Automatic fault recovery system for dual-redundancy network of industrial switch

Abstract

The invention relates to the technical field of communication networks, and particularly discloses an automatic fault recovery system for a dual-redundancy network of an industrial switch. The system comprises a fault sensing and feature extraction module, a fault mode recognition and confidence evaluation module and a self-adaptive decision and switching control module. The link state is monitored in real time, multi-dimensional features are extracted, the fault type is identified based on a fault mode classification model, the confidence coefficient is evaluated, and finally intelligent switching decision making is carried out in combination with a dynamic threshold value and a network context. According to the scheme, continuous faults and instantaneous interference can be effectively distinguished, the error switching rate is reduced, orderly transmission of upper-layer control signals is guaranteed in the switching process, and the reliability and stability of an industrial network are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of communication networks, and particularly relates to an industrial switch dual-redundancy network fault automatic recovery system. BACKGROUND

[0002] In the field of industrial automation and intelligent manufacturing, industrial Ethernet is the core network connecting controllers, sensors, actuators and other key devices, and its reliability and stability directly determine the continuity and efficiency of the production process. Network redundancy technology is a key means to ensure high availability of industrial networks, aiming to provide backup paths to achieve fast switching when the main communication link fails, thereby avoiding production interruption.

[0003] Dual-redundancy network architecture is a widely used high-reliability solution in industrial switches. This architecture usually contains two independent physical links, which realize automatic monitoring and switching of the main and backup links through specific protocols (such as Rapid Spanning Tree Protocol RSTP, ring network protocol, etc.), to ensure continuous and smooth network communication.

[0004] In the prior art, the redundancy switching mechanism of industrial switches mainly relies on periodic detection of link status. However, this mechanism faces challenges in actual industrial environments: the industrial field environment is complex, with electromagnetic interference, mechanical vibration and other factors that can easily cause transient jitter or temporary interruption of the link. Existing systems often lack the ability to distinguish such transient faults, and are prone to unnecessary redundancy switching.

[0005] Such frequent switching or out-of-order switching caused by misjudgment not only causes oscillation of the network topology, but more seriously causes out-of-order, delay or loss of upper-layer control signals. In high-speed, precise continuous production lines, any abnormality of control signals can directly cause device misoperation, production rhythm disorder, and ultimately lead to unplanned shutdown of the entire production line, causing significant economic losses. Therefore, how to accurately distinguish network faults and avoid misjudgment caused by transient interference has become a technical problem to be solved in industrial high-reliability networks. SUMMARY

[0006] The purpose of the present application is to provide an industrial switch dual-redundancy network fault automatic recovery system to solve the problem of frequent misjudgment of redundancy network caused by insufficient ability to distinguish transient link faults in the prior art, which further causes abnormal transmission of upper-layer control signals and production interruption.

[0007] The present application provides an industrial switch dual-redundancy network fault automatic recovery system, which is deployed in at least two industrial switches constituting a dual-redundancy network architecture, and the system comprises: The fault perception and feature extraction module is used for real-time monitoring of the primary link and the backup link in the dual-redundancy network, collecting original state parameters of the link layer, and performing multi-dimensional feature extraction on the collected original state parameters to generate a feature vector representing the health degree of the link. The fault mode recognition and confidence evaluation module is used for receiving the feature vector generated by the fault perception and feature extraction module, performing mode recognition on the current link state based on a preset fault mode classification model, outputting the recognized fault mode category, and calculating the confidence score of the recognition result. The adaptive decision and switching control module is used for receiving the fault mode category and the corresponding confidence score output by the fault mode recognition and confidence evaluation module, combining a preset switching strategy library and the current network running context, executing switching decision logic, and generating a final link control instruction to decide whether to maintain the current link, start switching to the backup link, or perform a link recovery operation.

[0008] Further, the fault perception and feature extraction module specifically includes a physical layer probe unit and a feature engineering unit.

[0009] The physical layer probe unit synchronously collects physical layer signal quality indicators of the primary link and the backup link at a sampling period greater than 1 millisecond. The indicators include but are not limited to received signal strength indication, carrier sensing signal quality, bit error rate, and existence of link pulse signal.

[0010] The feature engineering unit processes the time series data collected by the physical layer probe unit, and the extracted feature dimensions include: Within a time window of 100 milliseconds, the number of abnormal points with signal quality indicators less than a first preset threshold is counted; the sliding average and standard deviation of the signal quality indicators in the time window are calculated; Detecting a complete jump event from normal to abnormal and then to normal, and recording the duration of the event; Combining the above statistical values, calculated values, and event duration to form a multi-dimensional feature vector.

[0011] Further, the fault mode recognition and confidence evaluation module has a trained fault mode classification model built in.

[0012] The training process of the model is based on a large amount of feature vector data set collected historically and labeled with fault types, and the fault mode categories that can be recognized at least include: Persistent physical interruption, intermittent electromagnetic interference, port hardware aging attenuation, and no fault normal jitter.

