Fault detection method for multiple communication channels and self-healing control device

Abstract

The invention relates to the technical field of communication, and provides a multi-communication channel fault detection method and a self-healing control device, and the method achieves the three-dimensional description of a complex network fault through the fusion processing of multi-dimensional fault characterization parameters, and provides a comprehensive and structured data basis for the subsequent analysis. Accurate identification and positioning of a fault mode are realized through a space-time convolutional neural network, and a cascade risk of the fault mode is determined to avoid a secondary fault caused by an isolation operation; and finally, the state of a non-target channel is evaluated through a residual fault index, and channel reconfiguration and parameter optimization of the space-time convolutional neural network are triggered in combination with accumulation of the number of abnormal events. Therefore, the communication network can actively identify and process faults in a complex and changeable environment, the fault influence range is effectively reduced, the stability and reliability of the communication network are continuously improved, and the conversion from passive fault handling to active network management is realized.

Description

Technical Field

[0001] The present application relates to the field of communication technology, and in particular to a fault detection method and a self-healing control device for multiple communication channels. Background Art

[0002] In the development of communication network technology, fault detection in multiple communication channels has always been a core link in ensuring network reliability. Early fault detection mainly relied on manual inspections and simple threshold judgments. As network scale expands and service complexity increases, the detection method based on single-parameter thresholds has gradually evolved to multi-parameter analysis, achieving fault location by integrating multiple physical layer indicators.

[0003] In recent years, machine learning technology has begun to be applied to network fault detection. By training historical fault data with machine learning methods, the classification and identification of multi-channel faults can be achieved. Furthermore, convolutional neural networks and recurrent neural networks in deep learning technology have been introduced. Among them, convolutional neural networks have shown advantages in spatial feature extraction, while recurrent neural networks are good at processing time series data. The combination of the two has enabled fault detection to develop from single-channel analysis to multi-channel spatiotemporal correlation analysis.

[0004] While some solutions employ multi-parameter analysis, they only treat each parameter independently or simply weight them, failing to consider the temporal and spatial correlations between them. Furthermore, traditional machine learning methods rely on manual feature extraction, making it difficult to capture underlying fault characteristics. Furthermore, existing self-healing solutions often rely solely on fault detection results when performing channel isolation or reconfiguration, failing to fully assess the cascading risks associated with isolation operations. Most solutions fail to address how to detect and identify complex channel fault patterns and detect residual faults after faulty channels are isolated. Summary of the Invention

[0005] In response to the deficiencies of the prior art, the present application provides a multi-communication channel fault detection method and a self-healing control device.

[0006] In a first aspect, the present application provides a fault detection method for multiple communication channels, the method comprising: obtaining fault characterization parameters of multiple channels at a communication node, the fault characterization parameters including delay jitter, signal-to-noise ratio, and signal quality of each channel, aligning the fault characterization parameters with a time window according to a channel index, and fusing them into a fault feature tensor;

[0007] The fault feature tensor is processed through a spatiotemporal convolutional neural network to determine the fault mode and output the confidence probability of the fault mode. The fault mode includes local faults and global faults. When the confidence probability is high, a channel fault mark is generated based on the fault feature tensor. The target channel is determined based on the channel fault mark, and a detection signal is injected into the target channel. The bit error rate of the detection signal is used to determine the trigger signal isolation instruction.

[0008] Receive signal isolation instructions and isolate the target channel, while reacquiring the fault characterization parameters of the non-target channel. The residual fault index of the non-target channel is comprehensively judged based on the residual of the delay jitter and the degradation of the signal quality. Abnormal events are marked based on the abnormal judgment of the residual fault index and the number of abnormal events is accumulated. Channel reconfiguration is performed based on the number of abnormal events and the parameters of the spatiotemporal convolutional neural network are iteratively optimized.

[0009] As an optional implementation, the fusion logic of the fault feature tensor includes:

[0010] Acquire multi-channel fault characterization parameters at the communication node and perform normalization on the fault characterization parameters;

[0011] Construct a spatiotemporal alignment matrix with channel index as columns and acquisition time within the time window as rows, and fill the normalized fault characterization parameters into the spatiotemporal alignment matrix;

[0012] According to the historical contribution of each fault characterization parameter to fault judgment, a weight coefficient is assigned to each fault characterization parameter through the attention mechanism to form a weighted matrix. The weighted matrix and the spatiotemporal alignment matrix are subjected to a tensor product operation to obtain the fault feature tensor.

[0013] As an optional implementation, the triggering logic of the signal isolation instruction includes:

[0014] Determine the target channel based on the channel fault flag and dynamically adjust the detection signal parameters according to the fault mode and severity. The detection signal parameters include signal strength, frequency, and modulation mode.

[0015] The bit error rate of the detection signal is analyzed by the sequential probability ratio test method, and a bit error rate threshold is set. The bit error judgment result is obtained by comparing the bit error rate of the detection signal with the bit error rate threshold;

[0016] Determine the cascading risk of signal isolation based on the propagation risk, configure the risk threshold, compare the cascading risk of signal isolation with the risk threshold, and obtain the isolation risk result. Make a comprehensive judgment based on the error judgment result and the isolation risk result to trigger the signal isolation instruction.

[0017] As an optional implementation manner, the fault mode judgment sub-logic includes:

[0018] The fault feature tensor is processed through a spatiotemporal convolutional neural network, which includes a feature extraction layer, an attention enhancement layer, and a classification decision layer;

[0019] The confidence probability of the fault mode is output through the spatiotemporal convolutional neural network, and the entropy and divergence of the confidence probability are calculated to determine the degree of certainty of the fault mode;

[0020] The judgment threshold with high confidence probability is dynamically adjusted according to historical fault data and the current network operation status, and the confidence probability of the fault mode is compared with the judgment threshold to determine the fault mode.

[0021] As an optional implementation manner, the channel fault flag generation sub-logic includes:

[0022] When the confidence probability is high, the correlation coefficient between the fault sub-features and the fault mode of each channel in the fault feature tensor is calculated to obtain the fault impact matrix;

[0023] The causal relationship between the fault sub-features of each channel in the fault feature tensor is analyzed through a causal inference algorithm. Based on the causal relationship, the direct fault source and the indirect impact channel are distinguished to generate the fault propagation path.

[0024] A channel fault label including channel index, fault mode, severity and propagation risk is generated according to the fault impact matrix and fault propagation path.

[0025] As an optional implementation, the optimization logic of the parameters of the spatiotemporal convolutional neural network includes:

[0026] Set a multi-level data threshold, compare the accumulated number of abnormal events with the multi-level data threshold, and obtain the level of the number of abnormal events;

[0027] The parameter optimization granularity is determined based on the level of abnormal events. The parameter optimization granularity includes local parameter optimization and global parameter optimization.

[0028] The parameters of the spatiotemporal convolutional neural network are optimized according to the parameter optimization granularity, and the parameter optimization effect of the spatiotemporal convolutional neural network is evaluated based on the validation set to determine whether the parameters of the spatiotemporal convolutional neural network should be optimized again.

[0029] As an optional implementation manner, the judgment sub-logic of the residual fault index of the non-target channel includes:

[0030] Re-acquire the fault characterization parameters of the non-target channel, process the delay jitter through empirical mode decomposition to obtain the delay jitter residual, and determine the degree of signal quality degradation based on historical fault data;

[0031] Assign dimensional weights to the residual of delay jitter and the degree of signal quality degradation based on the current network operating status;

[0032] The residual error of delay jitter and the degradation degree of signal quality are weighted and summed according to the dimension weight to comprehensively judge the residual fault index of the non-target channel.

[0033] As an optional implementation, historical delay jitter is learned through a long short-term memory network to obtain a delay jitter prediction curve under normal conditions. The actual delay jitter of the non-target channel is then re-acquired and the actual delay jitter is subtracted from the predicted curve point by point to obtain a residual sequence of delay jitter. The residual sequence of delay jitter is decomposed into multiple intrinsic mode functions through empirical mode decomposition to obtain the residual of delay jitter.

[0034] The judgment sub-logic of the signal quality degradation degree includes:

[0035] Obtain historical signal quality and calculate the mean and standard deviation of the historical signal quality to determine the fluctuation range of normal signal quality;

[0036] Re-obtain the actual signal quality of the non-target channel, calculate the average value and variance of the actual signal quality, and compare the average value of the actual signal quality with the fluctuation range of the normal signal quality;

[0037] When the average value of the actual signal quality is not within the fluctuation range of the normal signal quality, the boundary ratio is calculated. The boundary ratio is obtained by solving the difference between the average value of the actual signal quality and the boundary value of the fluctuation range of the normal signal quality, and dividing the difference by the boundary value of the fluctuation range of the normal signal quality. The degree of abnormal fluctuation is 0.

[0038] When the average value of the actual signal quality is within the normal signal quality fluctuation range, the degree of fluctuation anomaly is calculated based on the difference between the variance of the actual signal quality and the variance of the historical signal quality. The boundary ratio is set to 0, and the boundary ratio and the degree of fluctuation anomaly are weighted summed to determine the signal quality degradation degree.

[0039] As an optional implementation manner, the execution sub-logic of the channel reconfiguration includes:

[0040] Based on the residual fault index of non-target channels, the fault correlation degree between each non-target channel is calculated to form a fault correlation matrix. The fault correlation matrix is ​​processed by the spectral clustering algorithm to divide the non-target channels into multiple channel clusters.

[0041] In each channel cluster, the residual fault index is sorted to determine the reconfiguration priority. At the same time, reinforcement learning is used to generate the reconfiguration operation sequence for each channel cluster with channel performance as the reward target.

[0042] Channel reconfiguration of each channel cluster is performed in series-parallel mode in combination with the fault correlation degree and the reconfiguration operation sequence.

