An electrical fireproof current-limiting protection method and device based on artificial intelligence numerical control

Abstract

This invention relates to an electrical fire prevention and current limiting protection method and device based on artificial intelligence numerical control. The method extracts millisecond, second, and minute-level transient, periodic, and inertial features from current, voltage, and temperature signals through parallel lightweight multi-scale convolution. After residual processing, a multi-scale residual feature map is generated and uniformly input into an attention encoder to adaptively highlight early signs of operational condition drift. A gated recurrent unit is used to integrate the potential energy of the weighted residual sequence, achieving real-time quantification of progressive structural disturbances under operational conditions. Combined with trend judgment logic, drift precursors are identified, driving online parameter freezing of the main prediction network and dynamic calibration of the top-level risk mapping. After the model completes pre-adaptive adjustment, a lightweight adaptation package is generated and uploaded to promote online incremental evolution of the model both locally and in the cloud. This technology significantly improves the sensitivity of electrical circuit drift warning and the adaptive capability of model risk prediction in actual operation.

Description

Technical Field

[0001] This invention relates to the field of edge-side intelligent analysis and operating condition risk prediction technology in electrical fire protection and current limiting protection systems, and in particular to an electrical fire protection and current limiting protection method and device based on artificial intelligence numerical control. Background Technology

[0002] Current electrical fire protection and current limiting systems widely employ AI-based intelligent analysis modules for real-time monitoring and risk prediction of electrical circuit operation data. Their core functions typically include anomaly identification of signals such as current, voltage, and temperature; extraction of overheating and arcing features; and early warning and alarm processing of over-limit trends. With the increasing complexity of smart power distribution and new energy scenarios, edge-side devices need to continuously process massive amounts of non-stationary operating condition data on-site to achieve rapid response to potential fire risks and dynamic model optimization. Therefore, electrical fire protection and current limiting systems have introduced various risk identification algorithms based on deep learning or statistical modeling, gradually forming an intelligent early warning technology system characterized by multi-source signal coupling, distributed computing, and online learning.

[0003] Currently, mainstream operating condition monitoring and model adaptation technologies in the industry can be broadly categorized as follows: One category involves building large-scale risk analysis models in the cloud, periodically transmitting data back through edge devices, and centrally training and distributing model weights. Representative methods include cloud-edge collaborative solutions based on federated learning and knowledge transfer. Another category involves embedding lightweight anomaly detection modules locally. Common algorithms include statistical distribution tests under sliding windows, PCA reconstruction error, EWMA control charts, and simple RNN or LSTM sequence model adaptive adjustment. These solutions aim to adapt to the long-term drift of electrical line operating conditions over time through anomaly detection or online fine-tuning, thereby enhancing the system's robustness and timely response capabilities. In recent years, some high-end systems have introduced advanced methods such as few-sample incremental learning and dynamically gated neural networks to reduce retraining costs and improve edge real-time processing efficiency.

[0004] However, existing methods still face several limitations in practical applications. First, methods relying on statistical significance detection or distribution change measurement often lag in their response to operating condition drift. Drift is only identified when there is a significant jump in the histogram statistical characteristics of the input signal or the model output error, making it difficult to capture gradual, low-amplitude structured trend anomalies in a timely manner. Second, conventional solutions generally rely on preset thresholds, historical baseline samples, or cloud data to form a global reference. However, in actual power systems, different lines have different load structures, ambient temperatures, and operating habits, making it difficult to establish a universal benchmark, which can easily lead to misjudgments and omissions. Third, although some methods support online fine-tuning of model parameters, they usually require periodic full retraining or complex learning processes, consuming large amounts of computational resources and making it difficult to run efficiently in resource-constrained edge MCU hardware environments. In addition, some emerging federated optimization and knowledge transfer methods rely heavily on sufficient bandwidth and cloud resources, and their performance is limited in industrial scenarios with poor field communication and limited data feedback.

[0005] In complex electrical operating scenarios, operating condition drift does not always manifest as abrupt, distributed changes. More often, it presents as weak, continuous, and progressive structural disturbances with multi-temporal and spatial scale coupling characteristics. For example, temperature rise inertia caused by long-term load increases, phase coupling distortion from localized arcs, and feature fine-tuning due to chronic equipment aging are all difficult to identify promptly by existing high-threshold, global comparative detection methods. Traditional mechanisms have low sensitivity to these early-stage drift signals, leading to lags in model parameter updates and delayed risk indicator calibration. This can easily cause missed early intervention windows, affecting the overall effectiveness of electrical fire prevention and posing certain safety hazards.

[0006] Therefore, a novel technological approach is urgently needed that, without relying on historical benchmark models, global statistical distributions, manual experience thresholds, or cloud collaboration, can independently achieve early, highly sensitive detection of electrical condition drift on edge devices and enable real-time, selective adaptive updates of the model. This technology should possess the following characteristics: the ability to extract "budding" features of long-term trend shifts from multi-scale signal residuals; the establishment of dedicated monitoring channels for structural disturbances at different time scales; rapid local optimization of model parameters under limited edge computing power, avoiding full retraining and large-scale computation; and complete local closed-loop operation without relying on a large number of additional training samples or external teacher models, ensuring data security and real-time performance. Through this approach, it is expected to significantly improve the early detection and proactive calibration capabilities of electrical condition drift, providing fundamental support for high-reliability electrical protection and intelligent risk perception, and effectively addressing the shortcomings of existing technologies in early response to abnormal operating conditions. Summary of the Invention

[0007] This application provides an electrical fire prevention and current limiting protection method and device based on artificial intelligence numerical control, which aims to solve one of the problems or problems of the prior art mentioned in the background art.

[0008] This application provides an electrical fire protection and current limiting method based on artificial intelligence numerical control, specifically including: The digital signals of current, voltage, and temperature are acquired, and the three digital signals are respectively input into three multi-scale convolution branches with different receptive field lengths to generate transient residual maps, periodic residual maps, and inertial residual maps.

[0009] The transient residual map, periodic residual map, and inertial residual map are normalized and then input into the attention encoder. The residual sequence is weighted using a dynamic weight allocation mechanism to generate a weighted residual sequence.

[0010] A drift potential energy integrator with memory decay characteristics is constructed based on the weighted residual sequence. The drift potential energy integral value is generated by updating the hidden state by inputting the weighted residual sequence into the gated loop unit.

[0011] If the integral value of the drift potential energy exceeds a preset soft threshold and shows a monotonically increasing trend within the maintenance time window, it is determined that the drift budding stage has been entered and a self-trigger control signal is generated; otherwise, the current model parameter state is maintained and the acquisition of digital signals is resumed.

[0012] In response to the self-triggered control signal, the bottom feature extraction layer of the main prediction network is frozen, and the neuron cluster in the top risk mapping layer is activated, so that the drift potential energy integral value is used as a gating signal input to the neuron cluster.

[0013] Based on the gating signal, the activation threshold and sensitivity coefficient within the neuron cluster are automatically adjusted to perform a pre-calibration operation on the overheating probability prediction index and the arc occurrence tendency prediction index, generating calibrated risk mapping parameters.

[0014] The risk prediction logic of the local AI intelligent analysis core module is updated using the calibrated risk mapping parameters, and the occurrence time, duration, dominant scale, and parameter increment of this drift initiation event are packaged into a lightweight adaptation package.

[0015] During communication idle periods, the lightweight adaptation package is uploaded to the cloud platform for archiving, and the incremental parameter data contained in the lightweight adaptation package provides a reference for the knowledge distillation process of subsequent similar line models, so as to complete the online incremental learning closed loop of the edge-side model.

[0016] An electrical fire protection and current limiting device based on artificial intelligence numerical control, comprising: The signal acquisition module is used to acquire digital current signals, digital voltage signals, and digital temperature signals, and inputs the three digital signals into three multi-scale convolution branches with different receptive field lengths to generate transient residual maps, periodic residual maps, and inertial residual maps. The multi-scale convolution module is used to normalize the transient residual map, periodic residual map and inertial residual map and input them into the attention encoder. The residual sequence is weighted by a dynamic weight allocation mechanism to generate a weighted residual sequence. The attention encoding module is used to construct a drift potential energy integrator with memory decay characteristics based on the weighted residual sequence. By inputting the weighted residual sequence into the gated loop unit to perform hidden state updates, the drift potential energy integral value is generated. The gated loop unit module is used to determine whether the integral value of the drift potential energy exceeds the preset soft threshold and shows a monotonically increasing trend within the maintenance time window. If the condition is met, it is determined to enter the drift budding stage and generate a self-trigger control signal; otherwise, the current model parameter state is maintained and the acquisition of digital signals is returned to continue. The main prediction network module is used to freeze the bottom feature extraction layer of the main prediction network in response to the self-triggered control signal, and activate the neuron cluster in the top risk mapping layer so as to input the drift potential energy integral value as a gating signal to the neuron cluster. The calibration module is used to automatically adjust the activation threshold and sensitivity coefficient within the neuron cluster based on the gating signal, so as to perform a pre-calibration operation on the overheating probability prediction index and the arc occurrence tendency prediction index, and generate calibrated risk mapping parameters.

[0017] The electrical fire protection and current limiting protection method and device based on artificial intelligence numerical control provided in this application have the following beneficial effects: (1) In response to the problem of delayed early warning caused by operating condition drift in existing electrical fire protection current limiting systems, traditional methods often rely on historical data modeling, global distribution statistics, or cloud-based collaborative calibration. These methods are difficult to adapt to complex dynamic environments and have high response delays. In particular, they are prone to missed detections or false triggers in gradual change scenarios such as frequent load switching and slow temperature accumulation. The online adaptation mechanism proposed in this application regards changes in operating conditions as "gradual structural disturbances" at the signal level. By constructing multi-scale residual feature maps, it directly captures the deviation patterns of the original current, voltage, and temperature signals relative to short-term local averages. This achieves early perception of drift precursors without any historical benchmark comparison or external intervention. Compared with traditional detection methods based on sliding window KS test or PCA reconstruction error, this approach avoids dependence on the assumption of stable distribution, significantly improves the identification sensitivity and response speed under non-steady-state operating conditions, and advances the detection time of the drift initiation stage by hundreds of milliseconds, effectively overcoming the risk warning blind spot caused by model output lag.

[0018] (2) Furthermore, this scheme introduces a two-layer perception structure composed of an attention encoder and a gated recurrent unit, which can focus on and accumulate time segments with typical abnormal evolution characteristics such as continuous unidirectional shift, enhanced low-frequency oscillation, and phase coupling distortion from the residual sequence, and realize the energy accumulation and dynamic judgment of the disturbance trend through a "drift potential energy integrator" with memory decay characteristics. This design not only significantly reduces the overall computational complexity of the algorithm and ensures that all operations can be executed in real time on resource-constrained edge MCUs, but also avoids the poor adaptability problem caused by manually setting fixed thresholds. When the judgment enters the drift bud stage, the system only activates the neuron clusters in the top layer of the main prediction network, and uses the drift potential energy as a gating signal to dynamically adjust the activation threshold and sensitivity coefficient of the risk mapping layer, thereby completing local adaptive calibration without changing the underlying feature extraction capability or interrupting the main task inference, which significantly improves the robustness and continuous service availability of the AI ​​model under unknown conditions.

[0019] (3) This mechanism also constructs a lightweight closed-loop feedback path: the key parameters of each drift event are packaged into a "lightweight adaptation package" and asynchronously uploaded to the cloud platform for subsequent cross-device knowledge distillation and model optimization reference. This not only ensures the independent operation capability of the local system, but also provides valuable edge experience data for global model iteration. The entire process does not require additional labeled sample input, nor does it require the introduction of teacher models or federated learning frameworks. It completely eliminates the dependence on cloud feedback loops and truly realizes the "edge-driven, self-sufficient" intelligent evolution mode. In summary, this technical path not only significantly improves the accuracy of risk prediction and real-time response of electrical fire protection systems in complex real-world scenarios, but also enhances the long-term operational stability and deployment scalability of AI modules, providing practical technical support for building a highly reliable, low-latency, and adaptive edge intelligent protection system.