[0013] When running, the module inputs the real-time input feature vector into the classification model, and the model outputs the probability distribution corresponding to each fault mode category.

[0014] The confidence evaluation unit then selects the class corresponding to the highest probability value in the probability distribution as the recognition result, and maps the highest probability value after normalization to an integer between 0 and 100 as the confidence score of this recognition.

[0015] Further, the decision logic of the adaptive decision and switching control module makes hierarchical decisions according to the combination of the fault mode category and the confidence score.

[0016] The module maintains a switching strategy library containing multiple decision rules. The decision process first matches the fault mode category. If the recognition result is a persistent physical interruption and the confidence score is greater than 95, a control instruction to switch to the backup link is immediately generated.

[0017] If the recognition result is intermittent electromagnetic interference, a delay decision mechanism is started: The mechanism compares the confidence score of the current recognition result with the dynamically adjusted confidence threshold, and starts an observation timer with a configurable duration. Before the observation timer expires, new recognition results and confidence scores are continuously received; If the recognition result is intermittent electromagnetic interference and the average confidence score is less than the dynamic confidence threshold during the observation period, it is determined to be transient interference, and an instruction to maintain the current link is generated; If the recognition result changes to a persistent physical interruption or the average confidence score is consistently greater than the dynamic confidence threshold during the observation period, it is determined to be a potential hard failure risk, and a preventive switching instruction is generated.

[0018] Further, the dynamic confidence threshold is not a fixed value, but is dynamically calculated by the network context perception unit according to the current network load rate and the priority of critical business flows.

[0019] The network context perception unit monitors the traffic passing through the current link in real time, calculates the percentage of occupied bandwidth as the network load rate, and identifies the data flows belonging to the pre-set critical business list.

[0020] The calculation formula of the dynamic confidence threshold is as follows: ; represents the calculated dynamic confidence threshold. represents the base threshold, which is set to 80. represents the network load rate coefficient, which is set to 0.1. represents the real-time network load rate, with a value range of 0 to 100. represents the critical business flow coefficient, which is set to 2. represents the number of currently active critical business flows.

[0021] When the network load rate is high or there is a critical service flow, the calculated dynamic confidence threshold will be raised, so that the system is more cautious about the switching decision, avoiding unnecessary switching during high load or critical service.

[0022] Further, the adaptive decision and switching control module further comprises a switching execution and state synchronization unit.

[0023] This unit is responsible for coordinating the relevant switches in the dual-redundant network to perform specific port start-stop, routing table update and address resolution table refresh operations after generating the switching control instruction.

[0024] At the same time, this unit ensures that within a short time window before and after the switching action is executed, the control messages traversing the network are cached and sequentially rearranged to ensure the transmission order consistency of the upper-layer control signals, especially the continuous control instruction frames based on the transmission control protocol.

[0025] After the switching is completed, the unit synchronizes the new network topology state and the switching event log to the network management system and the log server through the in-band or out-of-band management channel.

[0026] Further, the system further comprises an offline training and model updating module. This module periodically collects the feature vectors generated by the fault perception and feature extraction module and the final fault determination results recorded by the adaptive decision and switching control module to form new training samples.

[0027] When the number of new samples accumulates to a certain amount, or the system detects that the recognition accuracy of the existing fault mode classification model for new interference patterns decreases to below a preset threshold, the module automatically triggers the model retraining process.

[0028] The retraining process uses an incremental learning algorithm to train the new samples based on the original model, generates an updated fault mode classification model version, and after verification, seamlessly updates to the fault mode recognition and confidence evaluation module in a hot deployment manner.

[0029] Further, as an embodiment of the present application, the industrial switch dual-redundant network fault automatic recovery system runs in the form of embedded software on the network processor and coprocessor of the industrial switch.

[0030] The fault perception and feature extraction module is deployed in the data plane and directly processes physical layer data by the coprocessor.

[0031] The fault mode recognition and confidence evaluation module and the adaptive decision and switching control module are deployed in the control plane and executed by the main network processor.

[0032] The modules interact with each other at high speed through a shared memory and a message queue mechanism.

[0033] Compared with the prior art, the application has the beneficial effects that: 1、The application fundamentally changes the traditional link fault detection method based on simple threshold discrimination by introducing a fault mode recognition and confidence evaluation mechanism. The system can distinguish between persistent hardware faults and transient environmental disturbances and other different fault modes, and make a comprehensive judgment combined with the quantitative confidence score, thereby greatly reducing the misjudgment rate of transient link jitter and avoiding frequent switching and topology shock of the redundant network caused by misjudgment.

[0034] 2、The adaptive decision mechanism adopted by the application introduces a dynamic confidence threshold and a delay decision, so that the switching decision is no longer an isolated and transient reaction, but an intelligent process combined with real-time network context and continuous observation. The mechanism can automatically increase the switching decision threshold when there is high network load or critical business flow, thereby enhancing the decision robustness and reliability of the system in complex industrial scenarios and ensuring the stability of network behavior during the critical production period.

