An adaptive fault diagnosis method, device, equipment and medium of a circuit breaker

Abstract

The application provides an adaptive fault diagnosis method, device and equipment of a circuit breaker and a medium, wherein the method comprises: in response to a diagnosis instruction, loading multi-source signals to be diagnosed; calling a preset first model and a second model respectively to perform parallel diagnosis on the multi-source signals; in the process of parallel diagnosis, real-time monitoring of the current available computing power is performed, and if the available computing power is lower than a first threshold, hierarchical computing power scheduling is performed; when the available computing power is greater than or equal to a second threshold, the computing load of at least one model is reduced to maintain double-model parallel diagnosis; when the available computing power is less than the second threshold, all computing power resources occupied by one of the models are released, and the other model is driven to complete the remaining diagnosis; and based on the diagnosis results of the first model and / or the second model, the fault type of the circuit breaker is determined. Through the above method, the application embodiments help to solve the technical problem that the diagnosis accuracy of some fault types is low.

Description

Technical Field

[0001] This application relates to the field of high-voltage circuit breaker fault diagnosis technology, and more specifically, to an adaptive fault diagnosis method, apparatus, equipment, and medium for circuit breakers. Background Technology

[0002] High-voltage circuit breakers are core equipment in power systems, responsible for control and protection functions. The reliability of their operating mechanisms and electrical control circuits directly affects the safe operation of the power grid. Online fault diagnosis is an important means of timely detecting potential circuit breaker problems and preventing sudden power outages.

[0003] Currently, fault diagnosis of high-voltage circuit breakers mainly relies on single-type monitoring signals. However, the information coverage of a single signal source is limited, and relying solely on a single signal is insufficient to comprehensively cover the various common fault types of circuit breakers, resulting in low diagnostic accuracy for some faults. Summary of the Invention

[0004] In view of this, embodiments of this application provide an adaptive fault diagnosis method, apparatus, device, and medium for circuit breakers to solve the technical problem that the accuracy of diagnosis of some fault types is low due to incomplete coverage of information from a single monitoring signal.

[0005] In a first aspect, embodiments of this application provide an adaptive fault diagnosis method for a circuit breaker, the method comprising: In response to a diagnostic command, a multi-source signal to be diagnosed is loaded; the multi-source signal includes at least the current signal of the opening and closing coils and the dual-point vibration signal distributed on the closing coil side and the operating mechanism side; The multi-source signals are diagnosed in parallel by calling the preset first and second models respectively; During parallel diagnostics, the available computing power is monitored in real time. If the available computing power is lower than the first threshold, then the following tiered computing power scheduling is performed: When the available computing power is greater than or equal to the second threshold, the computational load of at least one model is reduced to maintain parallel diagnosis of the two models; and when the available computing power is less than the second threshold, all computing power resources occupied by one model are released, and the other model is driven to complete the remaining diagnosis based on all computing power resources; the second threshold is less than the first threshold; Based on the diagnostic results of the first model and / or the second model, the fault type of the circuit breaker is determined.

[0006] In a feasible implementation, the first model is a random forest model based on structured physical features, and the second model is a convolutional gated recurrent unit fusion network based on the original time series signal. The computational power consumption of the first model per unit time is lower than that of the second model per unit time. The parallel diagnosis of the multi-source signals by respectively calling the preset first model and the second model includes: The first model is used as the resident diagnostic model to continuously diagnose the multi-source signals; the second model is used as the auxiliary diagnostic model and is activated according to a preset period or in response to the current available computing power being greater than or equal to the first threshold, so as to achieve parallel diagnosis. When the hierarchical computing power scheduling is executed, the following is performed first: release all computing power resources occupied by the second model, and drive the first model to complete the remaining diagnosis based on all computing power resources.

[0007] In one feasible implementation, a preset first model and a second model are invoked respectively to perform parallel diagnosis of the multi-source signals, including: The multi-source signals are preprocessed and feature extracted to obtain a structured feature set, which is then input into the first model to obtain the diagnostic results of the first model. The current signal and each vibration signal from the multi-source signals are treated as independent single-source signals and input into the second model respectively to obtain the diagnostic results of the second model.

[0008] In one feasible implementation, the multi-source signals are processed to obtain a structured feature set, including: The current signal is denoised by sliding window smoothing filter and ensemble empirical mode decomposition, and the extreme points of the denoised current waveform and their corresponding time parameters are extracted to construct a coil current signal feature set. For each vibration signal, the short-time energy method is used to extract time-domain features, and the wavelet packet decomposition method is used to extract time-frequency domain frequency band energy features to construct a vibration signal feature set. The structured feature set is obtained by merging the coil current signal feature set with the multi-channel vibration signal feature set and then standardizing the result.

[0009] In one feasible implementation, the second model includes a data-level fusion branch and a feature-level fusion branch; The data-level fusion branch is used to concatenate the current signal and each of the vibration signals in the channel dimension into a multi-channel signal, and to perform joint convolution on the multi-channel signal through a one-dimensional convolutional neural network to achieve multi-source information fusion at the original data level. The feature-level fusion branch is used to take the current signal and each of the vibration signals as independent inputs, extract deep features through their respective one-dimensional convolutional neural networks, and then splice and fuse them at the feature level.

[0010] In one feasible implementation, determining the fault type of the circuit breaker based on the diagnostic results of the first model and / or the second model includes: If the diagnostic results of the first model and the second model are obtained respectively, and the fault type diagnostic results of the two diagnostic results are consistent, then the average of the confidence scores output by the first model and the second model is taken as the final confidence score, and the fault type is output. If the fault types of the two diagnostic results are inconsistent, the diagnostic result with higher confidence is selected as the final output and sent to the management terminal for review, and the review result of the management terminal is obtained. If only the diagnostic result of the first model is obtained, the diagnostic result of the first model is directly used as the final output, and the product of the confidence level of the first model and the preset compensation coefficient is used as the final confidence level, wherein the compensation coefficient is greater than 1.

[0011] In one feasible implementation, the method further includes: The circuit breaker's opening and closing coil current signals and the dual-point vibration signals are collected under normal and various fault conditions to obtain the original sample set. The various fault conditions include: poor coil circuit contact, abnormal coil control voltage, stuck opening electromagnet core, abnormal gap of opening electromagnet core, mechanism failure to move, fatigue of opening and closing springs, loose anchor bolts, and stuck transmission mechanism. A generative adversarial network with a gradient penalty term is used to augment the original sample set, generating simulated samples that meet the preset consistency requirements with the real sample distribution, thus constructing a simulated sample set. The first model and the second model are trained using the original sample set and the simulation sample set, respectively. During the training process, multiple model structure combinations and multiple hyperparameter configurations are used for training. The model structure with the highest diagnostic accuracy and the corresponding hyperparameters are selected through cross-validation as the final configuration of the first model and the second model.

[0012] Secondly, embodiments of this application also provide an adaptive fault diagnosis device for a circuit breaker, the device comprising: The response module is used to load the multi-source signals to be diagnosed in response to the diagnostic command; the multi-source signals include at least the current signal of the opening and closing coils and the dual-point vibration signal distributed on the closing coil side and the operating mechanism side; The diagnostic module is used to call the preset first model and second model respectively to perform parallel diagnosis of the multi-source signal; The monitoring module is used to monitor the current available computing power in real time during the parallel diagnostic process; The scheduling module is used to perform the following tiered computing power scheduling if the available computing power is lower than a first threshold: When the available computing power is greater than or equal to the second threshold, the computational load of at least one model is reduced to maintain parallel diagnosis of the two models; and when the available computing power is less than the second threshold, all computing power resources occupied by one model are released, and the other model is driven to complete the remaining diagnosis based on all computing power resources; the second threshold is less than the first threshold; The fault determination module is used to determine the fault type of the circuit breaker based on the diagnostic results of the first model and / or the second model.

[0013] In a feasible implementation, the first model is a random forest model based on structured physical features, and the second model is a convolutional gated recurrent unit fusion network based on the original time series signal. The computational power consumption of the first model per unit time is lower than that of the second model per unit time. The diagnostic module is used to perform parallel diagnostics on the multi-source signals by respectively calling a preset first model and a second model, including: The first model is used as the resident diagnostic model to continuously diagnose the multi-source signals; the second model is used as the auxiliary diagnostic model and is activated according to a preset period or in response to the current available computing power being greater than or equal to the first threshold, so as to achieve parallel diagnosis. When the hierarchical computing power scheduling is executed, the following is performed first: release all computing power resources occupied by the second model, and drive the first model to complete the remaining diagnosis based on all computing power resources.

[0014] In one feasible implementation, the diagnostic module is used to perform parallel diagnostics on the multi-source signals by respectively calling a preset first model and a second model, including: The multi-source signals are preprocessed and feature extracted to obtain a structured feature set, which is then input into the first model to obtain the diagnostic results of the first model. The current signal and each vibration signal from the multi-source signals are treated as independent single-source signals and input into the second model respectively to obtain the diagnostic results of the second model.

[0015] In one feasible implementation, the diagnostic module is used to process the multi-source signals to obtain a structured feature set, including: The current signal is denoised by sliding window smoothing filter and ensemble empirical mode decomposition, and the extreme points of the denoised current waveform and their corresponding time parameters are extracted to construct a coil current signal feature set. For each vibration signal, the short-time energy method is used to extract time-domain features, and the wavelet packet decomposition method is used to extract time-frequency domain frequency band energy features to construct a vibration signal feature set. The structured feature set is obtained by merging the coil current signal feature set with the multi-channel vibration signal feature set and then standardizing the result.

[0016] In one feasible implementation, the second model includes a data-level fusion branch and a feature-level fusion branch; The data-level fusion branch is used to concatenate the current signal and each of the vibration signals in the channel dimension into a multi-channel signal, and to perform joint convolution on the multi-channel signal through a one-dimensional convolutional neural network to achieve multi-source information fusion at the original data level. The feature-level fusion branch is used to take the current signal and each of the vibration signals as independent inputs, extract deep features through their respective one-dimensional convolutional neural networks, and then splice and fuse them at the feature level.

[0017] In one feasible implementation, the fault determination module is used to determine the fault type of the circuit breaker based on the diagnostic results of the first model and / or the second model, including: If the diagnostic results of the first model and the second model are obtained respectively, and the fault type diagnostic results of the two diagnostic results are consistent, then the average of the confidence scores output by the first model and the second model is taken as the final confidence score, and the fault type is output. If the fault types of the two diagnostic results are inconsistent, the diagnostic result with higher confidence is selected as the final output and sent to the management terminal for review, and the review result of the management terminal is obtained. If only the diagnostic result of the first model is obtained, the diagnostic result of the first model is directly used as the final output, and the product of the confidence level of the first model and the preset compensation coefficient is used as the final confidence level, wherein the compensation coefficient is greater than 1.

