Alarm signal reasoning and adaptive processing optimization method and system based on large model

Abstract

The invention discloses an alarm signal reasoning and self-adaptive processing optimization method and system based on a large model. The method comprises the following steps: collecting and preprocessing signal data; intelligent alarm signal reasoning based on a large model, wherein the intelligent alarm signal reasoning comprises input data preparation and embedded representation, large model time sequence modeling based on Transform, graph reasoning and equipment association modeling, dynamic knowledge injection and incremental learning and anomaly detection and fault reasoning; an event association and feedback learning mechanism based on a large model comprises a signal event association and adaptive disposal optimization method, an automatic alarm collection and execution mechanism and a feedback optimization and continuous learning mechanism; the system comprises a data processing module, a signal reasoning module and a feedback learning module. According to the method, the automation degree and the real-time response capability of power grid alarm signal processing are remarkably improved, power grid equipment faults and emergencies can be effectively handled, and efficient, stable and safe operation of a power grid in a complex environment is ensured.

Description

Technical Field

[0001] This invention relates to the field of power system automation, and in particular to a method and system for optimizing alarm signal reasoning and adaptive handling based on a large model. Background Technology

[0002] With the continuous expansion of the power grid and the advancement of integrated dispatching, the monitoring and management of power grid equipment has become increasingly complex. Power grid equipment is typically monitored centrally at the dispatching station, where dispatchers need to process tens of thousands of alarm signals daily. These alarm signals involve multiple dimensions, including equipment operating status, telemetry data, and fault information, and are often accompanied by complex anomalies. In this context, traditional alarm signal processing still relies on manual intervention. Dispatchers need to identify, record, track, and eliminate each signal individually. This manual processing mode is not only inefficient but also prone to errors such as misjudgment and omissions. Especially during power grid equipment operation, the volume and variety of alarm signal data are enormous. Manual processing cannot handle the large amount of information in real time and leads to delayed responses and untimely processing, thus affecting the safe and stable operation of the power grid. Furthermore, when power grid equipment failures occur frequently, the complexity and similarity of alarm signals further exacerbate the difficulty of manual judgment, revealing the shortcomings of traditional monitoring systems. Therefore, existing technologies cannot efficiently and accurately process massive alarm signals, nor can they provide real-time, intelligent fault diagnosis and handling solutions. This results in excessive workload for power grid dispatchers and fails to meet the demands of modern power grids for efficient, safe, and intelligent management. A new solution is urgently needed to improve the efficiency and accuracy of signal identification and processing.

[0003] The invention disclosed in CN119341188B presents a data block matching and weight calculation method based on historical abnormal event monitoring information of the power grid, involving the identification and monitoring technology of abnormal events in the power grid. This method, through weight setting and frequency analysis of historical data, combined with real-time collected monitoring information for event matching, can effectively identify abnormal events in the power grid, improve the safety and stability of power grid operation, and reduce manual intervention and improve response efficiency by pushing event containers to staff. However, this method also has some shortcomings. Although it can utilize historical data for event identification, its processing mainly relies on data blocks and frequency calculation, resulting in poor flexibility in identifying new, complex, or unknown events, and an inability to effectively respond to dynamic changes and sudden events in power grid equipment. Furthermore, event matching and alarm response still rely on manual judgment and manual threshold setting, which may lead to system response delays, failure to promptly address real-time problems of power grid equipment, and impact on the rapid fault repair and dispatching of the power grid.

[0004] The invention disclosed in CN119884700A presents a smart grid monitoring and alarm method based on real-time grid data. This method combines ANN and LSTM prediction models for equipment fault prediction, effectively extracting static and time-series features to improve the prediction accuracy of grid equipment faults. Through this method, the system can assess the health status of the grid in real time and provide early warnings of potential equipment faults, thereby improving the reliability and security of grid operation. However, this method also has some shortcomings. Although it can improve prediction accuracy by combining static and time-series data, relying on historical data and model training may not provide timely responses to sudden or atypical faults. Furthermore, while it can provide early warning information, the system still relies on subsequent human judgment and operation, lacking fully automated decision-making capabilities and making it difficult to respond in real time to complex dynamic changes in the grid. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide a method and system for alarm signal reasoning and adaptive handling optimization based on a large model.

[0006] Technical solution: The alarm signal inference and adaptive handling optimization method based on a large model described in this invention includes the following steps:

[0007] (1) Signal data acquisition and preprocessing, including data acquisition, data removal and cleaning, and missing value filling;

[0008] (2) Intelligent alarm signal reasoning based on large model, including input data preparation and embedding representation, large model time series modeling based on Transformer, graph reasoning and device association modeling, dynamic knowledge injection and incremental learning, and anomaly detection and fault reasoning.

