Solid state semiconductor electromagnetic circuit breaker

Abstract

This invention discloses a solid-state semiconductor electromagnetic circuit breaker, relating to the field of power system protection and control technology. It includes a state-aware acquisition module, a feature extraction and modeling module, a dynamic criterion identification module, an adaptive protection control module, a real-time monitoring and secondary verification module, and a closed-loop optimization and update module. The state-aware acquisition module acquires the transient voltage and current waveforms in the main circuit of the circuit breaker in real time, constructing a multi-dimensional electrical feature matrix including the current rise rate, voltage change rate, current peak value, and duration. This invention achieves refined classification of current disturbances through multi-feature fusion modeling and dynamic identification mechanisms. Combined with adaptive protection and delayed discrimination, it avoids false tripping. Furthermore, it improves system stability through real-time monitoring and model self-learning, significantly enhancing the intelligence and reliability of the circuit breaker in complex power grids.

Description

Technical Field

[0001] This invention relates to the field of power system protection and control technology, specifically to solid-state semiconductor electromagnetic circuit breakers. Background Technology

[0002] Solid-state semiconductor electromagnetic circuit breakers are composite circuit breakers that integrate solid-state power devices (such as IGBTs and MOSFETs) with traditional electromechanical structures. They are designed to achieve efficient, rapid, and safe control of current in power systems. Utilizing the nanosecond-level switching capabilities of solid-state devices, these circuit breakers achieve immediate current limiting and pilot interruption of short-circuit or overload currents. Subsequently, an electromagnetic drive mechanism completes the mechanical disconnection of the main circuit, ensuring the circuit breaker's breaking insulation distance and safety. Compared to traditional purely mechanical circuit breakers, they offer faster response speeds, higher switching frequencies, lower arc energy, and stronger intelligent control capabilities. They are widely used in industrial, rail transit, and new energy power distribution systems with high requirements for power supply continuity, safety, and intelligent management.

[0003] The existing technology has the following shortcomings: Under high-voltage, light-load operating conditions, due to the relatively small system load current, solid-state semiconductor electromagnetic circuit breakers primarily rely on rapid identification of current surges to determine the presence of short-circuit or fault risks. However, in such operating scenarios, the system is often subject to voltage disturbances caused by grid fluctuations, switching transients, or load capacitor charging and discharging. These disturbances may generate instantaneous, large current spikes, forming transient pulse current signals with high amplitude but extremely short duration. Because the characteristics of such pulses are highly similar to the initial state of a real short circuit in terms of current waveform, if the current detection and identification mechanism inside the circuit breaker fails to effectively distinguish between normal transients and actual faults, it is highly likely to misjudge the aforementioned disturbances as short-circuit events. This can lead to the solid-state device erroneously triggering premature conduction or shutdown, which in turn triggers the electromagnetic mechanism to perform circuit breaker tripping operations, causing unplanned power outages and load disconnections. This problem is highly concealed and sudden in high-voltage, light-load scenarios. If not effectively suppressed, it can not only affect the continuity of system power supply but also potentially lead to a series of safety and stability risks, such as frequent malfunctions, abnormal equipment responses, and even control system logic chaos.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a solid-state semiconductor electromagnetic circuit breaker that achieves refined classification of current disturbances through multi-feature fusion modeling and dynamic identification mechanisms. Combined with adaptive protection and delayed discrimination, it avoids false tripping and improves system stability through real-time monitoring and model self-learning. This significantly enhances the intelligence and reliability of the circuit breaker in complex power grids, thereby solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a solid-state semiconductor electromagnetic circuit breaker, comprising a state sensing and acquisition module, a feature extraction and modeling module, a dynamic criterion recognition module, an adaptive protection control module, a real-time monitoring and secondary verification module, and a closed-loop optimization and update module. The state-aware acquisition module collects the voltage and current transient waveforms in the main circuit of the circuit breaker in real time and constructs a multi-dimensional electrical feature matrix that includes the current rise rate, voltage change rate, current peak value and duration. The feature extraction and modeling module extracts transient pulse current features based on the feature matrix, calculates the correlation between voltage and current changes, and generates multi-feature composite fingerprints. The dynamic criterion recognition module takes the composite fingerprint input and trains the dynamic criterion neural network model, performs pattern recognition, and determines whether it is a real short circuit fault. The adaptive protection control module, based on the output of the dynamic criterion neural network model, if it identifies a real short-circuit fault, executes dual-threshold adaptive protection control, drives the solid-state device to conduct and triggers the electromagnetic drive mechanism to trip; if it identifies a transient disturbance, it enters the delayed protection judgment. The real-time monitoring and secondary verification module continuously monitors the changes in main circuit voltage and current during the delay phase and feeds the data back to the neural network model for secondary verification. The closed-loop optimization and update module updates the neural network model parameters based on the secondary verification results and stores new fingerprint samples, realizing model self-optimization and sample closed-loop update, thereby improving the accuracy of transient disturbance identification and the operational stability of the circuit breaker.

[0007] Preferably, the state-aware data acquisition steps are as follows: The real-time acquired transient waveform signals of main circuit voltage and current are preprocessed to complete baseline drift correction, power frequency interference filtering, signal noise suppression and fixed time window segmentation. Extract the current rise rate, voltage change rate, peak current amplitude, current duration, time interval from waveform start point to peak point, and waveform integral value within each preset time window. Based on the sliding time window, the correlation coefficient between current disturbance and voltage disturbance, the maximum position of the cross-correlation function, and the length of the synchronous rise segment of voltage and current are calculated to construct a structured multidimensional feature matrix. The generated structured feature matrix, along with the corresponding timestamps and state transition identifiers, is uploaded to the central processing unit as input for subsequent recognition models.

[0008] Preferably, the feature extraction and modeling steps are as follows: After the feature matrix is ​​constructed, the maximum current rise rate, peak current amplitude, current disturbance duration, maximum voltage change rate and its direction are extracted, and all features are standardized. Calculate the correlation indices between voltage and current changes, including Pearson correlation coefficient, peak position of cross-correlation function, duration of co-variance, and signal morphology similarity; Feature vectors are constructed based on eigenvalues ​​and statistics to generate multi-feature composite fingerprints that include timestamps and perturbation energy estimates. Multi-feature composite fingerprints are stored in a fingerprint data buffer and their integrity is verified, serving as the input for subsequent recognition models.

