic record of the target audio system. The causation prediction data and the extended diagnostic records are encapsulated in an accompanying correlation to obtain a diagnostic conclusion package for the target audio system. Based on preset report generation rules, the diagnostic conclusion package is converted into a natural language description to obtain a preliminary analysis report of the target audio system. The preliminary analysis report is validated for compliance with regulations to obtain the intelligent analysis report of the target audio system.

[0014] To address the aforementioned problems, the present invention also provides a machine learning-based audio system fault analysis system, the system comprising: An anomaly feature mining module is used to mine anomaly features of the audio features of the target audio system through unsupervised clustering, obtain anomaly feature clusters of the target audio system, and mark the anomaly feature clusters as fault samples to be verified. The fault feature reconstruction module is used to decouple the fault features of the fault sample to be verified, and reconstruct the fault mode based on the decoupled features to obtain the enhanced fault feature set of the target audio system. The diagnostic engine construction module is used to perform knowledge distillation training on a preset first deep diagnostic network based on the enhanced fault feature set, and transfer the fault judgment knowledge learned in the first deep diagnostic network to a preset second diagnostic architecture to obtain the embedded diagnostic engine of the target audio system. An anomaly evidence evaluation module is used to perform feature matching on the operating parameter stream of the target audio system based on the embedded diagnostic engine, and evaluate the deviation between the matching result and the enhanced fault feature set to obtain multidimensional anomaly evidence of the target audio system. The fault root cause tracing module is used to trace the root cause of the multidimensional abnormal evidence to obtain the fault location information and cause prediction data of the target audio system. The analysis report generation module is used to structurally integrate the fault location information and the cause prediction data to obtain an intelligent analysis report of the target audio system.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention uses unsupervised clustering to deeply mine the audio anomaly features of a target audio system. After fault feature decoupling and pattern reconstruction, an enhanced fault feature set is formed. Based on knowledge distillation, the transfer of fault judgment knowledge is completed, and an embedded diagnostic engine is constructed. This effectively improves the recognition accuracy of fault features and the scene adaptability of the diagnostic engine, ensuring the efficient advancement of the fault diagnosis process.

[0016] 2. This invention uses an embedded diagnostic engine to perform feature matching and deviation assessment of operating parameter streams, obtain multi-dimensional anomaly evidence, complete accurate root cause tracing, clarify fault location information and cause prediction data, and generate an intelligent analysis report through structured integration. This provides a comprehensive and reliable basis for the rapid handling of audio system faults, improves the overall quality of fault analysis work, and reduces the relevant costs of system operation and maintenance. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a machine learning-based audio system fault analysis method according to an embodiment of the present invention. Figure 2 A functional block diagram of a machine learning-based audio system fault analysis system provided in an embodiment of the present invention; The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] This application provides a machine learning-based method for audio system fault analysis. The execution entity of this machine learning-based audio system fault analysis method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the machine learning-based audio system fault analysis method can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0020] Reference Figure 1 The diagram shown is a flowchart illustrating a machine learning-based audio system fault analysis method according to an embodiment of the present invention. In this embodiment, the machine learning-based audio system fault analysis method includes: S1. Anomaly feature mining is performed on the audio features of the target audio system using unsupervised clustering to obtain anomaly feature clusters of the target audio system, and the anomaly feature clusters are marked as fault samples to be verified. In this embodiment of the invention, the step of mining abnormal features of the audio features of the target audio system through unsupervised clustering to obtain abnormal feature clusters of the target audio system, and marking the abnormal feature clusters as fault samples to be verified, includes: Acquire multi-channel raw audio signals of the target audio system while it is in operation; The audio features of the multi-channel raw audio signal are subjected to noise filtering to obtain the cleaned feature set of the target audio system; Region identification is performed on the data density distribution in the cleaning feature set, and cluster centers are extracted based on a preset neighborhood radius parameter to obtain candidate cluster centers for the target audio system. Based on the candidate cluster centers, the cleaning feature set is divided into feature spaces to obtain the abnormal feature clusters of the target audio system; The feature data in the abnormal feature cluster are labeled with attributes to obtain the fault sample to be verified of the target audio system.

[0021] The system acquires multi-channel raw audio signals of the target audio system while it is in operation. Multiple audio acquisition devices, each corresponding to a sound-generating component and a signal transmission component of the audio system, are used to synchronously acquire the audio signals generated by each component within a preset fixed sampling time. The sampling frequency of each audio acquisition device is set to be consistent with the rated operating frequency of the audio system, and finally, the multi-channel raw audio signals of the target audio system are obtained.

[0022] The audio features of the multi-channel raw audio signal are subjected to noise filtering. A time-domain filtering method is used to remove signal waves that exceed the normal operating frequency band of the audio system. At the same time, interference signals with amplitudes lower than a preset amplitude threshold are removed from the signal, and the effective frequency band signals of the audio features are retained, finally obtaining the cleaned feature set of the target audio system.

[0023] The data density distribution in the cleaned feature set is region identified, and cluster centers are extracted based on a preset neighborhood radius parameter. Different regions are divided according to the distribution density of each data point in the cleaned feature set. The preset neighborhood radius parameter is the maximum range covering the audio feature data distribution under normal operating conditions of the audio system. The location with the most concentrated data distribution in each divided region is selected as the core point, which is the candidate cluster center of the target audio system.

[0024] Based on the candidate cluster centers, the cleaned feature set is partitioned into feature spaces. The distance from each data point in the cleaned feature set to each candidate cluster center is calculated. The data points are assigned to the feature space regions corresponding to the nearest candidate cluster centers. All parts that deviate from the normal data distribution feature space regions are integrated to finally obtain the abnormal feature clusters of the target audio system.