[0020] (4) Furthermore, this method decouples the multi-scale residual feature extraction, attention weighting, and state update process of the gated recurrent unit into a reusable lightweight operator sequence on the edge computing device, significantly reducing the matrix operation dimension and intermediate feature cache requirements in real-time signal processing. This shortens the calculation cycle of the drift potential energy integral value and reduces memory usage. Simultaneously, by selectively freezing the bottom feature extraction layer and activating the top neuron cluster for local parameter calibration only after the drift bud is determined, redundant operations of gradient backpropagation and weight updates across the entire network in traditional online learning are avoided. This effectively controls the processor's dynamic power consumption and improves the system's long-term continuous operation capability under conditions without external intervention. These improvements enable this method to maintain the sensitivity of early drift identification and the timeliness of risk prediction calibration while possessing higher hardware computing efficiency and lower resource consumption, making it suitable for large-scale deployments in edge electrical fire protection scenarios. Attached Figure Description

[0021] Figure 1 This is the main flowchart of an electrical fire prevention and current limiting protection method based on artificial intelligence numerical control.

[0022] Figure 2 This is a sub-flowchart of an electrical fire prevention and current limiting protection method based on artificial intelligence numerical control.

[0023] Figure 3 This is another sub-flowchart of an electrical fire prevention and current limiting protection method based on artificial intelligence numerical control. Detailed Implementation

[0024] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0025] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0026] It should be noted that multi-scale convolutional branches refer to a set of parallel convolutional operation channels with different receptive field lengths, each of which can independently extract local temporal features of the signal at its corresponding time scale. The attention encoder is a sequence feature selection structure based on a multi-head self-attention mechanism. It can perform global correlation calculation on the input residual sequence and focus on segments with drift precursor features by adjusting weights. The dynamic weight allocation mechanism is implemented internally within the attention encoder. Based on the statistical characteristics of the sequence itself, it calculates the importance weights of residual segments at different time points in real time, replacing the fixed weighting method.

[0027] The weighted result from the above processing is fed into a building block called the drift potential energy integrator. This integrator simulates the potential energy accumulation effect in a physical system, energizing the weighted residual characteristics of continuous time steps. By introducing memory decay characteristics, it selectively and gradually forgets historical information, ensuring that the latest trend is responded to while stale noise is suppressed. The core computational component of this integrator is a gated recurrent unit, a basic unit of a recurrent neural network containing input gates, forget gates, and output gates. It is responsible for recursively updating its internal state in the time dimension and linearly mapping the state information to a scalar value representing the degree of operational drift.

[0028] Furthermore, in this application, the main prediction network refers to a deployed, complete neural network model used for predicting overheating probability and arcing tendency. This network is divided into two layers: a bottom feature extraction layer responsible for extracting general waveform features from the raw current, voltage, and temperature signals; and a top risk mapping layer responsible for mapping the extracted features to specific risk probability indicators. Within this top risk mapping layer, there is also a special cluster of neurons. Unlike the fixed-parameter neurons in the network, their activation thresholds and sensitivity coefficients can be dynamically reconfigured online upon receiving a self-triggered control signal, thereby achieving pre-calibration of the risk prediction logic.

[0029] like Figure 1 As shown, this application provides an electrical fire prevention and current limiting protection method and device based on artificial intelligence numerical control, specifically including: S1: Acquire digital signals of current, voltage, and temperature, and input the three digital signals into three multi-scale convolution branches with different receptive field lengths to generate transient residual maps, periodic residual maps, and inertial residual maps.

[0030] S2: After normalizing the transient residual map, periodic residual map, and inertial residual map, input them into the attention encoder. Use the dynamic weight allocation mechanism to perform weighted calculation on the residual sequence to generate a weighted residual sequence.

[0031] S3: Construct a drift potential energy integrator with memory decay characteristics based on the weighted residual sequence. The drift potential energy integral value is generated by performing hidden state updates by inputting the weighted residual sequence into the gated loop unit.

[0032] S4: Determine whether the integral value of the drift potential energy exceeds the preset soft threshold and shows a monotonically increasing trend within the maintenance time window. If the condition is met, it is determined that the drift budding stage has been entered and a self-trigger control signal is generated. Otherwise, the current model parameter state is maintained and the acquisition of digital signals is resumed.

[0033] S5: In response to the self-triggered control signal, the bottom feature extraction layer of the main prediction network is frozen, and the neuron clusters in the top risk mapping layer are activated so that the drift potential integral value is used as a gating signal input to the neuron clusters.

[0034] S6: Based on the gating signal, the activation threshold and sensitivity coefficient inside the neuron cluster are automatically adjusted to perform a pre-calibration operation on the overheating probability prediction index and the arc occurrence tendency prediction index, and generate calibrated risk mapping parameters.

[0035] S7: Update the risk prediction logic of the local AI intelligent analysis core module using the calibrated risk mapping parameters, and package the occurrence time, duration, dominant scale and parameter increment of this drift initiation event into a lightweight adaptation package.

[0036] S8: During communication idle periods, the lightweight adaptation package is uploaded to the cloud platform for archiving, and the incremental parameter data contained in the lightweight adaptation package provides a reference for the knowledge distillation process of subsequent similar line models, so as to complete the online incremental learning closed loop of the edge side model.

[0037] Step S1: Acquire digital signals of current, voltage, and temperature. Input these three digital signals into three multi-scale convolution branches with different receptive field lengths to generate transient residual maps, periodic residual maps, and inertial residual maps. Specifically, this includes: S1.1: Acquire the digital signals of current, voltage, and temperature after analog-to-digital conversion, and map the three digital signals to three independent signal buffer queues to generate transient signal streams, periodic signal streams, and inertial signal streams to be processed.

[0038] The current digital signal, voltage digital signal, and temperature digital signal output by the signal acquisition module and processed by analog-to-digital conversion are used as the input conditions for this step.

[0039] The digital current signal is written into the transient signal buffer queue according to the sampling frequency and timestamp information. The ring storage structure ensures that the new sampled value can cover the oldest data, realizing the continuous update of the millisecond-level transient data stream.

[0040] The voltage digital signal is mapped to a periodic signal buffer queue, and the buffer length is set to an integer multiple of the number of sampling points corresponding to the load period, so as to retain multiple complete periodic signal segments for subsequent convolution feature extraction.

[0041] The temperature digital signal is mapped to an inertial signal buffer queue, the buffer depth is set to the number of sampling points required to cover minute-level temperature rise changes, and the data order is maintained by time indexing to reflect the inertial trend of long-term temperature changes.

[0042] During the enqueueing process of each buffer queue, a timestamp alignment operation is performed. The sampling deviation between different sensing channels is eliminated through interpolation or delay compensation mechanism, so that the three signals have strict time synchronization in the subsequent processing stage.

[0043] Through the above mapping and caching processing, the three input digital signals are converted into transient signal streams, periodic signal streams, and inertial signal streams to be processed, respectively, realizing the structured separation of multi-time-scale data streams and providing a synchronous and continuous original signal reference for the multi-scale convolution operation in step S1.2.

[0044] The transient residual plot corresponds to the millisecond-level signal waveform deviation structure (such as the residual deviation of current spikes or voltage surges), the periodic residual plot corresponds to the voltage / current phase distortion or harmonic fluctuation residual within the second-level load cycle, and the inertial residual plot corresponds to the cumulative deviation of the minute-level temperature field or long-term load ramp-up.

[0045] S1.2: Based on transient signal flow, periodic signal flow and inertial signal flow, lightweight convolution kernel groups with different receptive field lengths at the millisecond, second and minute levels are respectively called to perform sliding window convolution operation to generate transient original feature map, periodic original feature map and inertial original feature map representing local temporal patterns.

[0046] Based on transient, periodic, and inertial signal streams, when performing sliding window convolution processing using lightweight convolutional kernel groups with different receptive field lengths at the millisecond, second, and minute levels, each signal stream is segmented into fixed-length slices according to the window length of its corresponding receptive field, ensuring that the time interval covered by each slice matches the preset receptive field. The segmented signal data is input to the input of the convolutional kernel group, and a one-dimensional convolution operation is performed based on the kernel size and stride parameters to extract local temporal pattern features at each time scale. Channel aggregation is performed on the feature tensors output by convolution, merging the multi-channel features extracted by different convolutional kernels at the same time scale to form a single-scale original feature map. Edge effect compensation processing is applied to the generated transient, periodic, and inertial original feature maps respectively, by adding zero-padding or mirror padding values ​​to the boundaries of the convolution results to eliminate truncation distortion caused by the convolution window at signal endpoints. A weight normalization algorithm is used to scale-correct the convolution weights of each original feature map to ensure that the numerical distribution of the convolution output at different time scales is within a uniform range. Through the above processing method, the transient signal stream, periodic signal stream and inertial signal stream in the previous step are transformed into transient original feature maps, periodic original feature maps and inertial original feature maps that can accurately represent the local time series patterns of their respective time scales, thereby achieving the accuracy and comparability of multi-scale time series feature extraction.

[0047] S1.3: Perform local mean pooling operations on the transient original feature map, periodic original feature map, and inertial original feature map respectively to generate transient local reference maps, periodic local reference maps, and inertial local reference maps corresponding to each time scale.

[0048] S1.4: The transient original feature map and the transient local reference map are differentially processed using the point-by-point subtraction algorithm, the periodic original feature map and the periodic local reference map are differentially processed, and the inertial original feature map and the inertial local reference map are differentially processed to generate transient residual map, periodic residual map and inertial residual map that characterize the signal deviation from the structure.

[0049] The input transient raw feature map, periodic raw feature map, and inertial raw feature map, along with the corresponding transient local reference map, periodic local reference map, and inertial local reference map, constitute a paired dataset for differential processing.

[0050] The transient paired dataset is subjected to differential calculation based on the point-by-point subtraction algorithm. The transient original feature value at each spatial location is numerically subtracted from the transient local reference value at the corresponding location to form a transient residual matrix.

[0051] The difference calculation is performed on the periodic paired dataset based on the point-by-point subtraction algorithm. The pixel values ​​in the original periodic feature map are subtracted from the corresponding pixel values ​​in the local periodic reference map to form the periodic residual matrix.

[0052] The inertial pairing dataset is subjected to differential calculation based on the point-by-point subtraction algorithm. The original inertial feature map and the local inertial reference map are subtracted at the corresponding positions to form the inertial residual matrix.

[0053] The transient residual matrix, periodic residual matrix, and inertial residual matrix are mapped to transient residual map, periodic residual map, and inertial residual map, respectively, and time-series labels are assigned to them to preserve the characteristic deviation structure at each time scale.

[0054] By using differential processing, the original feature map and local benchmark map results from the previous step are transformed into multi-scale residual feature maps that can quantify the signal deviation structure, enabling sensitive capture of subtle signal deviations in subsequent early identification of drift buds.

[0055] For example, in an industrial power distribution branch with a rated voltage of 220V and a rated current of 30A, the signal acquisition module acquires digital signals of current, voltage, and temperature at a sampling frequency of 10kHz. The transient raw feature map extracted by the transient convolution kernel group contains 128×64 matrix data, the periodic raw feature map output by the periodic convolution kernel group is a 256×128 matrix, and the inertial raw feature map output by the inertial convolution kernel group is a 64×32 matrix. Local mean pooling operations generate corresponding transient local reference maps, periodic local reference maps, and inertial local reference maps, respectively. In the point-by-point subtraction operation, the transient residual value is calculated according to the formula...