[0018] In one feasible implementation, the device further includes: The acquisition module is used to acquire the opening and closing coil current signals and the dual-point vibration signals of the circuit breaker under normal conditions and various fault conditions to obtain the original sample set. The various fault conditions include: poor contact of the coil circuit, abnormal coil control voltage, jamming of the opening electromagnet core, abnormal gap of the opening electromagnet core, failure of the mechanism to move, fatigue of the opening and closing springs, loose anchor bolts, and jamming of the transmission mechanism. The enhancement module is used to perform data augmentation on the original sample set using a generative adversarial network with a gradient penalty term, generate simulation samples that meet the preset consistency requirements with the real sample distribution, and construct a simulation sample set. The training module is used to train the first model and the second model using the original sample set and the simulation sample set, respectively. During the training process, the model is trained with multiple model structure combinations and multiple hyperparameter configurations. The model structure and corresponding hyperparameters with the highest diagnostic accuracy are selected through cross-validation as the final configuration of the first model and the second model.

[0019] Thirdly, embodiments of this application also provide an electronic device, including: a processor, a storage medium, and a bus, wherein the storage medium stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor communicates with the storage medium via the bus, and the processor executes the machine-readable instructions to perform the steps of the method as described in any one of the first aspects.

[0020] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the method as described in any one of the first aspects.

[0021] The embodiments of this application provide an adaptive fault diagnosis method, device, equipment, and medium for circuit breakers. By collecting multi-source signals, including at least the opening and closing coil current signal and the dual-point vibration signal, and calling the first model and the second model respectively for parallel diagnosis, it overcomes the problem that the accuracy of diagnosis of some fault types is low due to incomplete information coverage of a single signal source.

[0022] Furthermore, by monitoring available computing power in real time during parallel diagnosis and implementing tiered computing power scheduling when available computing power falls below a first threshold, dynamic management of diagnostic resources is achieved. Specifically, when available computing power drops to between the first and second thresholds, frequency reduction sampling and feature compression reduce the computational load while maintaining parallel diagnosis of the two models, avoiding inference interruptions caused by computing power fluctuations. When computing power further drops below the second threshold, one model is released and the other model is driven to continue completing the diagnosis, ensuring that the diagnostic task can still be completed normally under extreme computing power conditions.

[0023] Compared with existing diagnostic schemes that require fixed computing power, the embodiments of this application can dynamically adjust the diagnostic strategy based on the real-time computing power status of the device without increasing hardware costs, thereby maximizing diagnostic accuracy while ensuring real-time diagnostic performance.

[0024] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0025] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 A flowchart of an adaptive fault diagnosis method for a circuit breaker provided in an embodiment of this application is shown.

[0027] Figure 2 A schematic diagram of the structure of a generative adversarial network provided in an embodiment of this application is shown.

[0028] Figure 3 A schematic diagram of the network structure of the discriminator provided in an embodiment of this application is shown.

[0029] Figure 4 A schematic diagram of the network structure of the generator provided in an embodiment of this application is shown.

[0030] Figure 5 The opening and closing current waveforms and short-time vibration energy diagrams provided in the embodiments of this application are shown.

[0031] Figure 6 The diagram shows the frequency band energy distribution of two vibration signals under normal and different fault states provided in the embodiments of this application.

[0032] Figure 7 A schematic diagram of the structure of an electronic device provided in an embodiment of this application is shown. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0034] High-voltage circuit breakers are core equipment in power systems, responsible for control and protection functions. The reliability of their operating mechanisms and electrical control circuits directly affects the safe operation of the power grid. Online fault diagnosis is an important means of timely detecting potential circuit breaker problems and preventing sudden power outages.

[0035] Currently, fault diagnosis of high-voltage circuit breakers mainly relies on single-type monitoring signals. However, the information coverage of a single signal source is limited, and relying solely on a single signal is insufficient to comprehensively cover the various common fault types of circuit breakers, resulting in low diagnostic accuracy for some faults.

[0036] Based on this, embodiments of this application provide an adaptive fault diagnosis method, apparatus, device, and medium for circuit breakers, which are described below through embodiments.

[0037] To facilitate understanding of this embodiment, a detailed description of an adaptive fault diagnosis method for circuit breakers disclosed in this application embodiment will be provided first. For example... Figure 1 As shown, it includes the following steps: Step 101: In response to the diagnostic command, load the multi-source signal to be diagnosed; the multi-source signal includes at least the current signal of the opening and closing coils and the dual-point vibration signal distributed on the closing coil side and the operating mechanism side.

[0038] The diagnostic command can be triggered by the user, for example, when maintenance personnel need to test a circuit breaker, they can issue a diagnostic command through a host computer or mobile terminal; it can also be triggered automatically by the system according to a preset cycle, for example, the diagnostic process can be started automatically at a set time interval; it can also be triggered by an event, for example, the diagnostic command can be automatically triggered after the circuit breaker completes a switching action.

[0039] Upon receiving a diagnostic command, this embodiment loads the multi-source signals to be diagnosed. These multi-source signals include at least two types: one is the current signal of the opening and closing coils, and the other is a dual-point vibration signal. The opening and closing coil current signal refers to the waveform data of the current flowing through the electromagnet coil over time during the circuit breaker's opening or closing operation. This signal reflects the working state of the coil drive circuit and the electromagnet itself. The dual-point vibration signal refers to vibration signals collected at two different monitoring locations, one at the closing coil side and the other at the operating mechanism side. The vibration signal at the closing coil side mainly reflects the mechanical vibration generated by the electromagnet's attraction and impact; the vibration signal at the operating mechanism side mainly reflects the collision and movement information of components such as the transmission mechanism and contact system. By simultaneously collecting vibration signals at two locations, the mechanical vibration characteristics during the circuit breaker's operation can be captured from different spatial dimensions, providing a richer information basis for subsequent fault diagnosis.

[0040] In actual data acquisition, the current signal and vibration signal need to be synchronized in time to ensure accurate time correspondence between the different signals. Synchronous acquisition can be achieved through hardware triggering, such as using the same synchronous clock source to trigger all sensors to start acquiring data simultaneously; or through software calibration, such as synchronizing the signals by aligning timestamps after acquisition. Regardless of the method used, the final multi-source signal includes a time-aligned current signal and two vibration signals from the same action process.

[0041] It should be noted that the embodiments of this application do not limit the specific loading method of multi-source signals. In some scenarios, the signals can be collected by sensors in real time and directly transmitted to the diagnostic module; in other scenarios, the signals can be pre-stored in local or remote storage media and read from the storage media when diagnosis is needed.

[0042] Step 102: Call the preset first model and second model respectively to perform parallel diagnosis on the multi-source signal.

[0043] After loading the multi-source signal to be diagnosed, this application embodiment will call the preset first model and second model respectively to perform parallel diagnosis of the multi-source signal.

[0044] It should be noted that the first and second models are two independent diagnostic models with different structures, employing different algorithmic principles to process the input signals. In one possible implementation, the first model could be a traditional machine learning-based diagnostic model, taking structured features extracted from multi-source signals as input and outputting diagnostic conclusions through a classifier; the second model could be a deep learning-based diagnostic model, taking the raw time-series signal as input and automatically extracting features through a network to output diagnostic conclusions. However, this is only an example. In practical applications, the first and second models can also employ other types of algorithms; for example, the first model could use a rule-based expert system, and the second model could use an ensemble learning model.

[0045] Two models are deployed on the same edge computing device, sharing the device's CPU, memory, and other hardware computing resources. During parallel diagnostics, the two models run simultaneously, performing inference calculations independently, but sharing the limited computing power of the same device. Due to the differences in their structures and algorithms, the two models differ in their running speed and computing resource requirements. For example, models based on traditional machine learning typically require less computation, while models based on deep learning often require more matrix operations and memory bandwidth. Through parallel diagnostics, this embodiment of the application can simultaneously obtain the independent judgment results of the two models within a single diagnostic cycle, providing a more diverse information foundation for subsequent comprehensive judgment.

[0046] Taking one application scenario of this application embodiment as an example, the first model typically adopts a lightweight model with relatively low computational overhead and can run continuously; the second model typically adopts a high-precision model with relatively high computational overhead and participates in diagnosis as needed. The two models complement each other in the diagnosis process: the first model provides fast and stable baseline judgment, while the second model provides fine-grained verification judgment. This parallel design can balance diagnostic accuracy and computational efficiency.

[0047] The input data formats received by the two models can be the same or different. For example, the first model may require preprocessed and feature-extracted structured data as input, while the second model may directly receive the raw time-series signal waveform as input. In the embodiments of this application, before calling the two models, the multi-source signals can be processed according to their respective needs to ensure that each model can obtain an input format suitable for its algorithm characteristics.

[0048] Step 103: During the parallel diagnostic process, monitor the current available computing power in real time.

[0049] The available computing power monitored here refers to the remaining usable computing resources of the device performing the diagnostic task at the current moment, including but not limited to CPU utilization, memory usage, and GPU utilization. In actual operation, in addition to running diagnostic tasks, edge computing devices may also undertake other tasks such as data acquisition, communication transmission, and log recording. These tasks dynamically consume the device's computing power, causing the available computing resources for diagnosis to constantly change. Therefore, this embodiment of the application adopts a real-time monitoring method to continuously acquire the current status of available computing power during the diagnostic process.

[0050] In one possible implementation, available computing power can be monitored through periodic sampling. For example, the CPU idle rate, remaining memory, and GPU utilization of the current device can be obtained at set time intervals. The sampling period can be set according to actual needs; a shorter period results in higher monitoring accuracy, but also increases the consumption of device performance. In another possible implementation, the currently available computing resources can be queried through the operating system interface before each invocation of the model for inference to determine the subsequent scheduling strategy. The monitored available computing power is used for scheduling judgments in subsequent steps to determine whether computing power scheduling measures need to be implemented and, if so, to what extent.

[0051] For example, CPU idle rate is used as a quantitative indicator of available computing power. A higher idle rate indicates more abundant remaining computing power, while a lower idle rate indicates more strained computing power. Therefore, the first threshold can be set to a higher idle rate value, and the second threshold can be set to a lower idle rate value. For example, using CPU idle rate as a measure of available computing power, the first threshold can be set to 30%, and the second threshold can be set to 10%. Subsequent step 104 is then executed based on the relationship between the current CPU idle rate and the first and second thresholds. It can be seen that the second threshold of 10% is less than the first threshold of 30%, and the computing power strain corresponding to the second threshold is higher. Therefore, it's not just about reducing computing power for both models, but about releasing all computing power resources occupied by one model and driving the other model to complete the remaining diagnostics based on all available computing power resources. It can be seen that the second threshold is only activated when the computing power strain is higher.