[0025] The feature data in the abnormal feature cluster are labeled with attributes. The labeling content includes the location of the audio acquisition device corresponding to the feature data, the signal acquisition time, and the feature space region to which it belongs. After the labeling is completed, a structured data set is formed. This structured data set is the fault sample to be verified for the target audio system.

[0026] The beneficial effects are as follows: the acquisition of multi-channel raw audio signals can comprehensively cover various audio features under the operating state of the target audio system; noise filtering can effectively remove irrelevant interference signals, ensuring the purity and effectiveness of the cleaned feature set; regional identification and extraction of cluster centers for the data density distribution of the cleaned feature set can accurately locate the feature distribution pattern; feature space division based on candidate cluster centers can efficiently separate abnormal feature clusters; attribute labeling of the feature data of abnormal feature clusters can form structured fault samples to be verified, providing accurate and reliable sample support for subsequent fault feature decoupling and pattern reconstruction, and improving the scientificity and rationality of the overall fault analysis process.

[0027] S2. Decouple the fault features of the fault sample to be verified, and reconstruct the fault mode based on the decoupled features to obtain the enhanced fault feature set of the target audio system. In this embodiment of the invention, the step of decoupling the fault features of the fault sample to be verified and reconstructing the fault mode based on the decoupled features to obtain the enhanced fault feature set of the target audio system includes: Multivariate covariance analysis was performed on the fault samples to be verified to obtain the feature subspace of the target audio system; Physical mechanism mapping is performed on the feature subspace to obtain the physical semantic feature set of the target audio system; Based on the physical semantic feature set, the conditional generative adversarial framework of the target sound system is trained adversarially to obtain the synthetic feature generator of the target sound system. The synthetic feature generator is used to generate samples to obtain the synthetic feature sample set of the target audio system; The physical semantic feature set and the synthetic feature sample set are mixed and optimized to obtain the enhanced fault feature set of the target audio system.

[0028] Multivariate covariance analysis is performed on the fault sample to be verified to extract the correlation information of all feature dimensions in the fault sample to be verified, clarify the dependency and independence relationships between different feature dimensions, and divide the set of feature dimensions that are independent of each other and can completely characterize the fault characteristics. This set of feature dimensions is the feature subspace of the target audio system.

[0029] Physical mechanism mapping is performed on the feature subspace, and each feature dimension in the feature subspace is mapped one-to-one with the actual physical components, operation process and fault evolution mechanism of the target audio system. The physical meaning of each feature dimension is clarified, forming a feature set with clear physical orientation. This feature set is the physical semantic feature set of the target audio system.

[0030] Based on the physical semantic feature set, the conditional generative adversarial framework of the target audio system is trained adversarially. The physical semantic feature set is input into the generation part of the conditional generative adversarial framework, which produces simulated fault feature samples. The simulated fault feature samples and real samples in the physical semantic feature set are then input into the discrimination part, which judges the authenticity of the samples and feeds back to the generation part. The generation part adjusts the sample generation strategy based on the feedback, and the discrimination part optimizes the discrimination criteria simultaneously. The two continue to adjust against each other until the simulated fault feature samples produced by the generation part can completely match the feature distribution of the physical semantic feature set. At this point, the synthetic feature generator of the target audio system is obtained.

[0031] The synthetic feature generator generates samples by taking the feature attributes of the physical semantic feature set as input conditions, driving the synthetic feature generator to continuously produce simulated samples that conform to the distribution law of fault features of the target audio system. All the produced simulated samples are classified and integrated according to feature dimensions to form a well-structured sample set, which is the synthetic feature sample set of the target audio system.

[0032] The physical semantic feature set and the synthetic feature sample set are mixed and optimized. Duplicate feature samples in the two sets are removed, missing fault feature dimensions in the two sets are added, and the distribution density of feature samples in each dimension is adjusted so that the mixed feature set can fully cover all kinds of fault modes of the target audio system, and finally an enhanced fault feature set of the target audio system is obtained.

[0033] The beneficial effects are as follows: performing multivariate covariance analysis on the fault samples to be verified can accurately decouple fault features; the resulting feature subspace can completely characterize fault characteristics; mapping the feature subspace to a physical mechanism can give features a clear physical orientation; the constructed physical semantic feature set has a clear fault association logic; and adversarial training on a conditional generative adversarial framework based on the physical semantic feature set can obtain a stable synthetic feature generator. This generator can produce simulated samples that conform to the distribution law of fault features. The resulting synthetic feature sample set can effectively expand the coverage of fault samples. Hybrid optimization of the physical semantic feature set and the synthetic feature sample set can eliminate redundant samples and supplement missing feature dimensions. The final enhanced fault feature set has a comprehensive and accurate fault characterization capability, providing solid and reliable feature support for the subsequent construction of the diagnostic engine and accurate fault identification.

[0034] S3. Based on the enhanced fault feature set, a knowledge distillation training is performed on the preset first deep diagnostic network, and the fault judgment knowledge learned in the first deep diagnostic network is transferred to the preset second diagnostic architecture to obtain the embedded diagnostic engine of the target audio system. In this embodiment of the invention, the step of training a preset first deep diagnostic network with knowledge distillation based on the enhanced fault feature set, and transferring the fault judgment knowledge learned in the first deep diagnostic network to a preset second diagnostic architecture to obtain the embedded diagnostic engine of the target audio system includes: Based on the enhanced fault feature set, teacher training data containing feature samples and corresponding fault labels is constructed; The teacher training data is input into a preset first deep diagnostic network for parameter optimization training to obtain the convergent first deep diagnostic network of the target audio system. The logical relationship between deep feature representations and the final classification layer output is extracted from the first deep diagnostic network that has been trained and converged, thereby obtaining the knowledge abstraction rules of the target audio system; Based on the knowledge abstraction rules, the preset second diagnostic architecture is structurally adapted, and the second diagnostic architecture is fine-tuned and trained using the teacher training data to obtain the embedded diagnostic engine of the target audio system.