[0056] in, These are the original eigenvalues. The local reference value is used; the calculation of periodic and inertial residuals also follows this formula. Taking the transient residual calculation as an example, the original eigenvalue is 2.35, the local reference value is 2.10, and the calculated transient residual R is 0.25, which is marked as a deviation of 0.25 at this position in the matrix. Through this point-by-point subtraction process, the transient residual map successfully captures the millisecond-level weak shift of the current waveform at the corresponding time scale, the periodic residual map reveals the subtle changes in voltage fluctuations within the load cycle, and the inertial residual map records the slow deviation of the temperature rise curve at the minute scale. The verification results show that this differential processing significantly improves the comprehensive sensitivity of the subsequent drift bud detection algorithm to multi-scale signal deviations in this scenario, ensuring the timeliness and accuracy of the model update triggering conditions.

[0057] S1.5: The generated transient residual map, periodic residual map, and inertial residual map are encapsulated into a multi-scale residual feature dataset in a unified format to serve as the standardized input object for the attention encoder.

[0058] Step S2: After normalizing the transient residual map, periodic residual map, and inertial residual map, input them into the attention encoder. Utilize a dynamic weight allocation mechanism to perform weighted calculations on the residual sequence, generating a weighted residual sequence. Specifically, this includes: S2.1: Obtain the transient residual map, periodic residual map, and inertial residual map as the original input data. Perform zero-mean unit variance standardization on the three residual feature maps respectively to eliminate the dimensional differences between signals at different time scales and generate a standardized residual feature matrix.

[0059] S2.2: Construct an attention encoder based on the standardized residual feature matrix, and use a multi-head self-attention mechanism to calculate the correlation scores between residual segments in consecutive time steps to generate an initial attention weight distribution map representing local temporal dependencies.

[0060] Based on the standardized residual feature matrix output by step S2.1, the construction interface of the attention encoder pre-deployed in the edge computing environment is called to uniformly load the standardized residual matrices of transient, periodic and inertial time scales into the input channel of the encoder to ensure that subsequent correlation calculations are performed under the same dimensional framework.

[0061] The number of heads and the embedding dimension of each head are set in the multi-head self-attention mechanism. The embedding dimension is dynamically allocated according to the feature complexity of the signal at each time scale, so that the residual segments at different time scales can be computed in parallel in their respective attention subspaces.

[0062] In the multi-head self-attention computation process, for the residual feature matrix at any time scale, a query matrix, a key matrix, and a numerical matrix are constructed. The original correlation score matrix between each time step is generated by matrix multiplication, and scaling is performed on the matrix to stabilize the gradient value.

[0063] The original correlation score matrix after scaling is normalized by applying the Softmax function along the time dimension to concentrate the weights of high-correlation segments and suppress low-correlation noise segments, thus forming a preliminary time-step correlation distribution.

[0064] The normalized correlation distributions of each head are merged, and a splicing operation is performed to form a global initial attention weight distribution map. After fusion, the numerical range of the weight distribution is adjusted through a linear mapping layer to adapt it to subsequent dynamic weight allocation operations.

[0065] Through the above construction and calculation methods, the standardized residual feature matrix of the previous step is transformed into an initial attention weight distribution map that characterizes the local temporal dependencies within continuous time steps, thereby achieving the ability to structurally focus on the precursor features of possible continuous unidirectional shifts and phase coupling distortions.

[0066] For example, in the edge AI analysis module of an electrical fire protection current limiting protection device, the embedding dimension of the transient residual feature matrix is ​​set to 64, the embedding dimension of the periodic residual feature matrix is ​​set to 32, the embedding dimension of the inertial residual feature matrix is ​​set to 16, the number of heads in the multi-head self-attention mechanism is set to 4, and the weight parameters of the query and key matrices are optimized through online iteration to ensure that the element values ​​of the relevance score matrix are distributed in the range of [-2, 2]. For the transient residual matrix, a scaling factor is set to... After calculating the relevance score matrix, it is Softmax normalized over time steps to obtain the probability distribution matrix, where the normalized weights of highly relevant segments remain stable between 0.25 and 0.35. The periodic and inertia residual matrices are processed by scaling factors corresponding to the embedding dimensions and then subjected to the same normalization process. Finally, the normalized weight distributions of the four attention points are concatenated to form a global initial weight distribution map. The peak weights in high-risk periods are significantly higher than those in low-risk periods. The output attention weight map can effectively focus on drift precursor features when used for subsequent dynamic weight allocation operations. Verification results show that the weight distribution map can maintain stable discrimination performance even under noise interference conditions.

[0067] S2.3: Perform dynamic weight allocation based on the initial attention weight distribution map. Use a soft maximum function to perform probability mapping on residual sequence segments with continuous unidirectional shift trend and phase coupling distortion mode to generate a normalized dynamic attention weight vector focusing on drift precursor features.

[0068] S2.4: Perform element-wise weighted fusion operations on the standardized residual feature matrix using the normalized dynamic attention weight vector to amplify weak but discriminative long-term trend shift patterns and generate a weighted residual sequence containing enhanced drift precursor information.

[0069] Under the input conditions, the expansion sub-step receives the normalized dynamic attention weight vector and the standardized residual feature matrix as processing objects. The weight vector has been generated by the soft maximum probability mapping of the preceding sub-step and has the ability to focus on continuous in-direction offset and phase coupling distortion modes. The standardized residual feature matrix contains residual signal components of three time scales: transient, periodic and inertial.

[0070] The normalized dynamic attention weight vector and the standardized residual feature matrix are multiplied element-wise on the same index dimension. A vectorized multiplication operator is used to implement the weighted processing of each matrix element according to the one-to-one correspondence between time step and scale channel. This multiplication operation is performed in parallel on the hardware using the SIMD instruction set to reduce latency.

[0071] The product result is used as an intermediate fusion matrix. Secondary normalization is performed based on the prior importance parameter of the time-scale channel. The weighted components are proportionally adjusted within each scale to ensure that the feature contributions of the multi-scale are balanced and that the data is not distorted due to excessive weight in a certain channel. This importance parameter is dynamically updated by the statistical results of long-term drift monitoring.

[0072] The normalized intermediate fusion matrix is ​​input into the linear combination operator, and the weighted components at different time scales are accumulated and fused in the feature dimension direction to form an enhanced residual feature vector. This vector can highlight the weak but discriminative long-term trend shift pattern in the signal and retain cross-scale phase coupling information.

[0073] A feature masking mechanism is used to mask redundant components of the enhanced residual feature vector. The masking configuration is based on the analysis results of historical drift events, and high-noise feature segments that are prone to misjudgment are removed, thereby generating a weighted residual sequence containing enhanced drift precursor information and having a high signal-to-noise ratio.

[0074] By employing the aforementioned element-wise weighted fusion and scale equalization processing methods, the standardized residual feature matrix from the previous step is transformed into a high-quality weighted residual sequence containing amplified information about drift precursors, thereby significantly enhancing the early drift signal and providing accurate input for the subsequent state update of the drift potential energy integrator.

[0075] For example, in a three-phase power distribution circuit edge protection device, the transient, periodic, and inertial residual feature matrices are 128×16, 64×8, and 32×4 dimensions, respectively, and the length of the normalized dynamic attention weight vector is consistent with the total number of elements in each matrix. During element-wise multiplication, hardware-level SIMD parallel instructions are used to perform multiplication of the 128×16 matrix with its corresponding weight vector, the periodic matrix with its weight vector, and the inertial matrix with its weight vector, achieving a parallel throughput of 512 elements per cycle. The scale importance parameters are set to 0.4, 0.35, and 0.25, respectively. After adjusting these parameters, the product results are accumulated and fused along the feature dimensions to obtain an enhanced residual feature vector of length 128. In this vector, the masking rule blocks the noise frequency bands (such as high-frequency spikes in transient response) determined by historical drift analysis. The signal-to-noise ratio of the output weighted residual sequence at the input of the subsequent drift potential energy integrator is significantly improved compared with the original residual input, enabling early identification of the nascent stage of the operating condition drift, which is about 150 milliseconds earlier than traditional sliding detection.

[0076] S2.5: Perform time-series smoothing filtering on the weighted residual sequence to remove high-frequency random noise interference and retain low-frequency structured drift bud signals, so as to output a high-purity weighted residual sequence that is finally used to drive the state update of the drift potential energy integrator.

[0077] like Figure 2 As shown, step S3: Construct a drift potential energy integrator with memory decay characteristics based on the weighted residual sequence. This is achieved by inputting the weighted residual sequence into a gated recurrent unit to perform hidden state updates, thereby generating the drift potential energy integral value. Specifically, this includes: S3.1: Obtain the weighted residual sequence after processing by the attention encoder, and initialize the hidden state vector and memory decay coefficient matrix of the gated loop unit based on the computing power constraint of the edge computing device, so as to construct the initial architecture of the drift potential energy integrator with memory decay characteristics.

[0078] A high-purity weighted residual sequence processed by an attention encoder is obtained as the initial input data source for the drift potential energy integrator, ensuring that high-frequency random noise has been removed and low-frequency structured drift information is retained. Based on the computing power constraints of the edge computing device, the dimension and parameter precision of the hidden state vector of the gated loop unit are determined to meet real-time requirements and control resource consumption. An initial value matrix of the hidden state vector is constructed, initialized using a zero vector or normalized reference vector, ensuring it maintains a consistent feature space mapping with the weighted residual sequence at the initial moment. A memory decay coefficient matrix is ​​established, where each element is jointly determined by the computing power load assessment result and the estimated rate of change of operating conditions, and its value is limited to between 0 and 1 to achieve controllable adjustment of the decay degree of historical state information. The hidden state vector and the memory decay coefficient matrix are bound to the same state update architecture, forming an initial architecture for the drift potential energy integrator with memory decay characteristics, providing adjustable long-term memory characteristics for the subsequent time-series calculation process of the input gated unit. This initialization process transforms the weighted residual sequence from the previous step into the input state of a gated cyclic unit that can operate efficiently in a limited computing environment, thereby enabling the stable startup of the drift potential energy integrator and its ability to capture long-term trends.

[0079] For example, in a residential branch circuit with a rated voltage of 220V and a rated current of 15A, the high-purity weighted residual sequence output after processing by the attention encoder has a length of 128 time steps. Each time step contains 3-dimensional feature values, corresponding to transient, periodic, and inertial drift precursor components, respectively. The maximum available memory of the edge computing device MCU is 512KB, and the computing power assessment results show that the processing budget for each time step does not exceed 0.2 milliseconds. Therefore, the hidden state vector dimension is set to 16, and the parameter precision is represented using fixed-point 16-bit. During the initialization phase, the hidden state vector is set as an all-zero matrix. The element values ​​of the memory decay coefficient matrix are calculated based on a load change rate of 0.05 (unit change rate per second) and a computing power load factor of 0.8, using the following formula:

[0080] in, The memory decay coefficient, For the rate of load change, The computational load factor is used. Substituting the load change rate of 0.05 and the computational load factor of 0.8 into the formula, the memory decay coefficient is 0.96. This coefficient is used to initialize the global element values ​​of the decay matrix, achieving a smooth decay of historical state information. After the above processing, the initial architecture of this drift potential energy integrator can operate stably under low latency conditions when the input gating unit is connected to the weighted residual sequence, and significantly improves the ability to perceive the cumulative trend of progressive structural disturbances under the working condition. The verification results show that the detection lead of the model in the drift initiation stage in this loop is improved by several time steps compared with the traditional method.