[0052] In the specific judgment process, if the current CPU idle rate is 25%, which is lower than the first threshold of 30% but higher than the second threshold of 10%, the system determines that the computing power is slightly insufficient and executes scheduling measures to reduce the computing load of the model; if the current CPU idle rate is only 5%, which is lower than both the first threshold of 30% and the second threshold of 10%, the system determines that the computing power is severely insufficient, and executes the release of all computing power resources occupied by one model to drive the other model to complete the remaining diagnosis.

[0053] The aforementioned first threshold of 30% and second threshold of 10% are applicable to the computing power scheduling needs in general industrial scenarios. In some application scenarios with high requirements for real-time diagnostics, such as when a circuit breaker needs to complete fault determination and trigger protection operations within a very short action cycle, the system has a lower tolerance for computing power fluctuations. Therefore, it is necessary to increase the first and second thresholds for triggering scheduling accordingly to ensure that scheduling can be intervened immediately in the event of any slight drop in computing power, thereby avoiding diagnostic timeouts. In this case, the first threshold can be set to 85%, which means that hierarchical computing power scheduling is triggered when the CPU idle rate is lower than 85%, and the second threshold can be set to 45%, which means that the computing power of one of the models is released when the CPU idle rate is lower than 45%. Through this more sensitive threshold configuration, the system can execute scheduling measures in advance when the computing power is still at a relatively high level, thereby maximizing the guarantee of timely completion of diagnostic tasks in high real-time scenarios. At the same time, the second threshold of 45% is still lower than the first threshold of 85%, and the large interval between the two thresholds also provides sufficient buffer space for hierarchical scheduling. Both threshold configuration schemes mentioned above can be flexibly selected according to the real-time requirements of specific application scenarios, and the threshold values ​​can be calibrated by running standard test samples on the platform in advance and combining them with the deadline of the diagnostic task.

[0054] In another possible implementation, in addition to monitoring computing resource utilization, the diagnostic progress of each model can also be monitored. For example, diagnostic progress can be represented by time, recording the elapsed time from the start of inference to the current moment for each model, or estimating the time required for the model to complete the diagnosis. This information can indirectly reflect whether the current computing power is sufficient: when the actual diagnostic time of a model significantly exceeds its normal level, it indicates that the diagnostic progress is lagging, and even if the CPU utilization has not yet reached the preset threshold, it may indicate that the device is currently under heavy load.

[0055] Alternatively, the diagnostic progress can be determined based on the stage at which the model is processing data. Taking a deep learning model processing time-series signals as an example, the input signals are typically fed into the network segment by segment in chronological order. After processing each data segment, the internal state advances once. In such scenarios, stage markers or intermediate state flags can be set between the various processing nodes of the model. By reading the current values ​​of these flags, the number of data segments processed by the model or its current processing stage can be determined, thus confirming the diagnostic progress. This approach helps determine whether the overall diagnostic can be completed on schedule, and even before the final results are available, it allows for early prediction of potential computational bottlenecks based on the current stage.

[0056] For models that complete inference in one go, the diagnostic process typically lacks a step-by-step progression, making it difficult to judge progress using stage markers. In such cases, a timing method can be used as an auxiliary reference. This involves recording the cumulative time taken by the model from startup to the current moment and comparing it with the model's average time in historical runs. When the cumulative time deviates from the normal range, it is used as a reference signal that computing power may be insufficient.

[0057] In practical applications, the aforementioned monitoring methods can be used individually or in combination, depending on the model type, to more comprehensively reflect the computing power status during the diagnostic process. For example, the available computing power can be dynamically adjusted based on the model's diagnostic progress. Specifically, when the diagnostic progress of a model lags behind expectations, it indicates that the current computing power may be insufficient to support two models running at normal speeds simultaneously. In this case, the evaluation value of available computing power can be lowered accordingly to trigger the scheduling mechanism earlier. Conversely, when the diagnostic progress of each model is faster than expected, it indicates that the current computing power is sufficient, and the evaluation value of available computing power can be appropriately increased to avoid prematurely triggering unnecessary scheduling. In this way, the diagnostic progress is no longer just a reference information but directly participates in the quantitative evaluation of available computing power, making scheduling decisions more closely aligned with the actual operational status of the diagnostic task.

[0058] Step 104: If the available computing power is lower than the first threshold, then perform the following tiered computing power scheduling: When the available computing power is greater than or equal to the second threshold, the computational load of at least one model is reduced to maintain parallel diagnosis of the two models; and when the available computing power is less than the second threshold, all computing power resources occupied by one model are released, and the other model is driven to complete the remaining diagnosis based on all computing power resources; the second threshold is less than the first threshold.

[0059] After real-time monitoring of available computing power during parallel diagnostics, if the current available computing power is lower than a first threshold, this embodiment of the application will execute hierarchical computing power scheduling. The first threshold here can be understood as a warning line preset by the system. When the available computing power is lower than this value, it means that the current computing power resources can no longer support the two models to run normally in the current state, which may lead to risks such as inference stuttering and timeouts. Therefore, the scheduling mechanism needs to be triggered.

[0060] The core of tiered computing power scheduling lies in adopting different response strategies based on different levels of computing power stress, rather than shutting down a certain model indiscriminately.

[0061] When the available computing power is below the first threshold but still greater than or equal to the second threshold, it is considered a case of mild computing power strain. In this scenario, this embodiment of the application maintains parallel diagnosis of two models by reducing the computational load of at least one model. There are several ways to reduce the computational load. For example, the input signal can be downsampled, i.e., the sampling rate of the signal involved in the calculation can be reduced, reducing the amount of data to be processed, thereby shortening the inference time. Another example is that the input features can be compressed, such as dimensionality reduction of structured features, retaining the main components while discarding the secondary components, reducing the dimension of the model's input layer. Yet another example is that for deep learning models, the inference computation can be reduced, such as skipping some redundant network layers, reducing the number of convolutional kernels, or pruning intermediate feature maps. Yet another example is that for ensemble learning models, the number of base learners involved in inference can be reduced, such as using only some decision trees for voting in a random forest, thereby reducing computational overhead. The above methods can be used individually or in combination according to the actual situation, ensuring the real-time nature of the diagnosis while maintaining diagnostic accuracy as much as possible.

[0062] When available computing power further decreases below the second threshold, it indicates a severe computing power shortage. In this scenario, simply reducing the computational load of the models is insufficient to solve the problem, because even if the computational load of each model is reduced, the simultaneous operation of two models still incurs unavoidable concurrency overhead. Therefore, this embodiment releases all computing resources occupied by one model, allocating these resources to the other model, which then completes the remaining diagnostics using all available computing resources. Since the computational overhead of a single model is lower than the overhead of running two models simultaneously, and with the release of all resources from the other model, the remaining model can obtain more abundant computing resources, thereby ensuring that the diagnostic task can be completed within a reasonable time. It should be noted that releasing model resources means unloading the model from memory, stopping its inference process, or placing it in a hibernation state, returning all CPU, memory, and other resources it occupies to the system for use by the other model.

[0063] The above two-level scheduling strategy constitutes a complete response system: when computing power is slightly strained, the computational load is reduced to "throttle" and maintain the parallel operation of the two models; when computing power is severely strained, the models are released to "make room" and ensure the efficient operation of a single model. This hierarchical design enables the embodiments of this application to make differentiated responses according to different computing power conditions, avoiding the use of the same processing method under any circumstances of insufficient computing power.

[0064] Step 105: Based on the diagnostic results of the first model and / or the second model, determine the fault type of the circuit breaker.

[0065] After completing the above diagnostic process, the embodiments of this application will determine the fault type of the circuit breaker based on the diagnostic results of the first model and / or the second model.

[0066] Specifically, there are several different scenarios for obtaining diagnostic results. In some scenarios, both models complete their respective diagnoses during the scheduling process; in this case, the diagnostic results output by each model can be obtained. In other scenarios, if one model is released during the scheduling process, it may not have completed its diagnosis at the time of release, and only the other model will output a valid diagnostic result. Another scenario is that the two models have different diagnostic progress, with one completing first while the other is still running; in this case, the diagnostic result of the completed model can be obtained first as a reference.

[0067] Regardless of the scenario described above, the embodiments of this application can determine the fault type of the circuit breaker based on the actual diagnostic results obtained. When only the diagnostic result of one model is obtained, that result can be directly output as the final fault type. When both models have completed their diagnoses, the judgments of the two models can be combined to make the final decision. This design ensures that the diagnostic process can output valid fault type conclusions under different scheduling results, and will not fail to provide results due to the release of a certain model.

[0068] For example, in one specific implementation, different processing logics can be adopted depending on the actual diagnostic results obtained.

[0069] At this time, determining the fault type of the circuit breaker based on the diagnostic results of the first model and / or the second model includes: Scenario 1: If the diagnostic results of the first model and the second model are obtained respectively, and the fault type diagnostic results of the two diagnostic results are consistent, then the average of the confidence scores output by the first model and the second model is taken as the final confidence score, and the fault type is output.

[0070] When both the first and second models have completed the diagnosis, and the fault types output by the two models are consistent, it indicates that the two models' judgments on the current fault type corroborate each other. In this case, the fault type can be directly output as the final diagnosis result. For the confidence level, the average of the confidence levels output by the two models can be used as the final confidence level. Since the two models have made judgments on the same fault type from different perspectives, combining their confidence levels can more objectively reflect the reliability of the diagnostic conclusion.

[0071] Scenario 2: If the fault types of the two diagnostic results are inconsistent, the diagnostic result with higher confidence is selected as the final output and sent to the management terminal for review, thereby obtaining the review result from the management terminal.

[0072] When both models complete the diagnosis but output inconsistent fault types, it indicates a discrepancy in their judgments of the current signal. In this case, the diagnosis result with higher confidence can be selected as the final output. For example, if the first model outputs fault type A with a confidence level of 90%, and the second model outputs fault type B with a confidence level of 75%, then fault type A should be selected as the final result. Furthermore, since the discrepancy between the two models suggests that the current diagnostic condition may be complex or subject to interference, the selected diagnosis result, along with the original diagnostic information from both models, can be sent to the management system for review and confirmation by operations personnel. The review result returned by the management system will serve as the final basis for determining the fault type.

[0073] Scenario 3: If only the diagnostic result of the first model is obtained, the diagnostic result of the first model is directly used as the final output, and the product of the confidence level of the first model and the preset compensation coefficient is used as the final confidence level, wherein the compensation coefficient is greater than 1.