[0035] Based on the enhanced fault feature set, teacher training data containing feature samples and corresponding fault labels is constructed. Each feature sample in the enhanced fault feature set is matched with its corresponding audio system fault type label. The label content corresponds one-to-one with the actual fault type of the audio system. The feature samples and corresponding labels are combined in a fixed format to form a well-structured and complete dataset, which is the teacher training data for the target audio system.

[0036] The teacher training data is input into a preset first deep diagnostic network for parameter optimization training. The teacher training data is continuously input into the first deep diagnostic network, allowing the network to continuously learn the correspondence between feature samples and fault labels. The consistency rate between the fault judgment result output by the network and the sample label is checked simultaneously. When the consistency rate continuously reaches a preset stable threshold and no longer fluctuates, the training is stopped. At this time, the converged first deep diagnostic network of the target audio system is obtained.

[0037] The logical relationship between deep feature representation and final classification layer output is extracted from the first deep diagnostic network that has been trained and converged. The processing flow of the first deep diagnostic network for different fault features is sorted out, the transformation path of deep features within the network is clarified, and the correspondence between different transformation paths and the fault results output by the final classification layer is summarized. The set of these rules is the knowledge abstraction rule of the target audio system.

[0038] Based on the knowledge abstraction rules, the preset second diagnostic architecture is structurally adapted, and the second diagnostic architecture is fine-tuned using the teacher training data. The internal feature processing nodes of the second diagnostic architecture are adjusted according to the knowledge abstraction rules to ensure that the feature processing logic of the architecture is fully matched with the knowledge abstraction rules. Then, the teacher training data is input into the adjusted second diagnostic architecture in batches for targeted training. When the fault judgment accuracy of the architecture meets the preset application standards, the training is stopped, and the embedded diagnostic engine of the target audio system is obtained.

[0039] The beneficial effects are as follows: constructing teacher training data based on enhanced fault feature sets can provide high-quality data support with clear fault orientation for subsequent diagnostic network training. Inputting teacher training data into the first deep diagnostic network for parameter optimization training can enable the network to fully learn the correspondence between fault features and fault types, forming a convergent network with stable fault judgment capabilities. The knowledge abstraction rules extracted from the convergent network can condense the core logic of fault judgment, providing a clear basis for the cross-architecture transfer of diagnostic knowledge. Based on the knowledge abstraction rules, structural adaptation and fine-tuning training of the second diagnostic architecture can achieve efficient transfer of fault judgment knowledge. The resulting embedded diagnostic engine has both accurate fault discrimination capabilities and good scenario adaptability, providing reliable technical support for real-time feature matching and anomaly assessment of subsequent operating parameter streams.

[0040] S4. Based on the embedded diagnostic engine, feature matching is performed on the operating parameter stream of the target audio system, and the deviation between the matching result and the enhanced fault feature set is evaluated to obtain multidimensional abnormal evidence of the target audio system. In this embodiment of the invention, the step of performing feature matching on the operating parameter stream of the target audio system based on the embedded diagnostic engine, and evaluating the deviation between the matching results and the enhanced fault feature set to obtain multidimensional anomaly evidence of the target audio system includes: Based on the feature primitives of the embedded diagnostic engine, online feature distillation is performed on the operating parameter stream of the target audio system to obtain the real-time feature primitive sequence of the target audio system. The consistency between the real-time feature primitive sequence and the enhanced fault feature set is measured to obtain the mismatch tensor of the target audio system; Based on the attention map of the embedded diagnostic engine, the mismatch tensor is weighted and focused to construct the joint deviation index matrix of the target audio system; Based on a preset fault mode knowledge graph, the joint deviation index matrix is ​​decomposed into anomaly contribution to obtain a structured anomaly mode description of the target audio system. By encapsulating the structured anomaly pattern description, the temporal context of the real-time feature primitive sequence, and the attention map into an evidence chain, multidimensional anomaly evidence of the target audio system is obtained.

[0041] The attention map based on the embedded diagnostic engine is used to weight and focus the mismatch tensor to construct a joint deviation index matrix for the target audio system, including: Based on the feature activation history recorded by the embedded diagnostic engine during the knowledge distillation training process, the discrimination importance of different feature primitives in the enhanced fault feature set under various fault modes is statistically analyzed to obtain the static attention weight distribution of the target audio system. Based on the temporal context information and instantaneous statistical characteristics of the real-time feature primitive sequence, the static attention weight distribution is dynamically modulated to obtain a dynamic attention map of the target audio system. Based on the dynamic attention map, redundancy suppression is performed on the mismatch tensor to obtain the focusing abnormal signal component of the target audio system; The focused abnormal signal components are aggregated along the feature dimension and time dimension at multiple scales, and the weights in the dynamic attention map are used as aggregation coefficients to construct the joint deviation index matrix of the target audio system.

[0042] Based on the feature primitives of the embedded diagnostic engine, online feature distillation is performed on the operating parameter stream of the target audio system. The feature primitives of the embedded diagnostic engine are used as the reference feature template. The feature components of the operating parameter stream of the target audio system are extracted time-by-time. The effective parameter features that match the reference feature template are retained, and redundant parameter information that is irrelevant to fault determination is removed. The extracted effective parameter features are arranged in chronological order to form an ordered feature sequence. This feature sequence is the real-time feature primitive sequence of the target audio system.

[0043] The consistency of the real-time feature primitive sequence with the enhanced fault feature set is measured. Each feature component in the real-time feature primitive sequence is compared with the corresponding fault feature template in the enhanced fault feature set dimension by dimension. The feature consistency difference in each feature dimension is recorded. The differences in all dimensions are structured and integrated according to the feature dimension and the time dimension to form a data set in tensor form. This data set is the mismatch tensor of the target audio system.