[0043] In a second aspect, the present application provides a multi-communication channel fault detection and self-healing control device, the device comprising: a device body and a control module;

[0044] The control module includes an acquisition unit, a processing unit and a control unit. The acquisition unit is used to acquire multi-channel fault characterization parameters at the communication node through sensors, fuse them into a fault feature tensor, and simultaneously acquire the operating status of the device body;

[0045] The processing unit is configured to process the fault feature tensor through a spatiotemporal convolutional neural network to determine a fault mode and output a confidence probability of the fault mode; when a high confidence probability is determined, generate a channel fault mark based on the fault feature tensor; determine a target channel based on the channel fault mark; inject a detection signal into the target channel; determine a trigger signal isolation instruction based on a bit error rate of the detection signal; and determine an impact of the channel fault on the device body based on the operating status of the device body;

[0046] The control unit is used to receive a signal isolation instruction and control the device body to isolate the target channel, while reacquiring the fault characterization parameters of the non-target channel, and comprehensively judging the residual fault index of the non-target channel based on the residual of the delay jitter and the degradation degree of the signal quality, marking abnormal events based on the abnormal judgment of the residual fault index and accumulating the number of abnormal events, executing channel reconfiguration and iteratively optimizing the parameters of the spatiotemporal convolutional neural network based on the number of abnormal events, and controlling the device body based on the execution result of the channel reconfiguration and the impact of the channel fault on the device body.

[0047] Compared with the existing technology, the beneficial effects of the present application are: by fusing multi-dimensional fault characterization parameters, a three-dimensional characterization of complex network faults is achieved, providing a comprehensive and structured data basis for subsequent analysis; and accurate identification and positioning of fault modes are achieved through spatiotemporal convolutional neural networks, while judging the cascading risk of fault modes to avoid secondary faults caused by isolation operations; finally, the status of non-target channels is evaluated by the residual fault index, and the cumulative number of abnormal events is combined to trigger channel reconfiguration and parameter optimization of the spatiotemporal convolutional neural network; thereby enabling the communication network to actively identify and handle faults in a complex and changing environment, effectively reducing the scope of fault impact, continuously improving the stability and reliability of the communication network, and realizing the transformation from passive response to faults to active network management.

[0048] In communication networks, fault characterization parameters such as delay jitter, signal-to-noise ratio, and signal quality each reflect the channel operating status from different dimensions. Normalizing these fault characterization parameters can eliminate dimensional differences and ensure the integrity and accuracy of fault characteristic information. Constructing the normalized parameters into a time-space alignment matrix can unify the fault characteristics of different channels at different time points, making it easier to explore the changing patterns of channel faults in time series and the spatial interactions between different channels.

[0049] The three-dimensional convolution kernel of the spatiotemporal convolutional neural network can perform convolution operations on the fault feature tensor in both time and space dimensions, which enables the spatiotemporal convolutional neural network to capture the continuous changes of the fault features of a single channel over time. At the same time, it can also identify the spatial correlation patterns of fault features between multiple channels, thereby effectively distinguishing local faults from global faults and improving the accuracy of fault mode judgment; the confidence probability of the fault mode output by the spatiotemporal convolutional neural network can timely discover the uncertainty of the judgment, avoid misjudgment due to close probabilities, and provide a reliable metric for the fault mode judgment results; based on the fault feature tensor, a channel fault mark is generated, the target channel is determined according to the channel fault mark, and a detection signal is injected into the target channel. The bit error rate of the detection signal is used to determine the trigger signal isolation instruction, thereby enhancing the detection targeting of different types of faults and improving the accuracy of fault judgment based on the bit error rate.

[0050] After isolating the target channel, the network status changes, and new faults or residual faults may appear in the non-target channel. The fault characterization parameters of the non-target channel are re-acquired, and the residual fault index of the non-target channel is comprehensively determined based on the residual error of the delay jitter and the degradation of the signal quality. This can accurately identify the potential fault state of the non-target channel after the target channel is isolated, providing a reliable basis for subsequent decision-making. Abnormal events are marked and the number of abnormal events is accumulated based on the abnormal event number. Channel reconfiguration is performed based on the number of abnormal events and the parameters of the spatiotemporal convolutional neural network are iteratively optimized to continuously improve the fault identification ability and ensure the effectiveness and adaptability of the fault detection method. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be derived from these drawings without inventive work. Among them:

[0052] Figure 1 A flowchart of a method for fault detection of multiple communication channels provided in an embodiment of the present application;

[0053] Figure 2 A trigger logic diagram of a signal isolation instruction for a multi-communication channel fault detection method provided in an embodiment of the present application;

[0054] Figure 3 A sub-logic diagram for judging the degradation degree of signal quality of the fault detection method for multiple communication channels provided in an embodiment of the present application;

[0055] Figure 4This is a device structure diagram of the multi-communication channel fault detection and self-healing control device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0056] In order to make the objectives, technical solutions and advantages of the embodiments of the present application more obvious and easy to understand, the technical solutions in the embodiments of the present application are clearly and completely described below in conjunction with the drawings in the specification. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments.

[0057] Example 1

[0058] like Figure 1 As shown, a method flow chart of a multi-communication channel fault detection method is provided for an embodiment of the present application, the method comprising:

[0059] S1. Obtain multi-channel fault characterization parameters at the communication node. The fault characterization parameters include the delay jitter, signal-to-noise ratio, and signal quality of each channel. Align the fault characterization parameters with the time window according to the channel index and fuse them into a fault feature tensor.

[0060] Specifically, the fusion logic of the fault feature tensor includes:

[0061] Acquire multi-channel fault characterization parameters at the communication node. These parameters include each channel's delay jitter, signal-to-noise ratio, and signal quality, and perform normalization on the fault characterization parameters.

[0062] Construct a spatiotemporal alignment matrix with channel index as columns and acquisition time within the time window as rows, and fill the normalized fault characterization parameters into the spatiotemporal alignment matrix;

[0063] According to the historical contribution of each fault characterization parameter to fault judgment, a weight coefficient is assigned to each fault characterization parameter through the attention mechanism to form a weighted matrix. The weighted matrix and the spatiotemporal alignment matrix are subjected to a tensor product operation to obtain the fault feature tensor.

[0064] In communication networks, the original numerical ranges of fault-characterizing parameters such as delay jitter, signal-to-noise ratio (SNR), and signal quality vary greatly. Delay jitter is measured in milliseconds, while SNR values ​​are typically in the tens. If these parameters are directly integrated, the larger parameter will dominate the subsequent analysis, obscuring information about other parameters and failing to accurately reflect channel fault characteristics. In actual communication node deployments, sensors continuously acquire delay jitter, SNR, and signal quality for each channel. For delay jitter, when the network is in a stable state, historical delay jitter is used to calculate its mean and standard deviation. As the network operates, when new services are added or traffic surges occur, the fluctuation range of the current delay jitter is monitored and the normalization threshold is dynamically adjusted based on the fluctuation. For example, if network traffic suddenly doubles at a certain time, resulting in increased delay jitter, the normalization threshold is appropriately relaxed to avoid misidentifying jitter caused by normal traffic growth as a fault characteristic.

[0065] The signal-to-noise ratio and signal quality are normalized using the Z-score normalization method. This involves obtaining historical signal quality over a period of time, calculating its mean and standard deviation, and then re-obtaining the actual signal quality. The normalized signal quality = (actual signal quality - mean) / standard deviation. This normalized signal quality conforms to a standard normal distribution with a mean of 0 and a standard deviation of 1. During the normalization process, if data is missing, it is supplemented by interpolation based on neighboring node data to ensure the data integrity of the fault characterization parameters. By dynamically adjusting the normalization threshold, the normalization process can closely track changes in the network operating status. Even when the network environment changes frequently, data can still be accurately processed to reduce misjudgments. At the same time, the normalization process puts parameters of different magnitudes on the same scale, laying the foundation for accurate extraction of fault features.

[0066] Data acquisition from multiple communication channels has spatiotemporal characteristics. Different channels have different states at the same time, and the state of the same channel may also change at different times. To comprehensively analyze channel faults, it is necessary to uniformly arrange the fault characterization parameters of each channel according to the time window and channel index, and construct a matrix structure that can intuitively display the spatiotemporal relationship to facilitate subsequent exploration of the spatiotemporal correlation between channels. When constructing the spatiotemporal alignment matrix, the number of columns of the spatiotemporal alignment matrix is ​​determined by the number of channels in the communication network, and the number of rows of the spatiotemporal alignment matrix is ​​determined by the acquisition time within the set time window. For example, if there are 20 channels in the communication network, the time window is set to 1 minute, and the data acquisition interval is 1 second, then the spatiotemporal alignment matrix will have 20 columns and 60 rows.

[0067] The normalized fault characterization parameters are sequentially filled into the corresponding positions of the time-space alignment matrix. During the data filling process, if the data acquisition of a certain channel at a certain time fails, the bidirectional LSTM network is used for prediction and filling. The bidirectional LSTM network uses the information of the fault characterization parameters before and after the channel to learn the time series change law of the fault characterization parameters, thereby predicting the fault characterization parameters of the missing time. At the same time, the time series data is Fourier transformed to convert the time domain signal into the frequency domain signal, and the frequency and amplitude and other features in the frequency domain are extracted. The Fourier transform can be used to discover certain periodic interference signal frequencies, which are very important for judging the fault. The extracted frequency domain features are combined with the original time domain data to form a richer feature set, which is then filled into the space-time alignment matrix, so that the space-time alignment matrix contains more comprehensive channel status information; the space-time alignment matrix clearly presents the distribution of each channel fault feature in the time and space dimensions. By observing the data change trend in the matrix, it is possible to quickly discover whether there are abnormal fluctuations in certain channels and whether the abnormal fluctuations are propagating between channels. The frequency domain features can enrich the feature dimensions of the space-time alignment matrix, so that the space-time alignment matrix can not only reflect the time changes of the data, but also reflect the frequency characteristics of the signal, thereby improving the ability to identify complex fault modes.

[0068] The importance of fault characterization parameters such as delay jitter, signal-to-noise ratio, and signal quality to fault diagnosis varies across different network operation scenarios and fault modes. For faults caused by network congestion, delay jitter is more critical, while for faults caused by signal interference, signal-to-noise ratio and signal quality play a greater role. Therefore, a dynamic weighting method is required to adjust the weight coefficients of each parameter based on actual conditions. Historical fault data and corresponding fault modes are used as training samples. During the training process, the historical contribution of each fault characterization parameter to fault diagnosis under different fault modes is learned. Delay jitter plays a key role in fault diagnosis in fault cases.