[0074] When only the first model outputs a diagnostic result (e.g., the second model is released during scheduling and fails to complete the diagnosis), the diagnostic result of the first model can be used as the final output. However, due to the lack of verification information from the second model, the confidence level of the first model's output alone may have some uncertainty. In this case, the confidence level of the first model's output can be multiplied by a preset compensation coefficient to obtain the final confidence level. This compensation coefficient is greater than 1, and its purpose is to appropriately correct the confidence level of the first model to reflect that the confidence level may need to be moderately increased without the verification of the second model. The specific value of the compensation coefficient can be set according to the actual application scenario, for example, determined based on the difference between the accuracy rate of the first model's diagnosis alone and the accuracy rate of the joint diagnosis of the two models in historical data.

[0075] The above-mentioned processing logics cover various result acquisition scenarios that may occur during the dual-model diagnosis process, ensuring that the embodiments of this application can output valid diagnostic conclusions and corresponding confidence levels under any scheduling result.

[0076] In another possible implementation, when the second model is released during the hierarchical computing power scheduling process, it does not mean that the second model did not generate any valuable information in the previous diagnostic work. Before being released, the second model may have already processed some of the input data and generated corresponding intermediate feature vectors. Although these intermediate features are insufficient to support the second model in independently outputting a complete diagnostic result, they still contain some fault-related information and can be provided to the first model as auxiliary information.

[0077] Based on this, embodiments of this application also provide a method for transferring intermediate features from a second model to a first model. Specifically, for scenario three above, when the second model is released, the intermediate feature vectors generated during the completed processing phase can be extracted. These intermediate feature vectors reflect the second model's understanding of local information in the input signal and belong to a different feature space than the structured features used by the first model. To fuse the two, a pre-trained mapping network can be used to transform the intermediate feature vectors of the second model to the feature space of the first model, obtaining an equivalent feature vector. This equivalent feature vector is then concatenated and fused with the structured features currently extracted by the first model, enabling the first model to utilize the intermediate information from the second model as a supplement in subsequent diagnostic processes.

[0078] The training of the aforementioned mapping network can be completed offline. Specifically, using historical sample sets, the intermediate feature vectors output by the second model when processing each sample, and the structured feature vectors extracted by the first model when processing the same sample, are extracted. These two are used as training data pairs, and the mapping network is trained with the goal of minimizing the difference between the mapped equivalent features and the original structured features of the first model. Once trained, this mapping network can be used directly in the diagnostic process without online updates. In this way, even when the second model is released, the first model can still obtain some intermediate information from the second model through the mapping network, thereby compensating to some extent for the information loss caused by the lack of complete diagnostic results from the second model.

[0079] In this context, the compensation coefficient used to determine the final diagnostic result can be set based on the quantity or quality of the transferred intermediate feature vectors. For example, if the second model has processed a large number of data segments and generated rich intermediate features before being released, it indicates a greater amount of information transferred to the first model. In this case, the compensation coefficient can be relatively large because the first model has acquired more supplementary information, and its confidence level can be more fully corrected. Conversely, if the second model has only processed a small number of data segments and has fewer transferable intermediate features, the compensation coefficient should be reduced accordingly. In this way, the compensation coefficient is no longer a fixed preset value, but is dynamically determined based on the actual amount of information that can be transferred from the second model to the first model, making the final confidence level assessment more objective and accurate.

[0080] The embodiments of this application provide an adaptive fault diagnosis method, device, equipment, and medium for circuit breakers. By collecting multi-source signals, including at least the opening and closing coil current signal and the dual-point vibration signal, and calling the first model and the second model respectively for parallel diagnosis, it overcomes the problem that the accuracy of diagnosis of some fault types is low due to incomplete information coverage of a single signal source.

[0081] Furthermore, by monitoring available computing power in real time during parallel diagnosis and implementing tiered computing power scheduling when available computing power falls below a first threshold, dynamic management of diagnostic resources is achieved. Specifically, when available computing power drops to between the first and second thresholds, frequency reduction sampling and feature compression reduce the computational load while maintaining parallel diagnosis of the two models, avoiding inference interruptions caused by computing power fluctuations. When computing power further drops below the second threshold, one model is released and the other model is driven to continue completing the diagnosis, ensuring that the diagnostic task can still be completed normally under extreme computing power conditions.

[0082] Compared with existing diagnostic schemes that require fixed computing power, the embodiments of this application can dynamically adjust the diagnostic strategy based on the real-time computing power status of the device without increasing hardware costs, thereby maximizing diagnostic accuracy while ensuring real-time diagnostic performance.

[0083] In one feasible implementation, the first model is a random forest model based on structured physical features, and the second model is a convolutional gated recurrent unit fusion network based on the original time-series signal. The computational power consumption of the first model per unit time is lower than that of the second model per unit time.

[0084] The parallel diagnosis of the multi-source signals by respectively calling the preset first model and the second model includes: The first model is used as the resident diagnostic model to continuously diagnose the multi-source signals; the second model is used as the auxiliary diagnostic model and is activated according to a preset period or in response to the current available computing power being greater than or equal to the first threshold, so as to achieve parallel diagnosis.

[0085] When the hierarchical computing power scheduling is executed, the following is performed first: release all computing power resources occupied by the second model, and drive the first model to complete the remaining diagnosis based on all computing power resources.

[0086] In this embodiment, the first model can be a random forest model based on structured physical features, and the second model can be a convolutionally gated recurrent unit fusion network based on raw time-series signals. The computation process of the random forest model mainly involves traversing decision trees and statistical voting, with relatively small computational load and low computational power consumption per unit time. In contrast, the convolutionally gated recurrent unit fusion network involves a large number of convolution operations and recurrent neural network state updates, requiring more matrix multiplications and memory read / write operations, resulting in higher computational power consumption per unit time. This difference in computational power consumption between the two models provides a reasonable operational basis for hierarchical computational power scheduling. When it is necessary to reduce the computational load or release resources, prioritizing the second model with higher computational power consumption can more effectively alleviate computational pressure.

[0087] Based on the characteristics of the aforementioned models, a combination of resident and auxiliary models can be used when calling the first and second models for parallel diagnosis. The first model, as the resident diagnostic model, continuously diagnoses multi-source signals. Because the first model consumes low computing power, even long-term operation will not significantly burden the device's computing power, thus it can remain active at all times, ensuring a usable diagnostic output under any circumstances. The second model, as the auxiliary diagnostic model, does not need to be constantly running; instead, it starts at a preset period or in response to sufficient available computing power. Starting at a preset period means waking up the second model to perform diagnosis at set time intervals; starting in response to sufficient available computing power means that the second model is only activated to participate in parallel diagnosis when the detected available computing power exceeds a first threshold. These two activation conditions can be used individually or in combination. This combination of resident and auxiliary models ensures uninterrupted basic diagnostic capabilities while preventing high-computing-power-consuming models from continuously occupying system resources when resources are scarce.

[0088] When performing tiered computing power scheduling and needing to release model resources, priority is given to releasing all computing resources occupied by the second model, driving the first model to complete the remaining diagnostics based on the remaining computing resources. This priority setting is based on the following considerations: although the second model has higher accuracy, its computing power consumption is correspondingly higher, and releasing the second model can free up more computing power resources; while the first model has lower computing power consumption, and even if it takes over all diagnostic tasks after releasing the second model, it can use the freed-up resources to complete the remaining diagnostics more quickly. In addition, when releasing the second model, it is important to save its generated intermediate features (if needed) so that they can be used to assist the first model's judgment in possible subsequent situations.

[0089] In one feasible implementation, a preset first model and a second model are invoked respectively to perform parallel diagnosis of the multi-source signals, including: The multi-source signals are preprocessed and feature extracted to obtain a structured feature set, which is then input into the first model to obtain the diagnostic results of the first model. The current signal and each vibration signal in the multi-source signals are treated as independent single-source signals and input into the second model to obtain the diagnostic results of the second model.

[0090] When the first and second models are called to perform parallel diagnosis of multi-source signals, the two models receive input data in different formats.

[0091] For the first model, this embodiment preprocesses and extracts features from the multi-source signals to obtain a structured feature set, which is then input into the first model. The preprocessing and feature extraction referred to here means extracting features with clear physical meaning from the original acquired current and vibration waveforms, such as the peak value, peak occurrence time, and duration of the current waveform, as well as statistical quantities such as the energy value, peak value, and frequency band distribution of the vibration signal. These features are then organized into a fixed-dimensional feature vector, i.e., the structured feature set. Since the structured feature set is a condensation and abstraction of the original signal, the data volume is much smaller than the original waveform, thus the first model processes it quickly and is suitable as a continuously running, resident diagnostic model. The type and dimension of features extracted can be set according to the actual application scenario and the algorithm characteristics of the first model.

[0092] For the second model, in this embodiment, the current signal and each vibration signal from the multi-source signals are treated as independent single-source signals and input separately into the second model. Unlike the first model, the second model does not require pre-extraction of structured features; instead, it directly receives the original time-series signal waveform as input. The current signal is input as an independent single-source signal, and the two vibration signals are also input as independent single-source signals. Internally, the second model uses its own network structure to extract and fuse features from each signal. Since the original time-series signal contains complete waveform details and has a large amount of data, the second model can learn more refined features from it, but it also requires more computational resources.

[0093] It should be noted that although the inputs of the first and second models differ in form, they originate from the same source: multi-source signals loaded based on the same diagnostic instruction. The input of the first model is structured features extracted from the multi-source signals, while the input of the second model is the original time-series waveform from the multi-source signals. Both analyze the same set of signals from different granularities and perspectives, and their respective diagnostic outputs can complement and corroborate each other. During the parallel diagnostic process of the first and second models, each model independently processes its own input data format without interference. This application does not limit the specific input formats of the first and second models, as long as they can be adapted to the algorithmic characteristics of their respective models.

[0094] In one feasible implementation, the multi-source signals are processed to obtain a structured feature set, including: The current signal is denoised using sliding window smoothing filtering and ensemble empirical mode decomposition, and the extreme points of the denoised current waveform and their corresponding time parameters are extracted to construct a coil current signal feature set. For each vibration signal, the time domain features are extracted using the short-time energy method, and the time-frequency domain frequency band energy features are extracted using the wavelet packet decomposition method to construct a vibration signal feature set. The coil current signal feature set and the multi-channel vibration signal feature set are merged and standardized to obtain the structured feature set.

[0095] For the current signal of the opening and closing coils, denoising can be performed sequentially using sliding window smoothing filtering and ensemble empirical mode decomposition. Sliding window smoothing filtering suppresses high-frequency random noise in the signal. Its principle is to use the average value of each sampling point and its several neighboring points as the filtering result for that point, effectively removing glitches and spikes. Ensemble empirical mode decomposition further separates the noise-dominant and signal-dominant components in the signal. This method, by adding Gaussian white noise and performing multiple empirical mode decompositions, effectively overcomes the mode aliasing problem in traditional empirical mode decomposition, decomposing the signal into multiple intrinsic mode function components. After removing the high-frequency noise-dominant component, the remaining components are superimposed and reconstructed to obtain the denoised current signal. After denoising, key extreme points in the current waveform and their corresponding time parameters can be extracted. These extreme points reflect the critical action moments of the electromagnet and coil during the opening and closing process; for example, the current peak point corresponds to the moment the iron core strikes and trips, and the starting point of the current rising edge corresponds to the moment the iron core starts.