[0044] Based on the attention map of the embedded diagnostic engine, the mismatch tensor is weighted and focused to construct the joint deviation index matrix of the target audio system. The attention map of the embedded diagnostic engine marks the importance level of each feature primitive in the fault identification process. According to the importance level, the difference data at different positions in the mismatch tensor are assigned corresponding weight values ​​to strengthen the difference signals corresponding to key fault features and weaken the difference signals corresponding to irrelevant features. The weighted difference data is aggregated along the feature dimension and time dimension at multiple scales to finally form the joint deviation index matrix of the target audio system.

[0045] Based on a pre-defined fault mode knowledge graph, the joint deviation index matrix is ​​decomposed into anomaly contribution. The pre-defined fault mode knowledge graph contains the correspondence between various fault modes and characteristic anomalies. The values ​​in the joint deviation index matrix are matched one by one with the fault modes in the knowledge graph. The characteristic anomaly contribution degree corresponding to each fault mode is calculated. The matching results and contribution degree information are structured and organized to form a textual description. This description is the structured anomaly mode description of the target audio system.

[0046] The structured anomaly pattern description, the temporal context of the real-time feature primitive sequence, and the attention graph are encapsulated into an evidence chain. The structured anomaly pattern description is used as the core evidence, the temporal context of the real-time feature primitive sequence is used as the temporal dimension of the evidence, and the attention graph is used as the weight dimension of the evidence. The three are linked and integrated according to a preset evidence chain format, so that the information of the three is mutually corroborated and logically coherent, and finally multi-dimensional anomaly evidence of the target audio system is obtained.

[0047] Based on the feature activation history recorded by the embedded diagnostic engine during the knowledge distillation training process, the activation frequency and duration of each feature primitive in the judgment process of various fault modes are sorted out. The higher the activation frequency and the longer the activation duration, the higher the discrimination importance of the feature primitive. According to the comprehensive statistical results of activation frequency and activation duration, each feature primitive in the enhanced fault feature set is assigned a corresponding weight value to form the static attention weight distribution of the target audio system.

[0048] Based on the temporal context information and instantaneous statistical characteristics of the real-time feature primitive sequence, the static attention weight distribution is dynamically modulated. The temporal context information is the sequential relationship of the feature primitives in the time dimension, and the instantaneous statistical characteristics are the distribution state of the feature primitives at each moment. If a feature primitive has a close temporal relationship at the current moment and its distribution state matches the fault mode well, the weight value of the feature primitive is increased; otherwise, it is decreased. Finally, the dynamic attention map of the target audio system is obtained.

[0049] Based on the dynamic attention map, redundancy suppression is performed on the mismatch tensor. All mismatch tensor parts corresponding to feature primitives with weight values ​​lower than a preset weight threshold in the dynamic attention map are removed, and only the mismatch tensor parts corresponding to feature primitives with weight values ​​higher than the preset weight threshold are retained. These retained mismatch tensor parts are the focus abnormal signal components of the target audio system.

[0050] The focused abnormal signal components are aggregated at multiple scales along the feature dimension and time dimension, and the weights in the dynamic attention map are used as aggregation coefficients. The feature dimension aggregation is grouped and integrated according to the category of feature primitives, and the time dimension aggregation is segmented and integrated according to the preset time interval. The aggregation calculation of each group and segment uses the weight value of the corresponding feature primitive as the coefficient. The integrated aggregation results are arranged in order according to the feature dimension and time dimension to construct the joint deviation index matrix of the target audio system.

[0051] The beneficial effects are as follows: Online feature distillation of the operating parameter stream based on the feature primitives of the embedded diagnostic engine effectively retains relevant feature information for fault diagnosis, eliminates redundant parameter content, and generates a real-time feature primitive sequence with clear temporal correlation. Consistency measurement between the real-time feature primitive sequence and the enhanced fault feature set accurately captures the feature differences between the two. The resulting mismatch tensor fully reflects feature deviations. Weighted focusing of the mismatch tensor is achieved using the attention map of the embedded diagnostic engine. A static attention weight distribution is obtained through statistical feature activation history, and dynamic modulation is completed by combining the temporal context and instantaneous statistical characteristics of the real-time feature primitive sequence. The generated dynamic attention map can accurately mark key fault diagnosis features. Redundancy suppression using the attention graph can eliminate irrelevant anomalous signals while retaining focused anomalous signal components. Multi-scale aggregation of focused anomalous signal components along the feature and time dimensions enables multi-dimensional integration of anomalous features. The constructed joint deviation index matrix can comprehensively reflect the distribution of anomalies. Based on a pre-defined fault mode knowledge graph, the joint deviation index matrix is ​​decomposed into anomaly contribution, forming a structured anomalous pattern description with a clear fault indication. The structured anomalous pattern description, the temporal context of the real-time feature primitive sequence, and the attention graph are encapsulated into an evidence chain. The resulting multi-dimensional anomalous evidence contains information in the feature, time, and weight dimensions, possessing comprehensiveness and credibility, and providing accurate and solid evidence support for subsequent fault root cause tracing.

[0052] S5. Perform root cause analysis on the multidimensional abnormal evidence to obtain the fault location information and cause prediction data of the target audio system. In this embodiment of the invention, the step of tracing the multidimensional abnormal evidence to obtain fault location information and cause prediction data of the target audio system includes: Subgraph isomorphic matching is performed on the abnormal pattern features encapsulated in the multidimensional abnormal evidence to obtain the initial hypothesis space of the target audio system; Based on the temporal priority and statistical dependence of the feature primitive sequences in the multidimensional anomaly evidence, conflict resolution is performed on the initial hypothesis space to obtain the refined hypothesis network of the target audio system. Using the nodes in the refined hypothesis network as node states, the relationships in the refined hypothesis network as edge weights, and the deviation index in the multidimensional anomaly evidence as observation evidence, a dynamic neural causal graph of candidate fault source nodes in the target audio system is constructed to obtain the causal confidence score and explanatory contribution of the target audio system. Based on the causal confidence score and explanatory contribution, the candidate fault source nodes are optimized and ranked to obtain the fault location information of the target audio system. By performing link backtracking on the candidate fault source nodes, the cause prediction data of the target audio system is obtained.