[0009] Preferably, the dynamic criterion identification steps are as follows: After normalizing the multi-feature composite fingerprint, it is input into the dynamic criterion neural network model, performs forward propagation calculation, outputs intermediate feature results through nonlinear activation function, and obtains classification probability values ​​in the output layer. A confidence threshold control mechanism is set up so that when the classification probability is higher than the preset threshold, a high-confidence recognition result is output; otherwise, the current event is marked as pending confirmation. The event type is determined based on the classification results. If it is a real short-circuit fault, protection control is triggered. If it is a non-fault disturbance, a delayed judgment is entered. If it is a pending confirmation state, it is recorded and continuously monitored. Each model input, recognition result, confidence level, and response information is recorded in the event history database for subsequent model training and accuracy evaluation.

[0010] Preferably, the adaptive protection control steps are as follows: Extract the classification results and corresponding confidence values ​​of the dynamic criterion neural network model. If it is determined to be a real short circuit fault and the confidence value is higher than the strong trigger threshold, immediately drive the solid-state device to conduct and link the electromagnetic drive mechanism to trip. If the disturbance is determined to be non-fault disturbance or the confidence level is between the strong trigger threshold and the weak hold threshold, then the delayed protection judgment stage is entered, continuous monitoring is started and the observation period is set. During the observation period, voltage and current waveforms are continuously collected, the latest feature matrix is ​​constructed, and the data is fed back to the dynamic criterion neural network model to perform secondary judgment. If the judgment result is a non-fault disturbance, the warning status is cleared; if the judgment result is a real short-circuit fault and the confidence level is higher than the strong trigger threshold, the trip control process is initiated.

[0011] The preferred real-time monitoring and secondary verification steps are as follows: When the dynamic criterion neural network model identifies a non-fault disturbance state or a state awaiting confirmation, a delayed protection judgment mechanism is activated, and a fixed-duration delayed judgment time window is set. Within the delayed judgment time window, voltage and current signals in the main circuit are continuously acquired at a fixed frequency. The current rise rate, voltage change rate, current duration, current peak value, and disturbance energy are extracted to construct a complete multidimensional feature matrix. The constructed multidimensional feature matrix is ​​input into the dynamic criterion neural network model for secondary recognition, and the confidence value corresponding to the recognition result is used to determine whether to execute the trip control action or restore the circuit breaker to the normal monitoring state. All sampled data, reconstructed feature matrix, identification results, and control response status within the delay judgment period are uniformly written into the event record database to complete a complete closed-loop delay protection judgment process.

[0012] The preferred closed-loop optimization and update steps are as follows: After completing the delay protection judgment, multi-feature composite fingerprints and response information are extracted based on the secondary recognition results of the dynamic criterion neural network model, training samples are constructed and stored in the sample database; The sample database is structured and classified, duplicate samples are deleted, and new samples are marked as recent events. Based on the newly added samples, the parameters and structure of the neural network model are optimized, and the network weights, biases and number of hidden layer nodes are fine-tuned by incremental learning. The optimized neural network model is used in the subsequent recognition process, and the changes in recognition accuracy, response latency and confidence before and after model optimization are recorded to form a traceable record of model performance evolution.

[0013] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention achieves high-resolution real-time perception of electrical waveforms through a state-aware acquisition module. Combined with multi-feature fusion modeling and a dynamic neural network criterion recognition mechanism, it constructs a refined classification capability for disturbance behavior. Based on this, an adaptive protection control mechanism and delayed discrimination logic work together to ensure that rapid tripping is only performed when a fault actually occurs, thereby preventing the system from being erroneously disconnected due to transient interference. Simultaneously, the system's real-time monitoring and model self-learning capabilities continuously update the recognition model based on changes in operating status, giving it adaptive evolution capabilities and long-term stability. Overall, this invention realizes the transformation of circuit breakers from "static judgment" to "dynamic intelligent recognition," improving the reliability, accuracy, and intelligence level of the device in complex power grid environments while ensuring power supply continuity, demonstrating significant technological advancement and engineering practical value. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0015] Figure 1 This is a schematic diagram of the solid-state semiconductor electromagnetic circuit breaker of the present invention. Detailed Implementation

[0016] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0017] This invention provides, for example Figure 1 The solid-state semiconductor electromagnetic circuit breaker shown includes a state sensing and acquisition module, a feature extraction and modeling module, a dynamic criterion recognition module, an adaptive protection control module, a real-time monitoring and secondary verification module, and a closed-loop optimization and update module. The state perception acquisition module collects voltage and current transient waveform signals in the main circuit of the solid-state semiconductor electromagnetic circuit breaker in real time, and constructs a multi-dimensional electrical feature matrix containing feature dimensions such as current rise rate, voltage change rate, current peak value, and current duration, which is used to form the basis for instantaneous state perception of the current working state of the circuit breaker. To accurately identify transient disturbances in the main circuit of a solid-state semiconductor electromagnetic circuit breaker, high-resolution real-time acquisition and data modeling of key electrical parameters under its operating state are required to construct a foundation of timely and complete electrical characteristic data. This process includes the following specific steps: During circuit breaker operation, transient current and voltage waveforms in the main circuit are acquired by current sampling units and voltage sampling units deployed on both sides of the main current path and busbar, respectively. The current sampling units can employ high-bandwidth Hall sensors or high-speed current transformers, while the voltage sampling units can utilize high-impedance voltage divider circuits combined with optically isolated measurement circuits to safely acquire high-voltage signals. The sampling frequency should meet the requirements for capturing sub-microsecond pulse changes; a sampling rate of no less than 10MHz is recommended to ensure the capture of the rapid and abrupt changes in the initial stage of the current pulse. Simultaneously, the data acquisition nodes must possess high-precision analog-to-digital conversion capabilities (such as a 12-bit or higher ADC) and low-latency communication capabilities to ensure high timing synchronization and integrity in each acquisition.

[0018] The acquired transient current and voltage waveforms undergo preprocessing, including baseline drift correction, power frequency filtering, noise suppression, and discrete-time window segmentation, to extract representative transient disturbance response segments. Based on this, feature quantities are calculated according to a preset time granularity (e.g., a 100μs window) to construct a multi-dimensional electrical feature vector. Feature dimensions include, but are not limited to, the maximum rate of current rise (di / dt), the rate of change of voltage (dv / dt), the maximum peak amplitude of the current waveform, the duration of the current disturbance, the time interval between the initial and peak points of the waveform, and the waveform integral value (reflecting energy characteristics). Each feature dimension is quantified using a unified physical unit and time scale to ensure comparability and universality in subsequent identification and discrimination processes.