[0035] 3、The application considers the orderedness guarantee of the control signal at the switching execution level, and through the message caching and sequential rearrangement technology, ensures that the transmission of upper-layer key control instructions does not appear out of order or loss during network redundancy switching, directly maintains the integrity of the control logic in the continuous production process, and prevents equipment misoperation or production rhythm disorder caused by network switching.

[0036] 4、The application has self-evolution ability, and through the offline training and model updating module, the system can continuously learn new fault features from the running environment and dynamically update its core recognition model. This enables the system to adapt to changing industrial environments and new disturbances, maintain high-precision fault discrimination ability for a long time, realize the leap from static protection to dynamic adaptation, and improve the reliability and maintainability throughout the life cycle. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 is the overall technical scheme architecture schematic diagram of the industrial switch dual-redundancy network fault automatic recovery system proposed by the application; Figure 2 is the core principle framework schematic diagram of the fault mode recognition and confidence evaluation module in the application; Figure 3 is the logic flow framework diagram of the adaptive decision and switching control module in the application; Figure 4 is the multi-level interaction relationship schematic diagram of the dynamic confidence threshold calculation and network context perception in the application; Figure 5is a self-evolution logical framework diagram of the offline training and model updating module in the present application. DETAILED DESCRIPTION

[0038] Embodiment 1: This embodiment details the specific implementation of the industrial switch dual-redundancy network fault automatic recovery system. Please refer to the attached Figures 1 to 5 The system is deployed in at least two industrial switches that constitute a dual-redundancy network architecture.

[0039] The system runs in the form of embedded software on the network processor and co-processor of the industrial switch.

[0040] The fault perception and feature extraction module is deployed in the data plane and directly processes physical layer data by the co-processor.

[0041] The fault mode identification and confidence evaluation module and the adaptive decision and switching control module are deployed in the control plane and executed by the main network processor.

[0042] High-speed data interaction between modules is achieved through shared memory and message queue mechanisms.

[0043] The shared memory area is divided into multiple ring buffers, each corresponding to a specific type of data structure, such as raw sample data buffer, feature vector buffer, and control instruction buffer.

[0044] The message queue adopts a priority queue mechanism to ensure that high-priority fault events or switching instructions are processed first.

[0045] When the system starts, each module loads configuration parameters from non-volatile memory, including sampling period, time window length, various preset threshold values, and model parameter files, and completes the initialization of shared memory and message queues.

[0046] The fault perception and feature extraction module is the data source and feature processing center of the system. This module specifically includes a physical layer probe unit and a feature engineering unit.

[0047] The physical layer probe unit directly interacts with the physical layer chip driver interface of the switch to synchronously collect physical layer signal quality indicators of the primary link and backup link in the dual-redundancy network at a fixed sampling period greater than 1 millisecond.

[0048] These indicators include received signal strength indication, carrier sense signal quality, bit error rate, and existence of link pulse signals.

[0049] The value range of received signal strength indication is -120 decibels milliwatts to -40 decibels milliwatts, which is directly read from the physical layer chip register. Carrier sense signal quality is an integer value from 0 to 100, representing signal clarity.

[0050] The bit error rate is calculated by the ratio of the number of error frames received in a unit of time to the total number of frames, with the unit being parts per million.

[0051] The presence of the link pulse signal is a Boolean value, obtained by detecting the level of a specific state pin of the physical layer chip, and a high level indicates a normal link pulse, and a low level indicates a lost link pulse.

[0052] The physical layer probe unit packs the four groups of indicators obtained each time the sampling is performed, together with the time stamp accurate to the microsecond level, into a data record, and writes it into the raw sampling data buffer in the shared memory through direct memory access.

[0053] The buffer is designed as a double buffering structure, when one buffer is full, the physical layer probe unit automatically switches to another buffer for continuous writing, and sends a data ready notification to the feature engineering unit through the message queue.

[0054] After receiving the data ready notification, the feature engineering unit reads the historical sampling data within the specified time window from the raw sampling data buffer.

[0055] The length of the time window is configured as 100 milliseconds, corresponding to about 100 sampling records. The feature engineering unit performs multi-dimensional feature extraction on these time series data.

[0056] The first feature is the abnormal point statistics.

[0057] The feature engineering unit compares the received signal strength indicator, carrier sense signal quality and bit error rate in each sampling record with the first threshold value respectively.

[0058] For the received signal strength indicator, the first threshold value is set to -85 decibel milliwatts; For the carrier sense signal quality, the first threshold value is set to 30; For the bit error rate, the first threshold value is set to 100.

[0059] If any of the indicators is less than the corresponding first threshold value, the sampling point is marked as an abnormal point. The feature engineering unit counts the total number of abnormal points within the 100 millisecond time window.

[0060] The second feature is the calculation of the sliding statistics.

[0061] The feature engineering unit calculates the sliding average and standard deviation of the received signal strength indicator and the carrier sense signal quality within the time window respectively.