[0069] During network operation, the status changes of each channel are monitored in real time. Once the initial signs of a certain fault mode are detected, that is, the signal-to-noise ratio of a channel suddenly drops, the distribution of weight coefficients is immediately adjusted. At this time, the weight coefficient of the signal-to-noise ratio of the channel is increased, and the weight coefficients of other parameters are appropriately reduced, so that subsequent analysis can focus more on changes in the signal-to-noise ratio. In addition, the weight is adjusted according to the service type of the network. When the network carries services with high real-time requirements, the weight coefficient of delay jitter will increase accordingly, thus forming a weighted matrix; the attention mechanism realizes dynamic adjustment of the weight coefficient, which can closely fit the actual operation status and fault characteristics of the network. When facing a complex and changing network environment, it can always highlight the most valuable parameters for fault judgment, so that the fused fault characteristics can more accurately reflect the nature of the fault, effectively improving the accuracy and timeliness of fault detection.

[0070] Performing a tensor product operation on the weighted matrix and the spatiotemporal alignment matrix can fuse the weight information of each fault characterization parameter with the fault feature information in the spatiotemporal alignment matrix, thereby obtaining a fault feature tensor. This fault feature tensor contains the distribution information of the fault in the time and space dimensions, as well as the importance information of each fault characterization parameter, providing a more comprehensive and accurate basis for subsequent fault analysis and processing. The tensor product operation process is to multiply each element in the weighted matrix with the corresponding element in the spatiotemporal alignment matrix to obtain a new tensor, namely the fault feature tensor. In this way, the weight coefficient of each fault characterization parameter is integrated into the spatiotemporal alignment matrix, so that the fault feature tensor can more comprehensively reflect the characteristics of the fault.

[0071] The fused fault feature tensor combines the weight coefficients of the comprehensive fault characterization parameters and the fault feature information in the spatiotemporal alignment matrix, providing richer and more accurate information for subsequent fault analysis and processing. By analyzing the fault feature tensor, it is possible to more accurately judge the fault mode, generate channel fault marks, and trigger signal isolation instructions, thereby improving the accuracy of the entire fault detection. The fault feature tensor is the input data for subsequent steps such as fault mode judgment and channel fault mark generation. Its quality directly affects the accuracy and reliability of these steps. Therefore, the fusion of high-quality fault feature tensors is crucial to the effectiveness of the entire multi-communication channel fault detection method.

[0072] S2. Process the fault feature tensor through a spatiotemporal convolutional neural network to determine the fault mode and output the confidence probability of the fault mode. The fault mode includes local faults and global faults. When it is judged to have a high confidence probability, a channel fault mark is generated based on the fault feature tensor. The target channel is determined based on the channel fault mark, and a detection signal is injected into the target channel. The bit error rate of the detection signal is used to determine the trigger signal isolation instruction.

[0073] Furthermore, the fault mode judgment sub-logic includes:

[0074] The fault feature tensor is processed through a spatiotemporal convolutional neural network, which includes a feature extraction layer, an attention enhancement layer, and a classification decision layer;

[0075] The confidence probability of the fault mode is output through the spatiotemporal convolutional neural network, and the entropy and divergence of the confidence probability are calculated to determine the degree of certainty of the fault mode;

[0076] The judgment threshold with high confidence probability is dynamically adjusted according to historical fault data and the current network operation status, and the confidence probability of the fault mode is compared with the judgment threshold to determine the fault mode.

[0077] The fault feature tensor contains complex fault information of multiple channels in time and space dimensions. Traditional methods are difficult to effectively extract and analyze these high-dimensional and dynamic features. The spatiotemporal convolutional neural network can automatically extract the spatiotemporal features in the data, which is suitable for mining the fault mode laws in the fault feature tensor; construct a spatiotemporal convolutional neural network including feature extraction layer, attention enhancement layer and classification decision layer. The feature extraction layer uses the three-dimensional convolution kernel to perform convolution operations on the fault feature tensor along the channel dimension, time dimension and feature dimension. When multiple channel data of a communication node are detected to be abnormal, the three-dimensional convolution kernel will perform fault analysis at different time points and different channels. The attention enhancement layer calculates the importance weight of each feature dimension based on the features output by the feature extraction layer. When the spatiotemporal convolutional neural network detects a fault related to the signal-to-noise ratio, it increases the weight of the signal-to-noise ratio feature dimension, making the spatiotemporal convolutional neural network more focused on the feature of this dimension. Through the three-dimensional convolution kernel and the attention enhancement layer, the spatiotemporal convolutional neural network can efficiently extract the spatiotemporal correlation features hidden in the fault feature tensor, highlight the key fault features, reduce the interference of irrelevant information, and improve the accuracy and efficiency of fault mode judgment.

[0078] The confidence probability of the fault mode output by the spatiotemporal convolutional neural network only reflects the possibility of the spatiotemporal convolutional neural network's judgment on a certain fault mode, but cannot reflect the certainty of the judgment result. Calculating the entropy value and divergence of the confidence probability can measure the degree of chaos and dispersion of the confidence probability distribution, thereby judging the certainty of the fault mode judgment; the classification decision layer outputs the confidence probability of the fault mode being a local fault and a global fault; then the entropy value and divergence of the confidence probability are calculated. The higher the entropy value, the more uniform the probability distribution and the greater the uncertainty of the judgment, while the divergence reflects the difference between the current probability distribution and the ideal distribution. When it is detected that the confidence probabilities of two fault modes in a certain judgment are close, the entropy value will be higher, indicating that the certainty of this judgment is low; through the calculation of entropy value and divergence, the reliability of the fault mode judgment result can be intuitively evaluated. When the confidence probabilities are close or the distribution is chaotic, the uncertainty of the judgment can be discovered in time to avoid wrong decisions caused by unreliable judgments.

[0079] The network's operating state is constantly changing, and the distribution and manifestation of fault characteristics vary across time periods. A fixed high-confidence probability judgment threshold cannot adapt to network state changes, easily leading to misjudgments or missed judgments. Historical fault data is obtained, which records relevant information for each fault detection, including the fault mode, confidence probability, and network operating state. The network operating state includes current traffic and service type. When a new fault feature tensor is input into the spatiotemporal convolutional neural network, fault data from similar scenarios is filtered from the historical fault data based on the current network traffic volume and the real-time requirements of the carried service. When the network is transmitting high-traffic video services, historical fault data from similar high-traffic and video-carrying scenarios is filtered. The distribution of confidence probabilities for accurately determining fault modes in this data is analyzed, and the high-confidence probability judgment threshold is dynamically adjusted. The confidence probability of the fault mode output by the spatiotemporal convolutional neural network is compared with the adjusted judgment threshold to determine the final fault mode.

[0080] Dynamically adjusting the threshold enables fault mode judgment to closely follow the actual operation of the communication network, maintain high accuracy under different network conditions, reduce misjudgments and missed judgments caused by unreasonable judgment thresholds, and accurately judge the fault mode. It provides the premise for generating channel fault markers based on the fault feature tensor, making subsequent operations such as determining the target channel and injecting detection signals more targeted and effective.

[0081] Furthermore, the generation sub-logic of the channel fault flag includes:

[0082] When the confidence probability is high, the correlation coefficient between the fault sub-features and the fault mode of each channel in the fault feature tensor is calculated to obtain the fault impact matrix;

[0083] The causal relationship between the fault sub-features of each channel in the fault feature tensor is analyzed through a causal inference algorithm. Based on the causal relationship, the direct fault source and the indirect impact channel are distinguished to generate the fault propagation path.

[0084] A channel fault tag including channel index, fault mode, severity and propagation risk is generated according to the fault impact matrix and fault propagation path.

[0085] Each channel has different levels of involvement and impact when a fault occurs. Therefore, the degree of correlation between each channel and the fault mode needs to be quantified to accurately identify direct fault sources and indirect impact channels. When a fault mode is determined to be of high confidence probability, a correlation analysis is performed between the fault sub-features of each channel in the fault feature tensor and the determined fault mode to calculate the correlation coefficient. For a certain local fault mode, the similarity between the delay jitter, signal-to-noise ratio, and signal quality characteristics of each channel and the characteristics of the local fault mode is calculated to obtain the correlation coefficient. The correlation coefficients of all channels are arranged into a matrix to form a fault impact matrix. If a channel has a high correlation coefficient in the fault impact matrix, it means that the channel is closely associated with the current fault mode. The fault impact matrix clearly shows the degree of influence of each channel on the fault mode, providing a quantitative basis for subsequently distinguishing direct fault sources and indirect impact channels, avoiding blind judgment of fault channels and improving the accuracy of fault location.

[0086] The occurrence and propagation of faults in communication networks often have a causal relationship. Knowing only the degree of correlation between each channel and the fault mode is not enough to fully understand the fault situation. It is necessary to clarify the fault propagation path so that effective fault isolation and repair measures can be taken. The causal relationship between the fault sub-features of each channel in the fault feature tensor is analyzed through a causal inference algorithm. The causal inference algorithm determines whether there is a causal relationship between the fault sub-features of different channels by analyzing the changing trends and interdependencies of each parameter in the fault feature tensor. If it is found that the signal-to-noise ratio of a core channel A decreases, the signal quality of channels B and C directly connected to it also decreases, and there is a chronological order. At the same time, after eliminating other interference factors, the causal inference algorithm can infer that there is a causal relationship between the decrease in the signal-to-noise ratio of channel A and the decrease in the signal quality of channels B and C. During the analysis process, the causal relationship inferred by the causal inference algorithm is verified and corrected in combination with the topological structure of the communication network to ensure that the results conform to the actual network connection and data transmission logic.