[0019] After extracting the aforementioned feature dimensions, the dynamic correlation between current and voltage disturbances needs to be calculated. This process quantifies the coupling degree between voltage and current changes by simultaneously calculating indicators such as correlation coefficient, peak position of cross-correlation function, and length of signal co-rising segment within a sliding time window. This operation further enhances the data model's ability to perceive the nature of disturbances, enabling the system to not only rely on single-channel current characteristics for judgment but also to verify the rationality of current abrupt changes through voltage change trends. All quantified indicators will form a structured feature matrix, with each row representing a disturbance event and each column representing a specific electrical feature dimension, collectively constituting a complete electrical state description of the circuit breaker within the current operating cycle.

[0020] To achieve dynamic sensing of circuit breaker operating status, each completed electrical feature matrix should be recorded in real time and uploaded to the central processing unit for subsequent identification model calls and judgments. Simultaneously, auxiliary information such as timestamps and state transition information before and after event triggering should be retained during feature matrix construction as a timing reference for dynamic identification input. The generation of the feature matrix not only provides a high-quality input data foundation for subsequent transient pulse identification algorithms but also supports the normalized modeling of transient events under different operating scenarios through data structuring and standardization, exhibiting high adaptability and scalability.

[0021] Through the above steps, high-fidelity modeling and real-time data feature construction of the main circuit operation status of solid-state semiconductor electromagnetic circuit breakers can be achieved, laying the foundation for subsequent fault judgment and malfunction suppression strategies based on neural networks, and effectively improving the operational stability and identification accuracy of circuit breakers in complex disturbance environments.

[0022] This step aims to provide a data foundation and sensing support for the dynamic and accurate identification of the operating status of solid-state semiconductor electromagnetic circuit breakers. When operating under high voltage and light load conditions, the main circuit current signal of a solid-state semiconductor electromagnetic circuit breaker is inherently small and highly stable. However, once the system is affected by disturbances such as grid fluctuations, load switching, or switching transients, it is highly susceptible to generating transient pulse current signals with a sharp increase in amplitude but extremely short duration. These disturbance currents are remarkably similar in time-domain morphology to the fault current at the initial stage of a short circuit. Without in-depth analysis of their physical characteristics, they are highly likely to be misjudged as real faults, leading to malfunctions of the circuit breaker. Therefore, this step involves high-frequency, low-delay real-time sampling of the voltage and current signals of the circuit breaker's main circuit and extracting representative transient response features, including multiple dimensions such as the current rise rate, voltage change rate, current peak value, and current disturbance duration, thereby constructing a multi-dimensional electrical feature matrix. This matrix not only fully reflects the electrical state at the moment of the disturbance event, but also provides refined and structured feature inputs for subsequent fault discrimination models. This enables the circuit breaker to dynamically adjust its protection response strategy based on the sensing results, enhancing its identification accuracy and operational stability under nonlinear disturbance scenarios. Therefore, this data acquisition and feature construction step is not only the starting point of the protection decision-making process, but also a fundamental core link in ensuring the intelligent and adaptive operation capabilities of the circuit breaker.

[0023] The feature extraction and modeling module extracts key feature parameters of transient pulse current based on a multi-dimensional electrical feature matrix, calculates the numerical correlation index between voltage change and current change, and generates multi-feature composite fingerprint information representing the current transient pulse characteristics. To further enhance the ability to identify transient pulse current disturbances under high voltage and light load conditions, after constructing the electrical feature matrix, the collected feature data needs to be analyzed and structured in depth to generate representative and identifiable transient feature information. This process includes the following steps: Based on the constructed electrical feature matrix, data within each time period is segmented and extracted, focusing on the critical time window at the initial stage of pulse disturbances. Within this time window, the derivative of the current waveform is calculated to obtain its maximum rise rate, along with the peak current amplitude and duration of the current disturbance. For voltage channel data, its variation trend within the same time window is analyzed to obtain the maximum voltage change rate and direction. To improve data stability, all calculation results are standardized to ensure consistent dimensions and scale for feature data under different measurement conditions, laying the foundation for subsequent correlation modeling and feature fusion.

[0024] To investigate the relationship between voltage and current changes, various numerical correlation indices are calculated to reveal the coupling strength and response matching between the two. These indices include not only the traditional Pearson correlation coefficient but also multidimensional indicators such as the peak position of the cross-correlation function, the duration of zero-delay coordinated changes, and signal morphological similarity. These quantitative correlation indices help determine whether a current disturbance occurs in conjunction with a synchronous voltage disturbance, thus eliminating the risk of misjudgment caused by purely noise interference or isolated current abrupt changes. In actual calculations, the voltage-current signals are converted to a unified time axis, and a sliding window method is used for dynamic correlation matching to ensure that the results possess time-adaptive properties and noise suppression capabilities.

[0025] After extracting the key electrical parameters and quantifying the voltage-current relationship, feature vectors are formed using a unified data structure to construct a multi-feature composite fingerprint representing the current disturbance event. This composite fingerprint includes not only each current and voltage feature value but also its derived statistical features, such as average, variance, slope variation, and symmetry, to improve the overall accuracy and robustness of the identification. By integrating multiple physical layer information, mathematical layer structure, and statistical layer description, the sensitivity and accuracy of subsequent identification algorithms to abnormal patterns can be significantly improved. To enhance representational capabilities, each disturbance event's fingerprint vector is accompanied by a corresponding timestamp and an estimated disturbance energy value, ensuring that the fingerprint not only has lateral distinguishing capabilities but also vertical temporal localization capabilities.

[0026] The constructed multi-feature composite fingerprint information is temporarily stored in a fingerprint data buffer as the standard input format for subsequent classification and recognition, and for input to the neural network model. This feature fingerprint can not only be used for real-time decision recognition, but can also be synchronously written into a transient event history sample library as a data source for model iterative training. After the fingerprint is constructed, the fingerprint information is checked for completeness to ensure that there are no missing feature dimensions, no numerical anomalies, or out-of-bounds terms, so as to guarantee the stability and accuracy of subsequent recognition logic processing.