[0053] The formulas for the causal confidence score and the explanatory contribution are as follows: , ; In the formula, For the candidate fault source node number Causal confidence scores for each node For the candidate fault source node number The node and the first of the observed evidence nodes The strength of causal influence between each piece of evidence. The first of the observed evidence nodes Deviation index of each evidence node, The temporal distance between the candidate fault source node and the observed evidence node in the refined hypothesis network. The preset temporal decay coefficient, For the first The candidate fault source node and the first The prior association degree of each of the historical failure modes in the failure propagation knowledge graph. For the first The significance score of each of the historical failure modes in the current operating cycle. The preset prior fusion coefficients, For the first The total number of observation evidence nodes connected to each of the candidate fault source nodes. For the first The total number of historical failure modes associated with each of the candidate failure source nodes. For the first The observed evidence node for the first The explanatory contribution of each of the candidate fault source nodes.

[0054] Subgraph isomorphic matching is performed on the abnormal pattern features encapsulated in the multidimensional abnormal evidence. Various fault modes in the preset fault mode knowledge graph are transformed into standardized subgraph structures. At the same time, the abnormal pattern features in the multidimensional abnormal evidence are extracted and transformed into feature subgraphs. The feature subgraphs are compared with the standardized subgraphs node by node and edge by edge. The set of fault hypotheses corresponding to the fully matched standardized subgraphs is the initial hypothesis space of the target audio system.

[0055] Based on the temporal priority and statistical dependence of the feature primitive sequences in the multidimensional anomaly evidence, conflict resolution is performed on the initial hypothesis space. Temporal priority refers to the order in which feature primitives appear in the time dimension, and statistical dependence refers to the degree of correlation between different feature primitives. Each fault hypothesis in the initial hypothesis space is verified, and fault hypotheses that conflict with the temporal priority and statistical dependence of the feature primitive sequences are eliminated. Fault hypotheses that perfectly match the two are retained and structurally integrated. The integrated fault hypothesis network is the refined hypothesis network of the target audio system.

[0056] Using the nodes in the refined hypothesis network as node states, the relationships in the refined hypothesis network as edge weights, and the deviation index in the multidimensional anomaly evidence as observed evidence, a dynamic neural causal graph of candidate fault source nodes in the target audio system is constructed. The candidate fault source nodes in the refined hypothesis network are taken as core nodes, and the corresponding observed evidence nodes are taken as peripheral nodes. The edge weights characterize the causal association strength between core nodes and peripheral nodes, and the deviation index is used as the state identifier of peripheral nodes. Through this association method, a causal mapping relationship between nodes is established to form a dynamic neural causal graph. Based on the node association logic of this graph, the causal confidence score and explanatory contribution of the target audio system are obtained.

[0057] Based on the causal confidence score and explanatory contribution, the candidate fault source nodes are optimized and ranked. The preset ranking criterion is that the higher the comprehensive value of the causal confidence score and explanatory contribution, the higher the priority of the candidate fault source node. According to this criterion, all candidate fault source nodes are ranked by priority. The candidate fault source node with the highest priority in the ranking result and the corresponding fault association information are the fault location information of the target audio system.

[0058] Link backtracking is performed on the candidate fault source nodes. Starting from the high-priority candidate fault source nodes determined by the ranking, the backtracking is carried out along the causal relationship links between nodes in the refined hypothesis network. The complete path of the fault propagation from the candidate fault source nodes to each observation evidence node is sorted out. All related nodes, relationships between nodes and corresponding status information on the path are recorded. The collection formed by integrating these recorded contents is the cause prediction data of the target sound system.

[0059] The causal influence strength value between candidate fault source nodes and observed evidence nodes comes from a refined hypothesis network, which is formed by resolving conflicts of multidimensional anomaly evidence. The relationship between nodes directly corresponds to the specific content of the causal influence strength value.

[0060] The deviation index of the observed evidence nodes comes from the joint deviation index matrix, which is obtained by weighted focusing and multi-scale aggregation of the mismatch tensor through dynamic attention graph.

[0061] The temporal distance between candidate fault source nodes and observed evidence nodes comes from the refined hypothesis network, and the difference in their order is determined based on the temporal position of the nodes in the network.

[0062] The temporal decay coefficient is a preset fixed value used to measure the degree of influence of temporal distance on causal relationships.

[0063] The prior correlation between candidate fault source nodes and historical fault modes comes from the fault propagation knowledge graph, and the specific content is determined based on the correlation between nodes and historical fault modes recorded in the graph.

[0064] The significance score of historical failure modes is derived from statistical results within the current operating cycle and is determined based on the frequency of occurrence and scope of impact of historical failure modes within the cycle.

[0065] The prior fusion coefficient is a preset fixed value used to measure the degree of influence of prior association on causal confidence.

[0066] The total number of observation evidence nodes connected to the candidate fault source node is obtained from the refined hypothesis network by counting the number of observation evidence nodes that are directly associated with the node.

[0067] The total number of historical fault modes associated with the candidate fault source node is obtained from the fault propagation knowledge graph by counting the number of historical fault modes associated with the node in the graph.