[0009] (3) Event correlation and feedback learning mechanism based on large model, including signal event correlation and adaptive handling optimization method, automated alarm collection and execution mechanism, and feedback optimization and continuous learning mechanism.

[0010] Furthermore, the data acquisition in step (1) includes:

[0011] (1.1) Real-time telemetry and teleindication data are collected through the SCADA system and set as S S ={s1, s2, ..., s n}, where s i S represents the device status and telemetry data at a certain moment. s As the core data source, it connects with field devices and uses communication protocols to transmit data to the monitoring center in real time;

[0012] (1.2) Obtain high-precision, time-synchronized voltage / current phasors, frequency, and rate of change of frequency via PMU; set as Sp ={v i i i f i , Δf i}, where v i and i i f represents the voltage and current phasors, respectively. i For frequency, Δf i For the frequency change rate, the protection device and fault recorder provide the system with detailed fault event records, protection action timing, and fault waveforms, set as S. Pr = {ej, tj, wj}, where e j Indicates the type of fault event, t j For the fault timestamp, w j This is fault waveform data; data from the equipment condition monitoring system, set as S. Eq ={o k ,pd k , t k v k}, where o k This represents oil chromatography data, pd k For partial discharge, t k v represents the equipment temperature. k For mechanical vibration data, the energy management system is set to S. E ={Tg,Ps,Mn}, where Tg represents the power grid topology, Ps is the dispatch plan, Mn is the network model information, and the maintenance management system is set to S. C ={hm,pm,dm,rm}, where hm is the equipment maintenance history, pm is the maintenance plan, dm is the defect record, and rm is the test report;

[0013] (1.3) The meteorological information system provides environmental data and is set as follows: Where θ is temperature. For humidity, v w For wind speed, r w For rainfall, l w Lightning information is aggregated and processed to form a multi-dimensional signal dataset, Statal.

[0014] Furthermore, step (2) of input data preparation and embedding representation includes obtaining a dataset matrix S through data acquisition and preprocessing. total Its dimension is 7×d;

[0015] For the input signal Stotal, we first perform feature embedding on it, assuming each signal s t ∈R d Let be the signal feature vector at time step t, representing the feature space of a single signal, and d be the dimension of the feature. Then the embedding process is as follows:

[0016]

[0017] Where Embed(·) represents the feature embedding layer, and the output is a vector of dimension h, where h is the hidden layer dimension of the model. The entire input signal matrix Sttotal is transformed into an embedding matrix:

[0018]

[0019] The entire dataset Stotal is a T×d matrix, where T is the number of time steps, d is the feature dimension of each signal, and h is the dimension of the embedded signal.

[0020] Furthermore, step (2) of large-scale temporal modeling based on Transformer includes:

[0021] The Transformer model is used to perform global dependency modeling on the power grid alarm signal sequence to capture potential temporal patterns in the signal.

[0022] Assume the input signal sequence X∈R T×h The embedded matrix input X is mapped to a query, key, and value through a linear transformation:

[0023] Q = XW Q K = XW K V = XW V

[0024] Where WQ, WK, WV∈R h×hk It is a learnable weight matrix, h k It refers to the dimensions of query, key, and value;

[0025] The relevance between query Q and key K is calculated, and the similarity matrix S∈R is obtained by dot product. T×T :

[0026]

[0027] The similarity matrix is ​​normalized to ensure that the sum of the weights is 1.

[0028]

[0029] Where the weight matrix α represents the correlation between each query signal and all key signals, S ij These are the elements in the similarity matrix S;

[0030] The calculated attention weights α are applied to the value matrix V, and the final output representation is obtained by weighted summation:

[0031] A=αV

[0032] Where A∈R T×hk The positional encoding P∈R is obtained by weighting the value matrix with attention weights. T×h The position encoding is used to represent the position of each signal in the sequence, and the position encoding is performed using the following formula:

[0033]

[0034] Where i is the time step index in the sequence, j is the feature dimension index, and h is the embedding dimension. This positional encoding uses sine and cosine functions to encode the positional information of each time step, ensuring that positional information of different dimensions has different scales.

[0035] After position encoding, the input signal X is added to the position code P to form an input with timing information:

[0036] X input =X+P

[0037] The input is then passed to the Transformer's multi-layer encoder to ensure that the model can understand the temporal order of the signals.