[0027] The core function of this step is to perform in-depth feature analysis and pattern extraction on the constructed multi-dimensional electrical feature matrix to form a structured feature expression that accurately describes the current transient pulse current event, thereby providing high-quality and discriminable data input for subsequent fault identification models. Under the high-voltage, light-load operation conditions of solid-state semiconductor electromagnetic circuit breakers, instantaneous pulse current disturbances exhibit high uncertainty and complexity. Their electrical behavior is not only characterized by a rapid and steep rise in the current waveform but is also often accompanied by a coordinated response of voltage disturbances. Traditional fault identification methods rely on a single feature dimension (such as current amplitude or rise slope) for judgment, which is easily affected by occasional disturbances, measurement noise, and load nonlinearity, resulting in a high false recognition rate. To address this issue, this step extracts multiple key feature parameters (such as current rise rate, voltage change rate, current peak value, and disturbance duration) from the time-series data of current and voltage, and further quantifies the numerical correlation between voltage and current changes (such as Pearson coefficient, cross-correlation function, and coordinated rise time segment) to construct a composite feature vector that comprehensively characterizes the disturbance behavior. These features not only capture the instantaneous physical characteristics of the pulse but also establish the interaction relationships between various physical quantities of the disturbance, greatly enhancing the dimensionality and reliability of the discrimination. Finally, the above information is fused to generate a multi-feature composite fingerprint, which serves as the identification input format. This provides a refined and reliable data foundation for intelligent identification algorithms such as neural networks, effectively distinguishing electrical disturbances from real faults and improving the overall circuit breaker's judgment accuracy and malfunction suppression capability in complex transient scenarios. Therefore, this step not only occupies a crucial position in the data processing flow but also plays a key role in enhancing the robustness and intelligent decision-making capabilities of the identification system.

[0028] The dynamic criterion recognition module inputs multi-feature composite fingerprints into the trained dynamic criterion neural network model, and performs feature comparison and pattern recognition operations based on historical transient pulse sample data to classify whether the transient pulse current is caused by a real short circuit fault. To accurately determine whether a transient pulse current originates from a real short-circuit fault, the previously constructed multi-feature composite fingerprint needs to be used as input. This fingerprint is then processed by a trained dynamic criterion neural network model for feature matching and pattern recognition, thereby achieving intelligent classification of the event's nature. The specific steps are as follows: The input data is prepared by encoding and normalizing the constructed multi-feature composite fingerprint according to the unified format used during model training. This ensures that the dimensions, order, and numerical range of the input data are consistent with the structural requirements of the neural network model. During normalization, the mean and standard deviation of each feature dimension from historical sample data are used for standard transformation, thereby enhancing the consistency and resilience of the input data. Furthermore, to improve the model's stability under boundary conditions, the input data also includes auxiliary features such as event timestamps, state transition information, and perturbation energy calculations, further strengthening the model's ability to perceive complex scenes.

[0029] A criterion structure for a neural network recognition model is established. This model is preferably a feedforward neural network with adaptive learning capabilities, comprising an input layer, several hidden layers, and an output layer. The number of nodes in the input layer is consistent with the dimension of the multi-feature composite fingerprint. The hidden layers employ nonlinear activation functions (such as ReLU) to enhance the model's nonlinear representation of the feature space. The output layer is a two-node structure, representing two labels: "real short-circuit fault" and "non-fault disturbance." During network training, the model is pre-supervised using large-scale historical transient pulse sample data covering various operating conditions and disturbance types. The network weights are continuously adjusted through backpropagation to ultimately obtain the optimal discrimination boundary for the current application scenario.

[0030] Nonlinear activation functions are mathematical functions that introduce nonlinear relationships into neural networks. Their main function is to break the linear combination limitations between each layer of the neural network, thereby enabling the model to fit nonlinear feature mapping relationships. In the transient current identification task of solid-state semiconductor electromagnetic circuit breakers, the input multi-feature composite fingerprint contains multiple dimensions such as current rise rate, voltage change rate, duration, and energy, which essentially constitute a feature space with strong nonlinear coupling relationships. If only linear activation functions (such as the identity function) are used in the model, no matter how the network structure is stacked, the overall result is still a linear mapping, which cannot learn complex pattern boundaries and therefore cannot effectively distinguish between real faults and non-fault disturbances. Therefore, introducing nonlinear activation functions in the hidden layers is particularly crucial. Commonly used nonlinear activation functions include ReLU (Rectified LinearUnit), Sigmoid, Tanh, and LeakyReLU. Among them, the ReLU function is particularly suitable for handling high-dimensional sparse features because it maintains linearity in the positive interval and outputs zero in the negative interval. It is computationally simple and effectively avoids the gradient vanishing problem. In implementation, the specific steps include: first, determining the number of output nodes for each hidden layer; second, performing a weighted summation on the weighted input of each neuron and adding a bias term; and third, passing the weighted result to a nonlinear activation function (such as ReLU) for transformation, with the output serving as the input for the next layer. In this way, the network can learn more complex feature boundaries and classification patterns, thereby achieving accurate differentiation between short-circuit faults and transient pulses in a high-dimensional perturbation fingerprint feature space.

[0031] The currently collected multi-feature composite fingerprint is input into the neural network model, and forward propagation computation is performed. Internally, the neural network processes the input features through weighted weighting and activation functions, outputting intermediate feature representations layer by layer, and obtaining the probability values ​​corresponding to the two types of fault labels at the output layer. A confidence threshold control mechanism is introduced into the model structure design: only when the output probability of a certain classification label is higher than a set threshold (e.g., 90%) is it considered a high-confidence classification result; if the output probabilities of all labels do not exceed the threshold, a clear judgment result is not output, and the current event is marked as "pending confirmation." This mechanism effectively avoids the model making incorrect classifications on marginal or noisy samples, thereby improving the overall reliability of the judgment.

[0032] The event type is initially determined based on the output of the neural network model: if the output is "real short-circuit fault," this determination is used as the trigger for subsequent protection strategies; if the output is "non-fault disturbance," the delayed protection determination process begins; if the output is "pending confirmation," the event is marked and subsequent waveform changes are continuously observed. For each model identification result, regardless of whether it is used as the trigger for control commands, its feature input, discrimination output, confidence level, and actual response must be recorded and written into the event history database for subsequent model retraining and accuracy evaluation.