[0068] The formula for calculating the causal confidence score is used to assess the credibility of candidate fault source nodes as the true fault source of the target audio system. It comprehensively considers the influence of causal impact strength, deviation index, and temporal distance, while also incorporating the prior association and significance of historical fault modes to achieve an accurate evaluation of candidate fault source nodes. The formula for calculating the explanatory contribution is used to assess the degree to which a single observational evidence node supports the determination that its corresponding candidate fault source node is the true fault source. By comparing the causal impact strength and deviation index product of this evidence node with other evidence nodes, the specific role of a single evidence node in the fault determination process is clarified, providing a quantitative basis for determining fault location information.

[0069] The beneficial effects are as follows: Subgraph isomorphic matching of abnormal pattern features in multidimensional abnormal evidence can accurately match fault patterns and form an initial hypothesis space, laying a solid foundation for subsequent root cause tracing. Conflict resolution of the initial hypothesis space based on the temporal priority and statistical dependence of feature primitive sequences can effectively eliminate conflicting fault hypotheses, resulting in a logically rigorous and refined hypothesis network. A dynamic neural causal graph is constructed using the nodes, relationships, and deviation indices of the refined hypothesis network, which can quantify the causal confidence score and explanatory contribution of candidate fault source nodes, providing a quantitative basis for fault source determination. Based on the above scores and contributions, candidate fault source nodes are optimized and ranked, accurately pinpointing the fault location and obtaining fault location information. Link backtracking of candidate fault source nodes can clearly identify the fault propagation path and obtain causal inference data. The entire process achieves accurate root cause tracing of faults, providing a comprehensive and reliable basis for the generation of intelligent analysis reports.

[0070] S6. The fault location information and the cause prediction data are structured and integrated to obtain an intelligent analysis report of the target audio system.

[0071] In this embodiment of the invention, the step of structurally integrating the fault location information and the cause prediction data to obtain an intelligent analysis report of the target audio system includes: Logically associate the component identifiers and fault type codes in the fault location information with the dominant cause paths in the cause prediction data to obtain standardized diagnostic entries for the target audio system. The standardized diagnostic items are supplemented with information from the list of influencing factors in the causal inference data to obtain an extended diagnostic record of the target audio system. The causation prediction data and the extended diagnostic records are encapsulated in an accompanying correlation to obtain a diagnostic conclusion package for the target audio system. Based on preset report generation rules, the diagnostic conclusion package is converted into a natural language description to obtain a preliminary analysis report of the target audio system. The preliminary analysis report is validated for compliance with regulations to obtain the intelligent analysis report of the target audio system.

[0072] The component identifiers and fault type codes in the fault location information are logically associated with the dominant cause paths in the cause prediction data. The specific faulty parts corresponding to the component identifiers, the fault categories corresponding to the fault type codes, and the fault propagation chains corresponding to the dominant cause paths are bound one-to-one. The fault type of each faulty part and the complete path that causes the fault are clarified. These are integrated into a set of items with a unified structure and clear content. This set of items is the standardized diagnostic items for the target audio system.

[0073] The standardized diagnostic items are supplemented with information from the list of influencing factors in the cause prediction data. The environmental conditions, equipment operating time, and component wear and tear in the list of influencing factors are added to the cause description section of each standardized diagnostic item. This ensures that the information in each diagnostic item covers all dimensions of the fault-related factors. The information set formed after the supplementation is completed is the extended diagnostic record of the target audio system.

[0074] The causal inference data and the extended diagnostic records are encapsulated in an accompanying association manner. Each extended diagnostic record is matched with corresponding complete causal inference data. The two are combined according to a preset encapsulation format to ensure that the correlation between the extended diagnostic records and the causal inference data is not lost. The combined overall data set is the diagnostic conclusion package of the target audio system.

[0075] Based on preset report generation rules, the diagnostic conclusion package is converted into a natural language description. The preset report generation rules include four fixed chapters: fault overview, fault cause, influencing factors, and propagation path, as well as specific expression specifications for each chapter. According to these rules, the structured data in the diagnostic conclusion package is converted into fluent and coherent natural language text, and the converted text is the preliminary analysis report of the target audio system.

[0076] The preliminary analysis report is then subjected to a standardization check. The check criteria include three dimensions: uniformity of terminology, coherence of logical structure, and completeness of information content. The terminology used must conform to the general standards in the field of audio system fault diagnosis, the logical structure must be arranged in an orderly manner according to the preset chapters, and the information content must contain no missing fault-related elements. Any content that does not meet the check criteria is corrected, and after the correction is completed, an intelligent analysis report of the target audio system is obtained.

[0077] The beneficial effects are as follows: By logically associating component identifiers and fault type codes in fault location information with dominant causal paths in cause prediction data, standardized diagnostic items with clear correspondences can be formed, ensuring the regularity and relevance of diagnostic information. Supplementing the standardized diagnostic items with the list of influencing factors in cause prediction data can enrich the dimensions of diagnostic content, forming an extended diagnostic record covering all related elements of the fault. Encapsulating the cause prediction data and extended diagnostic records with accompanying associations can integrate complete and coherent diagnostic information, forming a logically closed-loop diagnostic conclusion package. Based on preset report generation rules, the diagnostic conclusion package is converted into a preliminary analysis report with natural language description, which can improve the readability and understandability of the report. Standardization verification of the preliminary analysis report can ensure that the terminology used in the report is consistent and the logical structure is rigorous. The final intelligent analysis report is comprehensive, accurate and practical, providing a clear and reliable basis for the subsequent handling, operation and maintenance optimization and prevention of similar faults in the target audio system.

[0078] like Figure 2 The diagram shown is a functional block diagram of a machine learning-based audio system fault analysis system provided in an embodiment of the present invention.

[0079] The machine learning-based audio system fault analysis system 100 described in this invention can be installed in an electronic device. Depending on the functions implemented, the machine learning-based audio system fault analysis system 100 may include an anomaly feature mining module 101, a fault feature reconstruction module 102, a diagnostic engine construction module 103, an anomaly evidence evaluation module 104, a fault root cause tracing module 105, and an analysis report generation module 106. The modules described in this invention can also be referred to as units, which are a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, stored in the memory of the electronic device.