[0087] Based on the causal relationship derived from the causal inference algorithm, each channel is classified. Channels whose fault characteristic changes cause changes in the fault characteristics of other channels and whose own fault causes cannot be directly explained by other channels are determined to be direct fault sources. In the above example, channel A is determined to be the direct fault source, and other channels whose fault characteristic changes are caused by the direct fault source are classified as indirectly affected channels, including channels B and C. During the differentiation process, the topology of the communication network and data transmission logic are again referenced to ensure the accuracy of the classification results. At the same time, the classification information and causal relationship of each channel are recorded to form a detailed fault analysis record; this clearly distinguishes between direct fault sources and indirectly affected channels, allowing operation and maintenance personnel to quickly identify the core of the fault, concentrate resources on handling key issues, avoid wasting time and energy on indirectly affected channels, effectively improve fault repair efficiency, and reduce losses caused by network failures.

[0088] Starting from the direct fault source, the fault propagation path is gradually derived based on the established causal relationships and channel classification information. Starting from the direct fault source, the affected indirect channels are identified as the next node in the propagation path. The indirect channels are then analyzed to see if they have triggered other channel faults, and so on, to construct a complete fault propagation path. Preferably, starting from channel A, channels B and C are connected in sequence. If a fault in channel B causes a problem in channel D, channel D is also included in the fault propagation path. The fault propagation path is displayed as a directed graph, where nodes represent channels and directed edges represent fault propagation directions and causal relationships. This allows operators to intuitively understand the fault propagation process. The directed graph also annotates the fault characteristic changes and time nodes of each channel, providing more detailed information for fault analysis and handling. This intuitive fault propagation path diagram enables operators to quickly grasp the overall picture and spread trend of the fault, facilitating the development of targeted fault isolation measures, such as severing connections on critical propagation paths to prevent further spread of the fault. It also helps assess the scope of the fault's impact on the communication network, providing a reference for subsequent network repair and optimization.

[0089] To accurately identify faulty channels and record fault-related information during troubleshooting for subsequent operations and analysis, a channel fault tag containing detailed information is generated. The fault impact matrix and fault propagation path are combined to generate a channel fault tag, which integrates and encodes the channel index, fault mode, severity, and propagation risk of each channel. The channel index uniquely identifies each channel in the communication network. Specifying the channel index when generating the channel fault tag allows for precise location of the faulty channel, facilitating subsequent troubleshooting and analysis of that specific channel. Severity is assessed by combining the correlation coefficient in the fault impact matrix and the range of the fault propagation path. Channels with high correlation coefficients and located at critical locations along the fault propagation path are considered to have high fault severity. Conversely, channels with low correlation coefficients and limited impact within a small, localized area, without significantly impacting primary services, are considered to have a low severity rating. Furthermore, the severity assessment is refined and refined based on indicators such as the duration of network service interruption and the number of affected users, categorizing severity into low, medium, and high levels. The specific standards and characteristics corresponding to each level are clearly defined to facilitate accurate assessment.

[0090] At the same time, based on the structure and trend of the fault propagation path, the possibility of the fault spreading to other nodes is analyzed. If the fault propagation path tends to extend to the core area of ​​the network or important business channels, and if these areas or channels are affected, large-scale business interruptions will occur, then the propagation risk is high. At the same time, if the network topology lacks redundancy and the current load is high, the fault can easily trigger a chain reaction due to local channel failure, resulting in increased propagation risk. The propagation risk is also divided into three levels: low, medium, and high, and the judgment criteria for each level are clearly defined. Low risk means that the fault will basically not spread, medium risk means that there is a certain possibility of spread but it is controllable, and high risk means that the fault will spread rapidly and cause serious consequences. Based on the above content, a unified channel fault tag can be generated, which can comprehensively and accurately convey fault information, improve the efficiency and accuracy of information transmission, and avoid information omissions or misunderstandings. Operations and maintenance personnel can quickly understand the full picture of the fault through the tag, make reasonable decisions, and improve the coordination and efficiency of fault handling.

[0091] Specifically, if Figure 2 As shown, the trigger logic of the signal isolation instruction includes:

[0092] Determine the target channel based on the channel fault flag and dynamically adjust the detection signal parameters according to the fault mode and severity. The detection signal parameters include signal strength, frequency, and modulation mode.

[0093] The bit error rate of the detection signal is analyzed by the sequential probability ratio test method, and a bit error rate threshold is set. The bit error judgment result is obtained by comparing the bit error rate of the detection signal with the bit error rate threshold;

[0094] Determine the cascading risk of signal isolation based on the propagation risk, configure the risk threshold, compare the cascading risk of signal isolation with the risk threshold, and obtain the isolation risk result. Make a comprehensive judgment based on the error judgment result and the isolation risk result to trigger the signal isolation instruction.

[0095] Different fault modes and severities have different impacts on the channel. In order to detect fault conditions more accurately, it is necessary to adjust the detection signal parameters according to the characteristics of the target channel to improve the accuracy and effectiveness of detection; based on the generated channel fault mark, determine the target channel that needs further detection. When the channel fault mark shows that a certain channel is a serious local fault, it is determined as the target channel. According to the fault mode and severity, adjust the signal strength, frequency and modulation method of the detection signal. For faults caused by severe signal interference, appropriately increase the signal strength of the detection signal, select a detection frequency different from the interference frequency, and adopt a modulation method with strong anti-interference ability. For faults caused by insufficient bandwidth, adjust the frequency and modulation method of the detection signal to adapt to the channel bandwidth; adjusting the detection signal parameters according to the characteristics of the target channel can make the detection signal better adapt to the channel fault condition, improve the fault detection capability, and obtain more accurate fault information.

[0096] The bit error rate of the detection signal is an important indicator for measuring the transmission quality of the communication channel. By analyzing the bit error rate of the detection signal, it is possible to determine whether there is a fault in the target channel and the severity of the fault, providing a key basis for whether to isolate the channel; after the adjusted detection signal is injected into the target channel, the bit error rate of the detection signal is analyzed by the sequential probability ratio test method to compare with the pre-set bit error rate threshold. When the bit error rate is detected to be continuously greater than the bit error rate threshold and reaches a certain detection time requirement, it is determined that there is a fault in the target channel, and the error judgment result is a severe error; if the bit error rate is not greater than the bit error rate threshold or does not reach the detection time, it is judged to be a minor error; the sequential probability ratio test method can quickly analyze and judge the bit error rate while ensuring the accuracy of detection, avoid waiting for a long time for detection results, and improve the timeliness of fault handling.

[0097] In a communication network, isolating a channel can trigger a chain reaction in other channels, leading to cascading failures. Therefore, it is necessary to comprehensively consider the error judgment results and the cascading risk of signal isolation to avoid larger-scale network failures caused by blindly isolating channels. Based on the previously generated fault propagation path and propagation risk assessment results, analyze the chain reaction that will be triggered if signal isolation is implemented on a specific channel. Preferably, in a certain fault propagation path, channel A is the direct fault source, and channels B and C are affected by it. If channel A is isolated, it is necessary to evaluate whether it will cause new faults in channels B and C due to the loss of upstream signals, and whether this fault will further propagate to other channels. Based on the topology of the communication network and service dependencies, determine the number of channels that will be affected by isolating a channel, the scope of service interruption, and the possibility of fault propagation. If a channel is a critical transmission node for multiple services, isolating the channel will cause a large number of services to be interrupted, and the cascading risk is high. Assessing cascading risks in advance can help operation and maintenance personnel fully understand the potential impact of isolation operations, avoid blind isolation that leads to the expansion of faults, and ensure network stability and reliability.

[0098] Network administrators set different levels of risk thresholds based on factors such as the importance of the communication network, business type, and service level agreement. For critical business networks, the risk threshold can be set lower to ensure that any isolation operations that may cause serious consequences are carefully evaluated. For non-critical business networks, the risk threshold can be appropriately increased to reduce the frequent triggering of isolation instructions due to minor risks. Risk thresholds are divided into multiple levels, including low risk thresholds, medium risk thresholds, and high risk thresholds. Reasonable risk threshold settings can balance network security and business continuity, make isolation decisions more scientific and reasonable, avoid excessive or insufficient isolation, and improve the efficiency and quality of network management.

[0099] By comparing the cascading risk of signal isolation with the pre-configured risk threshold, the risk level of the current isolation operation can be clarified, providing a quantitative basis for subsequent decision-making; the assessed cascading risk of signal isolation is compared with the configured risk threshold one by one. If all risk indicators are less than the low-risk threshold, the isolation risk result is judged to be low risk; if some indicators are greater than the low-risk threshold but do not reach the medium-risk threshold, it is judged to be medium risk; if some indicators are greater than the medium-risk threshold or even reach the high-risk threshold, it is judged to be high risk; clear isolation risk results provide a clear and objective basis for subsequent decision-making, enabling operation and maintenance personnel to take appropriate measures according to the risk level and improve the accuracy and effectiveness of decision-making.

[0100] The error judgment result reflects the communication quality of the current channel, while the isolation risk result reflects the negative impact of the isolation operation. Taking these two factors into consideration, it is possible to minimize the side effects of the isolation operation while ensuring the communication quality, and to optimize the network fault handling. According to the different combinations of the error judgment result and the isolation risk result, it is decided whether to trigger the signal isolation instruction. If the error judgment result shows that there are serious errors in the channel and the isolation risk result is low risk, the signal isolation instruction is triggered immediately. If the error judgment result shows that there are moderate errors and the isolation risk result is medium risk, further evaluation is carried out, including separation. Analyze the specific impact of the bit error on the service and whether there are backup channels available for switching, and then decide whether to trigger the command; if the bit error judgment result shows that there is only a slight bit error, but the isolation risk result is high risk, the signal isolation command will not be triggered temporarily, but other measures will be taken, including strengthening monitoring and adjusting channel parameters. In this way, the communication quality and isolation risk can be fully considered, and decision-making errors caused by a single factor can be avoided. The triggering of the signal isolation command is more scientific and reasonable, which can not only effectively handle faults but also reduce unnecessary impacts on the network. The triggered signal isolation command will guide subsequent network operations and directly affect the effectiveness of network fault handling and recovery speed.

[0101] S3. Receive the signal isolation instruction and isolate the target channel, while reacquiring the fault characterization parameters of the non-target channel, and comprehensively judge the residual fault index of the non-target channel based on the residual of the delay jitter and the degradation of the signal quality. Based on the abnormal judgment of the residual fault index, mark the abnormal events and accumulate the number of abnormal events. According to the number of abnormal events, perform channel reconfiguration and iteratively optimize the parameters of the spatiotemporal convolutional neural network.