[0033] This step involves inputting the extracted and constructed multi-feature composite fingerprint into the trained dynamic criterion neural network model. Leveraging the model's deep nonlinear feature extraction and pattern learning capabilities, it accurately classifies and determines the source of transient pulse current events, distinguishing whether they are caused by a real short-circuit fault, normal voltage disturbance, or non-fault disturbance. Under high-voltage, light-load conditions, the identification challenge for solid-state semiconductor electromagnetic circuit breakers lies in the extremely short waveform of transient current pulses on the time axis, which is highly similar in shape to the initial waveform of a real short-circuit fault. Relying on traditional rule matching or single threshold judgment is prone to misjudgment. This step constructs a neural network discrimination model, using feature vectors extracted from a large number of historical transient pulse events, such as current rise rate, voltage change rate, duration, and energy peak, as training samples. During the training phase, the backpropagation algorithm optimizes the network weights, enabling the model to automatically learn complex nonlinear feature mapping relationships, thus achieving high fault tolerance and strong generalization ability in the identification stage. When a new disturbance event occurs, its corresponding multi-feature composite fingerprint is used as input. Through forward propagation calculation of a neural network, the probability result of whether the disturbance event belongs to "real short circuit" or "non-fault disturbance" is output. Compared with the static threshold judgment method, this method does not rely on thresholds set by human experience and has dynamic, scalable, and adaptive characteristics, which can effectively cope with the diverse signal changes under different power grid structures, load types, and disturbance environments.

[0034] The adaptive protection control module executes a set of dual-threshold adaptive protection control logic based on the output of the dynamic criterion neural network model. If the identification result is a real short-circuit fault, it immediately drives the solid-state device to conduct and links the electromagnetic drive mechanism to perform the circuit breaker trip control operation; if the identification result is a non-fault transient disturbance, it enters the delayed protection judgment stage. To achieve both malfunction suppression and rapid fault isolation control of solid-state semiconductor electromagnetic circuit breakers, a set of dual-threshold adaptive protection control logic is needed, based on the classification results of the current disturbance event using a dynamic criterion neural network model. This ensures that the protection action possesses both fast response speed and accurate judgment. This control logic consists of the following steps, ensuring that the entire decision-making process is scientific, stable, and possesses adaptive adjustment capabilities.

[0035] Based on the output of the dynamic criterion neural network model, its classification label and corresponding confidence probability value are extracted. The output includes two dimensions: the first is the category judgment result of the disturbance event (i.e., "real short-circuit fault" or "non-fault disturbance"), and the second is the probability confidence value of the judgment result. The confidence value here is used to measure the model's confidence in its own output result, and is a normalized value obtained through the activation function output of the last layer of the model. To enhance the stability of the decision, a dual threshold judgment mechanism is introduced. The first threshold is a "strong trigger threshold," which means that when the confidence value is higher than this value (e.g., greater than 0.9), the system will unconditionally execute the control action corresponding to the current judgment result. The second threshold is a "weak hold threshold," which is used as the basis for entering the judgment mechanism. When the confidence value is between the two thresholds, the system does not immediately execute the action, but enters the next stage of continuous judgment process.

[0036] When the classification result is "true short-circuit fault" and its confidence level is higher than the strong trigger threshold, a rapid protection action is immediately executed. This initiates the conduction control logic of the solid-state power device and simultaneously triggers the electromagnetic drive mechanism to complete the main contact disconnection operation of the circuit breaker. This action chain employs a "solid-state device-led current limiting + mechanical mechanism-followed disconnection" approach. The former suppresses the short-circuit current within nanoseconds, while the latter completes the insulation isolation of the disconnection point within milliseconds, forming a double-layer protection barrier. Control commands are directly issued by the central processing unit. The solid-state device control signal is rapidly amplified by the drive circuit and sent to the power semiconductor gate. The electromagnetic mechanism provides the mechanical impact force required for disconnection through the energy storage unit. The entire circuit-breaking action is completed within milliseconds, effectively preventing further expansion of the short-circuit current and its impact on the system.

[0037] When the classification result is "non-fault disturbance" or the judgment confidence level is between the strong trigger threshold and the weak hold threshold, meaning the judgment result does not yet have sufficient credibility, the circuit breaker tripping action is not immediately executed; instead, the delayed protection judgment stage is entered. In this stage, the control logic does not trigger any hardware execution; it only initiates a continuous monitoring mechanism internally and sets a short observation period (e.g., 5-10 milliseconds). During this period, the voltage and current waveforms in the main circuit are dynamically and continuously sampled, and their latest feature matrix is ​​re-extracted to form an information flow of the current disturbance development trend. This processing method avoids over-response to non-persistent disturbances (such as surges, impulsive loads, and transient oscillations), thereby effectively reducing false tripping events caused by grid disturbances.

[0038] Real-time data generated during the delayed judgment phase is fed back to the aforementioned dynamic criterion neural network model for secondary classification. If the disturbance gradually subsides within the delay window and the classification result stabilizes as "non-fault disturbance," the system automatically clears the current protection warning state and returns to the normal operation monitoring state. If the judgment result tends towards "real short-circuit fault" and its confidence level rises above the strong trigger threshold, the trip control process is immediately initiated. This processing mechanism constructs a closed-loop control logic of "confirmation-delay-re-judgment," enabling the circuit breaker to proactively reserve dynamic decision space when facing boundary disturbance events. This avoids unnecessary power outages caused by atypical signals such as model misjudgment, data anomalies, or external interference, thereby improving the operational reliability and system stability of the circuit breaker in high-voltage, light-load scenarios.

[0039] This step aims to implement a dual-threshold adaptive protection control logic that balances responsiveness and stability based on the classification results of transient disturbance events by the dynamic criterion neural network model. This enables rapid tripping to disconnect the power supply in the event of a real fault, while suppressing malfunctions and delaying decisions in non-fault disturbance conditions, thus improving the reliability and intelligence of solid-state semiconductor electromagnetic circuit breakers in complex operating environments. Under high-voltage, light-load conditions, the waveform and timing characteristics of transient pulse currents are extremely similar to those of the initial stage of a short-circuit fault. Relying solely on the one-time output of the neural network model as the control trigger could easily lead to false trips in boundary judgment situations. Therefore, this step sets two different confidence thresholds: a "strong trigger threshold" and a "weak holding threshold," to achieve a dynamic, graded response to the model's output. When the identification result is a "real short-circuit fault" and the confidence level is higher than the strong trigger threshold, the system immediately executes tripping control. First, the solid-state device performs nanosecond-level current limiting, and then the electromagnetic drive mechanism performs mechanical disconnection of the main circuit, ensuring that the short-circuit current is promptly cut off, preventing further damage to the system and load. When the identification result is "non-fault disturbance" or the confidence level does not reach the strong triggering standard, the system enters a delayed judgment stage. By pausing protection actions and starting continuous monitoring, it further observes the development trend of the disturbance, providing more basis for subsequent tripping decisions. This control method breaks the rigid logic of traditional single threshold triggering. By introducing confidence quantification and hierarchical response strategies, it achieves flexible judgment and refined control of disturbance events. This not only significantly reduces the probability of malfunction but also enhances the adaptability of the circuit breaker under dynamic operating conditions, demonstrating strong engineering practical value and non-obviousness.