[0081] S3.2: Input the weighted residual sequence into the input gating unit in the initial architecture of the drift potential energy integrator according to the time step, and use the nonlinear activation function to filter the residual characteristics at the current time to generate the gating activation signal.

[0082] The high-purity weighted residual sequence, processed and smoothed by the attention encoder, is used as input data and sequentially fed into the input gating unit of the initial architecture of the drift potential energy integrator at set time steps. Element-wise multiplication of the residual feature vector at the current time step is performed according to the preset input gating weight matrix to form a weighted instantaneous residual response signal. A nonlinear activation function is used to transform the weighted residual response signal, limiting the output range and filtering out low-amplitude random disturbance components, thereby numerically enhancing the significance of the structural disturbance under the operating condition. An amplitude threshold is applied to the signal after activation function processing to remove invalid data points below the dynamic sensitivity threshold, retaining the effective signal components that characterize the instantaneous structural disturbance intensity at the current time step. A gating activation signal vector is constructed by combining the remaining effective signal components. This vector serves as the direct driving source for subsequent state updates of the drift potential energy integrator, establishing an effective mapping relationship between instantaneous disturbance characteristics and integral state variables.

[0083] Through the above stepwise processing method, the input weighted residual sequence is transformed into a quantitative gated activation signal with noise suppression effect at the current time step, so as to realize the accurate extraction and enhancement of the instantaneous structural disturbance intensity under the working condition.

[0084] For example, in an edge protector of an industrial power distribution line, the time step of the drift potential energy integrator is set to 10ms, the input gating weight matrix size is 32×32, the matrix element values ​​range from -0.5 to 0.5, the hyperbolic tangent function is selected as the nonlinear activation function, and the sensitivity threshold is set to 0.15. During execution, a high-purity weighted residual feature vector at a certain time step is obtained. After each element is multiplied by the weight matrix, the total amplitude is 0.42. After transformation by the hyperbolic tangent function, the output range is limited to -1 to 1, and the part below 0.15 is discarded, resulting in 18 effective signal components. The gating activation signal vector is calculated using the following formula:

[0085] in, For residual eigenvectors, For the input gate weight matrix, This is the gating activation signal vector. Substituting the actual values, we find that 18 elements in the gating activation signal vector are greater than 0.15, which constitute the subsequent state update unit of the effective driving source input drift potential energy integrator. The results show that under these parameter settings, the characteristic signals of instantaneous structural disturbances are significantly enhanced, and can provide a stable and highly discriminative input in the overall integral value calculation, effectively improving the accuracy of drift initiation identification.

[0086] S3.3: Based on the gating activation signal and the hidden state vector of the previous time step, a weighted operation is performed by applying a memory decay coefficient matrix through a forgetting gating mechanism to generate decayed historical state information that removes historical noise interference and retains long-term trend characteristics.

[0087] The gating activation signal generated by the input gating unit and the hidden state vector of the previous time step are loaded into the input buffer of the forget gating mechanism in the same processing thread.

[0088] Based on the preset value of the memory decay coefficient matrix, element-wise multiplication is performed on the hidden state vector to achieve exponential decay of historical state information.

[0089] The attenuated historical state information is fused with the current gating activation signal using matrix weighting operations, while maintaining a high weight allocation to low-frequency trend components during the fusion process.

[0090] The fusion results are subjected to noise suppression filtering to remove high-frequency random fluctuation components in order to retain the core pattern of long-term trend drift.

[0091] By generating historical state information after attenuation, long-term trend features are effectively preserved while historical noise interference is weakened, providing a clean historical background input for subsequent state cell update operations.

[0092] By using a forgetting gating mechanism and weighted calculation, the gating activation signal and historical state from the previous step are transformed into decayed historical state information containing long-term trend characteristics, thereby achieving smooth accumulation of progressive structural disturbances under operating conditions.

[0093] For example, in an industrial power distribution branch with a rated current of 32A, the drift potential energy integrator of the edge AI intelligent analysis module is set to a diagonally identical matrix with a diagonal element value of 0.85. The hidden state vector of the previous time step is set to 64 dimensions, and the amplitude range of the current gated activation signal is 0~1. When performing attenuation calculation, each dimension of the historical state value is multiplied by the corresponding coefficient of 0.85, and then fused with the current gated activation signal at a weight ratio of 0.6:0.4 to obtain the attenuated historical state information vector. For example, if the first dimension of the historical state value is 0.72 and the current gated activation signal component value is 0.55, the attenuation calculation formula is: (0.72×0.85)×0.6+0.55×0.4, and the calculated fused value is 0.5848. After being processed by a high-frequency noise suppression filter, this fused value stably retains the low-frequency trend signal characteristics. Test results show that, under the scenario of continuous monitoring for 20 minutes and load change frequency of less than 0.1Hz, the stability of the generated historical state information after decay is significantly improved, the drift trend characteristics are accurately preserved in the integral calculation, and the response delay of the final drift potential energy integral value is reduced to less than 150 milliseconds.

[0094] S3.4: Utilize the output gating unit to fuse the attenuated historical state information with the gating activation signal at the current moment, and perform a state cell update operation to generate the latest hidden state vector containing the cumulative amount of progressive structural disturbance under the current operating condition.

[0095] The decayed historical state information output by the forgetting gating mechanism and the gating activation signal at the current moment are obtained and used as two input vectors for fusion calculation.

[0096] The output gating unit in the drift potential energy integrator is invoked, and the historical state information after decay is proportionally adjusted through the gating weight coefficient, so that the long-term trend component and the short-term disturbance component are assigned different weights respectively.

[0097] The weighted and decayed historical state information and the unadjusted current gating activation signal are summed element-wise in the fusion operator of the output gating unit to form a fused state candidate vector.

[0098] A nonlinear activation function is applied to the candidate state vector of the fused state to perform amplitude shaping, so that small input disturbance signals can be smoothed without causing a sudden increase in the state vector, while maintaining the cumulative effect of large disturbance signals.

[0099] The processed fusion state candidate state vector is written into the state cell update pathway to replace the original hidden state vector, generating the latest hidden state vector containing the cumulative amount of progressive structural perturbation under the current operating condition.

[0100] Through the above processing method, the result of the previous step is transformed into new state data that simultaneously retains long-term trend and short-term disturbance information, thereby enabling the drift potential energy integrator to continuously accumulate and perceive the initial stage of operating condition drift.

[0101] For example, for an electrical branch with a rated current of 15A, the memory decay coefficient matrix of the drift potential energy integrator is set to retain the historical state with a weight of 0.92, and the amplitude of the current gating activation signal is 0.35. After fusion processing by the output gating unit, each element of the decayed historical state vector is multiplied by 0.92 and added to the corresponding element of the gating activation signal to form a candidate state vector. For example, a historical state element with an initial value of 0.50 becomes 0.50 × 0.92 + 0.35, or 0.81, after fusion. This candidate state vector, after activation by the hyperbolic tangent function, has a smoothed value of approximately 0.67, which replaces the original hidden state in the state cell update path. In the test, after performing this state update operation for 10 consecutive time steps, the cumulative value of the latest hidden state is significantly improved when dealing with continuous low-frequency drift, enabling the early output of the drift potential energy integral value for subsequent soft threshold determination, thereby enhancing the system's early warning capability for chronic load changes.

[0102] S3.5: Perform a linear mapping transformation on the latest hidden state vector to extract its scalar potential energy component, so as to generate a drift potential energy integral value that quantitatively reflects the degree of drift initiation of electrical circuit operating conditions.

[0103] The linear mapping operator is called on the latest hidden state vector obtained through the state cell update operation. The initial values ​​of the mapping weight matrix and the bias vector are determined by the statistical results of the model training phase to ensure computational stability under the constraints of edge computing resources.

[0104] Perform matrix multiplication on the mapping weight matrix and the latest hidden state vector, and generate an intermediate potential energy vector by accumulating the element-wise multiplications of the weights to preserve the contribution of each dimension of the state components to the drift potential energy.

[0105] A mapping bias vector is added to the intermediate potential energy vector for translation operation to correct potential zero-point drift and ensure that the output value is consistent with the preset threshold system.

[0106] The intermediate potential energy vector after translation correction is subjected to dimensionality reduction processing. A single scalar component is selected as the drift potential energy integral value. The dimensionality reduction method adopts global average pooling or specific position extraction mechanism to generate a scalar signal that can be used for subsequent threshold determination in a fixed manner.

[0107] The generated drift potential energy integral value is precisely truncated and quantized, retaining the necessary decimal places and converting it into an integer representation, thereby improving the processing efficiency in the edge storage and comparison stage. Through this processing method, the result of the previous step is transformed into a core technical indicator that quantitatively reflects the degree of drift in the electrical circuit operating condition, realizing efficient and low-latency early warning criterion output.

[0108] For example, on a distribution branch with a rated voltage of 220V and a rated current of 32A, the mapping weight matrix is ​​set to a size of 5×1, with element values ​​of 0.12, 0.08, 0.15, 0.10, and 0.05, and the bias vector is 0.02. The latest hidden state vector has a length of 5, and its components have values ​​of 0.35, 0.28, 0.40, 0.30, and 0.25, respectively. The scalar components of the intermediate potential energy vector are calculated using matrix multiplication, using the formula:

[0109] in This is the scalar potential energy component, i.e., the integral value of the drift potential energy. The weight matrix is ​​the first One element, The first hidden state vector Each component. Substituting the above values, we obtain... E = 0.12 × 0.35 + 0.08 × 0.28 + 0.15 × 0.40 + 0.10 × 0.30 + 0.05 × 0.25 = 0.042 + 0.0224 + 0.060 + 0.030 + 0.0125 = 0.1669. Adding the bias of 0.02, we get 0.1869. Quantizing this value and retaining three decimal places, we obtain the drift potential energy integral value of 0.187. During verification, when this value is compared with the soft threshold of 0.150 and meets the monotonically increasing trend judgment condition, the system generates a self-triggered control signal in advance. After the pre-parameter calibration of the model risk mapping layer, this control signal significantly improves the sensitivity and accuracy of overheat probability prediction and arc tendency prediction in the early stage of operating condition drift while keeping edge computing power consumption controllable.

[0110] like Figure 3 As shown, step S4: Determine whether the integral value of the drift potential energy exceeds a preset soft threshold and shows a monotonically increasing trend within the maintenance time window. If the condition is met, it is determined that the drift budding stage has been entered and a self-triggered control signal is generated; otherwise, the current model parameter state is maintained and the process returns to continue acquiring digital signals. Specifically, this includes: S4.1: Obtain the real-time drift potential energy integral value and the preset soft threshold output by the gated loop unit, and use numerical comparison logic to perform a size judgment operation on the real-time drift potential energy integral value and the preset soft threshold to generate an initial over-limit flag.

[0111] The preset soft threshold is not a fixed constant, but can be preset or dynamically adjusted based on the statistical distribution of historical drift potential energy integral values ​​or the offline calibration results of edge computing devices. For example, the preset soft threshold parameter can be automatically configured based on the rated current and the average potential energy integral under historical drift-free operating conditions, with an initial value set to 1.25.