[0096] Taking the opening and closing coil current signal as an example, six key features can be extracted, including the coil current start time, current peak value, peak value arrival time, current steady-state value, current falling edge start time, and total opening and closing action duration, forming a 6-dimensional coil current signal feature set, such as... Figure 5 As shown: This includes: a time axis (time / ms), current waveform values ​​(current / A), and short-time vibration energy values ​​(short-time vibration energy / V²). The figure shows the numerical records of the changes in the opening current and vibration signals over time during the circuit breaker's opening and closing operations (corresponding to the closing current waveform, opening current waveform, and short-time vibration energy, respectively). Figure 5 The current waveform and the corresponding values ​​of short-time vibration energy at different time points are presented in the form of row records.

[0097] As can be seen, the time gradually increases from 0 milliseconds. The current waveform data records the complete process of the current flowing through the coil changing with time during the opening and closing process, including the complete waveform shape of the current gradually rising from zero, reaching a peak value, and then falling back to zero. The short-time vibration energy data records the short-time energy value of the vibration signal at the corresponding moment, reflecting the change in the intensity of mechanical vibration during the circuit breaker operation.

[0098] Several key features can be extracted for subsequent diagnostics. These key features include the coil current initiation time (the moment when the current rises significantly from zero); the current peak value (the maximum amplitude reached during the rising phase of the current waveform); the peak arrival time (the time elapsed from the initiation time to reaching the current peak); the steady-state current value (the value of the current after the peak value when it enters a relatively stable phase); the current falling edge initiation time (the moment when the current begins to decrease from the steady state); and the total duration of the opening and closing actions (the entire duration from current initiation to current returning to zero). By identifying and calculating these features from the waveform data, a 6-dimensional coil current signal feature set can be constructed, serving as one of the input features for the subsequent diagnostic model.

[0099] Understandable Figure 5 The waveform data shown is presented in tabular form, but in practical applications, it can also be displayed as a waveform graph or other format, as long as it reflects the relationship between current and short-term vibration energy changes over time. Furthermore, the specific shape of the current waveform may differ depending on the circuit breaker model or fault condition, and the extracted characteristic values ​​will also vary accordingly.

[0100] For each vibration signal, features can be extracted from both the time domain and the time-frequency domain. In the time domain, the short-time energy method is used to extract time-domain features. This method divides the vibration signal into multiple short-time windows of fixed length, calculates the energy value within each window, and obtains a sequence of energy changes over time. Taking the circuit breaker vibration signal as an example, the vibration signal can be divided into 5 short-time windows according to the opening and closing sequence. Statistical features such as short-time energy, peak value, mean, standard deviation, kurtosis, and margin factor are calculated for each window, resulting in 30-dimensional time-domain features. The change in energy value reflects the intensity difference of the vibration signal at different times; for example, the energy peak corresponding to the contact collision time is usually significantly higher than at other times. In the time-frequency domain, wavelet packet decomposition is used to extract frequency band energy features. Taking wavelet packet decomposition as an example, the db4 wavelet is used to perform 3-level wavelet packet decomposition on the vibration signal, dividing the signal into 8 equal-width frequency bands, and calculating the frequency band energy and energy proportion of each band. The energy distribution across different frequency bands reflects the differences in vibration signals across different frequency components. For example, vibrations generated by mechanical collisions are typically concentrated in lower frequency bands, while high-frequency noise is distributed in higher frequency bands. By comparing the changes in energy across different frequency bands, different types of mechanical anomalies can be identified.

[0101] The 30-dimensional time-domain features and 8-dimensional time-frequency domain features of a single vibration signal are combined to obtain a 38-dimensional single-channel vibration signal feature set. The combined features of the two vibration signals result in a 76-dimensional vibration feature set. For example... Figure 6As shown, frequency band energy distribution diagrams are presented for two vibration signals under normal and different fault conditions: Figure 6 In the graph, the horizontal axis of each coordinate system represents frequency in kHz, indicating the frequency bands divided after wavelet packet decomposition of the signal. The vertical axis represents amplitude in V², indicating the energy value corresponding to each frequency band. The graph contains multiple curves, each corresponding to the frequency band energy distribution at different monitoring locations under different conditions. For example, Vib1 represents the signal collected by the vibration sensor installed at the first monitoring location, and Vib2 represents the signal collected by the vibration sensor installed at the second monitoring location.

[0102] The five distribution curves, "Normal 1" to "Normal 5," represent the vibration signal frequency band energy distribution curves corresponding to multiple closing operations of the circuit breaker under normal conditions. "Normal Closing" indicates the vibration signal frequency band energy distribution when the circuit breaker performs a closing operation under fault-free conditions. "Spring Fatigue During Closing" indicates the frequency band energy distribution when the circuit breaker performs a closing operation with fatigue failure of the opening and closing springs. "Loose Anchor Bolts During Closing" indicates the frequency band energy distribution when the circuit breaker performs a closing operation with loose anchor bolts. "Main Shaft Jamming During Closing" indicates the frequency band energy distribution when the circuit breaker performs a closing operation with a jamming fault in the transmission mechanism.

[0103] Figure 6 Furthermore, significant differences in energy values ​​across certain frequency bands can be observed under different conditions. For instance, at a certain frequency, the amplitude remains stable within a relatively small range under normal conditions, while the amplitude increases significantly in that frequency band under spring fatigue conditions. Also, the energy distribution curve under loose anchor bolt conditions exhibits different fluctuation patterns in certain frequency bands compared to the normal state. In addition, the energy distribution of two vibration signals across different frequency bands may show different responses to the same fault. These differences reflect the impact of different fault types on the vibration characteristics of circuit breakers at different monitoring locations.

[0104] Understandable Figure 6 The specific number of frequency bands, the energy values ​​of each frequency band, and the listed fault state types shown are for illustrative purposes only and do not constitute a limitation on the scope of protection of this application. In practical applications, the frequency band division method, the analyzed fault state types, and the energy distribution patterns under each state can be adjusted according to specific signal characteristics and diagnostic requirements.

[0105] After obtaining the coil current signal feature set and the two vibration signal feature sets respectively, the three are merged. For example, the 6-dimensional coil current feature set is merged with the 76-dimensional vibration feature set to obtain an 82-dimensional structured feature set. Z-score normalization is then applied to each of the 82 dimensions for standardization. The formula for Z-score normalization is as follows:

[0106] in, These are the original eigenvalues. This is the mean of the feature across all samples. The standard deviation of this feature. These are the standardized feature values. It's important to note that each dimension independently calculates its own mean and standard deviation, and then transforms all sample values ​​for that dimension using these values. The purpose of standardization is to eliminate the influence of different features due to differences in units or numerical ranges. For example, the unit of current features is amperes or milliseconds, while the unit of vibration features is energy or amplitude. Directly combining these could lead to features with larger values ​​dominating the model's judgment. Z-score standardization transforms all features into a standard normal distribution with a mean of 0 and a standard deviation of 1, ensuring that each feature has a similar influence in subsequent diagnosis.

[0107] While constructing the structured feature set, feature subsets of each individual source signal can also be retained separately, i.e., features of the current signal, the first vibration signal, and the second vibration signal can be retained separately. The purpose of this is to allow for more flexible use of the feature set in subsequent diagnostic processes. This means that the combined complete structured feature set can be used for comprehensive diagnosis, or only a subset of features from one signal can be used for single-source diagnosis, or features from different signal combinations can be used for fusion diagnosis. This makes the use of the feature set more flexible and adaptable to different diagnostic scenarios and model requirements.

[0108] It should be noted that the specific parameters in the above feature extraction method (such as the size of the sliding window, the number of short-time window divisions, the number of wavelet packet decomposition levels, and basis functions) can be selected and adjusted according to the actual signal characteristics and diagnostic requirements. This application embodiment does not impose specific limitations on this. For example, the number of short-time window divisions can be adjusted according to the total duration of the opening and closing operation, and the number of wavelet packet decomposition levels can be set according to the effective bandwidth of the signal.

[0109] In addition, different feature fusion strategies can be employed. For example, principal component analysis (PCA) can be used for dimensionality reduction fusion. PCA is performed on the concatenated features, and principal components whose cumulative contribution rate reaches a set threshold are selected. This eliminates feature redundancy and reduces feature dimensionality while retaining the main information. Linear discriminant analysis (LDA) can also be used for dimensionality reduction fusion. Using fault category labels as supervision information, the features are reduced in dimensionality under the criteria of maximizing inter-class distance and minimizing intra-class distance, resulting in a low-dimensional discriminant feature set.

[0110] In one feasible implementation, the second model includes a data-level fusion branch and a feature-level fusion branch.

[0111] The data-level fusion branch is used to concatenate the current signal and each of the vibration signals into a multi-channel signal in the channel dimension, and to perform joint convolution on the multi-channel signal through a one-dimensional convolutional neural network to achieve multi-source information fusion at the original data level.

[0112] The feature-level fusion branch is used to take the current signal and each of the vibration signals as independent inputs, extract deep features through their respective one-dimensional convolutional neural networks, and then splice and fuse them at the feature level.

[0113] In the embodiments of this application, the second model may include two different fusion structures: a data-level fusion branch and a feature-level fusion branch.

[0114] The data-level fusion branch is used to fuse multi-source information at the raw data level. Specifically, this branch concatenates the current signal and various vibration signals along the channel dimension into a multi-channel signal, and then uses this multi-channel signal as input to perform joint convolution through a one-dimensional convolutional neural network. In this way, the signals are combined together in the initial stage of entering the network, and the convolutional kernel can simultaneously extract the correlation features between different signals across channels. For example, at a certain point in time, a sudden change in the current waveform and a peak in the vibration signal may have a temporal correspondence, and data-level fusion can use convolution operations to simultaneously perceive this cross-signal temporal correlation. This method directly preserves the synchronization relationship between multi-source signals at the raw data level, and the convolutional network can autonomously learn and utilize these cross-signal correlations in the subsequent feature extraction process.

[0115] Feature-level fusion is used to fuse multi-source information at the feature level. Specifically, this branch treats the current signal and various vibration signals as independent inputs, each extracting deep features through an independent one-dimensional convolutional neural network. These deep features are then fused at the feature level before being input to subsequent fully connected and classification layers. Unlike data-level fusion, feature-level fusion allows each signal to be processed independently through its own convolutional network, forming a high-level abstract feature representation for each signal. These high-level features from different signal sources are then combined. This approach captures the inherent patterns of different signals in their respective feature spaces, allowing these patterns to be synthesized before decision-making.