[0062] The calculation of the sliding average uses the arithmetic average method.

[0063] The calculation of standard deviation is used to quantify the fluctuation of signal quality. The third feature is state transition event detection.

[0064] The feature engineering unit maintains a link state machine, with states including normal, abnormal, and unknown. The state machine transitions according to the abnormal point statistics: If 5 consecutive sampling points are marked as abnormal, the state transitions to abnormal; if 10 consecutive sampling points are not marked as abnormal, the state transitions back to normal.

[0065] When a complete cycle of state transition from normal to abnormal and back to normal is detected, the feature engineering unit records the duration of this event in milliseconds.

[0066] Finally, the feature engineering unit combines the six values of total abnormal points, average and standard deviation of received signal strength indicator, average and standard deviation of carrier sense signal quality, and state transition event duration to form a six-dimensional feature vector.

[0067] This feature vector is packaged into a data structure, written to the feature vector buffer in shared memory, and sent to the fault pattern recognition and confidence assessment module through the message queue.

[0068] The data structure contains the feature vector itself, the start timestamp of the time window in which the vector was generated, and the corresponding link identifier.

[0069] The fault pattern recognition and confidence assessment module is the intelligent discrimination core of the system.

[0070] Please refer to the attached Figure 2 This module contains a trained fault pattern classification model. This model is a multi-classification model based on gradient boosting decision tree algorithm, and its training process is based on a large set of feature vectors collected from history and labeled with fault types.

[0071] The fault type labels in the data set include at least persistent physical interruption, intermittent electromagnetic interference, port hardware aging attenuation, and normal jitter without fault.

[0072] The feature of persistent physical interruption is that the abnormal point statistics value is consistently high, the signal quality indicator mean is extremely low, and the standard deviation is small.

[0073] The feature of intermittent electromagnetic interference is that the abnormal point statistics value bursts periodically or randomly, the signal quality indicator mean decreases moderately, but the standard deviation is extremely large.

[0074] The feature of port hardware aging attenuation is that the signal quality indicator mean shows a slow and monotonic downward trend, and the abnormal point statistics value gradually increases.

[0075] The normal jitter without fault is shown as all eigenvalues fluctuate in a very small normal range.

[0076] The model file is loaded from storage to memory at system startup.

[0077] At runtime, the module fetches the feature vector sent by the feature engineering unit from the message queue and feeds it into the classification model for forward inference.

[0078] The model outputs a probability distribution vector, each element of which corresponds to the predicted probability of a fault mode category, and the sum of all elements is 1.

[0079] The confidence assessment unit is responsible for processing the output of the model.

[0080] First, the unit finds the maximum value in the probability distribution vector, which corresponds to the fault mode identified this time.

[0081] Subsequently, the confidence assessment unit normalizes and maps the maximum value to generate a confidence score.

[0082] The mapping function is designed as a linear function, mapping probability values from 0 to 1 to integer values from 0 to 100.

[0083] The specific calculation method is to multiply the probability value by 100 and round it to the nearest integer. For example, if the maximum probability value output by the model is 0.92, the confidence score is 92.

[0084] The confidence score reflects the degree of certainty of the model in the current recognition result.

[0085] The higher the score, the more likely the model believes that the current link state belongs to this fault mode.

[0086] The fault mode recognition and confidence assessment module encapsulates the identified fault mode category and the corresponding confidence score into a decision request message and sends it to the adaptive decision and switching control module through the message queue.

[0087] The message also contains the timestamp of the original feature vector and the link identifier.

[0088] The adaptive decision and switching control module is the command center of the system.

[0089] Please refer to the attached Figure 3 The module receives the decision request message from the fault mode recognition and confidence assessment module and executes complex decision logic based on the switching strategy library it maintains internally and the real-time network context.

[0090] The switching strategy library is a set of rules, each rule consisting of a fault mode category, a confidence condition, a context condition, and a corresponding action.

[0091] The core of the decision process is hierarchical judgment.

[0092] The module first parses the failure mode category in the decision request message.

[0093] If the identification result is persistent physical interruption and the confidence score is greater than 95, the decision logic matches an emergency switching rule.

[0094] This rule does not need to consider the network context, and the module immediately generates a control instruction to switch to the backup link.

[0095] The control instruction contains the target link identifier, the switching type is forced switching, and the high priority flag.

[0096] If the identification result is intermittent electromagnetic interference, the decision logic enters the delay judgment mechanism.

[0097] This mechanism involves two core components: Dynamic confidence threshold calculation unit and observation timer.

[0098] Please refer to the attached Figure 4 The dynamic confidence threshold is not a fixed value, but is calculated in real time by the network context perception unit.

[0099] The network context perception unit continuously monitors the traffic flow through the current main link.

[0100] It reads the port counter of the switch to obtain the number of bytes sent and received in the last 1 second, calculates the percentage of the theoretical bandwidth of the occupied port, and takes it as the real-time network load rate.