[0040] The real-time monitoring and secondary verification module continuously monitors the voltage and current waveform changes of the main circuit during the delayed protection judgment stage, and feeds the monitoring data back to the dynamic criterion neural network model in real time for secondary verification to confirm whether there is misidentification or state change. To further improve the accuracy of solid-state semiconductor electromagnetic circuit breakers in handling non-fault disturbances, avoid false tripping due to insufficient confidence in the primary classification, and ensure that protection actions are not delayed in response to real faults, a continuous monitoring and secondary verification mechanism is introduced in the delayed protection judgment stage. This mechanism can continue to collect the electrical characteristics of the main circuit and feed them back to the neural network model in real time for re-judgment even when the disturbance event is not yet fully clear, thereby achieving closed-loop optimization of dynamic confirmation and protection decision-making. This process includes the following steps: When the dynamic criterion neural network model outputs a classification result of "non-fault disturbance" or "pending confirmation state," a delayed protection judgment mechanism is activated, suspending all hardware control execution and setting a delayed judgment time window. This time window is generally set between several milliseconds and tens of milliseconds, and can be dynamically adjusted according to the system response capability and load characteristics. During this time period, the sampling unit continuously collects voltage and current signals in the main circuit of the circuit breaker at a high frequency in real time, ensuring that any minute trend changes during the disturbance are fully captured. The sampling frequency is maintained at the same level as in the previous sensing stage (e.g., 10MHz) to maintain the consistency of characteristic timing and the accuracy of data analysis.

[0041] The real-time acquired voltage and current signals undergo rapid preprocessing and feature reconstruction. Sliding segment analysis is performed according to a fixed time window to re-extract key parameters such as current rise rate, voltage change rate, current duration, peak value, and disturbance energy. These parameters are then combined with the changing trends within the current cycle (e.g., whether the current continues to increase or the voltage decreases rapidly) to construct a new multi-dimensional feature matrix. Compared to the composite fingerprint constructed during the initial identification, the feature data at this stage is more temporally continuous and trend-oriented, reflecting the evolution path of the disturbance. The key to this step is extracting "disturbance trend information," rather than just single-frame event morphological features, representing an upgrade from static judgment to dynamic tracking.

[0042] The newly constructed multidimensional feature matrix is ​​fed back into the previously trained dynamic criterion neural network model for secondary classification and recognition. During this secondary recognition, the neural network model utilizes the temporal difference patterns and disturbance trend characteristics learned during training to make a more robust judgment on whether the current disturbance has transformed into a real short-circuit fault, whether it has stabilized and subsided, or whether it is a persistent non-fault event. The secondary recognition is still based on the output confidence level: if the output result is "real short-circuit fault" and the confidence level is higher than the preset strong trigger threshold, the system immediately switches to a fast protection process, driving the solid-state device to conduct and triggering the electromagnetic drive mechanism to trip; if the recognition result is "non-fault disturbance" and the confidence level further increases, the delayed judgment process is automatically terminated and the circuit breaker is restored to normal monitoring status.

[0043] All sampling data, feature reconstruction results, model output results, and whether a trip was ultimately triggered within this delay judgment period are fully recorded and written to the event log database for subsequent model accuracy analysis and sample training data expansion. After recording is completed, the monitoring task for the standard operating cycle is restarted.

[0044] The purpose of this step is to further analyze the development trend of the disturbance by continuously monitoring the voltage and current waveforms of the main circuit when the initial identification result is "non-fault disturbance" or the confidence level is insufficient to trigger protection action. The real-time collected electrical parameters are then re-input into the dynamic criterion neural network model for a second judgment, thereby effectively verifying and correcting the initial judgment result, and ultimately confirming whether the nature of the disturbance has changed or whether the initial identification was a misjudgment. In the high-voltage, light-load operating environment faced by solid-state semiconductor electromagnetic circuit breakers, transient current disturbance signals are highly sudden and unstable. Some disturbances may initially exhibit atypical short-circuit characteristics. If an action decision is made immediately, it is easy to cause unnecessary tripping due to misjudgment, affecting the continuity of power supply. Therefore, this step designs a closed-loop control process of "delay-observation-feedback-secondary judgment". When entering the delayed protection judgment stage, the circuit breaker does not immediately perform a tripping action, but instead continuously samples the voltage and current waveforms at a high frequency and dynamically analyzes their rising trend, amplitude changes, oscillation behavior, and other characteristics, constructing a new multi-dimensional feature matrix through a sliding time window. Subsequently, these new features are input into the trained neural network model for a second identification and judgment. If the neural network outputs "real short-circuit fault" at this stage with a confidence level that meets the standard, it indicates that the fault, which was not explicitly expressed before, has gradually become apparent, and protection operations should be executed immediately. If the model outputs "non-fault disturbance" and the confidence level further increases, it indicates that the disturbance is decaying and is a transient phenomenon caused by normal power grid fluctuations or load impacts. The system will terminate the current protection judgment process and return to normal monitoring. Through this step, the circuit breaker not only has the ability to dynamically verify the judgment result but also realizes the perception and response to the development trend of disturbances. This improves the accuracy of judgment, effectively avoids false actions and protection lag, and enhances the intelligence, reliability, and engineering adaptability of the entire protection and control mechanism, demonstrating significant non-obviousness and creativity.