[0112] In the process of determining the drift potential energy integral value, the real-time drift potential energy integral value output by the gated loop unit in the preceding steps is used as the main comparison object, and the preset soft threshold of the edge device is used as the judgment benchmark input. A numerical comparison logic module is constructed to complete the over-limit judgment. The drift potential energy integral value is input to the first port of the numerical comparison unit, and the preset soft threshold parameter is input to the second port. The numerical deviation between the two is obtained through differential calculation, and the sign of the deviation is used as the identification basis for the instantaneous over-limit state. The numerical deviation is used as a condition input, and the relational operation node is called to perform a greater than or less than relational judgment to form a Boolean judgment signal. The Boolean judgment signal is passed to the status register, which sets the initial over-limit flag when the judgment signal is true and resets and clears the initial over-limit flag when the judgment signal is false. To meet the accuracy requirements of the comparison operation, a floating-point comparison method is used to avoid over-limit misjudgment caused by quantization error of fixed-point arithmetic. The sensitivity of the judgment is controlled by adjusting the tolerance coefficient of the floating-point comparison. The judgment is performed using the following mathematical expression:

[0113] in, This is the real-time drift potential energy integral value. To preset the soft threshold parameter, The initial over-limit flag is represented by a square bracket structure, which indicates a Boolean mapping of the relational operation result. Through the numerical comparison logic and flag generation process described above, the drift potential energy integral value from the previous step is transformed into an initial over-limit flag characterizing the instantaneous over-limit state, thus realizing the pre-trigger condition for early identification of the initial stage of operating condition drift.

[0114] For example, in the edge fire protection device of a 10kV distribution circuit, the real-time drift potential energy integral value output by the drift potential energy integrator is 12.56 after floating-point encoding, and the preset soft threshold parameter is configured as 9.80. The numerical comparison logic module calculates the difference between the two as 2.76, determines the sign as positive, outputs the truth signal from the relational operation node, and sets the status register to generate an initial over-limit flag. Under the condition that the tolerance coefficient is set to 0.05, even if the drift potential energy integral value fluctuates within the range of ±0.05, the judgment result remains stable. This flag is then used as a prerequisite to drive the slope calculation operation in S4.2, ultimately triggering the confirmation of a monotonically increasing trend within a continuous time window. This allows the system to respond in advance before the operating condition drift significantly affects the output of the prediction model, thereby significantly improving the accuracy and reliability of early warning.

[0115] S4.2: Based on the condition that the initial over-limit flag is true, read the drift potential energy integral value sequence of historical moments within the continuous maintenance time window from the circular buffer, and use the sliding difference algorithm to perform slope calculation operation on adjacent data points of the drift potential energy integral value sequence to generate a first-order difference sequence.

[0116] The system retrieves the state indicated by the initial out-of-limit flag generated in step S4.1 being true. It then calls the circular buffer interface module to read the continuously stored historical drift potential energy integral value sequence under the time window constraint. Data points are extracted in chronological order using the buffer address indexing mechanism to ensure sequence integrity and timestamp continuity. A sliding difference operation is performed on the drift potential energy integral value sequence, defining the sliding window span as the length of two adjacent time steps. A numerical subtractor is used to calculate the slope value of adjacent data points in the sequence. The slope is calculated using the following mathematical formula:

[0117] in, For the i-th slope value, Let i be the integral value of the drift potential energy. The timestamps are used as the corresponding timestamps. The calculated slope values ​​are sequentially stored in a first-order difference sequence storage structure, ensuring that the sequence elements correspond one-to-one with the data point positions of the original integral value sequence. A floating-point precision control strategy is set to ensure that the slope calculation results are not misjudged due to loss of numerical precision. This first-order difference sequence is used as the input data for the subsequent trend consistency determination in S4.3.

[0118] By using the sliding difference calculation method, the drift potential energy integral value sequence from the previous step is transformed into a first-order difference sequence representing the direction of change, thereby realizing a quantitative description of the early structural characteristics of the drift trend under the operating condition.

[0119] For example, in a certain electrical fire protection current limiting edge device, when the initial over-limit flag is true, a circular buffer continuously stores a sequence of historical drift potential energy integral values ​​with a maintenance time window length of 10 seconds, and its time step is set to 0.5 seconds. The sequence read from the buffer is [2.15,2.20,2.28,2.39,2.47,2.55,2.63,2.70,2.78,2.85,2.92,3.00,3.08,3.17,3.25,3.33,3.40,3.48,3.55,3.63], corresponding to timestamps from 0 to 9.5 seconds. The slope is calculated using the above formula; for example, the calculation process for the third slope value is as follows: S3 = (2.39 - 2.28) / (1.0 - 0.5), resulting in 0.22. The first-order difference sequence calculated sequentially is [0.10, 0.16, 0.22, 0.16, 0.16, 0.16, 0.14, 0.16, 0.14, 0.14, 0.16, 0.16, 0.18, 0.16, 0.16, 0.14, 0.16, 0.14, 0.16]. This sequence is stored in the subsequent trend consistency judgment module. The verification results show that all slopes are positive, indicating that the drift potential energy integral value shows a monotonically increasing trend within the maintenance time window, triggering the subsequent adaptive update process, which significantly improves the system's real-time detection and response capability for the initial stage of continuous drift in the operating condition.

[0120] It should be noted that the maintenance time window refers to the sliding time interval used to continuously observe whether the integral value of the drift potential energy shows a monotonically increasing trend.

[0121] S4.3: Receive the first-order difference sequence, and use the all-positive-item verification logic to perform a non-negativity traversal detection operation on each element in the first-order difference sequence to generate a trend consistency judgment result.

[0122] S4.4: Combining the initial over-limit flag and the trend consistency judgment result, the logic AND operation mechanism is used to perform a joint state confirmation operation on the two to generate the final trigger decision signal.

[0123] The initial over-limit flag signal from S4.1 and the trend consistency determination result from S4.3 are obtained as input conditions.

[0124] The initial out-of-limit flag signal and the trend consistency judgment result are used to construct a joint state function by performing logic and operation mechanism. Bit operation is used to realize synchronous judgment at the binary level, so that the two can be effectively combined within the same clock cycle.

[0125] Before the logical AND operation is executed, the initial over-limit flag and the trend consistency judgment result are mapped to a unified Boolean value encoding matrix to ensure the consistency of the dimensions and the consistency of the numerical expression of the judgment conditions.

[0126] By using a Boolean encoding matrix to invoke bitwise AND operation instructions within the hardware instruction set, the joint state confirmation is achieved through the following formal expression:

[0127] in, The Boolean code value representing the initial out-of-limit flag. Boolean encoded value representing the trend consistency determination result, logical AND operator. This means that the output is true only if both inputs are true.

[0128] The generated joint state results are subjected to precise flag assignment operations. Results that meet the conditions are marked as the final trigger decision signal established in the early stage of the characteristic operating condition drift, and this signal is written into the trigger register of the control module for subsequent calls.

[0129] Through logic and computation mechanisms, the immediate state of exceeding the limit in the previous step and the result of continuous trend consistency are seamlessly integrated and transformed into a precise final trigger decision signal, so as to achieve rigor and real-time judgment of the drift initiation stage.

[0130] For example, in an electrical circuit monitoring scenario, the initial over-limit flag is generated by real-time comparison of the drift potential energy integral value of 2.85 and the preset soft threshold of 2.50, with a Boolean code value of 1. The trend consistency judgment result comes from the non-negativity detection of the differential sequence of integral values ​​over five consecutive time steps [0.05, 0.07, 0.04, 0.06, 0.08], with a Boolean code value of 1. Inputting both into the joint state function 1&1 results in an output of 1, corresponding to a true state. After writing this true state code into the trigger register, subsequent step S4.5 generates a self-triggering control command through pulse code modulation, driving step S5 to freeze the bottom layer of the main prediction network and activate the risk mapping layer, enabling the model to respond instantly to the drift bud stage. Performance verification shows that under these conditions, the system trigger latency is 45 milliseconds, significantly improving the decision response speed compared to traditional threshold judgment methods, while maintaining the accuracy of trigger judgment while keeping computational power consumption low.

[0131] S4.5: In response to the final trigger decision signal, generate a self-trigger control command using pulse code modulation technology, or generate a hold command when the final trigger decision signal is false, to drive subsequent execution of model parameter freeze and activation operations or return to the signal acquisition loop.

[0132] After receiving the final trigger decision signal output from the preceding steps and the current control interface status data of the system, a pulse code modulation instruction generation process is constructed based on the true / false condition of the trigger decision signal. The final trigger decision signal is used as the input gating of the pulse code modulator. By extracting the high-level duration and low-level interval of this signal, a set of control variables determining the pulse width and period parameters is established. Based on the response requirements of the edge execution module, the frequency setting logic of the pulse code modulator is invoked to map the period parameter to the carrier frequency of the output channel. The pulse amplitude is then set to the level threshold required for subsequent model parameter freezing and activation operations through the amplitude modulation unit. Using the time width adjustment logic, the pulse width parameter in the control variable set is dynamically adjusted to a preset range according to the floating threshold configuration table, and combined with the carrier frequency to form a complete pulse training sequence. An output buffer synchronization mechanism is adopted to bind the pulse training sequence to a designated execution channel. The sequence is then converted into a binary control code stream by the instruction encoder to match the internal interface protocol of the high-speed CNC switch and the AI ​​core module. The generated binary control code stream is injected into the control signal bus. If the final trigger decision signal is false, a hold instruction code is extracted from the instruction template library, and the encoding and injection operations are performed in the same manner as described above, but without the pulse width adjustment step. This maintains the current model parameter state and returns to the signal acquisition loop. Through pulse code modulation technology and conditional branch control, the trigger decision signal from the previous step is accurately converted into a self-triggering control instruction or hold instruction with specific timing width and amplitude characteristics. This enables precise driving of subsequent model parameter freezing and activation operations, ensuring the system's response timing stability and control accuracy in the early stages of operational drift.

[0133] For example, in an electrical fire protection current limiting system, if the real-time drift potential energy integral value exceeds the soft threshold and shows a monotonically increasing trend within a 5-second maintenance window, the final trigger decision signal is set to true. The pulse code modulator receives this signal and extracts state information with a high-level duration of 2ms and a low-level interval of 3ms, forming a periodic variable of 5ms. The carrier frequency is set to 200Hz, and the amplitude is set to 3.3V to match the gating level requirements of the AI ​​core module. The pulse width is adjusted to 2.5ms by the adjustment logic and combined with the carrier frequency to generate a complete training sequence. During the encoding stage, the frequency and width parameters are combined into a binary instruction code, which is injected into the execution channel through the output buffer synchronization mechanism to drive the freezing of the bottom feature extraction layer of the main prediction network and the activation of the neuron clusters of the top risk mapping layer. In another operating condition, if the drift potential energy integral value is lower than the soft threshold and the trend consistency is determined to be false, the final trigger decision signal is set to false. The hold instruction code is extracted from the template library, and the amplitude and carrier frequency remain at their default values. After the instruction is injected through the control bus, the system returns to the signal acquisition loop. The pulse period calculation formula used in the above scenario is:

[0134] in, The duration of the high level. The low-level interval time is 3ms. In this embodiment, the high-level duration is 2ms, the low-level interval time is 3ms, and the period T is calculated to be 5ms. The output pulse sequence achieves precise control of model freezing and activation within the execution unit, significantly reducing system response delay and greatly improving parameter modulation efficiency in the early stages of drift.

[0135] Step S5: In response to the self-triggered control signal, freeze the bottom feature extraction layer of the main prediction network and activate the neuron clusters in the top risk mapping layer, so that the drift potential integral value is used as a gating signal input to the neuron clusters. Specifically, this includes: S5.1: Obtain the self-triggered control signal generated by the previous steps and the complete set of weight parameters of the current main prediction network. Based on the self-triggered control signal, perform gradient blocking operation on the bottom convolutional and pooling layers of the main prediction network, and set the weight update flag of the bottom feature extraction layer to a locked state to generate a frozen bottom feature extractor with fixed feature extraction capability, ensuring that the general electrical waveform feature extraction logic does not deviate during subsequent processing.

[0136] The self-triggered control signal generated by the preceding steps is used as the trigger source for the freeze operation, and the complete mirror data of the weight parameter set is called at the hardware register level of the main prediction network to ensure that the establishment of the freeze state has global consistency.