[0038] Furthermore, step (2) graph reasoning and device association modeling includes:

[0039] Let G = (V, E) be the equipment graph of the power grid, where V represents the set of equipment nodes, E is the set of edges between equipment, and each equipment vi ∈ V has an eigenvector, hi ∈ R. d It represents the state of the device, and the adjacency matrix A∈R is defined. V|×|V| And normalization is applied:

[0040]

[0041] in, It is the adjacency matrix after adding self-connections. Here, H is the degree matrix, and I is the identity matrix. Based on the Graph Convolutional Network (GCN), information transfer and node feature updates are performed on the device graph. Let H = [h1, h2, ..., hN] be the feature matrix of all nodes in the graph. The GCN update rule is expressed as:

[0042] H (l+1) =σ(A norm H (l) W (l) +H (l) )

[0043] Among them, W (1) σ is the learnable weight matrix of the first layer, and σ is the non-linear activation function;

[0044] A graph attention network (GAT) based on a global attention mechanism is introduced. The weights of the graph attention network based on the global attention mechanism are calculated in an adaptive manner, β:

[0045]

[0046] The update characteristics of node i are:

[0047]

[0048] Where 'a' is the attention weight of the learning process, || represents the concatenation operation, and h... i and h j Let i and j be the feature vectors of node i and node j respectively, and N(i) be the set of neighbors of node i.

[0049] Furthermore, step (2) dynamic knowledge injection and incremental learning includes:

[0050] Construct a knowledge graph K = {(h, r, t)} related to power grid equipment. In TransE++, embed the triples (h, r, t) in the knowledge graph as vectors h, r, t ∈ R. d And use the following optimization objectives:

[0051]

[0052] Where K represents all triples in the knowledge graph, h′ represents negative samples, and γ is a hyperparameter. A temporal attention mechanism is introduced, and the updated knowledge graph embedding is composed of temporal information z. t Weighted average yields:

[0053] h t+1 =h t +α·z t

[0054] Where α is the adaptive weight, z t This indicates the temporal feature added at time t.

[0055] Employing a model-independent meta-learning strategy, in MAML, given a new task T... t+1 For each task T t+1 The goal of updating the model parameters θ is to minimize the task loss function L(θ):

[0056]

[0057] Where α is the learning rate. Based on the current task T t+1 The gradient;

[0058] Incremental knowledge K t+1The triples (h′, r′, t′) in the original knowledge graph Kt are embedded into the existing knowledge graph. Given the embedding representation Ht of the original knowledge graph Kt and the embeddings h′, r′, t′ corresponding to the new knowledge triples (h′, r′, t′), the loss function is defined as:

[0059]

[0060] Here, || represents the Euclidean distance, and the goal of minimizing the loss function is to make the embedding of the new triplet as close as possible to its corresponding tail entity t′.

[0061] Furthermore, the signal event association and adaptive handling optimization method in step (3) includes:

[0062] The model dynamically correlates signal events with equipment states. Given the real-time state vector x(t) of the power grid equipment and historical data d, h In addition to the alarm signal type y(t), the system quantifies the correlation between the alarm signal and the equipment operating status by calculating the correlation degree A:

[0063]

[0064] Among them, f i It is the characteristic correlation function for each specific signal, w i Here, N is the importance weight of each feature, and N is the number of devices involved. The system assigns a priority P(t) to each signal. Assume... To optimize the signal priority, the system uses the following objective function to achieve adaptive optimization:

[0065]

[0066] Where C(t) is the disposal cost, R(t) is the expected risk of equipment recovery, and λ1, λ2, and λ3 are the corresponding weighting coefficients.

[0067] Furthermore, the automated alarm collection and execution mechanism in step (3) includes:

[0068] Assuming y(t) represents an alarm signal event at each time point t, the system uses a large model inference engine to instantly determine the type and urgency of the alarm signal and generate a matching handling plan s(t):

[0069]

[0070] Here, M represents the execution module, which is responsible for generating the corresponding equipment handling plan; during execution, the system can adjust the handling steps based on real-time feedback and equipment changes.

[0071] Furthermore, the feedback optimization and continuous learning mechanism in step (3) includes:

[0072] After each alarm signal is processed, the system adjusts its decision based on the results, setting the feedback reward R(t) as the indicator of whether the processing was successful.

[0073]

[0074] Where δ(k) is the feedback reward of device k at time point k, which is usually the recovery time of the device or the success rate of the repair, and γ is the discount factor.