[0101] At the same time, the unit deeply analyzes the data packets flowing through, and matches them with the preset key business flow feature table according to the five-tuple source address, destination address, source port, destination port, and protocol type of the Internet Protocol.

[0102] The key business flow feature table predefines several rules, such as traffic from a specific programmable logic controller address and a destination port of Transmission Control Protocol 502 port being identified as a key business flow.

[0103] The network context perception unit counts the number of currently active key business flows.

[0104] The calculation formula of the dynamic confidence threshold is as follows: ; represents the calculated dynamic confidence threshold. represents the base threshold, which is set to 80. Network Load Coefficient, set to 0.1. Real-time Network Load, ranging from 0 to 100. Key Traffic Coefficient, set to 2. Number of currently active key traffic flows.

[0105] This formula shows that when the network load L increases or the number of key traffic flows increases, the dynamic confidence threshold will increase accordingly.

[0106] For example, when the network load is 50 and there are 2 key traffic flows, the calculated .

[0107] The increase in the threshold means that the system is more cautious about handover decisions, requiring a higher confidence to determine that a handover is needed, thereby avoiding unnecessary and potentially disruptive handover operations during periods of high load or key traffic.

[0108] When starting the delay judgment, the adaptive decision and handover control module first calculates the initial dynamic confidence threshold according to the current network context.

[0109] At the same time, it starts a configurable observation timer, with a default duration of 500 milliseconds.

[0110] Before the observation timer expires, the module continuously receives new identification results and confidence scores sent by the fault mode identification and confidence evaluation module for the same link.

[0111] The module maintains a list of confidence scores within the observation period. The decision logic is triggered at the end of the observation period.

[0112] The module calculates the arithmetic mean of all confidence scores within the observation period and compares this mean with the dynamically recalculated confidence threshold at the end of the observation period.

[0113] In addition, the module also checks the consistency of the identification results within the observation period.

[0114] If the identification result is consistently intermittent electromagnetic interference within the observation period, and the mean of the confidence scores is less than the current dynamic confidence threshold, the system determines that the current link condition is transient interference and has no lasting impact on traffic transmission.

[0115] At this time, the module generates instructions to maintain the current link, which may include slight parameter adjustments such as temporarily increasing the receive sensitivity of the physical layer chip.

[0116] If the identification result changes to persistent physical interruption or the average confidence score is consistently greater than the dynamic confidence threshold during the observation period, the system determines that there is a potential hard failure risk or persistent strong interference in the link.

[0117] At this time, the module generates a preventive switchover instruction, which has a priority lower than that of the emergency switchover but triggers a smooth switchover process.

[0118] For the identification result of port hardware aging attenuation or normal jitter without failure, a monitoring and alarm rule is configured in the switchover strategy library.

[0119] For example, for the port hardware aging attenuation, if the confidence score thereof is greater than 80 for a plurality of times successively, the module does not immediately trigger the switchover but generates a serious-level alarm event, which is reported to the network management system through the management channel to prompt the maintenance personnel to perform preventive maintenance.

[0120] For the normal jitter without failure, the module only performs log recording and does not perform any link control action.

[0121] The switchover execution and state synchronization unit is responsible for converting the control instruction generated by the decision into actual network operation.

[0122] When receiving the switchover control instruction, the unit first coordinates the relevant switches in the dual-redundancy network. The coordination process is performed through the pre-established encrypted control channel between the switches.

[0123] The unit sends a port control command to the switch to be switched, which commands the port of the standby link to switch from the blocking state to the forwarding state and simultaneously commands the port of the original main link to switch from the forwarding state to the blocking state.

[0124] Subsequently, the unit triggers a route protocol convergence acceleration mechanism, such as immediately sending a route update packet and refreshing the media access control address table of the switch.

[0125] To ensure the transmission sequence consistency of the upper-layer control signals, in particular, the continuous control instruction frames based on the transmission control protocol, the unit enables a packet buffer queue before and after the execution of the switchover action.

[0126] 10 milliseconds before the start of the switchover, the system starts to temporarily store the transmission control protocol packets sent to the link in the queue. The switchover action is usually completed within 5 milliseconds.

[0127] After the completion of the switchover, the system sends the packets in the buffer queue from the new main link in the original sequence.

[0128] The buffer queue has a size limit, and when the queue is full, the oldest non-critical packet is discarded to ensure the timely transmission of new packets.

[0129] After the handover is completed, the handover execution and state synchronization unit encapsulates the new network topology state, including the current active master link identification, the state of each port, and detailed information such as the cause, time, involved failure mode, and confidence score of the current handover event, into a management information base trap message, and synchronizes it to the network management system and log server through the simple network management protocol or a custom management channel based on the transmission control protocol.

[0130] The offline training and model updating module is the key to the self-evolution of the system.

[0131] Please refer to the attached Figure 5 The module periodically collects operation data, usually with a 24-hour cycle.