[0102] Furthermore, the judgment sub-logic of the residual fault index of the non-target channel includes:

[0103] Re-acquire the fault characterization parameters of the non-target channel, process the delay jitter through empirical mode decomposition to obtain the delay jitter residual, and determine the degree of signal quality degradation based on historical fault data;

[0104] Assign dimensional weights to the residual of delay jitter and the degree of signal quality degradation based on the current network operating status;

[0105] The residual error of delay jitter and the degradation degree of signal quality are weighted and summed according to the dimension weight to comprehensively judge the residual fault index of the non-target channel.

[0106] After the target channel is isolated, the traffic distribution and signal transmission path of the communication network change, and the operating environment of the non-target channel changes accordingly. Previously normal non-target channels may experience potential faults due to carrying additional traffic or changes in signal interference, or residual effects related to the target channel fault may remain. It is necessary to re-acquire fault characterization parameters and perform targeted processing on delay jitter and signal quality to accurately assess the actual fault status of the non-target channel. The delay jitter, signal-to-noise ratio, and signal quality of the non-target channel are re-acquired, and the delay jitter is processed through empirical mode decomposition to obtain the delay jitter residual. At the same time, the degree of signal quality degradation is determined based on historical fault data. This allows the precise capture of subtle anomalies that occur in non-target channels after network status changes. Even potential faults that are not obvious at the beginning can be discovered through data analysis, effectively avoiding missed faults caused by simple and rough data processing, and improving the accuracy and timeliness of fault detection.

[0107] Under different operating conditions of communication networks, the importance of residual delay jitter and signal quality degradation in determining non-target channel faults varies. When the network carries services with extremely high real-time requirements, even slight delay jitter can lead to a significant degradation in user experience. In this case, the residual delay jitter dimension should have a higher weight in fault diagnosis. However, in service scenarios where data transmission is the primary focus and real-time requirements are lower, signal quality degradation may have a greater impact on data integrity and accuracy, and its dimension weight should be increased accordingly. Therefore, it is necessary to dynamically assign dimension weights based on the real-time operating status of the network to make the calculation of the residual fault index more accurate. By dynamically assigning dimension weights, the diversity and complexity of network operating conditions are fully considered, allowing the calculation of the residual fault index to closely track changes in the actual network conditions. The resulting residual fault index can more accurately reflect the actual fault risk of non-target channels under specific network conditions, improving the reliability and practicality of fault assessment results and providing a more valuable reference for subsequent decision-making.

[0108] Delay jitter residual and signal quality degradation reflect the fault status of non-target channels from different perspectives. A single metric cannot fully assess the channel fault risk. By weighting and summing the two, information from these two dimensions can be integrated to generate a quantitative index. This intuitively determines the residual fault status of non-target channels, providing a clear and specific basis for subsequent fault handling and network optimization decisions. Based on the assigned dimensional weights for delay jitter residual and signal quality degradation, the corresponding values ​​are weighted and calculated to obtain the residual fault index. This transforms complex multi-dimensional fault information into a single, quantifiable residual fault index, greatly facilitating operations and maintenance personnel to quickly understand the fault status of non-target channels, making fault assessment results more intuitive and clear, and improving the efficiency and accuracy of fault handling. The residual fault index directly determines whether to mark an abnormal event and configure the abnormal threshold. When the residual fault index exceeds the abnormal threshold, it is marked as an abnormal event and the number of abnormal events is accumulated. This data further influences subsequent channel reconfiguration decisions and triggers the optimization of spatiotemporal convolutional neural network parameters, and serves as a critical basis for subsequent operations in the entire fault handling process.

[0109] Furthermore, the historical delay jitter is learned through a long short-term memory network to obtain a prediction curve of the delay jitter under normal conditions. The actual delay jitter of the non-target channel is re-acquired, and the actual delay jitter is subtracted from the prediction curve point by point to obtain a residual sequence of the delay jitter. The residual sequence of the delay jitter is decomposed into multiple intrinsic mode functions through empirical mode decomposition to obtain the residual of the delay jitter.

[0110] During normal operation, the delay jitter of a communication channel typically follows certain patterns. These patterns are closely related to factors such as the network topology, service type, and traffic distribution. By learning from historical delay jitter, a prediction curve for delay jitter under normal conditions is derived. This provides a reliable benchmark for determining whether the current delay jitter is abnormal. Only by establishing an accurate prediction curve can deviations between actual delay jitter and normal conditions be detected, thereby detecting potential faults. Historical delay jitter of non-target channels is obtained and organized into a time series format. This data is then fed into a long-short-term memory (LSTM) network for training. The LSTM network, with its unique memory unit structure, effectively processes long-term dependencies in time series data and identifies patterns in delay jitter across different time periods and service scenarios. After multiple iterations of training, the LSTM network learns the inherent patterns of historical delay jitter and generates a delay jitter prediction curve for the channel. The LSTM network's powerful time series processing capabilities enable it to deeply analyze historical delay jitter and accurately capture complex patterns in the data. The generated prediction curve can truly reflect the changing trend of the channel's delay jitter under normal operating conditions, providing an accurate reference standard for subsequent comparison with actual data, greatly improving the accuracy and reliability of judging delay jitter anomalies.

[0111] By comparing the actual delay jitter obtained in real time with the predicted curve, the deviation between the current delay jitter and the normal state can be intuitively found. These deviations are presented in the form of residuals. By obtaining the residual sequence, potential fault information can be separated from normal delay jitter fluctuations, providing a data basis for further analysis of fault characteristics; the actual delay jitter of non-target channels is continuously obtained in real time, and the specific values ​​of each time point are recorded in chronological order. The actual delay jitter of each time point is subtracted point by point from the data of the corresponding time point of the predicted curve. If the actual delay jitter at a certain time point is large, the actual delay jitter is calculated. If the actual delay jitter is greater than the value at the corresponding point on the predicted curve, the residual is positive, indicating that the delay jitter at that point in time is greater than normal. Conversely, if the actual delay jitter is less than the value at the corresponding point on the predicted curve, the residual is negative, indicating that the delay jitter at that point in time is less than normal. In this way, a complete residual sequence of delay jitter is obtained, which records the difference between the actual delay jitter and the normal state at each time point. This residual sequence can clearly and intuitively present abnormal changes in delay jitter, providing rich and accurate data for subsequent in-depth analysis of fault characteristics, and improving the sensitivity and timeliness of fault detection.

[0112] The residual sequence of delay jitter contains fluctuations in multiple frequency components. These fluctuations may be caused by normal network fluctuations or abnormal changes caused by faults. The empirical mode decomposition algorithm can decompose the complex residual sequence into multiple intrinsic mode functions with different frequency characteristics. By analyzing these intrinsic mode functions, the components that truly reflect the fault characteristics can be screened out, thereby accurately determining whether there is an abnormality in the delay jitter and the specific circumstances of the abnormality, providing strong support for fault diagnosis. The residual sequence of delay jitter is processed by the empirical mode decomposition algorithm. The algorithm first processes the residual sequence, finds the maximum and minimum points in the sequence, and fits the upper and lower envelopes respectively using the cubic spline interpolation method. The mean of the upper and lower envelopes is calculated, and the mean is subtracted from the residual sequence to obtain the first preliminary intrinsic mode function component. The first component is then determined to meet the intrinsic mode function conditions, including a roughly equal number of extreme points and zero-crossing points and a local mean of zero. If not, the above process is repeated until a satisfactory intrinsic mode function is obtained. The same operation is then performed on the remaining residual sequence, sequentially decomposing multiple intrinsic mode functions. The frequency and amplitude characteristics of each intrinsic mode function are analyzed. Combined with historical fault data and network operation status, components with significant abnormal fluctuations and matching known fault characteristics are screened out. These components are combined to form the final extracted delay jitter residual for subsequent fault assessment.

[0113] The empirical mode decomposition algorithm can effectively separate key fault characteristic components from complex delay jitter residual sequences, removing interference from noise and irrelevant information such as normal network fluctuations. This greatly improves the accuracy and sensitivity of delay jitter fault diagnosis, enabling more precise identification of various types of delay jitter faults and providing a reliable basis for fault diagnosis for network operations and maintenance personnel. The extracted accurate delay jitter residuals are used to calculate the residual fault index of non-target channels. The accuracy of the residual fault index directly determines the reliability of the index in assessing non-target channel faults. Only accurate delay jitter residuals can ensure that the residual fault index truly reflects the channel fault condition, thereby influencing subsequent decisions such as whether to mark abnormal events, perform channel reconfiguration, and optimize spatiotemporal convolutional neural network parameters.

[0114] Furthermore, if Figure 3 As shown, the judgment sub-logic of the degradation degree of signal quality includes:

[0115] Obtain historical signal quality and calculate the mean and standard deviation of the historical signal quality to determine the fluctuation range of normal signal quality;

[0116] Re-obtain the actual signal quality of the non-target channel, calculate the average value and variance of the actual signal quality, and compare the average value of the actual signal quality with the fluctuation range of the normal signal quality;

[0117] When the average value of the actual signal quality is not within the fluctuation range of the normal signal quality, the boundary ratio is calculated. The boundary ratio is obtained by solving the difference between the average value of the actual signal quality and the boundary value of the fluctuation range of the normal signal quality, and dividing the difference by the boundary value of the fluctuation range of the normal signal quality. The degree of abnormal fluctuation is 0.

[0118] When the average value of the actual signal quality is within the normal signal quality fluctuation range, the degree of fluctuation anomaly is calculated based on the difference between the variance of the actual signal quality and the variance of the historical signal quality. The boundary ratio is set to 0, and the boundary ratio and the degree of fluctuation anomaly are weighted summed to determine the signal quality degradation degree.