[0045] The closed-loop optimization and update module updates the parameter structure of the dynamic criterion neural network model based on the secondary verification results, and stores the multi-feature composite fingerprint corresponding to the new identification results into the sample database, thereby realizing the online self-optimization of the criterion model and the closed-loop update of sample information, thereby improving the accuracy of transient current disturbance identification under high voltage and light load conditions and the overall stability of circuit breaker operation. To continuously improve the accuracy and adaptability of solid-state semiconductor electromagnetic circuit breakers in identifying transient current disturbances under high-voltage, light-load conditions, especially in the context of changing operating environments, evolving disturbance modes, and increased load behavior complexity, an online self-optimization mechanism of a dynamic criterion neural network model is introduced after completing delayed protection judgment and secondary verification. This is combined with a closed-loop update strategy for the sample library to achieve continuous learning and enhanced recognition capabilities of the protection criteria. This process includes the following steps, constituting a complete adaptive optimization path: After completing the delay protection judgment phase and obtaining the second classification result from the neural network model, the accuracy of the classification result is compared with the actual system response to determine whether the triggering action is reasonable and whether there is any deviation in the current model output or room for accuracy improvement. If the classification is confirmed to be a high-confidence correct judgment, the judgment result can be regarded as a valid sample source for subsequent model structure optimization and parameter updates. Based on this, multi-feature composite fingerprint data corresponding to the disturbance event is extracted, including high-dimensional information such as current rise rate, voltage change rate, disturbance duration, waveform energy, trend indicators, and second judgment confidence, and timestamps and final execution action identifiers are added to construct structured training samples.

[0046] The aforementioned sample information is stored in the sample database. To ensure the representativeness and diversity of the samples used for model learning, the sample database content is structured and redundancy checked before storage to avoid excessive storage space consumption or model overfitting due to large amounts of repetitive data. The sample database is categorized and managed according to tags such as disturbance type, identification label, triggering action, and confidence level, facilitating the subsequent extraction of samples for incremental training as needed. In addition, newly added samples will be marked as "recent events" to distinguish them from historical stable samples, enabling the model to adapt more quickly to the evolution of disturbance characteristics under the current power grid environment during optimization.

[0047] This process triggers a parameter structure optimization procedure for the dynamic criterion neural network model. Based on the principle of "lightweight online training," this optimization employs incremental learning with small batches of new samples as input. By fine-tuning network weights and bias parameters, it guides the model to establish a response mechanism to novel perturbation features without compromising the original model's convergence. Model optimization utilizes pre-saved training strategies and learning rate control mechanisms to prevent model deviation or oscillation. In certain situations, if new samples contain previously unseen perturbation patterns, a structure update mechanism can be triggered, automatically adjusting the number of hidden layer nodes or the distribution of connection weights, thereby expanding the model's fitting ability in high-dimensional feature spaces.

[0048] After the model parameters are optimized, the updated model is reloaded and used in the next round of disturbance identification. Simultaneously, the system compares and analyzes the model identification results before and after optimization, monitoring key performance indicators such as the model's misjudgment rate, identification latency, and confidence level changes, and recording these in the model performance log to form a complete model iteration history and performance tracking record. This approach, combining model updates with sample library expansion, constructs a closed-loop feedback mechanism of "identification-verification-learning-optimization," which not only improves the adaptability and accuracy of the dynamic discrimination model under high-voltage, light-load disturbance environments but also ensures the judgment stability and intelligent evolution capability of the solid-state semiconductor electromagnetic circuit breaker during long-term operation.

[0049] This step aims to achieve online adaptive optimization and continuous enhancement of the recognition capability of the dynamic criterion neural network model in solid-state semiconductor electromagnetic circuit breakers. By analyzing and providing feedback on the secondary verification results during the delayed judgment stage, the model structure and parameter configuration are continuously adjusted. Simultaneously, multi-feature composite fingerprint samples corresponding to the new recognition results are incorporated into the sample database, forming a dynamic discrimination system with closed-loop learning capabilities. Under high-voltage, light-load conditions, transient current disturbances exhibit high diversity and uncertainty. The disturbance waveform is not only affected by various factors such as load characteristics, grid fluctuations, and equipment status, but its evolution may also show non-linear development over time. Recognition models with fixed structures and static weights are difficult to adapt to such changes in the long term, easily leading to decreased judgment accuracy or frequent malfunctions. Therefore, this step, through a "judgment-while-learning" mechanism, confirms the accuracy of the current output of the criterion model after each disturbance event's secondary verification. If the model output is reliable and consistent with the actual system response, the multi-feature composite fingerprint of that disturbance event is written into the database as a new sample. While accumulating data, the model fine-tunes its parameters or optimizes its structure based on these new samples, such as adjusting weight distribution, updating connection strength, and increasing hidden layer capacity, thereby gradually building the model's ability to respond to new disturbance patterns. Simultaneously, by recording sample labels, temporal characteristics, and system feedback results, the sample library becomes a high-quality data source for model training and evaluation, forming a continuous feedback channel between the identification logic and real-world operating conditions. This step significantly improves the circuit breaker's adaptability to complex disturbance environments, ensuring that the protection strategy can continuously evolve with changes in the operating environment, effectively suppressing the false positive rate, and improving overall operational stability. Compared to traditional static threshold judgment logic, this step is highly intelligent, self-evolving, and engineering-practical, demonstrating significant non-obviousness and technological innovation.

[0050] Through the above scheme, the solid-state semiconductor electromagnetic circuit breaker can effectively distinguish between transient pulse currents caused by non-fault disturbances and current surges caused by actual short-circuit faults under high-voltage, light-load operating conditions, significantly reducing the risk of misidentification and malfunction. The device achieves high-resolution real-time perception of electrical waveforms through a state-aware acquisition module, and combines multi-feature fusion modeling and a dynamic neural network criterion recognition mechanism to construct a refined classification capability for disturbance behavior. Based on this, an adaptive protection control mechanism and delayed discrimination logic work together to ensure that rapid tripping is only performed when a fault actually occurs, thereby avoiding erroneous power outages due to transient interference. Simultaneously, the system's real-time monitoring and model self-learning capabilities continuously update the recognition model according to changes in operating status, giving it adaptive evolution capabilities and long-term stability. Overall, this invention realizes the transformation of circuit breakers from "static judgment" to "dynamic intelligent recognition," improving the reliability, accuracy, and intelligence level of the device in complex power grid environments while ensuring power supply continuity, demonstrating significant technological advancement and engineering practical value.

[0051] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0052] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0053] It should be noted that, in this document, the use of relational terms such as "first" and "second" is merely for distinguishing one entity or operation from another, and does not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0054] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0055] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0056] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0057] The units described as separate components may or may not be physically separate. The components shown as units 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.

[0058] 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.

[0059] The above description is merely a specific embodiment 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.