[0137] The self-triggered control signal is input into the gradient propagation path of the bottom convolutional and pooling layers. The gradient blocking operator is used to intercept the backpropagation link, so that these two types of layers no longer receive any gradient update signals in subsequent iterations.

[0138] Based on the state machine of the frozen execution unit, the weight update flag of the underlying feature extraction layer is switched from a writable state to a locked state, and the underlying feature extraction logic is thus switched to read-only mode and maintained at the current weight distribution.

[0139] After the locked state configuration is completed, the initial feature template of the underlying convolution and pooling operations is called to compare the consistency of the feature output before and after freezing, ensuring that the general electrical waveform feature extraction logic does not shift during the freezing period.

[0140] By combining gradient blocking and weight flag locking, the trigger signal from the previous step is transformed into a frozen-state low-level feature extractor with fixed feature extraction capabilities, thereby ensuring the stability of basic features and the reliability of prediction performance during subsequent local model parameter adjustments.

[0141] For example, an edge-side electrical fire protection current limiting model generates a self-triggered control signal when it detects a drift potential energy integral value of 0.842 that exceeds the soft threshold of 0.75. This signal is sent to the gradient path of the bottom convolutional layer to perform a gradient blocking operation, ensuring that the backpropagation gradient at this layer is always equal to 0. The weight update flag of the pooling layer is manually set to the lock status code 1, corresponding to register address 0x1A7C. The weight matrix dimension of the bottom feature extractor is 64×64, and the weights are calculated using the feature similarity formula before and after freezing. Verify consistency, where and These are the feature vectors before and after freezing, respectively. The calculated result is 0, indicating that the freezing operation did not change the features. The execution results show that during the subsequent parameter adjustment of the neuron cluster in the top-level risk mapping layer of the model, the output of the bottom-level feature extraction is stable and without drift. The risk prediction model's ability to capture arc occurrence tendency signals is significantly improved, and the prediction delay is maintained within 190 milliseconds, meeting the performance indicators of real-time fire protection.

[0142] S5.2: Read the neuron connection matrix of the top-level risk mapping layer of the main prediction network, locate the neuron clusters responsible for outputting overheating probability and arc tendency in the top-level risk mapping layer based on the output dimension of the frozen-state bottom-level feature extractor, switch the state register of the neuron clusters from dormant mode to active mode, generate neuron cluster instances in the state to be adjusted, and prepare for parameter modification to receive dynamic gating signals.

[0143] The complete neuron connection matrix of the top-level risk mapping layer of the main prediction network is obtained as the initial input. Combined with the output dimension information of the frozen-state bottom-level feature extractor, matrix dimension matching analysis is performed on the top-level risk mapping layer to locate the risk prediction functional substructure that corresponds one-to-one with the output dimension. A function label filtering algorithm based on the dimension matching results is executed to perform cross-validation of function labels on all neuron clusters, filtering out neuron clusters unrelated to the overheating probability output channel and the arc tendency output channel, retaining only target neuron clusters whose function labels contain both temperature risk and arc risk features. The state register of the located neuron cluster is loaded into the register access buffer, and the state switching control logic is called to change its mode bit from the dormant state flag to the active state flag. Simultaneously, the activation flag bit is modified in the register mapping table to ensure consistency between the physical register and software register states. The parameter configuration structure of the activated neuron cluster instance is read, and the parameter variability flag bit and dynamic gating signal receiving channel are preset, enabling the neuron cluster to receive gating signal input and adjust its internal parameters accordingly. The state register activation completes the transition of the neuron cluster from dormant to the adjustable state, preparing it for subsequent dynamic parameter modifications. By switching register states and pre-setting parameter variability, a controllable parameter link is established between the frozen bottom-level feature extractor output channel from the previous step and the neuron cluster in the top-level risk mapping layer that is in an unadjusted state, thereby realizing the construction of the local dynamic adaptability of the risk prediction output channel.

[0144] For example, in the core module of edge AI intelligent analysis, the risk mapping layer at the top of the main prediction network contains 256 neurons distributed across four functional clusters. Cluster 1 is responsible for overheat probability prediction, Cluster 2 for arc tendency prediction, and Clusters 3 and 4 are redundant detection and auxiliary feature clusters. The output dimension of the frozen-state bottom-level feature extractor is 128. The functional label filtering algorithm finds in the connection matrix that the input mappings of Clusters 1 and 2 both cover all 128-dimensional general feature vectors, and the label matching items are temperature risk and arc risk. After the register access buffer loads the state registers of Clusters 1 and 2, their mode bits are switched from 0 (dormant state) to 1 (active state) via control logic, and the register mapping table is updated synchronously. The variability flag in the parameter configuration structure is changed from 0 to 1, opening the dynamic gating signal receiving channel and allowing the input of normalized gating coefficients. For this scenario, the gating coefficient is calculated as follows:

[0145] in, This is the real-time drift potential energy integral value. This is the drift-free reference potential energy value. The maximum reference value for drift potential energy set for the system. This is the gating coefficient. This coefficient k is input to the first and second clusters in the active state via a hardware gating signal bus, initiating the dynamic parameter adjustment preparation process. In the verification test, after receiving the gating coefficient mapped from the drift potential energy integral value, the active neuron clusters can initialize their internal activation threshold and sensitivity coefficient within 50 milliseconds. This enables the risk prediction network to respond to the emerging trend of long-term drift in the next cycle, increasing its ability to detect overheating and arc risks in advance, and achieving rapid model adaptation in the early stages of operating condition drift.

[0146] S5.3: Obtain the drift potential energy integral value reflecting the degree of accumulation of progressive structural disturbance under the working condition, map the drift potential energy integral value to the normalized gating coefficient, and use the normalized gating coefficient to perform a linear transformation operation on the input weighting path of the neuron cluster in the state to be adjusted, so as to generate a gating modulation signal carrying the drift intensity information under the working condition, and establish a direct mapping relationship between drift energy and the activation level inside the model.

[0147] S5.4: Receive a gated modulation signal carrying information about the intensity of the operating condition drift. Based on the amplitude characteristics of the gated modulation signal, perform dynamic reconstruction calculations on the bias vector and activation slope parameters inside the neuron cluster, adjust the excitation threshold and response sensitivity of the neuron cluster, and generate a calibrated neuron cluster with pre-calibration characteristics, so that the risk prediction logic can respond in advance to weak but continuous operating condition drift trends.

[0148] The system receives the gated modulation signal carrying the drift intensity information from the preceding step S5.3 and uses this signal as the trigger condition for dynamic parameter adjustment of the neuron cluster in the top-level risk mapping layer. Based on the amplitude characteristics of the gated modulation signal, the parameter calculation unit is invoked to perform amplitude scaling on the bias vector within the neuron cluster. The normalized value of the gated signal is used as a scaling factor on the existing bias components to form a dynamic bias vector directly corresponding to the amplitude of the drift potential energy integral. The dynamic bias vector is then subjected to element-wise nonlinear reconstruction processing, using a hyperbolic tangent function to compress the bias components to a specified parameter domain range, ensuring the calibration process is stable and does not produce saturation distortion. The activation slope calculation unit is invoked, using the amplitude distribution curve of the gated modulation signal as input, to calculate the activation function slope correction factor. This factor is then applied element-wise to the activation function slope parameters of each risk output node in the neuron cluster, forming a dynamic slope vector with enhanced drift trend sensitivity. Based on the latest states of the dynamic bias vector and dynamic slope vector, an internal parameter refresh operation is performed on the neuron cluster. The adjusted parameters are injected into the forward propagation path of the neuron cluster, generating a calibrated neuron cluster instance with pre-calibration characteristics. This enables the neuron cluster to respond in advance and amplify weak but persistent operating condition drift signals during real-time inference. Through the joint adjustment of dynamic bias and slope parameters, the gated modulation signal from the previous step is transformed into adaptive threshold and sensitivity configuration data within the neuron cluster, thereby enhancing the proactive response capability of the risk prediction logic in the early stages of drift trends.

[0149] For example, in an industrial power distribution line protection device, the normalized amplitude of the received gated modulation signal is 0.68. The parameter calculation unit uses this value as a scaling factor to perform a scaling operation on the original bias vector [0.12, 0.15], resulting in a dynamic bias vector [0.0816, 0.102]. The hyperbolic tangent function is then applied to the above dynamic bias vector. <tanh>The output parameters are transformed and restricted to the [-1,1] domain, resulting in a compressed bias vector [0.0815, 0.1018]. The slope calculation unit is activated to read the amplitude distribution curve of the gated modulation signal and calculates a slope correction factor of 1.25. This factor is applied to the original slope values ​​[0.85, 0.92] of the two types of risk output nodes to generate a dynamic slope vector [1.0625, 1.15]. The above dynamic bias and dynamic slope are injected into the forward propagation path of the neuron cluster. When the inference is performed on the local MCU, when the change amplitude of the temperature feature vector is lower than the conventional statistical threshold, the risk prediction output already shows an increased overheating probability and an enhanced arcing tendency. Verification shows that after calibration, the response delay of the neuron cluster to the fault probability in the early stage of the drift trend is significantly reduced, and the adaptability is greatly improved.

[0150] S5.5: Integrates the frozen-state low-level feature extractor with the calibrated neuron cluster with pre-calibration characteristics, and reconnects the output of the calibrated neuron cluster to the risk indicator output interface of the main prediction network to complete the dynamic reorganization of the local topology, so as to generate an online updated version of the risk prediction model that adapts to the current working condition drift, and realizes the real-time adaptive evolution of the edge AI core module under the condition of no sample back transmission.

[0151] Step S6: Based on the gating signal, automatically adjust the activation threshold and sensitivity coefficient within the neuron cluster to perform a pre-calibration operation on the overheating probability prediction index and the arc occurrence tendency prediction index, generating calibrated risk mapping parameters. Specifically, this includes: S6.1: Obtain the drift potential energy integral value output from the previous step as the gating signal input, and use the linear mapping function to perform normalized scaling transformation on the gating signal to generate standardized gating weight coefficients that characterize the structural disturbance intensity under the current working condition, providing a quantitative benchmark for the subsequent dynamic adjustment of neuron parameters.

[0152] S6.2: Based on the standardized gating weight coefficient, nonlinear offset calculation is performed on the static bias term of the first type of neuron responsible for overheat probability prediction in the neuron cluster to generate a dynamic overheat prediction activation threshold with operating condition adaptive characteristics, thereby eliminating the reference drift error caused by the inertial change of line temperature rise.

[0153] The standardized gating weight coefficients, representing the intensity of structural disturbance under the current operating condition, are used as input. The static bias term of the first type of neuron responsible for overheat probability prediction in the neuron cluster is located and its current fixed value is loaded as the baseline bias parameter. Based on the requirement for dynamic adaptive adjustment, the standardized gating weight coefficients are input into the nonlinear offset calculation module. This module has a preset differentiable activation function structure, such as a hyperbolic tangent or sigmoid function, to achieve a smooth bias change curve. During nonlinear modulation, a modulated input signal is generated according to the multiplicative relationship between the gating weight coefficients and the baseline bias parameter. This signal is then converted into a bias increment value using a nonlinear mapping, thus forming differentiated adjustment amplitudes under different disturbance intensities. The modulated bias increment value is applied to the baseline bias parameter for superposition calculation, generating a dynamic overheat prediction activation threshold with operating condition adaptive characteristics. This threshold can automatically correct the model's internal judgment threshold when temperature rise inertia causes signal baseline shift. The generated dynamic activation threshold is written into the parameter register of the first type of neuron, making it effective immediately in subsequent risk probability prediction. By using this nonlinear offset calculation method, the standardized gating weight coefficients are mapped to the dynamic overheat prediction activation threshold, thereby achieving the expected technical effect of eliminating the reference drift error caused by the inertial change of line temperature rise.