[0080] In this embodiment, the functions of each module / unit are as follows: The abnormal feature mining module 101 is used to mine abnormal features of the audio features of the target audio system through unsupervised clustering, obtain abnormal feature clusters of the target audio system, and mark the abnormal feature clusters as fault samples to be verified. The fault feature reconstruction module 102 is used to decouple the fault features of the fault sample to be verified, and reconstruct the fault mode based on the decoupled features to obtain the enhanced fault feature set of the target audio system. The diagnostic engine construction module 103 is used to perform knowledge distillation training on a preset first deep diagnostic network based on the enhanced fault feature set, and transfer the fault judgment knowledge learned in the first deep diagnostic network to a preset second diagnostic architecture to obtain the embedded diagnostic engine of the target audio system. The abnormal evidence evaluation module 104 is used to perform feature matching on the operating parameter stream of the target audio system based on the embedded diagnostic engine, and evaluate the deviation between the matching result and the enhanced fault feature set to obtain multidimensional abnormal evidence of the target audio system. The fault root cause tracing module 105 is used to trace the root cause of the multidimensional abnormal evidence to obtain the fault location information and cause prediction data of the target audio system. The analysis report generation module 106 is used to perform structured integration of the fault location information and the cause prediction data to obtain an intelligent analysis report of the target audio system.

[0081] In the several embodiments provided by this invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0082] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0083] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0084] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0085] This application embodiment can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A machine learning-based method for fault analysis of audio systems, characterized in that, The method includes: S1. Anomaly feature mining is performed on the audio features of the target audio system using unsupervised clustering to obtain anomaly feature clusters of the target audio system, and the anomaly feature clusters are marked as fault samples to be verified. S2. Decouple the fault features of the fault sample to be verified, and reconstruct the fault mode based on the decoupled features to obtain the enhanced fault feature set of the target audio system. S3. Based on the enhanced fault feature set, a knowledge distillation training is performed on the preset first deep diagnostic network, and the fault judgment knowledge learned in the first deep diagnostic network is transferred to the preset second diagnostic architecture to obtain the embedded diagnostic engine of the target audio system. S4. Based on the embedded diagnostic engine, feature matching is performed on the operating parameter stream of the target audio system, and the deviation between the matching result and the enhanced fault feature set is evaluated to obtain multidimensional abnormal evidence of the target audio system. S5. Perform root cause analysis on the multidimensional abnormal evidence to obtain the fault location information and cause prediction data of the target audio system. S6. The fault location information and the cause prediction data are structured and integrated to obtain an intelligent analysis report of the target audio system.

2. The machine learning-based audio system fault analysis method as described in claim 1, characterized in that, The process of mining abnormal features from the audio features of the target audio system using unsupervised clustering to obtain abnormal feature clusters of the target audio system, and marking these abnormal feature clusters as fault samples to be verified, includes: Acquire multi-channel raw audio signals of the target audio system while it is in operation; The audio features of the multi-channel raw audio signal are subjected to noise filtering to obtain the cleaned feature set of the target audio system; Region identification is performed on the data density distribution in the cleaning feature set, and cluster centers are extracted based on a preset neighborhood radius parameter to obtain candidate cluster centers for the target audio system. Based on the candidate cluster centers, the cleaning feature set is divided into feature spaces to obtain the abnormal feature clusters of the target audio system; The feature data in the abnormal feature cluster are labeled with attributes to obtain the fault sample to be verified of the target audio system.

3. The machine learning-based audio system fault analysis method as described in claim 1, characterized in that, The process involves decoupling the fault features of the fault samples to be verified and reconstructing the fault modes based on the decoupled features to obtain an enhanced fault feature set for the target audio system, including: Multivariate covariance analysis was performed on the fault samples to be verified to obtain the feature subspace of the target audio system; Physical mechanism mapping is performed on the feature subspace to obtain the physical semantic feature set of the target audio system; Based on the physical semantic feature set, the conditional generative adversarial framework of the target sound system is trained adversarially to obtain the synthetic feature generator of the target sound system. The synthetic feature generator is used to generate samples to obtain the synthetic feature sample set of the target audio system; The physical semantic feature set and the synthetic feature sample set are mixed and optimized to obtain the enhanced fault feature set of the target audio system.

4. The machine learning-based audio system fault analysis method as described in claim 1, characterized in that, The process involves training a preset first deep diagnostic network using knowledge distillation based on the enhanced fault feature set, and then transferring the fault judgment knowledge learned in the first deep diagnostic network to a preset second diagnostic architecture to obtain an embedded diagnostic engine for the target audio system, comprising: Based on the enhanced fault feature set, teacher training data containing feature samples and corresponding fault labels is constructed; The teacher training data is input into a preset first deep diagnostic network for parameter optimization training to obtain the convergent first deep diagnostic network of the target audio system. The logical relationship between deep feature representations and the final classification layer output is extracted from the first deep diagnostic network that has been trained and converged, thereby obtaining the knowledge abstraction rules of the target audio system; Based on the knowledge abstraction rules, the preset second diagnostic architecture is structurally adapted, and the second diagnostic architecture is fine-tuned and trained using the teacher training data to obtain the embedded diagnostic engine of the target audio system.