[0119] During normal communication, signal quality fluctuates within a certain range. This range is affected by various factors, including network equipment performance, environmental interference, and signal transmission distance. Understanding the normal fluctuation range of signal quality is essential for determining whether and to what extent the current signal quality has deteriorated. Only by determining the normal fluctuation range can the real-time signal quality be compared with it to determine whether the signal quality is abnormal; obtain the historical signal quality of non-target channels, and perform statistical analysis on these historical signal qualities, and calculate the mean and standard deviation of the historical signal quality, where the mean represents the average level of signal quality under normal conditions, and the standard deviation reflects the degree of dispersion of historical signal quality. The fluctuation range of normal signal quality is determined with the mean as the center and combined with the standard deviation. Generally, the mean plus or minus twice the standard deviation is used as the boundary of the normal fluctuation range. The fluctuation range determined in this way can cover most signal quality fluctuations under normal conditions and provide a standard for subsequent judgment; through detailed analysis of historical signal quality data, the normal fluctuation range determined is highly reliable and reasonable. It provides a clear and definite reference standard for judging whether the real-time signal quality is abnormal, enabling operation and maintenance personnel to quickly and accurately identify changes in signal quality, avoid misjudgment due to lack of standards, and improve the accuracy of signal quality monitoring.

[0120] Comparing the actual signal quality obtained in real time with the determined normal signal quality fluctuation range is a key step in preliminarily judging whether the signal quality is abnormal. By comparing, the actual signal quality status can be quickly determined so that different calculation methods can be used to evaluate the degradation of the signal quality. Only by clarifying the relationship between the actual signal quality and the normal range can subsequent analysis be carried out in a targeted manner to improve the accuracy and efficiency of the evaluation. The actual signal quality of non-target channels is obtained in real time, the average value and variance of the actual signal quality are calculated, and the average value of the actual signal quality is compared with the boundary of the predetermined normal signal quality fluctuation range. If the actual average value exceeds the normal range, the signal quality of the non-target channels is analyzed. If the actual average value is between the upper and lower limits of the fluctuation range of normal signal quality, the actual signal quality is judged to be outside the fluctuation range of normal signal quality. If the actual average value is between the upper and lower limits of the fluctuation range of normal signal quality, the actual signal quality is judged to be within the fluctuation range of normal signal quality. According to different judgment results, different methods will be used to calculate the degradation degree of signal quality in the future. In this way, the signal quality status can be quickly classified to quickly identify whether the signal quality is abnormal, provide a clear direction for the subsequent targeted calculation of the degradation degree of signal quality, avoid unnecessary calculation and analysis, improve the efficiency of signal quality assessment, and timely discover potential signal quality problems.

[0121] When the average value of the actual signal quality is not within the range of normal signal quality, calculating the boundary ratio can quantify the degree to which the signal quality deviates from the normal state; when the average value of the actual signal quality is within the range of normal signal quality, by calculating the degree of fluctuation anomaly, the stability of the signal quality fluctuation can be evaluated. The boundary ratio and the degree of fluctuation anomaly are weighted and summed according to a certain weight, which can comprehensively and accurately quantify the degree of signal quality degradation and provide a comprehensive quantitative indicator for evaluating the fault condition of non-target channels; when the average value of the actual signal quality is not within the range of normal signal quality, the proportion of the excess part to the boundary value is calculated, that is, the boundary ratio, where the upper and lower limits of the fluctuation range of normal signal quality are the normal signal quality. The boundary value of the fluctuation range of the signal quality is preferably (110-100) / 100 if the upper limit of the normal signal quality range is 100 and the average value of the actual signal quality is 110. In this case, the value of the abnormal fluctuation degree is 0. When the average value of the actual signal quality is within the normal signal quality range, the difference between the variance of the actual signal quality and the variance of the historical signal quality is compared to evaluate the abnormal fluctuation degree. That is, a value that can reflect the abnormal fluctuation degree is determined by calculating the difference between the two. In this case, the boundary ratio is 0. The weight values ​​of the boundary ratio and the abnormal fluctuation degree are set according to the network operation status. The boundary ratio and the abnormal fluctuation degree are weighted and summed to obtain the final signal quality degradation degree.

[0122] This fully considers the changes in signal quality under different states and converts the degradation of signal quality into specific quantitative indicators. Whether it is the degree of deviation of signal quality beyond the normal range or the fluctuation stability within the normal range, it can be accurately evaluated through this method, providing detailed and accurate signal quality for fault diagnosis and network optimization, and improving the accuracy and comprehensiveness of non-target channel fault assessment. The obtained signal quality degradation degree is used to calculate the residual fault index of the non-target channel, which is an important component. Its accuracy directly affects the residual fault index's judgment of the overall fault condition of the non-target channel.

[0123] Furthermore, the execution sub-logic of the channel reconfiguration includes:

[0124] Based on the residual fault index of non-target channels, the fault correlation degree between each non-target channel is calculated to form a fault correlation matrix. The fault correlation matrix is ​​processed by the spectral clustering algorithm to divide the non-target channels into multiple channel clusters.

[0125] In each channel cluster, the residual fault index is sorted to determine the reconfiguration priority. At the same time, reinforcement learning is used to generate the reconfiguration operation sequence for each channel cluster with channel performance as the reward target.

[0126] Channel reconfiguration of each channel cluster is performed in series-parallel mode in combination with the fault correlation degree and the reconfiguration operation sequence.

[0127] Non-target channels in a communication network are not independent of each other when a fault occurs. There are complex correlations between them. A link failure can cause performance degradation in multiple channels connected to it. By calculating the degree of fault correlation and dividing the channels into clusters, channels with similar fault characteristics or mutual influence can be grouped together, facilitating subsequent targeted reconfiguration operations and improving fault handling efficiency and accuracy. After the target channel is isolated, the communication node continuously obtains the residual fault index of the non-target channel and calculates the fault correlation degree between each two non-target channels based on the residual fault index using the cosine similarity algorithm. In the specific operation, the residual fault index of each channel is regarded as a vector in a high-dimensional space, and their similarity is measured by calculating the cosine value of the angle between the vectors. The closer the cosine value is to 1, the higher the degree of fault correlation between the two channels. The calculated results of the correlation degree between all non-target channels are arranged into a matrix, forming a fault correlation matrix.

[0128] The fault correlation matrix is ​​then processed using a spectral clustering algorithm. This algorithm views the fault correlation matrix as a graph, where each channel corresponds to a node in the graph and the weights of the edges between nodes are determined by the degree of fault correlation. By analyzing the eigenvalues ​​and eigenvectors of the graph's Laplace matrix, the algorithm divides the nodes into clusters. This ensures that the weights of edges between nodes within the same cluster are larger, indicating a high degree of fault correlation, while the weights of edges between different clusters are smaller. This successfully divides non-target channels into multiple channel clusters. This allows the fault relationships between non-target channels to be clearly identified, breaking down the originally complex multi-channel fault handling problem into multiple relatively simple sub-problems. Operations and maintenance personnel can then develop reconfiguration strategies tailored to the characteristics of each channel cluster, avoiding blind operations and significantly improving the efficiency and specificity of fault handling.

[0129] Within each channel cluster, the severity of faults varies across channels, and different reconfiguration operations also have varying effects on channel performance. To maximize network performance restoration within limited time and resources, it is necessary to determine reconfiguration priorities and generate an optimal reconfiguration sequence, ensuring that channels with the greatest impact on network performance are prioritized. Within each assigned channel cluster, non-target channels are ranked based on their residual fault index. A higher residual fault index indicates a more severe fault in the non-target channel, and its reconfiguration priority is higher.

[0130] To generate a reconfiguration sequence for each channel cluster, reinforcement learning is used to target channel performance improvement. Reconfiguration operations, including adjusting channel bandwidth, changing modulation schemes, and switching transmission protocols, are defined. Initially, the reinforcement learning agent randomly selects a reconfiguration operation and applies it to the channel cluster. The agent then observes the performance changes after the operation, including whether throughput increases and latency decreases. Based on the performance changes, the agent is given corresponding reward or penalty signals. If the throughput of the channel cluster significantly improves after the operation, the agent receives a higher reward; conversely, if performance decreases, the agent receives a penalty. Through continuous trial and error and learning, the agent gradually adjusts its strategy to maximize long-term cumulative rewards, ultimately generating the optimal reconfiguration sequence for each channel cluster. Clear reconfiguration priorities ensure that the issues most impacting network performance are addressed first during the reconfiguration process, accelerating network recovery. The reconfiguration sequence generated through reinforcement learning is the optimal solution, derived through extensive experimentation and learning. It leverages the advantages of various reconfiguration operations, configuring the channel in the optimal order and manner, thereby more effectively improving channel performance.

[0131] The degree of fault correlation between different channel clusters varies. Some channel clusters are independent, and their reconfiguration operations will not affect each other. However, some channel clusters are strongly correlated, and the reconfiguration of one channel cluster will affect other channel clusters. To quickly complete channel reconfiguration while ensuring network stability, it is necessary to select an appropriate execution mode based on the degree of fault correlation. Based on the previously calculated fault correlation matrix, the degree of correlation between each channel cluster is analyzed. For independent channel clusters with low fault correlation, the reconfiguration operation is executed in parallel mode. That is, the reconfiguration process of these channel clusters is started simultaneously, and the channels within the channel cluster are configured according to their respective reconfiguration operation sequences.

[0132] For channel clusters with potential impact relationships and high fault correlation, reconfiguration is performed in serial mode. According to the pre-determined dependencies between channel clusters, the channel clusters are reconfigured in sequence. During the execution process, key performance indicators of the network, including network throughput, latency, and packet loss rate, are monitored in real time. Once a performance indicator drops below a pre-set threshold, the current reconfiguration operation is immediately suspended and a rollback mechanism is activated to restore the reconfiguration operation to the previous state to avoid further impact on the network. Channel reconfiguration is performed in serial and parallel mode, fully considering the actual relationship between channel clusters, which not only improves the efficiency of reconfiguration but also ensures the stability of the communication network. The parallel mode speeds up the processing of independent channel clusters, while the serial mode and rollback mechanism effectively prevent cascading failures caused by reconfiguration, ensuring the continuous and stable operation of the communication network during the reconfiguration process. After the successful completion of channel reconfiguration, the topology and performance status of the communication network change, which provides new network environment data for subsequent iterative optimization of the parameters of the spatiotemporal convolutional neural network. The changes in network performance will serve as an important reference for evaluating the parameter optimization effect of the spatiotemporal convolutional neural network.