[0132] The collected data includes two parts: First, the feature vectors generated by the fault perception and feature extraction module and their corresponding time stamps; Second, the final fault determination results recorded by the adaptive decision and handover control module, i.e., the actual actions and reasons executed by the system. These two parts of data are aligned through the association of time stamps to form a complete labeled training sample, with the label being the fault type determined by the system.

[0133] These new samples are stored in the local sample database.

[0134] The module is equipped with a model performance monitor.

[0135] The monitor periodically, for example, once a week, uses a subset of the recently collected sample data as a validation set in the background to evaluate the accuracy of the currently running fault mode classification model. Evaluation indicators include overall accuracy and recall rate for each type of fault.

[0136] When the monitor detects that the model's recognition accuracy for a certain type of new interference mode, such as a newly emerging periodic pulse interference, has fallen below the preset threshold of 70, or when the number of new samples accumulated is greater than 10,000, the module automatically triggers the model retraining process.

[0137] The retraining process is carried out in a separate sandbox environment.

[0138] The process first extracts all historical accumulated sample data from the sample database.

[0139] The training algorithm uses an incremental learning algorithm, specifically a variant of the online gradient boosting decision tree algorithm.

[0140] The algorithm takes the new sample data as the main learning object based on the original model parameters, while retaining some historical samples for playback to prevent catastrophic forgetting.

[0141] The training process continues for iterations until the performance of the model on an independent test set no longer improves or reaches a maximum number of iteration rounds.

[0142] After the training is completed, an updated fault pattern classification model file is generated. The new model file is first verified in a simulation test environment, and after passing the verification, it is seamlessly updated to the fault pattern recognition and confidence assessment module through hot deployment technology.

[0143] The hot deployment process involves atomic file replacement and reloading of in-memory model structures, ensuring that during the update period, the system can still use the old model for decision-making. After the new model is fully loaded and successfully initialized, traffic is smoothly switched to the new model, and the entire process is imperceptible to the business.

[0144] The operation of the entire system relies on precise timing synchronization. The high-precision timer inside the industrial switch provides a unified time reference for each module.

[0145] The sampling of the physical layer probe unit, the time window sliding of the feature engineering unit, the start and stop of the observation timer, and the time window of the message cache are all based on this clock source, ensuring the timing consistency of the entire data processing and decision-making chain, and avoiding analysis errors caused by time deviation.

[0146] The system also has perfect exception handling and degradation strategies.

[0147] For example, when the fault pattern recognition and confidence assessment module responds to a timeout, the adaptive decision and switching control module will automatically degrade to a backup decision logic based on simple threshold discrimination. This logic directly uses the abnormal point statistics value and signal strength average in the feature vector for quick judgment, although the accuracy is reduced, but it ensures the availability of the system at the most basic level.

[0148] All key state changes, decision logs, and exception events of all modules are recorded in the non-volatile log storage of the switch for post-analysis and auditing.

[0149] Embodiment 2: This embodiment provides another specific implementation of the industrial switch dual-redundant network fault automatic recovery system, the core of which is that the fault pattern recognition and confidence assessment module uses a different model architecture and inference mechanism, and the dynamic threshold calculation in the adaptive decision and switching control module introduces historical fault frequency as a new context factor.

[0150] In embodiment 2, the fault pattern classification model built-in the fault pattern recognition and confidence assessment module uses a hybrid architecture that combines one-dimensional convolutional neural networks and long short-term memory networks.

[0151] The one-dimensional convolutional neural network layer is used to automatically extract spatial features of the feature vector in the local time dimension, with a kernel size of 3, a step size of 1, and a kernel number of 32.

[0152] The long short-term memory network layer is used to capture the time sequence dependence of the features in the continuous time window sequence, with a hidden unit number of 64.

[0153] The input of the model is no longer a single time window feature vector, but a sequence of feature vectors of 10 consecutive time windows, forming a 10x6 matrix.

[0154] This structure enables the model to perceive the dynamic evolution process of the failure mode, such as distinguishing between continuous hardware aging and sudden transient interference that quickly recovers.

[0155] The training of the model is also based on massive historical data, but the data samples are organized in sequence form. During inference, the module maintains a first-in-first-out feature vector sequence buffer.

[0156] After each new feature vector is generated, it is pushed into the buffer, and the oldest vector is popped out, always maintaining a sequence length of 10 vectors.

[0157] This sequence is sent to the hybrid model, and the model outputs the failure mode probability distribution for the current time, i.e., the last time window of the sequence.

[0158] The processing method of the confidence assessment unit is the same as that of Embodiment 1, and the maximum probability value is selected for mapping to obtain the confidence score.

[0159] This sequential recognition method has higher accuracy for recognizing failures that are not obvious in a single time window but gradually clear over time, such as slow port corrosion or gradually increasing background noise interference.

[0160] The network context awareness unit of the adaptive decision and switching control module in Embodiment 2 is enhanced in function.