5. The machine learning-based audio system fault analysis method as described in claim 1, characterized in that, Based on the embedded diagnostic engine, feature matching is performed on the operating parameter stream of the target audio system, and the deviation between the matching results and the enhanced fault feature set is evaluated to obtain multidimensional anomaly evidence of the target audio system, including: Based on the feature primitives of the embedded diagnostic engine, online feature distillation is performed on the operating parameter stream of the target audio system to obtain the real-time feature primitive sequence of the target audio system. The consistency between the real-time feature primitive sequence and the enhanced fault feature set is measured to obtain the mismatch tensor of the target audio system; Based on the attention map of the embedded diagnostic engine, the mismatch tensor is weighted and focused to construct the joint deviation index matrix of the target audio system; Based on a preset fault mode knowledge graph, the joint deviation index matrix is ​​decomposed into anomaly contribution to obtain a structured anomaly mode description of the target audio system. By encapsulating the structured anomaly pattern description, the temporal context of the real-time feature primitive sequence, and the attention map into an evidence chain, multidimensional anomaly evidence of the target audio system is obtained.

6. The machine learning-based audio system fault analysis method as described in claim 5, characterized in that, The attention map based on the embedded diagnostic engine is used to weight and focus the mismatch tensor to construct a joint deviation index matrix for the target audio system, including: Based on the feature activation history recorded by the embedded diagnostic engine during the knowledge distillation training process, the discrimination importance of different feature primitives in the enhanced fault feature set under various fault modes is statistically analyzed to obtain the static attention weight distribution of the target audio system. Based on the temporal context information and instantaneous statistical characteristics of the real-time feature primitive sequence, the static attention weight distribution is dynamically modulated to obtain a dynamic attention map of the target audio system. Based on the dynamic attention map, redundancy suppression is performed on the mismatch tensor to obtain the focusing abnormal signal component of the target audio system; The focused abnormal signal components are aggregated along the feature dimension and time dimension at multiple scales, and the weights in the dynamic attention map are used as aggregation coefficients to construct the joint deviation index matrix of the target audio system.

7. The machine learning-based audio system fault analysis method as described in claim 1, characterized in that, The root cause tracing of the multidimensional abnormal evidence to obtain fault location information and cause inference data of the target audio system includes: Subgraph isomorphic matching is performed on the abnormal pattern features encapsulated in the multidimensional abnormal evidence to obtain the initial hypothesis space of the target audio system; Based on the temporal priority and statistical dependence of the feature primitive sequences in the multidimensional anomaly evidence, conflict resolution is performed on the initial hypothesis space to obtain the refined hypothesis network of the target audio system. Using the nodes in the refined hypothesis network as node states, the relationships in the refined hypothesis network as edge weights, and the deviation index in the multidimensional anomaly evidence as observation evidence, a dynamic neural causal graph of candidate fault source nodes in the target audio system is constructed to obtain the causal confidence score and explanatory contribution of the target audio system. Based on the causal confidence score and explanatory contribution, the candidate fault source nodes are optimized and ranked to obtain the fault location information of the target audio system. Link backtracking is performed on the candidate fault source nodes to obtain the inferred cause data of the target audio system.

8. The machine learning-based audio system fault analysis method as described in claim 7, characterized in that, The formulas for calculating the causal confidence score and the explanatory contribution are as follows: ， ； In the formula, For the candidate fault source node number Causal confidence scores for each node For the candidate fault source node number The node and the first of the observed evidence nodes The strength of causal influence between each piece of evidence. The first of the observed evidence nodes Deviation index of each evidence node, The temporal distance between the candidate fault source node and the observed evidence node in the refined hypothesis network. The preset temporal decay coefficient, For the first The candidate fault source node and the first The prior association degree of each of the historical failure modes in the failure propagation knowledge graph. For the first The significance score of each of the historical failure modes in the current operating cycle. The preset prior fusion coefficients, For the first The total number of observation evidence nodes connected to each of the candidate fault source nodes. For the first The total number of historical failure modes associated with each of the candidate failure source nodes. For the first The observed evidence node for the first The explanatory contribution of each of the candidate fault source nodes.

9. The machine learning-based audio system fault analysis method as described in claim 1, characterized in that, The structured integration of the fault location information and the cause prediction data to obtain the intelligent analysis report of the target audio system includes: Logically associate the component identifiers and fault type codes in the fault location information with the dominant cause paths in the cause prediction data to obtain standardized diagnostic entries for the target audio system. The standardized diagnostic items are supplemented with information from the list of influencing factors in the causal inference data to obtain an extended diagnostic record of the target audio system. The causation prediction data and the extended diagnostic records are encapsulated in an accompanying correlation to obtain a diagnostic conclusion package for the target audio system. Based on preset report generation rules, the diagnostic conclusion package is converted into a natural language description to obtain a preliminary analysis report of the target audio system. The preliminary analysis report is validated for compliance with regulations to obtain the intelligent analysis report of the target audio system.

10. A machine learning-based audio system fault analysis system, characterized in that, A system for implementing a machine learning-based audio system fault analysis method as described in claim 1, the system comprising: An anomaly feature mining module is used to mine anomaly features of the audio features of the target audio system through unsupervised clustering, obtain anomaly feature clusters of the target audio system, and mark the anomaly feature clusters as fault samples to be verified. The fault feature reconstruction module is used to decouple the fault features of the fault sample to be verified, and reconstruct the fault mode based on the decoupled features to obtain the enhanced fault feature set of the target audio system. The diagnostic engine construction module is used to perform knowledge distillation training on a preset first deep diagnostic network based on the enhanced fault feature set, and transfer the fault judgment knowledge learned in the first deep diagnostic network to a preset second diagnostic architecture to obtain the embedded diagnostic engine of the target audio system. An anomaly evidence evaluation module is used to perform feature matching on the operating parameter stream of the target audio system based on the embedded diagnostic engine, and evaluate the deviation between the matching result and the enhanced fault feature set to obtain multidimensional anomaly evidence of the target audio system. The fault root cause tracing module is used to trace the root cause of the multidimensional abnormal evidence to obtain the fault location information and cause prediction data of the target audio system. The analysis report generation module is used to structurally integrate the fault location information and the cause prediction data to obtain an intelligent analysis report of the target audio system.

Images