[0133] Specifically, the optimization logic of the parameters of the spatiotemporal convolutional neural network includes:

[0134] Set a multi-level data threshold, compare the accumulated number of abnormal events with the multi-level data threshold, and obtain the level of the number of abnormal events;

[0135] The parameter optimization granularity is determined based on the level of abnormal events. The parameter optimization granularity includes local parameter optimization and global parameter optimization.

[0136] The parameters of the spatiotemporal convolutional neural network are optimized according to the parameter optimization granularity, and the parameter optimization effect of the spatiotemporal convolutional neural network is evaluated based on the validation set to determine whether the parameters of the spatiotemporal convolutional neural network should be optimized again.

[0137] The different numbers of accumulated abnormal events reflect the different frequencies and severity of faults in the communication network. The degree of impact of abnormal events of different numbers on network performance varies. Therefore, it is necessary to set multi-level data thresholds to classify the number of abnormal events so as to adopt different parameter optimization granularity for different levels to achieve reasonable allocation of resources and maximize the optimization effect. Network administrators set multi-level data thresholds, including low thresholds, medium thresholds and high thresholds, based on the scale of the network, historical fault data and business requirements for network stability. The settings of these thresholds are not fixed, but are dynamically adjusted according to the actual situation of network operation. During network operation, the number of abnormal events is accumulated in real time. When the number of abnormal events is greater than a certain threshold, it is divided into corresponding levels. When the accumulated number of abnormal events is greater than the low threshold but less than the medium threshold, it is judged as a low level. When the accumulated number of abnormal events is greater than the medium threshold but less than the high threshold, it is judged as a medium level, and when it is greater than the high threshold, it is judged as a high level. This enables a more detailed assessment of the current fault status of the network. Different levels correspond to different parameter optimization granularities, so that optimization resources can be reasonably arranged according to the severity of the fault, avoiding over-optimization when there are fewer abnormal events or under-optimization when there are more abnormal events, thereby improving the pertinence and efficiency of parameter optimization.

[0138] The impact of abnormal events of different levels on network performance varies in scope and degree, and a unified parameter optimization approach cannot achieve optimal results. Therefore, it is necessary to determine the appropriate parameter optimization granularity based on the level of the abnormal events to ensure optimization results while reducing unnecessary computing resource consumption. When the level of abnormal events is low, it means that the frequency of network failures is low and the impact on overall network performance is relatively small. In this case, local parameter optimization is selected, that is, only the parameters in the spatiotemporal convolutional neural network that are closely related to the current abnormal event characteristics are adjusted. If the current abnormal event is mainly manifested as abnormal delay jitter in certain channels, only the parameters of the relevant layers in the network responsible for extracting and analyzing delay jitter features are optimized.

[0139] When the number of abnormal events is medium, it indicates that network failures occur frequently and have a significant impact on network performance. In this case, local parameter optimization is selected and the scope of parameter optimization is expanded. In addition to adjusting the parameters directly related to the current abnormal event, the surrounding parameters associated with these parameters are also appropriately optimized to further improve the network's fault detection capability. When the number of abnormal events is high, it means that failures occur frequently in the network and have a serious impact on network performance. In this case, global parameter optimization is selected to comprehensively adjust all parameters of the entire spatiotemporal convolutional neural network. By retraining the network, the structure and parameters of the spatiotemporal convolutional neural network are deeply optimized to adapt to complex and changing fault environments. Flexible selection of parameter optimization granularity based on the number of abnormal events allows the parameters of the spatiotemporal convolutional neural network to be optimized in the most reasonable way under different fault conditions. This not only avoids the waste of computing resources caused by large-scale global optimization when faults are rare, but also effectively improves network performance through comprehensive optimization when faults are severe, thereby improving the efficiency and effectiveness of parameter optimization.

[0140] After optimizing the parameters of the spatiotemporal convolutional neural network, it is necessary to evaluate the optimization effect to determine whether the optimization has achieved the expected goals, so as to determine whether it is necessary to adjust the optimization strategy or optimize again, so as to continuously improve the network's fault detection capability and accuracy; according to the determined parameter optimization granularity, select the corresponding optimization algorithm to adjust the parameters of the spatiotemporal convolutional neural network. When performing local parameter optimization, the stochastic gradient descent algorithm will be used to optimize a specific parameter subset to speed up the convergence speed. During the optimization process, the learning rate is dynamically adjusted according to the gradient information of the parameters, so that the parameters can be updated more quickly towards the optimal solution.

[0141] When performing global parameter optimization, a second-order optimization algorithm is used to more effectively search for the global optimal solution. During the optimization process, regularization techniques are used to prevent network overfitting and improve the generalization ability of the spatiotemporal convolutional neural network. After completing parameter optimization, the optimized spatiotemporal convolutional neural network is tested based on a validation set. The validation set contains various types of fault data and normal data that have not participated in the network training process. The data in the validation set is input into the spatiotemporal convolutional neural network to observe performance indicators such as the accuracy of the spatiotemporal convolutional neural network in judging fault modes, the accuracy of generating channel fault labels, and the rationality of triggering signal isolation instructions. The evaluated performance indicators are compared with the pre-set targets. If the performance indicators meet or exceed the targets, it means that the current parameter optimization has achieved good results and no further optimization is needed for the time being. If the performance indicators do not meet the targets, the reasons are analyzed and the parameter optimization strategy is adjusted, including changing the optimization algorithm, adjusting the initial parameter values, or changing the optimization granularity. Then, the parameters are optimized again until a satisfactory optimization effect is achieved.

[0142] Through rigorous parameter optimization and effect evaluation processes, the performance of the spatiotemporal convolutional neural network can be continuously improved, making it better adapt to the ever-changing fault conditions in the network. Continuous optimization and evaluation ensure the accuracy and reliability of the spatiotemporal convolutional neural network in fault detection, thereby improving the effectiveness of the entire fault detection. The optimized and performance-compliant spatiotemporal convolutional neural network will provide more accurate and reliable support for subsequent fault mode judgment, channel fault mark generation, and signal isolation instruction triggering, forming a virtuous circle and continuously improving the performance and effect of the entire multi-communication channel fault detection method.

[0143] Example 2

[0144] The embodiment of the present application provides a multi-communication channel fault detection and self-healing control device, the device structure includes: a device body and a control module;

[0145] The control module includes an acquisition unit, a processing unit and a control unit. The acquisition unit is used to acquire multi-channel fault characterization parameters at the communication node through sensors, fuse them into fault feature tensors, and simultaneously acquire the operating status of the device body.

[0146] Taking the smart street light as an example, the device body is the street light, and the device structure of the street light is as follows: Figure 4 As shown, multiple street lamps distributed at different intersections serve as communication nodes. Each communication node exchanges data through a wireless communication channel. The control module is integrated on the circuit board inside the street lamp to realize fault detection and self-healing control of the street lamp communication channel, while ensuring the normal operation of the street lamp.

[0147] During the operation of traffic lights, the sensors in the acquisition unit work continuously. After obtaining fault characterization parameters such as delay jitter, signal-to-noise ratio, and signal quality of each channel, the acquisition unit classifies and organizes the fault characterization parameters according to channel index and time window, using the number of the street light communication channel as the index identifier. The acquisition unit then normalizes these fault characterization parameters to eliminate dimensional differences and constructs a time-space alignment matrix, filling data with channel index as columns and time window as rows; and fuses them into a fault feature tensor to achieve deep extraction and integration of multi-channel fault features.

[0148] The acquisition unit is also equipped with sensors specifically used to monitor the operating status of the device body, including a current sensor to monitor the current size of the street lamp circuit and determine whether the street lamp is short-circuited or overloaded; a voltage sensor to monitor the power supply voltage to ensure that the voltage is stable within the normal operating range; an angle sensor to detect the inclination angle of the street lamp pole to prevent the pole from tipping over due to external force. These sensors obtain the operating status of the street lamp in real time and transmit the data to the acquisition unit to provide information for subsequent analysis.

[0149] The processing unit is used to process the fault feature tensor through a spatiotemporal convolutional neural network to determine the fault mode and output the confidence probability of the fault mode. When it is determined to be a high confidence probability, a channel fault mark is generated based on the fault feature tensor, the target channel is determined according to the channel fault mark, and a detection signal is injected into the target channel. The trigger signal isolation instruction is determined based on the bit error rate of the detection signal, and the impact of the channel fault on the device body is determined based on the operating status of the device body.

[0150] After receiving the fault feature tensor from the acquisition unit, the processing unit processes the fault feature tensor through a spatiotemporal convolutional neural network to determine the fault mode and output the confidence probability of the fault mode. The output fault mode is the confidence probability of a local fault or a global fault. When the judgment result is a high confidence probability, the processing unit generates a channel fault mark based on the fault feature tensor. According to the generated channel fault mark, the processing unit determines the target channel, injects a detection signal into the target channel, and determines the bit error rate of the detection signal. At the same time, the processing unit combines the operating status of the device body to determine the impact of the channel fault on the smart street light, including whether the street light has abnormal flickering and display errors, etc., and comprehensively determines whether to trigger the signal isolation instruction based on the judgment result of the bit error rate and the impact of the channel fault on the device body.

[0151] The control unit is used to receive signal isolation instructions and control the device body to isolate the target channel, while reacquiring the fault characterization parameters of the non-target channel, and comprehensively judging the residual fault index of the non-target channel based on the residual of the delay jitter and the degradation degree of the signal quality, marking abnormal events based on the abnormal judgment of the residual fault index and accumulating the number of abnormal events, executing channel reconfiguration and iteratively optimizing the parameters of the spatiotemporal convolutional neural network based on the number of abnormal events, and controlling the device body based on the execution results of the channel reconfiguration and the impact of the channel fault on the device body.