[0161] In addition to real-time monitoring of network load rate and number of key service flows, the unit also adds a historical failure frequency statistic.

[0162] The statistic maintains a sliding time window, such as the past 1 hour, and counts the number of occurrences of each type of failure mode identified for each link.

[0163] In particular, for intermittent electromagnetic interference, the frequency of occurrence is counted. The network context awareness unit takes the historical failure frequency as a new input factor.

[0164] The calculation formula of the dynamic confidence threshold is extended based on Embodiment 1: ; 、 、 、 、 、 The meaning of the same as example 1. Represents the historical failure frequency coefficient, set to 0.5. Represents the number of failure patterns identified on the current link within the last 1 hour, which is the same type as the current suspected failure.

[0165] For example, if the current identification result is intermittent electromagnetic interference, The value is the number of times the link is identified as intermittent electromagnetic interference within the past 1 hour.

[0166] This extension enables the decision logic to learn from historical experience.

[0167] If a link frequently reports the same type of intermittent interference within a short period of time, even if the confidence score of a single identification is not extremely high, the high frequency will raise the overall level of the dynamic threshold .

[0168] However, in the decision logic, for high-frequency intermittent interference, the system adopts a reverse strategy: Instead of switching immediately, a longer observation period, such as 2 seconds, is started, and more in-depth feature analysis is performed.

[0169] Because high-frequency transient interference may indicate potential problems such as loose physical connectors or the persistence of external interference sources, more careful observation is required.

[0170] On the contrary, if a link has a very clean history record, the first occurrence of a high-confidence persistent physical interruption identification, the system may trigger a preventive switching alarm when the confidence score is slightly less than 95, for example, 93, because it is a serious deviation from the historical baseline behavior of the link.

[0171] The switching execution and state synchronization unit of example 2 optimizes the packet caching strategy. For transmission control protocol packets, it not only caches data packets, but also deeply analyzes the transmission control protocol sequence number.

[0172] When retransmitting cached packets after switching is completed, the unit checks the maximum acknowledgment sequence number of successfully transmitted packets on the new link, and adjusts the starting sequence number of retransmission packets in the cache queue accordingly, avoiding repeated sending of packets that have been acknowledged by the opposite end, further optimizing retransmission efficiency and reducing redundant traffic.

[0173] Furthermore, for key traffic flows of the user datagram protocol, embodiment 2 introduces a temporary application of forward error correction coding.

[0174] When a handover is determined to be likely and the network context awareness unit identifies a key user datagram protocol flow, the handover execution unit inserts forward error correction redundant packets for this flow for a short window before and after the handover.

[0175] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting; it is not intended to exclude myriad other embodiments of the present application that other inventors can develop based on the same general inventive concepts embodied by the described embodiments. That is, although the present application is described in terms of particular embodiments and implementations, it is to be understood that the terminology used is for the purpose of descriptive clarity and that it is intended to be broadly construed to encompass all embodiments of the present application, with the scope of the present application to be defined by the appended claims and equivalents thereof.

[0176] While the embodiments of the application have been illustrated and described, it will be understood by those skilled in the art that various changes, modifications, substitutions, and alterations can be made therein without departing from the spirit and scope of the application, which is defined by the appended claims and their equivalents.

Claims

1. An automatic fault recovery system for dual-redundant network of industrial switches, characterized in that, Deployed in at least two industrial switches forming a dual-redundant network architecture, the system includes: The fault perception and feature extraction module is used to monitor the primary and backup links in a dual-redundant network in real time, collect the original state parameters of the link layer, and extract multi-dimensional features from the collected original state parameters to generate feature vectors that characterize the health of the links. The fault mode recognition and confidence assessment module is used to receive the feature vector generated by the fault perception and feature extraction module, perform pattern recognition on the current link state based on the preset fault mode classification model, output the identified fault mode category, and calculate the confidence score of the recognition result. The adaptive decision-making and handover control module receives the fault mode category and corresponding confidence score output by the fault mode recognition and confidence assessment module, combines the preset handover strategy library with the current network operating context, executes the handover decision logic, and generates the final link control command to decide whether to maintain the current link, initiate handover to the backup link, or perform link recovery operation.

2. The industrial switch dual-redundant network fault automatic recovery system according to claim 1, characterized in that, The fault perception and feature extraction module includes a physical layer probe unit and a feature engineering unit; The physical layer probe unit is used to synchronously collect physical layer signal quality indicators of the main link and the backup link with a sampling period of more than 1 millisecond. The physical layer signal quality indicators include received signal strength indication, carrier sense signal quality, bit error rate, and the presence of link pulse signal. The feature engineering unit is used to process the time-series data collected by the physical layer probe unit, and the extracted feature dimensions include: Within the time window, count the number of outliers where the signal quality index is less than the first preset threshold. Calculate the moving average and standard deviation of the signal quality index within this time window; Detect the complete transition event of the link status from normal to abnormal and back to normal, and record the duration of the event; The above statistical values, calculated values, and event duration are combined to form a multi-dimensional feature vector.