[0154] S6.3: Using standardized gating weight coefficients, gradient scaling update processing is performed on the connection weight matrix of the second type of neurons in the neuron cluster responsible for predicting the tendency of arc occurrence, so as to generate dynamic arc prediction sensitivity coefficients with high frequency feature sensitivity, thereby enhancing the model's ability to capture weak arc fault features.

[0155] Based on the standardized gating weight coefficient, this coefficient is bound to the connection weight matrix of the second type of neurons in the neuron cluster that are responsible for predicting the tendency of arc occurrence, forming a target weight set that can be updated for parameters.

[0156] The baseline gradient calculation before gradient scaling is performed on the target weight set. The original gradient matrix of the current arc prediction branch under real-time signal input conditions is generated by combining forward inference and backpropagation.

[0157] By multiplying the standardized gating weight coefficients element-wise with the original gradient matrix, a weighted scaling gradient matrix is ​​formed. The scaling factor of this matrix is ​​directly related to the degree of drift potential energy accumulation, thereby amplifying the gradient components in the high-frequency fault-sensitive direction during parameter update.

[0158] Matrix accumulation and fusion are performed on the weighted scaling gradient matrix. The scaling gradients within consecutive time steps are summed and low-frequency redundant components are suppressed through an adaptive regularization mechanism to ensure that the update direction is focused on the frequency domain of weak arc fault characteristics.

[0159] The weight update formula is based on the scaling gradient matrix:

[0160] in, To connect the weight matrix, For learning rate, It is the scaled gradient matrix modulated by the gated weight coefficients.

[0161] The high-frequency response index is calculated on the updated connection weight matrix, and its sensitization effect on arc characteristics is evaluated by the spectral energy ratio index. This index is then converted into a dynamic arc prediction sensitivity coefficient output.

[0162] By updating the gradient scaling and generating the sensitivity coefficients, the standardized gating weight coefficients from the previous step are transformed into enhanced parameters for high-frequency arc characteristics, achieving the expected technical effect of significantly improving the model's ability to capture weak arc fault characteristics in the early stages of drift.

[0163] For example, when the edge-side AI intelligent analysis core module operates on a low-voltage distribution line with periodic load fluctuations and occasional arcing faults, the gating weight coefficient generated after normalizing the drift potential energy integral value is 0.65. The current second-type neuron connection weight matrix is ​​a 64×128 floating-point matrix. The mean of the original gradient matrix in the latest time step is 0.002. This gradient matrix is ​​multiplied element-wise by 0.65 to obtain a scaled gradient matrix. Then, it is accumulated and summed over 10 time steps and L2 regularized to suppress low-frequency components. The regularization coefficient is set to 0.01. With a learning rate of 0.001, the above formula is updated. The weights were updated. After the weights were updated, the high-frequency response index was calculated. The energy ratio of the arc characteristic frequency band increased from 1.8 to 3.4, and the sensitivity coefficient increased from 0.5 to 0.92. This indicates that the model's ability to capture the characteristics of weak arc faults under this operating condition has been significantly improved. After outputting the dynamic arc prediction sensitivity coefficient and injecting it into the risk mapping layer, the prediction logic can identify the nascent stage of arc faults in advance and achieve early warning.

[0164] S6.4: The dynamic overheating prediction activation threshold and the dynamic arc prediction sensitivity coefficient are synchronously injected into the forward propagation path of the top-level risk mapping layer, and forward inference operations are performed on the real-time collected current, voltage and temperature feature vectors to generate an intermediate risk probability distribution vector after pre-calibration.

[0165] S6.5: Based on the intermediate state risk probability distribution vector, the parameter solidification operation is performed through the inverse constraint mechanism of the multi-task loss function to generate the final calibrated risk mapping parameters that adapt to the current operating condition drift, and complete the online incremental update of the risk prediction logic of the local AI intelligent analysis core module.

[0166] Step S7: Update the risk prediction logic of the local AI intelligent analysis core module using the calibrated risk mapping parameters, and package the occurrence time, duration, dominant scale, and parameter increment of this drift initiation event into a lightweight adaptation package. Specifically, this includes: S7.1: Obtain the calibrated risk mapping parameters generated by the previous steps and the drift potential energy integral value at the current moment. Based on the memory address mapping mechanism, write the calibrated risk mapping parameters into the risk prediction logic register of the AI ​​intelligent analysis core module to replace the original static threshold coefficient and generate an updated risk prediction logic instance with dynamic adaptability.

[0167] S7.2: Read the timestamp signal of the updated risk prediction logic instance, combine it with the pre-recorded duration data of the drift initiation stage and the dominant scale identifier, and use the time series alignment algorithm to synchronously verify the multi-source metadata to generate a standardized drift event metadata dataset containing the precise occurrence time, duration and dominant scale.

[0168] The timestamp signal of the updated risk prediction logic instance is obtained as the time base input. The duration data of the drift initiation stage and the dominant scale identifier are pre-stored in the drift event recording module to form a data set containing three core time-related meta-parameters.

[0169] The aforementioned time reference inputs are vectorized and combined with the duration data of the drift initiation stage to construct a multi-source time matrix for time series alignment. A dominant scale identifier is introduced into the matrix structure as an additional index field to support scale-related time offset correction.

[0170] A time series alignment algorithm is used to perform synchronization verification on a multi-source time matrix. Dynamic time warping (DTW) constraints and window limitation strategies are introduced during the alignment process to ensure that the occurrence time, duration and dominant scale parameters do not drift significantly within a given time window, thus generating a preliminary aligned time series synchronization matrix.

[0171] By applying differential consistency detection logic, row and column difference calculations are performed on the initially aligned time series synchronization matrix to eliminate abnormal deviations and retain standardized parameter combinations that conform to the characteristics of the drift initiation stage, forming a multi-dimensional vector containing the precise occurrence time, duration, and dominant scale.

[0172] By utilizing structured metadata encapsulation rules, the aforementioned multidimensional vectors are mapped to a standardized drift event metadata set, and unique event identifiers and integrity check codes are attached to achieve closed-loop generation of the metadata dataset.

[0173] By using multi-source time verification and structured encapsulation, the updated risk prediction logic instance time parameters and drift initiation stage metadata from the previous step are transformed into standardized drift event metadata that can be directly used for subsequent lightweight adaptation package packaging, thereby achieving unified management of event time characteristics and scale identifiers.

[0174] For example, in a certain edge electrical fire protection current limiting device, the timestamp signal detection value of the updated risk prediction logic instance is 1623456789012 milliseconds, the duration of the drift initiation stage is 480000 milliseconds, and the dominant scale identifier is "periodic scale". The time series alignment algorithm sets the DTW window size to 600000 milliseconds, and the multi-source time matrix formed during the alignment process is [[1623456789012,480000,'periodic scale']]. In the differential consistency detection logic, threshold control is applied to the difference between the occurrence time and the duration, and the thresholds are calculated by the following formulas:

[0175] in, The standard deviation of the timestamp sequence. Sampling frequency, The time of occurrence;

[0176] in, The standard deviation of the duration series. Sampling frequency, This represents the duration difference value. In actual calculations, the standard deviation of the timestamp sequence is 1000 milliseconds, the sampling frequency is 1Hz, resulting in an occurrence time of 3000 milliseconds; the standard deviation of the duration sequence is 500 milliseconds, resulting in a duration difference value of 1000 milliseconds. After passing the difference detection, the generated standardized drift event metadata dataset contains an occurrence time of 1623456789012 milliseconds, a duration of 480000 milliseconds, a dominant scale "periodic scale", and is appended with the event ID "EVT-20240610-001" and an integrity CRC16 checksum. This metadata dataset is directly called in the subsequent lightweight adaptation package packaging step, achieving a significant improvement in event timing accuracy and scale consistency.

[0177] S7.3: Extract the activation threshold increment and sensitivity coefficient increment of the neuron cluster in the top-level risk mapping layer from the calibrated risk mapping parameters. Use differential coding technology to compress the activation threshold increment and sensitivity coefficient increment to generate a sparse parameter increment vector that represents the model fine-tuning amplitude.

[0178] Based on the calibrated risk mapping parameters generated from the previous steps, the activation threshold increment and sensitivity coefficient increment data regions of the neuron clusters in the top-level risk mapping layer are located. The register readout mechanism is then invoked to accurately extract the two types of increment values. The extracted activation threshold increments are separated according to parameter category, and sign determination and amplitude encoding are performed at a preset quantization precision, mapping positive and negative adjustment to different sign bit identifiers. The sensitivity coefficient increments are discretized using the same quantization precision, converting continuous floating-point numbers into fixed-length integer codes for inclusion in the compression algorithm. When constructing the differential encoding input vector, the extracted increment values ​​are compared with the corresponding parameter values ​​from the previous version using an element-wise difference calculation. The difference calculation formula is:

[0179] in, Indicates the parameter value. and These represent the version indices before and after the update, respectively. This represents the incremental value. The difference result is input into a variable-length encoder, and the encoding word length is selected based on the absolute magnitude of the difference value. Differential values ​​with smaller absolute magnitudes are encoded using shorter word lengths, thereby maximizing compression efficiency in the sparse parameter sequence. The outputs of the differential encoding of the activation threshold increment and the sensitivity coefficient increment are concatenated into a sparse parameter increment vector, and a structure descriptor is added to the head of the vector to identify the parameter category and quantity. Through differential encoding and sparsification, the result of the previous step is transformed into a data structure that can be transmitted at low cost and has controllable computational overhead, achieving the expected effect of lightweight storage and transmission.

[0180] For example, in an edge-side risk prediction model for an electrical circuit, the top-level risk mapping layer contains 48 neurons responsible for overheat prediction and 32 neurons responsible for arc prediction. After adjustment in step S6, the average activation threshold of the overheat prediction neurons is increased by 0.015 units compared to the previous version, and the average sensitivity coefficient of the arc prediction is increased by 0.022 units. Element-wise difference calculations are performed on the specific values ​​of each neuron. For example, the threshold of the first overheat prediction neuron is increased from 1.200 to 1.215, corresponding to a difference value of 1.215 - 1.200 = 0.015; the sensitivity of the fifth arc prediction neuron is increased from 0.500 to 0.522, corresponding to a difference value of 0.522 - 0.500 = 0.022. All difference values ​​are processed with a quantization precision of 0.001 and converted to integer encoding before being input into the differential encoder. Differences with an absolute value not exceeding 0.025 use 5-bit encoding, while differences exceeding this range use 8-bit encoding. The resulting sparse parameter increment vector is 80 bytes long and includes complete activation thresholds and sensitivity coefficient category identifiers. This sparse parameter increment vector significantly reduces bandwidth consumption during cloud platform uploads and can be directly restored to accurate increment values ​​during subsequent model fine-tuning, enabling fast and low-cost knowledge distillation references.

[0181] S7.4: Receives a standardized drift event metadata dataset and a sparse parameter increment vector, binds the two together using a structured data encapsulation protocol and adds an integrity check code to generate a lightweight adaptation package data structure containing complete operating condition drift features and model adjustment information.

[0182] S7.5: Performs serialization operation on the lightweight adapter package data structure, converts it into a binary stream format and stores it in the send buffer queue of the communication module to complete the data preparation for the local model's online incremental learning closed loop and trigger the subsequent cloud platform upload process.

[0183] Step S8: During communication idle periods, upload the lightweight adaptation package to the cloud platform for archiving. The incremental parameter data contained in the lightweight adaptation package will serve as a reference for the knowledge distillation process of subsequent similar line models, thus completing the online incremental learning loop for the edge-side model. Specifically, this includes: S8.1: Obtain the status monitoring data of the current communication link, use the channel occupancy detection algorithm to evaluate the uplink bandwidth load in real time, and generate a time window signal that identifies the communication idle period.