[0116] Each of the two fusion branches has its own characteristics. Data-level fusion performs fusion at the raw data level, suitable for scenarios where there is strong temporal synchronization and interrelationship between multiple source signals, allowing convolutional networks to establish cross-signal connections early on. Feature-level fusion processes each signal independently before fusion, suitable for scenarios where the correlation between signals is not particularly strong, or where sufficient feature extraction needs to be completed within each signal first. In practical applications, one fusion branch can be selected based on specific data characteristics, hardware conditions, and diagnostic requirements, or different branches can be used at different times.

[0117] In one feasible implementation, the method further includes: The circuit breaker's opening and closing coil current signals and the dual-point vibration signals are collected under normal and various fault conditions to obtain the original sample set. The various fault conditions include: poor coil circuit contact, abnormal coil control voltage, stuck opening electromagnet core, abnormal gap of opening electromagnet core, mechanism failure to move, fatigue of opening and closing springs, loose anchor bolts, and stuck transmission mechanism.

[0118] A generative adversarial network with a gradient penalty term is used to augment the original sample set, generating simulated samples that meet the preset consistency requirements with the real sample distribution, thus constructing a simulated sample set.

[0119] The first model and the second model are trained using the original sample set and the simulation sample set, respectively. During the training process, multiple model structure combinations and multiple hyperparameter configurations are used for training. The model structure with the highest diagnostic accuracy and the corresponding hyperparameters are selected through cross-validation as the final configuration of the first model and the second model.

[0120] Before loading the multi-source signal to be diagnosed, the embodiments of this application also include a pre-training step, namely, pre-collecting a sample set and training the first model and the second model.

[0121] To obtain the sample data required for training, the current signals of the opening and closing coils and the vibration signals at two points can be collected from the circuit breaker under normal conditions and various fault conditions. The fault conditions can cover a variety of common circuit breaker fault types, such as poor coil circuit contact, abnormal coil control voltage, jamming of the opening electromagnet core, abnormal gap in the opening electromagnet core, failure of the mechanism to operate, fatigue of the opening and closing springs, loose anchor bolts, and jamming of the transmission mechanism. These fault types cover both electrical circuit faults and mechanical mechanism faults, providing a relatively comprehensive set of fault samples for model training.

[0122] After acquiring the original sample set, a Generative Adversarial Network with Gradient Penalty (WGAN-GP) can be used to augment the original sample set. The core idea of ​​WGAN is to enable the generator to gradually learn the distribution characteristics of real samples through adversarial training between the generator and the discriminator, thereby generating simulated samples that are basically consistent with the distribution of real samples. In the embodiments of this application, an improved WGAN is used, namely, a WGAN with gradient penalty.

[0123] When using generative adversarial networks (GANs) for data augmentation, the one-dimensional time-series signals in the original sample set can be preprocessed by normalization. This normalization maps the signal amplitude linearly to a predetermined interval (e.g., [-1, 1]) to eliminate the impact of differences in the dimensions and amplitudes of different signals on network training. The normalization formula is as follows:

[0124] in, The original signal amplitude, and These are the minimum and maximum values ​​of the signal sequence, respectively. This represents the normalized signal amplitude. Through this transformation, the amplitude of all signals is mapped to the interval [-1, 1], ensuring consistency in the numerical range across different signals. The processed signal can then be used as the training input for a generative adversarial network.

[0125] The network comprises two subnetworks: a generator and a discriminator. The generator receives random noise as input and outputs a one-dimensional time-series simulation signal with the same dimension as the real sample. The discriminator receives either a real sample or a generated sample as input and outputs a result indicating the sample's authenticity. During training, the generator attempts to generate simulated samples that can "fool" the discriminator, while the discriminator tries to distinguish between real and generated samples. The two networks continuously optimize through this adversarial process. The network structure diagram of the generative adversarial network is shown below. Figure 2 As shown, the generated samples (i.e., simulated samples) from generator G are input into discriminator D along with real samples. Discriminator D outputs a judgment result: true / false, used to distinguish between real samples and generated samples. Then, the judgment result from discriminator D is input into generator G to optimize parameters, thereby continuously optimizing the parameters of the generative adversarial network. The generated samples from generator G are based on received random noise.

[0126] The discriminator D employs a multi-layer one-dimensional convolutional network structure, using LeakyReLU as the activation function. The output layer does not have an activation function. Its network structure is as follows: Figure 3As shown, its input layer receives a feature map in the form of a two-dimensional matrix. For processing time-series signals, the one-dimensional time-series signal can first be converted into a two-dimensional representation through some transformation, such as arranging the signal segments into a two-dimensional matrix, or splicing the signals from multiple channels to form a two-dimensional input. The size of this two-dimensional matrix is ​​64×64, and the number of channels is 1, that is, the input layer specification is 64×64×1.

[0127] The input data is processed sequentially through multiple two-dimensional convolutional layers. Each convolutional layer uses LeakyReLU as the activation function to introduce non-linear transformation capabilities while avoiding neuron death. The role of the convolutional layers is to slide the convolutional kernel across the input data, extracting features from local regions, and gradually expanding the receptive field by stacking layers, abstracting simple features such as local edges and waveforms at lower levels into global semantic features at higher levels. As the number of network layers increases, the size of the feature map gradually decreases, while the number of channels gradually increases. For example, after the first convolutional layer, the feature map size becomes 32×32, and the number of channels increases to 128 (32×32×128); after the second convolutional layer, the feature map size becomes 16×16, and the number of channels increases to 256 (16×16×256); after the third convolutional layer, the feature map size becomes 8×8, and the number of channels increases to 512 (8×8×512); after the fourth convolutional layer, the feature map size becomes 4×4, and the number of channels increases to 1024 (4×4×1024). The gradual increase in the number of channels means that the network can extract richer and more abstract feature representations at deeper layers, while the gradual reduction in feature map size is achieved through downsampling via stride settings in convolutional operations, thereby reducing computational cost and expanding the receptive field. Finally, the last convolutional layer compresses the feature map size to 1×1×1, meaning the output is a scalar value representing the discriminator's score for classifying the input sample as real. This output layer does not have an activation function, outputting the unscaled raw score instead of a normalized probability value.

[0128] The parameters of each layer in the above network structure (such as kernel size, stride, number of channels, etc.) can be adjusted according to the actual application scenario and data type. In addition, the discriminator based on one-dimensional time series signals can also adopt a one-dimensional convolutional network structure. Its principle is similar to that of the two-dimensional convolutional discriminator, only differing in the dimension of the convolution operation.

[0129] The generator G employs a fully connected network with an upsampling structure. Its input is a low-dimensional random noise vector, such as 100-dimensional random noise following a standard normal distribution. This random noise vector is mapped to a higher-dimensional feature vector through a fully connected layer. In one exemplary structure, such as... Figure 4As shown, the feature vector of the input layer can be represented as 1×1×100 (1×100 noise), where 100 is the number of channels. After passing through the fully connected layer, its shape is adjusted to a tensor with a spatial dimension of 4×4 and a number of channels of 1024, i.e., 4×4×1024, through a reconstruction operation.

[0130] Subsequently, this tensor is upsampled through multiple transposed convolutional layers. Transposed convolution, in contrast to ordinary convolution, compresses the input data into smaller feature maps with more channels using a sliding window. Transposed convolution, however, expands these smaller feature maps into larger ones with fewer channels, thus generating a high-dimensional signal from low-dimensional features. For example, after the first transposed convolution, the feature map size expands to 8×8, and the number of channels decreases to 512 (8×8×512); after the second transposed convolution, the feature map size expands to 16×16, and the number of channels decreases to 256 (16×16×256); after the third transposed convolution, the feature map size expands to 32×32, and the number of channels decreases to 128 (32×32×128); after the final transposed convolution, the feature map size expands to 64×64, and the number of channels becomes 1 (64×64×1), resulting in a single-channel two-dimensional matrix. Each transposed convolutional layer can be followed by ReLU as the activation function to introduce nonlinear transformation capability. Finally, the output layer uses transposed convolution + Tanh activation function to map the output value to the [-1,1] interval, consistent with the numerical range of the original signal normalized before training, outputting a 64×64 two-dimensional matrix. Thus, the generator gradually transforms low-dimensional random noise into a simulated signal in two-dimensional matrix form. After flattening, a one-dimensional time-series simulation signal with the same dimension as the real sample is obtained. The parameters of each layer in the above network structure can be adjusted according to the actual application scenario and data type. A generator based on one-dimensional time-series signals can also use a one-dimensional transposed convolutional network structure, the principle of which is similar to the two-dimensional structure described above.

[0131] During training, a Wasserstein loss function with a gradient penalty term can be used to constrain the Lipschitz continuity of the discriminator, thereby improving training stability. This loss function... The expression is as follows:

[0132] in, This represents the distribution of the real samples, that is, the distribution that the original acquired signals follow.

[0133] This represents the distribution of the generated samples, that is, the distribution that the simulated signal output by the generator follows.

[0134] Indicates the distribution from the real sample The real sample obtained from sampling is the original one-dimensional time-series signal acquired.

[0135] Indicates the distribution of generated samples The generated sample (simulation signal) obtained by sampling.

[0136] Indicates from real samples With generated samples Samples obtained by uniform sampling along a straight path between the points are used to calculate the gradient penalty term. Indicates sample The distribution of .

[0137] This represents the discriminator's output score for the input sample. The discriminator aims to output a higher score for real samples and a lower score for generated samples.

[0138] Indicates that the discriminator is in The gradient of the derivative with respect to the input is the partial derivative of the discriminator output with respect to the input.

[0139] The L2 norm, used to represent this gradient, is used to measure the discriminator's performance. Sensitivity in the vicinity.

[0140] This is the gradient penalty coefficient, used to control the weight of the penalty term.

[0141] The formula consists of three parts, the first part... It is the discriminator's expectation of the generated sample score, Part Two. This is the discriminator's expectation of the score for the real sample. The difference between the two is an approximation of the Wasserstein distance, used to measure the difference between the distribution of the real samples and the distribution of the generated samples. The generator's goal is to minimize this difference, so that the discriminant score of the generated samples is close to the discriminant score of the real samples. Part Three This is a gradient penalty term used to constrain the L2 norm of the gradient of the discriminator on the interpolation path between real and generated samples to be close to 1, thereby satisfying the Lipschitz continuity constraint and improving the stability of training.

[0142] During training, the original sample set can be divided into a training set and a validation set according to a set ratio, and an optimizer can be used for iterative training. When the difference in feature distribution between the generated samples and the real samples is less than a preset threshold, and the network loss tends to stabilize, the network is considered to have converged. The trained generator outputs simulated samples in batches, which are then merged with the original sample set to construct an expanded sample set.