[0060] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A solid-state semiconductor electromagnetic circuit breaker, characterized in that, It includes a state-aware acquisition module, a feature extraction and modeling module, a dynamic criterion recognition module, an adaptive protection control module, a real-time monitoring and secondary verification module, and a closed-loop optimization and update module. The state-aware acquisition module collects the voltage and current transient waveforms in the main circuit of the circuit breaker in real time and constructs a multi-dimensional electrical feature matrix that includes the current rise rate, voltage change rate, current peak value and duration. The feature extraction and modeling module extracts transient pulse current features based on the feature matrix, calculates the correlation between voltage and current changes, and generates multi-feature composite fingerprints. The dynamic criterion recognition module takes the composite fingerprint input and trains the dynamic criterion neural network model, performs pattern recognition, and determines whether it is a real short circuit fault. The adaptive protection control module, based on the output of the dynamic criterion neural network model, if it identifies a real short-circuit fault, executes dual-threshold adaptive protection control, drives the solid-state device to conduct and triggers the electromagnetic drive mechanism to trip; if it identifies a transient disturbance, it enters the delayed protection judgment. The real-time monitoring and secondary verification module continuously monitors the changes in main circuit voltage and current during the delay phase and feeds the data back to the neural network model for secondary verification. The closed-loop optimization and update module updates the neural network model parameters based on the secondary verification results and stores new fingerprint samples, realizing model self-optimization and sample closed-loop update, thereby improving the accuracy of transient disturbance identification and the stability of circuit breaker operation.

2. The solid-state semiconductor electromagnetic circuit breaker according to claim 1, characterized in that, The state-aware data acquisition steps are as follows: The real-time acquired transient waveforms of main circuit voltage and current are preprocessed to complete baseline drift correction, power frequency interference filtering, signal noise suppression, and fixed time window segmentation. Extract the current rise rate, voltage change rate, peak current amplitude, current duration, time interval from waveform start point to peak point, and waveform integral value within each preset time window. Based on the sliding time window, the correlation coefficient between current disturbance and voltage disturbance, the maximum position of the cross-correlation function, and the length of the synchronous rise segment of voltage and current are calculated to construct a structured multidimensional feature matrix. The generated structured feature matrix, along with the corresponding timestamps and state transition identifiers, is uploaded to the central processing unit as input for subsequent recognition models.

3. The solid-state semiconductor electromagnetic circuit breaker according to claim 1, characterized in that, The feature extraction and modeling steps are as follows: After the feature matrix is ​​constructed, the maximum current rise rate, peak current amplitude, current disturbance duration, maximum voltage change rate and its direction are extracted, and all features are standardized. Calculate the correlation indices between voltage and current changes, including Pearson correlation coefficient, peak position of cross-correlation function, duration of co-variance, and signal morphology similarity; Feature vectors are constructed based on eigenvalues ​​and statistics to generate multi-feature composite fingerprints that include timestamps and perturbation energy estimates. Multi-feature composite fingerprints are stored in a fingerprint data buffer and their integrity is verified, serving as the input for subsequent recognition models.

4. The solid-state semiconductor electromagnetic circuit breaker according to claim 1, characterized in that, The dynamic criterion identification steps are as follows: After normalizing the multi-feature composite fingerprint, it is input into the dynamic criterion neural network model, and forward propagation calculation is performed. The intermediate feature results are output through the nonlinear activation function, and the classification probability value is obtained in the output layer. A confidence threshold control mechanism is set up so that when the classification probability is higher than the preset threshold, a high-confidence recognition result is output; otherwise, the current event is marked as pending confirmation. The event type is determined based on the classification results. If it is a real short-circuit fault, protection control is triggered. If it is a non-fault disturbance, a delayed judgment is entered. If it is a pending confirmation state, it is recorded and continuously monitored. Each model input, recognition result, confidence level, and response information is recorded in the event history database for subsequent model training and accuracy evaluation.

5. The solid-state semiconductor electromagnetic circuit breaker according to claim 1, characterized in that, The adaptive protection control steps are as follows: Extract the classification results and corresponding confidence values ​​of the dynamic criterion neural network model. If it is determined to be a real short circuit fault and the confidence value is higher than the strong trigger threshold, immediately drive the solid-state device to conduct and link the electromagnetic drive mechanism to trip. If the disturbance is determined to be non-fault disturbance or the confidence level is between the strong trigger threshold and the weak hold threshold, then the delayed protection judgment stage is entered, continuous monitoring is started and the observation period is set. During the observation period, voltage and current waveforms are continuously collected, the latest feature matrix is ​​constructed, and the data is fed back to the dynamic criterion neural network model to perform secondary judgment. If the judgment result is a non-fault disturbance, the warning status is cleared; if the judgment result is a real short-circuit fault and the confidence level is higher than the strong trigger threshold, the trip control process is initiated.

6. The solid-state semiconductor electromagnetic circuit breaker according to claim 1, characterized in that, The real-time monitoring and secondary verification steps are as follows: When the dynamic criterion neural network model identifies a non-fault disturbance state or a state awaiting confirmation, a delayed protection judgment mechanism is activated, and a fixed-duration delayed judgment time window is set. Within the delayed judgment time window, voltage and current signals in the main circuit are continuously acquired at a fixed frequency. The current rise rate, voltage change rate, current duration, current peak value, and disturbance energy are extracted to construct a complete multidimensional feature matrix. The constructed multidimensional feature matrix is ​​input into the dynamic criterion neural network model for secondary recognition, and the confidence value corresponding to the recognition result is used to determine whether to execute the trip control action or restore the circuit breaker to the normal monitoring state. All sampled data, reconstructed feature matrix, identification results, and control response status within the delay judgment period are uniformly written into the event record database to complete a complete closed-loop delay protection judgment process.

7. The solid-state semiconductor electromagnetic circuit breaker according to claim 1, characterized in that, The closed-loop optimization and update steps are as follows: After completing the delay protection judgment, multi-feature composite fingerprints and response information are extracted based on the secondary recognition results of the dynamic criterion neural network model, training samples are constructed and stored in the sample database; The sample database is structured and classified, duplicate samples are deleted, and new samples are marked as recent events. Based on the newly added samples, the parameters and structure of the neural network model are optimized, and the network weights, biases and number of hidden layer nodes are fine-tuned by incremental learning. The optimized neural network model is used in the subsequent recognition process, and the changes in recognition accuracy, response latency and confidence before and after model optimization are recorded to form a traceable record of model performance evolution.

Images