[0184] S8.2: Based on the time window signal, a lightweight adapter packet in the local storage area is invoked. The data serialization protocol is used to encapsulate the structured data containing the occurrence time, duration, dominant scale and parameter increment to generate a binary data stream to be transmitted.

[0185] S8.3: Receive the binary data stream to be transmitted, and use the encrypted transmission mechanism of the remote communication module to perform data packaging and sending operations, so as to securely upload the lightweight adapter package to the designated storage node of the cloud platform to form a cloud archive record.

[0186] S8.4: Extract incremental parameter data from cloud archive records, and use the knowledge distillation algorithm under the transfer learning framework to use the risk mapping parameters after edge calibration as teacher signals to construct feature distribution constraints for similar line models.

[0187] S8.5: Based on the feature distribution constraints, the loss function of the global model in the cloud is weighted and corrected. The gradient descent optimization strategy is used to perform fine-tuning and updating of the model parameters to generate a new generation of electrical fire protection current limiting protection model with stronger adaptability to working conditions and complete the online incremental learning closed loop.

[0188] This application also provides an electrical fire prevention and current limiting protection device based on artificial intelligence numerical control, the specific structure of which is as follows: The device includes a signal acquisition module, a multi-scale convolution module, an attention coding module, a gated recurrent unit module, a master prediction network module, and a calibration module. These modules are connected in sequence to form a closed loop for edge-side signal processing and analysis.

[0189] The signal acquisition module adopts a multi-channel synchronous sampling architecture, with three built-in independent analog-to-digital converters to perform parallel quantization of current, voltage, and temperature sensing signals respectively. The output of each channel is connected to a circular buffer queue of the corresponding time scale to ensure that millisecond-level transient data, second-level periodic data, and minute-level inertial data are timestamped before being enqueued. The module outputs three synchronized digital signal streams, which are then sent to the multi-scale convolution module.

[0190] The multi-scale convolution module is used to normalize the transient residual map, periodic residual map, and inertial residual map before inputting them into the attention encoder. It uses a dynamic weight allocation mechanism to perform weighted calculations on the residual sequence to generate a weighted residual sequence.

[0191] The attention encoding module is used to construct a drift potential energy integrator with memory decay characteristics based on the weighted residual sequence. By inputting the weighted residual sequence into the gated recurrent unit to perform hidden state updates, the drift potential energy integral value is generated.

[0192] The gated loop unit module is used to determine whether the drift potential energy integral value exceeds the preset soft threshold and shows a monotonically increasing trend within the maintenance time window. If the condition is met, it is determined to enter the drift initiation stage and generate a self-trigger control signal; otherwise, it maintains the current model parameter state and returns to continue acquiring digital signals.

[0193] After responding to the self-triggered control signal, the main prediction network module locks the weights of the bottom feature extraction layer through the gradient blocking mechanism, while activating the neuron cluster in the top risk mapping layer that is responsible for overheating probability prediction and arc tendency prediction, and mapping the drift potential energy integral value into a normalized gating signal and injecting it into the neuron cluster.

[0194] After receiving the gating signal, the calibration module performs nonlinear offset calculation and gradient scaling update on the bias vector and activation slope parameters of the neuron cluster, respectively, to generate the dynamic overheating prediction activation threshold and dynamic arc prediction sensitivity coefficient, complete the pre-calibration of the risk mapping parameters, and output the calibrated parameter set to the local AI intelligent analysis core module.

[0195] Through the collaboration of the above modules, the device can independently complete the entire closed-loop processing from multi-scale signal acquisition, residual feature extraction, drift potential energy integration, budding trigger determination to top-level parameter adaptive calibration at the edge, without relying on the cloud or historical baseline comparison.

[0196] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0197] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," "third," and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, "an" or "a" and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms "comprising" or "including" and similar terms mean that the element or object preceding "comprising" or "including" encompasses the element or object listed following "comprising" or "including" and its equivalents, and do not exclude other elements or objects. The "multiple" involved in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0198] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered 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.< / tanh>

Claims

1. An electrical fireproof current-limiting protection method based on artificial intelligence numerical control, characterized in that, include: S1: Acquire digital signals of current, voltage, and temperature, and input the three digital signals into three multi-scale convolution branches with different receptive field lengths to generate transient residual maps, periodic residual maps, and inertial residual maps. S2: After normalizing the transient residual map, periodic residual map and inertial residual map, input them into the attention encoder, and use the dynamic weight allocation mechanism to perform weighted calculation on the residual sequence to generate a weighted residual sequence; S3: Construct a drift potential energy integrator with memory decay characteristics based on the weighted residual sequence, and generate the drift potential energy integral value by inputting the weighted residual sequence into the gated loop unit to perform hidden state update; S4: Determine whether the integral value of the drift potential energy exceeds the preset soft threshold and shows a monotonically increasing trend within the maintenance time window. If the condition is met, determine that the drift budding stage has been entered and generate a self-trigger control signal. Otherwise, maintain the current model parameter state and return to continue collecting digital signals. S5: In response to the self-triggered control signal, freeze the bottom feature extraction layer of the main prediction network and activate the neuron cluster in the top risk mapping layer so as to input the drift potential energy integral value as a gating signal to the neuron cluster; S6: Based on the gating signal, automatically adjust the activation threshold and sensitivity coefficient inside the neuron cluster to perform a pre-calibration operation on the overheating probability prediction index and the arc occurrence tendency prediction index, and generate calibrated risk mapping parameters.

2. The electrical fireproof current-limiting protection method based on artificial intelligence numerical control according to claim 1, characterized in that, The process of adding steps S7-S8 after step S6 includes: S7: Update the risk prediction logic of the local AI intelligent analysis core module using the calibrated risk mapping parameters, and package the occurrence time, duration, dominant scale and parameter increment of this drift initiation event into a lightweight adaptation package; S8: During communication idle periods, the lightweight adaptation package is uploaded to the cloud platform for archiving, and the incremental parameter data contained in the lightweight adaptation package provides a reference for the knowledge distillation process of subsequent similar line models, so as to complete the online incremental learning closed loop of the edge-side model.

3. The electrical fireproof current-limiting protection method and device based on artificial intelligence numerical control according to claim 1, characterized in that, Step S3 specifically includes: The weighted residual sequence after processing by the attention encoder is obtained, and the hidden state vector and memory decay coefficient matrix of the gated loop unit are initialized based on the computing power constraint of the edge computing device to construct the initial architecture of the drift potential energy integrator with memory decay characteristics. The weighted residual sequence is sequentially input into the input gating unit in the initial architecture of the drift potential energy integrator according to the time step. The residual characteristics at the current time are filtered by a nonlinear activation function to generate a gating activation signal. Based on the gated activation signal and the hidden state vector of the previous time step, the memory decay coefficient matrix is ​​applied through the forgetting gating mechanism for weighted calculation to generate decayed historical state information that removes historical noise interference and retains long-term trend characteristics. The output gating unit is used to fuse the decayed historical state information with the gating activation signal at the current moment, and a state cell update operation is performed to generate the latest hidden state vector containing the cumulative amount of progressive structural disturbance under the current operating condition. A linear mapping transformation is performed on the latest hidden state vector to extract its scalar potential energy component, thereby generating the drift potential energy integral value.

4. The electrical fireproof current-limiting protection method based on artificial intelligence numerical control according to claim 1, characterized in that, Step S4 specifically includes: The real-time drift potential energy integral value and the preset soft threshold output by the gated loop unit are obtained. The magnitude of the real-time drift potential energy integral value and the preset soft threshold are determined by numerical comparison logic to generate an initial over-limit flag. Based on the condition that the initial over-limit flag is true, the drift potential energy integral value sequence of historical moments within the continuous maintenance time window is read from the circular buffer, and the slope calculation operation of adjacent data points is performed on the drift potential energy integral value sequence using the sliding difference algorithm to generate a first-order difference sequence. Receive the first-order difference sequence, and perform a non-negativity traversal detection operation on each element in the first-order difference sequence using all-positive term verification logic to generate a trend consistency judgment result; By combining the initial over-limit flag and the trend consistency determination result, a joint state confirmation operation is performed on the two using a logical AND operation mechanism to generate the final trigger decision signal; In response to the final trigger decision signal, the self-trigger control command is generated using pulse code modulation technology, or a hold command is generated when the final trigger decision signal is false, so as to drive subsequent execution of model parameter freezing and activation operations or return to the signal acquisition loop.

5. The electrical fireproof current-limiting protection method based on artificial intelligence numerical control according to claim 1, characterized in that, The attention encoder employs a multi-head self-attention mechanism and a soft maximum function to automatically identify and amplify the weighted residual sequence with features of continuous unidirectional offset and phase coupling distortion precursors.

6. The electrical fire prevention and current limiting protection method based on artificial intelligence numerical control according to claim 1, characterized in that, The drift potential energy integrator performs weighted fusion and forgetting gating processing on the current gating activation signal and the historical hidden state vector by setting a memory decay coefficient matrix, and generates the drift potential energy integral value by linear mapping.

7. The electrical fireproof current-limiting protection method based on artificial intelligence numerical control according to claim 1, characterized in that, When the integral value of the drift potential energy exceeds the preset soft threshold and all elements of the first-order difference sequence within the maintenance time window are non-negative, the self-trigger control signal is generated using a logical AND operation.

8. The electrical fireproof current-limiting protection method based on artificial intelligence numerical control according to claim 1, characterized in that, After receiving the self-triggered control signal, the bottom feature extraction layer of the main prediction network enters a frozen state through gradient blocking and weight flag locking, and only the neuron clusters in the top risk mapping layer are activated.

9. The electrical fireproof current-limiting protection method based on artificial intelligence numerical control according to claim 1, characterized in that, The normalized gating signal is used to linearly or nonlinearly reconstruct the bias terms and activation slopes within the neuron cluster, enabling the main prediction network to respond to the operating condition drift trend earlier than traditional statistical anomaly algorithms.

10. An electrical fire prevention and current limiting protection device based on artificial intelligence numerical control, comprising: The signal acquisition module is used to acquire digital current signals, digital voltage signals, and digital temperature signals, and inputs the three digital signals into three multi-scale convolution branches with different receptive field lengths to generate transient residual maps, periodic residual maps, and inertial residual maps. The multi-scale convolution module is used to normalize the transient residual map, periodic residual map and inertial residual map and input them into the attention encoder. The residual sequence is weighted by a dynamic weight allocation mechanism to generate a weighted residual sequence. The attention encoding module is used to construct a drift potential energy integrator with memory decay characteristics based on the weighted residual sequence. By inputting the weighted residual sequence into the gated loop unit to perform hidden state updates, the drift potential energy integral value is generated. The gated loop unit module is used to determine whether the integral value of the drift potential energy exceeds the preset soft threshold and shows a monotonically increasing trend within the maintenance time window. If the condition is met, it is determined to enter the drift budding stage and generate a self-trigger control signal; otherwise, the current model parameter state is maintained and the acquisition of digital signals is returned to continue. The main prediction network module is used to freeze the bottom feature extraction layer of the main prediction network in response to the self-triggered control signal, and activate the neuron cluster in the top risk mapping layer so as to input the drift potential energy integral value as a gating signal to the neuron cluster. The calibration module is used to automatically adjust the activation threshold and sensitivity coefficient within the neuron cluster based on the gating signal, so as to perform a pre-calibration operation on the overheating probability prediction index and the arc occurrence tendency prediction index, and generate calibrated risk mapping parameters.

Images