3. The industrial switch dual-redundant network fault automatic recovery system according to claim 2, characterized in that, The fault mode recognition and confidence assessment module has a built-in trained fault mode classification model. The fault mode classification model is used to receive feature vectors input in real time and output probability distributions corresponding to each fault mode category. The fault mode categories include at least continuous physical interruption, intermittent electromagnetic interference, port hardware aging and decay, and normal jitter without faults. The fault mode recognition and confidence assessment module also includes a confidence assessment unit, which selects the category corresponding to the highest probability value in the probability distribution as the recognition result, and normalizes the highest probability value to map it to an integer between 0 and 100 as the confidence score of this recognition.

4. The industrial switch dual-redundant network fault automatic recovery system according to claim 3, characterized in that, The adaptive decision-making and switching control module maintains a switching strategy library containing multiple decision rules; The decision-making logic of the adaptive decision-making and switching control module is based on a hierarchical judgment according to the combination of fault mode category and confidence score; If the identification result is a persistent physical interruption and the confidence score is greater than 95, then a control command to switch to the backup link is generated immediately. If the identification result is intermittent electromagnetic interference, then the delayed decision mechanism is activated.

5. The industrial switch dual-redundant network fault automatic recovery system according to claim 4, characterized in that, The delayed decision mechanism includes: The confidence score of the current identification result is compared with the dynamically adjusted confidence threshold, and a configurable observation timer is started at the same time. Continuously receive new recognition results and confidence scores before the observation timer expires; If, during the observation period, the identification result remains intermittent electromagnetic interference and the average confidence score is less than the dynamic confidence threshold, it is determined to be transient interference, and a command to maintain the current link is generated. If, during the observation period, the identification result changes to a persistent physical interruption or the mean confidence score remains greater than the dynamic confidence threshold, it is determined to be a potential hard fault risk, and a preventive handover instruction is generated.

6. The industrial switch dual-redundant network fault automatic recovery system according to claim 5, characterized in that, The dynamic confidence threshold is dynamically calculated by the network context awareness unit based on the current network load rate and the priority of critical service flows; The network context awareness unit is used to monitor the service traffic passing through the current link in real time, calculate the percentage of bandwidth it occupies as the network load rate, and identify the data flow that belongs to the preset key service list. The formula for calculating the dynamic confidence threshold is as follows: ； This represents the calculated dynamic confidence threshold. Represents the basic threshold. Represents the network load factor. Represents real-time network load rate. Represents the key business coefficient. This represents the number of currently active key business flows.

7. The industrial switch dual-redundant network fault automatic recovery system according to claim 6, characterized in that, The adaptive decision-making and switching control module also includes a switching execution and state synchronization unit; The switching execution and status synchronization unit is used to coordinate the relevant switches in the dual-redundant network to perform port start / stop, routing table update and address resolution table refresh operations after generating the switching control command. The switching execution and state synchronization unit is also used to cache and rearrange the control messages traversing the network within a short time window before and after the switching action is executed, so as to ensure the consistency of the transmission order of the upper-layer control signals. The switching execution and status synchronization unit is also used to synchronize the new network topology status and switching event log to the network management system and log server after the switching is completed.

8. The industrial switch dual-redundant network fault automatic recovery system according to claim 7, characterized in that, It also includes offline training and model update modules; The offline training and model update module is used to periodically collect feature vectors generated by the fault perception and feature extraction module and the final fault judgment results recorded by the adaptive decision-making and switching control module to form new training samples. The offline training and model update module is also used to automatically trigger the model retraining process when a certain number of new samples are accumulated, or when the system detects that the recognition accuracy of the existing fault mode classification model for new interference modes has dropped below a preset threshold. The retraining process uses an incremental learning algorithm to train the model based on the original model, incorporating new samples to generate an updated version of the fault mode classification model, which is then updated to the fault mode recognition and confidence assessment module via hot deployment.

9. The industrial switch dual-redundant network fault automatic recovery system according to claim 1, characterized in that, The system is implemented as embedded software and runs on the network processor and coprocessor of the industrial switch. The fault perception and feature extraction module is deployed on the data plane and the physical layer data is directly processed by the coprocessor. The fault mode identification and confidence assessment module and the adaptive decision-making and switching control module are deployed on the control plane and executed by the main network processor. The modules interact with each other at high speed through shared memory and message queue mechanisms.

10. The industrial switch dual-redundant network fault automatic recovery system according to claim 9, characterized in that, During operation, the fault mode recognition and confidence assessment module maintains a first-in-first-out feature vector sequence buffer. The fault mode classification model is a hybrid architecture model, and its input is a sequence of feature vectors from multiple consecutive time windows in the buffer. The hybrid architecture model combines a one-dimensional convolutional neural network and a long short-term memory network to extract the spatial features of feature vectors and capture their temporal dependencies in a continuous time window sequence, so as to output the probability distribution of the failure mode at the current moment.

Images