[0152] When the control unit receives the signal isolation command from the processing unit, it immediately controls the isolation device, namely the relay, inside the smart street light to isolate the target channel, cut off the communication connection of the faulty channel, and prevent the fault from spreading further. While isolating the target channel, the control unit initiates the data re-acquisition process for the non-target channel, and the collaborative acquisition unit re-acquires the fault characterization parameters of the non-target channel. The control unit processes the delay jitter through empirical mode decomposition to obtain the residual of the delay jitter, and at the same time determines the degree of signal quality degradation based on historical fault data. Then, according to the current operating status of the smart street light, the control unit assigns dimensional weights to the residual of the delay jitter and the degree of signal quality degradation, takes a weighted sum of the two, and comprehensively determines the residual fault index of the non-target channel.

[0153] Based on the abnormal conditions of the residual fault index, the control unit marks abnormal events and accumulates the number of abnormal events. According to the number of abnormal events, it determines whether to perform channel reconfiguration and iteratively optimizes the parameters of the spatiotemporal convolutional neural network. The control unit also performs comprehensive control of the street lights based on the execution results of the channel reconfiguration and the impact of the channel fault on the street lights. If the street lights still display abnormal conditions after channel reconfiguration, the control unit checks the hardware connections and control circuits of the street lights and performs corresponding troubleshooting and repairs. Through the collaborative work of various units, the fault detection and self-healing control device for multiple communication channels can achieve rapid detection, accurate judgment, effective isolation and intelligent self-healing of communication channel faults, while ensuring the normal operation of smart street lights.

[0154] Since the principle of solving the problem by the device in the embodiment of the present application is similar to the method described above in the embodiment of the present application, the implementation of the device refers to the implementation of the method, and the repeated parts will not be repeated.

Claims

1. A multi-communication channel fault detection method, characterized in that: include: Acquire multi-channel fault characterization parameters at the communication node. These parameters include each channel's delay jitter, signal-to-noise ratio, and signal quality. Align the fault characterization parameters with the time window by channel index and fuse them into a fault feature tensor. The fault feature tensor is processed through a spatiotemporal convolutional neural network to determine the fault mode and output the confidence probability of the fault mode. The fault mode includes local faults and global faults. When the confidence probability is high, a channel fault mark is generated based on the fault feature tensor. The target channel is determined based on the channel fault mark, and a detection signal is injected into the target channel. The bit error rate of the detection signal is used to determine the trigger signal isolation instruction. Receive signal isolation instructions and isolate the target channel, while reacquiring the fault characterization parameters of the non-target channel. The residual fault index of the non-target channel is comprehensively judged based on the residual of the delay jitter and the degradation of the signal quality. Abnormal events are marked based on the abnormal judgment of the residual fault index and the number of abnormal events is accumulated. Channel reconfiguration is performed based on the number of abnormal events and the parameters of the spatiotemporal convolutional neural network are iteratively optimized.

2. The multi-communication channel fault detection method according to claim 1, wherein: The fusion logic of the fault feature tensor includes: Acquire multi-channel fault characterization parameters at the communication node and perform normalization on the fault characterization parameters; Construct a spatiotemporal alignment matrix with channel index as columns and acquisition time within the time window as rows, and fill the normalized fault characterization parameters into the spatiotemporal alignment matrix; According to the historical contribution of each fault characterization parameter to fault judgment, a weight coefficient is assigned to each fault characterization parameter through the attention mechanism to form a weighted matrix. The weighted matrix and the spatiotemporal alignment matrix are subjected to a tensor product operation to obtain the fault feature tensor.

3. The multi-communication channel fault detection method according to claim 2, wherein: The trigger logic of the signal isolation instruction includes: Determine the target channel based on the channel fault flag and dynamically adjust the detection signal parameters according to the fault mode and severity. The detection signal parameters include signal strength, frequency, and modulation mode. The bit error rate of the detection signal is analyzed by the sequential probability ratio test method, and a bit error rate threshold is set. The bit error judgment result is obtained by comparing the bit error rate of the detection signal with the bit error rate threshold; Determine the cascading risk of signal isolation based on the propagation risk, configure the risk threshold, compare the cascading risk of signal isolation with the risk threshold, and obtain the isolation risk result. Make a comprehensive judgment based on the error judgment result and the isolation risk result to trigger the signal isolation instruction.

4. The method for detecting a fault in multiple communication channels according to claim 3, wherein: The judgment sub-logic of the failure mode includes: The fault feature tensor is processed through a spatiotemporal convolutional neural network, which includes a feature extraction layer, an attention enhancement layer, and a classification decision layer; The confidence probability of the fault mode is output through the spatiotemporal convolutional neural network, and the entropy and divergence of the confidence probability are calculated to determine the degree of certainty of the fault mode; The judgment threshold with high confidence probability is dynamically adjusted according to historical fault data and the current network operation status, and the confidence probability of the fault mode is compared with the judgment threshold to determine the fault mode.

5. The method for detecting a fault in multiple communication channels according to claim 4, wherein: The generation sub-logic of the channel fault mark includes: When the confidence probability is high, the correlation coefficient between the fault sub-features and the fault mode of each channel in the fault feature tensor is calculated to obtain the fault impact matrix; The causal relationship between the fault sub-features of each channel in the fault feature tensor is analyzed through a causal inference algorithm. Based on the causal relationship, the direct fault source and the indirect impact channel are distinguished to generate the fault propagation path. A channel fault tag including channel index, fault mode, severity and propagation risk is generated according to the fault impact matrix and fault propagation path.

6. The method for detecting a fault in multiple communication channels according to claim 5, wherein: The optimization logic of the parameters of the spatiotemporal convolutional neural network includes: Set a multi-level data threshold, compare the accumulated number of abnormal events with the multi-level data threshold, and obtain the level of the number of abnormal events; The parameter optimization granularity is determined based on the level of abnormal events. The parameter optimization granularity includes local parameter optimization and global parameter optimization. The parameters of the spatiotemporal convolutional neural network are optimized according to the parameter optimization granularity, and the parameter optimization effect of the spatiotemporal convolutional neural network is evaluated based on the validation set to determine whether the parameters of the spatiotemporal convolutional neural network should be optimized again.

7. The method for detecting a fault in multiple communication channels according to claim 6, wherein: The judgment sub-logic of the residual fault index of the non-target channel includes: Re-acquire the fault characterization parameters of the non-target channel, process the delay jitter through empirical mode decomposition to obtain the delay jitter residual, and determine the degree of signal quality degradation based on historical fault data; Assign dimensional weights to the residual of delay jitter and the degree of signal quality degradation based on the current network operating status; The residual error of delay jitter and the degradation degree of signal quality are weighted and summed according to the dimension weight to comprehensively judge the residual fault index of the non-target channel.

8. The method for detecting a fault in multiple communication channels according to claim 7, wherein: The system uses a long short-term memory network to learn historical delay jitter and obtain a delay jitter prediction curve under normal conditions. It then re-obtains the actual delay jitter of the non-target channel and subtracts the actual delay jitter from the prediction curve point by point to obtain a delay jitter residual sequence. This residual sequence is then decomposed into multiple intrinsic mode functions (IMFs) using empirical mode decomposition to obtain the delay jitter residual. The judgment sub-logic of the signal quality degradation degree includes: Obtain historical signal quality and calculate the mean and standard deviation of the historical signal quality to determine the fluctuation range of normal signal quality; Re-obtain the actual signal quality of the non-target channel, calculate the average value and variance of the actual signal quality, and compare the average value of the actual signal quality with the fluctuation range of the normal signal quality; When the average value of the actual signal quality is not within the fluctuation range of the normal signal quality, the boundary ratio is calculated. The boundary ratio is obtained by solving the difference between the average value of the actual signal quality and the boundary value of the fluctuation range of the normal signal quality, and dividing the difference by the boundary value of the fluctuation range of the normal signal quality. The degree of abnormal fluctuation is 0. When the average value of the actual signal quality is within the normal signal quality fluctuation range, the degree of fluctuation anomaly is calculated based on the difference between the variance of the actual signal quality and the variance of the historical signal quality. The boundary ratio is set to 0, and the boundary ratio and the degree of fluctuation anomaly are weighted summed to determine the signal quality degradation degree.

9. The method for detecting a fault in multiple communication channels according to claim 8, wherein: The execution sub-logic of the channel reconfiguration includes: Based on the residual fault index of non-target channels, the fault correlation degree between each non-target channel is calculated to form a fault correlation matrix. The fault correlation matrix is ​​processed by the spectral clustering algorithm to divide the non-target channels into multiple channel clusters. In each channel cluster, the residual fault index is sorted to determine the reconfiguration priority. At the same time, reinforcement learning is used to generate the reconfiguration operation sequence for each channel cluster with channel performance as the reward target. Channel reconfiguration of each channel cluster is performed in series-parallel mode in combination with the fault correlation degree and the reconfiguration operation sequence.

10. A multi-communication channel fault detection and self-healing control device, used to implement the multi-communication channel fault detection method according to any one of claims 1 to 9, characterized in that: include: Device body and control module; The control module includes an acquisition unit, a processing unit and a control unit. The acquisition unit is used to acquire multi-channel fault characterization parameters at the communication node through sensors, fuse them into a fault feature tensor, and simultaneously acquire the operating status of the device body; The processing unit is configured to process the fault feature tensor through a spatiotemporal convolutional neural network to determine a fault mode and output a confidence probability of the fault mode; when a high confidence probability is determined, generate a channel fault mark based on the fault feature tensor; determine a target channel based on the channel fault mark; inject a detection signal into the target channel; determine a trigger signal isolation instruction based on a bit error rate of the detection signal; and determine an impact of the channel fault on the device body based on the operating status of the device body; The control unit is used to receive a signal isolation instruction and control the device body to isolate the target channel, while reacquiring the fault characterization parameters of the non-target channel, and comprehensively judging the residual fault index of the non-target channel based on the residual of the delay jitter and the degradation degree of the signal quality, marking abnormal events based on the abnormal judgment of the residual fault index and accumulating the number of abnormal events, executing channel reconfiguration and iteratively optimizing the parameters of the spatiotemporal convolutional neural network based on the number of abnormal events, and controlling the device body based on the execution result of the channel reconfiguration and the impact of the channel fault on the device body.

Images