[0075] The alarm signal inference and adaptive handling optimization system based on a large model described in this invention includes:

[0076] The data processing module is used for signal data acquisition and preprocessing, including data acquisition, data removal and cleaning, and missing value filling.

[0077] The signal reasoning module is used for intelligent alarm signal reasoning based on large models, including input data preparation and embedding representation, large model time series modeling based on Transformer, graph reasoning and device association modeling, dynamic knowledge injection and incremental learning, and anomaly detection and fault reasoning.

[0078] The feedback learning module is used for event correlation and feedback learning mechanisms based on large models, including signal event correlation and adaptive handling optimization methods, automated alarm collection and execution mechanisms, and feedback optimization and continuous learning mechanisms.

[0079] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0080] (1) By introducing innovative signal data acquisition and preprocessing technologies and intelligent alarm signal inference methods based on large models, these problems were effectively solved. In terms of signal data acquisition and preprocessing, the system acquires real-time operating data of power grid equipment from multiple core data sources such as SCADA and PMU through multi-source data acquisition technology, including key parameters such as voltage, current, and frequency. Noise in the signal is removed by moving average filtering and exponential smoothing, and missing values ​​are filled by KNN, ensuring the stability and integrity of the data, providing high-quality input data for subsequent inference, and improving the accuracy and stability of the processing;

[0081] (2) In terms of intelligent reasoning, a large-scale Transformer structure is adopted, which captures long-range dependencies between signals through a self-attention mechanism. This enables global dependency modeling of power grid alarm signals, ensuring that each signal is accurately interpreted within its context. Simultaneously, the combination of Graph Neural Networks (GNNs) and Graph Attention Networks (GATs) allows the system to model the topological relationships of power grid equipment more flexibly and accurately. To avoid the gradient vanishing problem in deep networks, residual connections are added to each layer of the graph convolutional network, ensuring stable information transmission and further enhancing the model's deep learning capabilities.

[0082] (3) Through incremental learning and meta-learning techniques, the system can adapt to new equipment fault modes and state changes in real time, continuously optimize reasoning and handling strategies, and improve the intelligent management level of power grid equipment. This invention significantly improves the automation, intelligence and adaptability of power grid alarm signal processing; it significantly improves the automation level and real-time response capability of power grid alarm signal processing, and can effectively cope with power grid equipment faults and emergencies, ensuring the efficient, stable and safe operation of the power grid in complex environments. Attached Figure Description

[0083] Figure 1 This is a flowchart of the present invention;

[0084] Figure 2 This is a flowchart of signal data acquisition and preprocessing.

[0085] Figure 3 Flowchart for intelligent alarm signal reasoning;

[0086] Figure 4 This is a flowchart of the event association and feedback learning mechanism. Detailed Implementation

[0087] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0088] like Figure 1 As shown, this invention proposes an alarm signal reasoning and adaptive handling optimization method based on a large model. It aims to improve the efficiency and accuracy of power grid equipment fault handling through intelligent data acquisition and preprocessing, signal reasoning based on a large model, and event correlation optimization. This method effectively reduces manual intervention and ensures the safe and stable operation of the power grid by automatically analyzing alarm signals, dynamically associating equipment status, and generating optimal handling solutions in real time.

[0089] 1. Signal data acquisition and preprocessing

[0090] In this application, data preprocessing provides high-quality input data to the model through three main steps: multi-source data acquisition, noise reduction, and missing value imputation. First, multi-dimensional power grid monitoring data is collected from systems such as SCADA, PMU, and protection devices, and organized into a standardized time series matrix. Next, moving average and exponential smoothing methods are used to remove fluctuations and noise from the signal, ensuring data stationarity. Simultaneously, the KNN imputation algorithm is used to fill in missing values, and prediction is performed based on a weighted average of similar data points, such as... Figure 2 As shown.