[0143] Through the above methods, the generated simulation samples and real samples meet the preset consistency requirements in terms of distribution, thereby effectively expanding the number of training samples and alleviating the problems of difficulty in obtaining fault samples and insufficient sample size.

[0144] After obtaining the original sample set and the simulation sample set, the first model and the second model are trained respectively.

[0145] For the first model, structured feature sets can be extracted from the original sample set and the simulation sample set to form training samples, which are then divided into training and test sets in an 8:2 ratio. During training, various machine learning algorithms can be trained and validated separately. For example, different algorithms such as Random Forest, Support Vector Machine, K-Nearest Neighbors, Decision Tree, and LightGBM (Lightweight Gradient Boosting Machine) can be selected as candidates. For each algorithm, a combination of cross-validation and grid search is used to optimize its hyperparameters. Different algorithms have different hyperparameter configurations. For example, Support Vector Machine requires optimization of the penalty coefficient and kernel parameters, Random Forest requires optimization of the number of decision trees, maximum depth, and minimum number of split samples, and K-Nearest Neighbors requires optimization of the nearest neighbor K value and distance metric. By inputting the training samples into each algorithm for fitting training and evaluating the diagnostic accuracy of each algorithm on the test set, the algorithm that performs best on the structured feature set and its corresponding hyperparameters are selected as the final configuration of the first model.

[0146] During training, different feature fusion strategies can be employed to process the structured feature sets. For example, the coil current signal feature set and the vibration signal feature sets from various sources can be directly concatenated in terms of dimension to form a complete fused feature set, retaining all original feature information. Alternatively, principal component analysis (PCA) can be used for dimensionality reduction fusion. PCA is performed on the concatenated features, selecting principal components whose cumulative contribution rate reaches a set threshold, thus eliminating feature redundancy and reducing feature dimensionality while retaining key information. Linear discriminant analysis (LDA) can also be used for dimensionality reduction fusion. Using fault category labels as supervisory information, the features are dimensionality reduced under the criteria of maximizing inter-class distance and minimizing intra-class distance, resulting in a low-dimensional discriminant feature set. The feature sets obtained from each fusion strategy are then input into various machine learning algorithms for training and testing. By comparing the diagnostic accuracy under different fusion strategies and algorithm combinations, the optimal fusion strategy and algorithm combination are selected. Experimental results show that among various candidate algorithms, the random forest algorithm performs best in single-source signal diagnosis. For example, the single-source diagnostic accuracy of current signals using random forest can reach over 90%, and the diagnostic accuracy of two vibration signals fused using random forest can reach over 95%. Furthermore, after multi-source feature fusion, the combined diagnostic accuracy of the direct splicing fusion strategy and the random forest algorithm can reach over 99%, with a small standard deviation, indicating that the combination has good accuracy and stability.

[0147] For the second model, the original time-series signals from both the original and simulated sample sets can be used as inputs for training. During training, various deep learning models adapted to the time-series signals can be constructed as candidates. For example, a one-dimensional convolutional neural network can be constructed, employing a wide convolutional kernel design, containing multiple one-dimensional convolutional layers and pooling layers. The convolutional operation slides along the time dimension to extract local waveform features of the signal, and finally, a fully connected layer and a Softmax classification layer output the classification result. Alternatively, a convolutional gated recurrent unit fusion network can be constructed, connecting multiple gated recurrent units after the one-dimensional convolutional layers. Each hidden layer can be set to 128 dimensions. The convolutional layers are responsible for extracting spatial local features of the signal, while the gated recurrent units are responsible for transmitting state information along the time dimension to capture long-range temporal dependencies in the signal. A time-frequency convolutional network can also be constructed, first converting the one-dimensional time-series signal into a two-dimensional time-frequency graph through a short-time Fourier transform, and then using a two-dimensional convolutional network to extract image-level features from the time-frequency graph. For each of the above model structures, different multi-source information fusion strategies can be combined for training. For example, a data-level fusion strategy can be used, where the current signal and various vibration signals are concatenated into a multi-channel signal at the channel dimension and input as a whole into the network, allowing the network to process multiple signals simultaneously at the raw data level. Alternatively, a feature-level fusion strategy can be used, where the current signal and various vibration signals are treated as independent inputs, each extracting deep features through an independent feature extraction branch, and then concatenating and fusing them at the feature level. During training, the expanded sample set is divided into training, validation, and test sets proportionally, for example, in a 7:2:1 ratio. Iterative training is performed using an optimizer and loss function, and an early stopping strategy is used to prevent overfitting.

[0148] Comparisons during training reveal that different candidate network structures and fusion strategies prioritize different aspects of diagnostic accuracy and training efficiency. One-dimensional convolutional neural networks (CNNs) are simple and train quickly, but their accuracy is relatively limited when handling complex temporal dependencies. Convolutionally gated recurrent unit (CRU) fusion networks introduce gated recurrent unit (ROU) layers after convolutional layers, effectively capturing the temporal dependencies of signals and achieving a good balance between diagnostic accuracy and training efficiency; for example, the diagnostic accuracy can reach over 98% when fusing two vibration signals. Time-frequency convolutional networks convert one-dimensional time-series signals into two-dimensional time-frequency maps for image-level feature extraction, fully utilizing time-frequency domain information and achieving a diagnostic accuracy of over 99%, but training time is correspondingly increased. Considering both diagnostic accuracy and training efficiency, the CNN can be selected as the core network structure of the second model, combined with a feature-level fusion strategy for fusing multi-source information. After model construction, the trained model is configured in the diagnostic process; inputting the multi-source time-series signals of the target circuit breaker outputs the diagnostic results and corresponding confidence levels.

[0149] It should be noted that the specific descriptions of the algorithm selection, network structure, and fusion strategy for the first and second models described above are only one specific implementation of this application. In practical applications, the first model can also adopt other types of machine learning models, such as support vector machines and gradient boosting trees, depending on the requirements; the second model can also select other network structures or fusion strategies according to specific computing power conditions and accuracy requirements. For example, in scenarios with sufficient computing power, a time-frequency convolutional network can be selected to pursue higher diagnostic accuracy, or in scenarios with limited computing power, a one-dimensional convolutional neural network can be selected to reduce inference latency.

[0150] Furthermore, the training of the first model and the second model can be performed independently or share different representations of the same set of training data. The first model uses a structured feature set, while the second model uses raw time-series signals. This separate training approach ensures that each model achieves optimal diagnostic performance under its respective input format and algorithm characteristics. Since the first and second models are trained independently based on their respective optimal configurations during the training process, they maintain independent model structures and parameters in the final diagnostic architecture.

[0151] After determining the final configuration of the first and second models, this embodiment of the application pre-places the two models in the diagnostic system for subsequent online diagnostic processes.

[0152] In a feasible implementation, the original sample set of this application embodiment can be obtained in the following manner.

[0153] First, a multi-source signal acquisition platform is established. For the opening and closing coil current signal, a Hall effect current sensor can be used for acquisition; this type of sensor can accurately capture rapid current changes during the transient process of opening and closing. For the circuit breaker vibration signal, a piezoelectric accelerometer can be used, installed at two different monitoring locations on the circuit breaker body, for example, one on the side of the operating mechanism housing and the other on the transmission mechanism base, thus simultaneously acquiring two independent vibration signals. During the acquisition process, an appropriate signal sampling rate can be set to ensure that signal details during the opening and closing operation are captured. Simultaneously, an oscilloscope can be used with the synchronous acquisition unit to synchronize the timestamps of the multi-source signals, ensuring an accurate time correspondence between the current signal and the two vibration signals.

[0154] After the data acquisition platform is built, fault simulation experiments can be conducted for common fault types of circuit breakers. Taking a high-voltage circuit breaker with a spring-operated mechanism as an example, standardized simulation experiments can be carried out for common fault types in power systems, while a normal operating state can be set as a control group. For example, fault types such as poor coil circuit contact, abnormal coil control voltage, jamming of the tripping electromagnet core, abnormal gap of the tripping electromagnet core, failure of the mechanism to move, fatigue of the opening and closing springs, loose anchor bolts, and jamming of the transmission mechanism can be simulated. For each fault state, different severity of fault scenarios can be simulated by adjusting the mechanism parameters or setting fault conditions to ensure the realism and coverage of the fault simulation.

[0155] After completing the above fault simulation experiments, opening and closing experiments can be conducted under the same experimental environment for each fault state and normal state. Multiple sets of synchronous opening and closing data can be collected for each state, for example, no fewer than 56 sets. Finally, multiple sets of valid raw opening and closing data are accumulated. Each set of data includes the opening and closing coil current timing signal, two synchronous vibration timing signals, and corresponding state labels, thus completing the construction of the original sample set.

[0156] It should be noted that the above description of the sensor model, sampling rate, fault type, and sample quantity is only one specific implementation of this application. In practical applications, different specifications of sensors, different sampling rate settings, and different fault simulation schemes can be selected according to the type and model of the circuit breaker under test and the field conditions, as long as a raw sample set containing current signals and multiple vibration signals with status labels can be obtained.

[0157] Based on the same technical concept, this application also provides an adaptive fault diagnosis device for a circuit breaker, the device comprising: The response module is used to load the multi-source signals to be diagnosed in response to the diagnostic command; the multi-source signals include at least the current signals of the opening and closing coils and the dual-point vibration signals distributed on the closing coil side and the operating mechanism side.

[0158] The diagnostic module is used to call the preset first model and second model respectively to perform parallel diagnosis of the multi-source signals.

[0159] The monitoring module is used to monitor the current available computing power in real time during the parallel diagnostic process.

[0160] The scheduling module is used to perform the following tiered computing power scheduling if the available computing power is lower than a first threshold: When the available computing power is greater than or equal to the second threshold, the computational load of at least one model is reduced to maintain parallel diagnosis of the two models; and when the available computing power is less than the second threshold, all computing power resources occupied by one model are released, and the other model is driven to complete the remaining diagnosis based on all computing power resources; the second threshold is less than the first threshold.

[0161] The fault determination module is used to determine the fault type of the circuit breaker based on the diagnostic results of the first model and / or the second model.

[0162] In one feasible implementation, the first model is a random forest model based on structured physical features, and the second model is a convolutional gated recurrent unit fusion network based on the original time-series signal. The computational power consumption of the first model per unit time is lower than that of the second model per unit time.

[0163] The diagnostic module is used to perform parallel diagnostics on the multi-source signals by respectively calling a preset first model and a second model, including: The first model is used as the resident diagnostic model to continuously diagnose the multi-source signals; the second model is used as the auxiliary diagnostic model and is activated according to a preset period or in response to the current available computing power being greater than or equal to the first threshold, so as to achieve parallel diagnosis.

[0164] When the hierarchical computing power scheduling is executed, the following is performed first: release all computing power resources occupied by the second model, and drive the first model to complete the remaining diagnosis based on all computing power resources.

[0165] In one feasible implementation, the diagnostic module is used to perform parallel diagnostics on the multi-source signals by respectively calling a preset first model and a second model, including: The multi-source signals are preprocessed and feature extracted to obtain a structured feature set, which is then input into the first model to obtain the diagnostic results of the first model.

[0166] The current signal and each vibration signal from the multi-source signals are treated as independent single-source signals and input into the second model respectively to obtain the diagnostic results of the second model.

[0167] In one feasible implementation, the diagnostic module is used to process the multi-source signals to obtain a structured feature set, including: The current signal is denoised by sliding window smoothing filter and ensemble empirical mode decomposition, and the extreme points of the denoised current waveform and their corresponding time parameters are extracted to construct a coil current signal feature set.

[0168] For each vibration signal, the short-time energy method is used to extract time-domain features, and the wavelet packet decomposition method is used to extract time-frequency domain frequency band energy features to construct a vibration signal feature set.

[0169] The structured feature set is obtained by merging the coil current signal feature set with the multi-channel vibration signal feature set and then standardizing the result.

[0170] In one feasible implementation, the second model includes a data-level fusion branch and a feature-level fusion branch.

[0171] The data-level fusion branch is used to concatenate the current signal and each of the vibration signals into a multi-channel signal in the channel dimension, and to perform joint convolution on the multi-channel signal through a one-dimensional convolutional neural network to achieve multi-source information fusion at the original data level.

[0172] The feature-level fusion branch is used to take the current signal and each of the vibration signals as independent inputs, extract deep features through their respective one-dimensional convolutional neural networks, and then splice and fuse them at the feature level.

[0173] In one feasible implementation, the fault determination module is used to determine the fault type of the circuit breaker based on the diagnostic results of the first model and / or the second model, including: If the diagnostic results of the first model and the second model are obtained respectively, and the fault type diagnostic results of the two diagnostic results are consistent, then the average of the confidence scores output by the first model and the second model is taken as the final confidence score, and the fault type is output.

[0174] If the fault types of the two diagnostic results are inconsistent, the diagnostic result with higher confidence is selected as the final output and sent to the management terminal for review, thereby obtaining the review result from the management terminal.

[0175] If only the diagnostic result of the first model is obtained, the diagnostic result of the first model is directly used as the final output, and the product of the confidence level of the first model and the preset compensation coefficient is used as the final confidence level, wherein the compensation coefficient is greater than 1.

[0176] In one feasible implementation, the device further includes: The acquisition module is used to acquire the opening and closing coil current signals and the dual-point vibration signals of the circuit breaker under normal conditions and various fault conditions to obtain the original sample set. The various fault conditions include: poor contact of the coil circuit, abnormal coil control voltage, jamming of the opening electromagnet core, abnormal gap of the opening electromagnet core, failure of the mechanism to move, fatigue of the opening and closing springs, loose anchor bolts, and jamming of the transmission mechanism.

[0177] The enhancement module is used to perform data augmentation on the original sample set using a generative adversarial network with a gradient penalty term, generating simulation samples that meet the preset consistency requirements with the real sample distribution, and constructing a simulation sample set.

[0178] The training module is used to train the first model and the second model using the original sample set and the simulation sample set, respectively. During the training process, the model is trained with multiple model structure combinations and multiple hyperparameter configurations. The model structure and corresponding hyperparameters with the highest diagnostic accuracy are selected through cross-validation as the final configuration of the first model and the second model.

[0179] Figure 7 A schematic diagram of an electronic device provided in this application embodiment includes: a processor 701, a storage medium 702, and a bus 703. The storage medium 702 stores machine-readable instructions executable by the processor 701. When the electronic device runs the adaptive fault diagnosis method for a circuit breaker as described in the embodiment, the processor 701 communicates with the storage medium 702 via the bus 703, and the processor 701 executes the machine-readable instructions to perform the steps as described in the embodiment.

[0180] In this embodiment, the storage medium 702 may also execute other machine-readable instructions to perform other methods as described in the embodiment. For details on the specific execution steps and principles, please refer to the description of the embodiment, which will not be repeated here.

[0181] This application also provides a computer-readable storage medium storing a computer program that is executed by a processor to perform the steps as described in the embodiments.

[0182] In this embodiment, the computer program, when run by the processor, can also execute other machine-readable instructions to perform other methods as described in the embodiments. For details on the specific execution steps and principles, please refer to the description of the embodiments, which will not be repeated here.

[0183] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the coupling or direct coupling or communication connection shown or discussed may be through some communication interface; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0184] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0185] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0186] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0187] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An adaptive fault diagnosis method for circuit breakers, characterized in that, The method includes: In response to a diagnostic command, a multi-source signal to be diagnosed is loaded; the multi-source signal includes at least the current signal of the opening and closing coils and the dual-point vibration signal distributed on the closing coil side and the operating mechanism side; The multi-source signals are diagnosed in parallel by calling the preset first and second models respectively; During parallel diagnostics, the available computing power is monitored in real time. If the available computing power is lower than the first threshold, then the following tiered computing power scheduling is performed: When the available computing power is greater than or equal to the second threshold, the computational load of at least one model is reduced to maintain parallel diagnosis of the two models; and when the available computing power is less than the second threshold, all computing power resources occupied by one model are released, and the other model is driven to complete the remaining diagnosis based on all computing power resources; the second threshold is less than the first threshold; Based on the diagnostic results of the first model and / or the second model, the fault type of the circuit breaker is determined.

2. The method according to claim 1, characterized in that, The first model is a random forest model based on structured physical features, and the second model is a convolutional gated recurrent unit fusion network based on the original time series signal. The computational power consumption of the first model per unit time is lower than that of the second model per unit time. The parallel diagnosis of the multi-source signals by respectively calling the preset first model and the second model includes: Using the first model as a permanent diagnostic model, continuous diagnosis is performed on the multi-source signals; The second model is used as an auxiliary diagnostic model, and the second model is activated according to a preset period or in response to the current available computing power being greater than or equal to the first threshold, so as to achieve parallel diagnosis. When the hierarchical computing power scheduling is executed, the following is performed first: release all computing power resources occupied by the second model, and drive the first model to complete the remaining diagnosis based on all computing power resources.

3. The method according to claim 1, characterized in that, Parallel diagnosis of the multi-source signals is performed by respectively calling the preset first model and the second model, including: The multi-source signals are preprocessed and feature extracted to obtain a structured feature set, which is then input into the first model to obtain the diagnostic results of the first model. The current signal and each vibration signal from the multi-source signals are treated as independent single-source signals and input into the second model respectively to obtain the diagnostic results of the second model.

4. The method according to claim 3, characterized in that, The multi-source signal is processed to obtain a structured feature set, including: The current signal is denoised by sliding window smoothing filter and ensemble empirical mode decomposition, and the extreme points of the denoised current waveform and their corresponding time parameters are extracted to construct a coil current signal feature set. For each vibration signal, the short-time energy method is used to extract time-domain features, and the wavelet packet decomposition method is used to extract time-frequency domain frequency band energy features to construct a vibration signal feature set. The structured feature set is obtained by merging the coil current signal feature set with the multi-channel vibration signal feature set and then standardizing the result.

5. The method according to claim 3, characterized in that, The second model includes a data-level fusion branch and a feature-level fusion branch; The data-level fusion branch is used to concatenate the current signal and each of the vibration signals in the channel dimension into a multi-channel signal, and to perform joint convolution on the multi-channel signal through a one-dimensional convolutional neural network to achieve multi-source information fusion at the original data level. The feature-level fusion branch is used to take the current signal and each of the vibration signals as independent inputs, extract deep features through their respective one-dimensional convolutional neural networks, and then splice and fuse them at the feature level.

6. The method according to claim 1, characterized in that, The determination of the fault type of the circuit breaker based on the diagnostic results of the first model and / or the second model includes: If the diagnostic results of the first model and the second model are obtained respectively, and the fault type diagnostic results of the two diagnostic results are consistent, then the average of the confidence scores output by the first model and the second model is taken as the final confidence score, and the fault type is output. If the fault types of the two diagnostic results are inconsistent, the diagnostic result with higher confidence is selected as the final output and sent to the management terminal for review, and the review result of the management terminal is obtained. If only the diagnostic result of the first model is obtained, the diagnostic result of the first model is directly used as the final output, and the product of the confidence level of the first model and the preset compensation coefficient is used as the final confidence level, wherein the compensation coefficient is greater than 1.

7. The method according to claim 1, characterized in that, The method further includes: The circuit breaker's opening and closing coil current signals and the dual-point vibration signals are collected under normal and various fault conditions to obtain the original sample set. The various fault conditions include: poor coil circuit contact, abnormal coil control voltage, stuck opening electromagnet core, abnormal gap of opening electromagnet core, mechanism failure to move, fatigue of opening and closing springs, loose anchor bolts, and stuck transmission mechanism. A generative adversarial network with a gradient penalty term is used to augment the original sample set, generating simulated samples that meet the preset consistency requirements with the real sample distribution, thus constructing a simulated sample set. The first model and the second model are trained using the original sample set and the simulation sample set, respectively. During the training process, multiple model structure combinations and multiple hyperparameter configurations are used for training. The model structure with the highest diagnostic accuracy and the corresponding hyperparameters are selected through cross-validation as the final configuration of the first model and the second model.

8. An adaptive fault diagnosis device for a circuit breaker, characterized in that, The device includes: The response module is used to load the multi-source signals to be diagnosed in response to the diagnostic command; the multi-source signals include at least the current signal of the opening and closing coils and the dual-point vibration signal distributed on the closing coil side and the operating mechanism side; The diagnostic module is used to call the preset first model and second model respectively to perform parallel diagnosis of the multi-source signal; The monitoring module is used to monitor the current available computing power in real time during the parallel diagnostic process; The scheduling module is used to perform the following tiered computing power scheduling if the available computing power is lower than a first threshold: When the available computing power is greater than or equal to the second threshold, the computational load of at least one model is reduced to maintain parallel diagnosis of the two models; and when the available computing power is less than the second threshold, all computing power resources occupied by one model are released, and the other model is driven to complete the remaining diagnosis based on all computing power resources; the second threshold is less than the first threshold; The fault determination module is used to determine the fault type of the circuit breaker based on the diagnostic results of the first model and / or the second model.

9. An electronic device, characterized in that, include: The device includes a processor, a storage medium, and a bus, wherein the storage medium stores machine-readable instructions executable by the processor, and when the electronic device is in operation, the processor communicates with the storage medium via the bus, and the processor executes the machine-readable instructions to perform the steps of the adaptive fault diagnosis method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the adaptive fault diagnosis method as described in any one of claims 1 to 7.

Images