[0091] 1.1 Data Acquisition

[0092] During the data acquisition phase, this system obtains real-time power grid monitoring data from multiple core data sources. First, it acquires real-time remote signaling and telemetry data through the SCADA system, including equipment switching status, protection action signals, and operating parameters such as voltage, current, power, frequency, and temperature. (Set as S...) S = {s1, s2, ..., sn}, where s i S represents the device status and telemetry data at a certain moment. s As the core data source, it connects to field devices and transmits data to the monitoring center in real time using communication protocols (such as Modbus, DNP3, IEC61850, etc.). Secondly, it acquires high-precision, time-synchronized voltage / current phasors, frequency, and rate of change of frequency (ROCOF) through a PMU (Phasor Measurement Unit). (Set as S) p ={v i i i f i , Δf i}, where v i and i i f represents the voltage and current phasors, respectively. i For frequency, Δf i For the frequency change rate, the PMU can synchronously sample at the microsecond level, providing accurate time synchronization data for dynamic process analysis and supporting rapid grid response. Furthermore, protection devices and fault recorders provide the system with detailed fault event records, protection action timing, and fault waveforms. (Set to S) Pr = {ej, tj, wj}, where e j Indicates the type of fault event, t j For the fault timestamp, w j This refers to fault waveform data. These devices, through high-precision sampling and recording, can provide accurate timestamps and waveform data when a fault occurs, aiding in the analysis of the fault's root cause. Data from the equipment condition monitoring system, including transformer oil chromatography, partial discharge, equipment temperature, mechanical vibration, and other information, is set as S. Eq ={o k ,pdk ,t k v k}, where o k This represents oil chromatography data, pd k For partial discharge, t k v represents the equipment temperature. k For mechanical vibration data, it provides equipment health status, helping to predict and prevent potential failures. The Energy Management System (EMS) provides the system with information such as grid topology, scheduling plans, and network models, set as S... E ={Tg,Ps,Mn}, where Tg represents the power grid topology, Ps is the dispatch plan, and Mn is the network model information, which helps manage the load allocation and optimized dispatch of the power grid. The Maintenance Management System (CMMS) provides equipment maintenance history, maintenance plans, defect records, and test reports, set as S. C ={hm,pm,dm,rm}, where hm is the equipment maintenance history, pm is the maintenance plan, dm is the defect record, and rm is the test report, ensuring effective support for the operation and maintenance management of power grid equipment. Finally, the meteorological information system provides environmental data, including temperature, humidity, wind speed, rainfall, and lightning information, set as... Where θ is temperature. Humidity, vw is wind speed, r w For rainfall, l w This lightning information helps analyze the impact of the external environment on the power grid, providing support for emergency dispatch and early warning. All this data is aggregated and processed using advanced data acquisition technologies, such as real-time data transmission, protocol conversion, data synchronization, and storage integration, forming a multi-dimensional signal dataset, Statotal.

[0093]

[0094] Each row corresponds to all features at a given time point, and the number of columns in the matrix corresponds to the time step n, providing a solid data foundation for subsequent large-scale model inference and intelligent decision-making.

[0095] 1.2 Data Removal and Cleaning

[0096] After data collection, the raw data may contain noise or outliers, requiring denoising and cleaning processes to improve data quality. In this stage, moving average filters and exponential smoothing are used to remove fluctuations and noise from the data, ensuring the accuracy of subsequent analysis. Simultaneously, a KNN prediction algorithm is added to fill in missing values.

[0097] Moving average filters are suitable for removing periodic fluctuations. In a moving average filter, we smooth the signal by applying a window to the signal sequence and averaging the data within each window. Assume that from S... totalExtract the voltage signal data v1, v2, ..., v n its smoothed value Its moving average can be calculated using the following formula:

[0098]

[0099] Where k is the window size, v j It is the original signal. It is an average voltage signal. This method can effectively remove high-frequency noise from the signal, and the filtered signal can better reflect the true operating status of the equipment.

[0100] The calculated average voltage signal is then subjected to exponential smoothing to remove prominent short-term fluctuations. The formula is as follows:

[0101]

[0102] Here, α is the smoothing factor, with a value ranging from 0 < α ≤ 1. A larger α assigns higher weight to more recent data points. The advantage of exponential smoothing is that it can handle abrupt changes in the data well and quickly reflect the latest signal changes.

[0103] 1.3 Missing value imputation

[0104] In practical applications, real-time data collected in power grid monitoring systems often contains missing values, which may be due to equipment failure, abnormal data acquisition, transmission errors, or other reasons. To ensure the accuracy and completeness of subsequent models, filling in these missing values ​​is crucial.

[0105] The KNN (K-Nearest Neighbors) imputation method fills in missing data points by finding neighboring data points that are similar to the missing data points. For a given missing value x... i Based on its similarity to other data points, we select the k closest neighbors and calculate the weighted average of these neighbors as the predicted value for the missing value.

[0106] Assume the missing values ​​are located in S total v in i Its predicted value v i Given by the following formula:

[0107]

[0108] Where k is the number of neighbors selected, v n It represents the neighbor data of the missing value, where N(i) represents the number of neighbor points most similar to v. i Find the k nearest neighbors. Fill in the missing values ​​by calculating the average of these neighbors.

[0109] Through this series of data processing steps, the system can significantly improve the quality and consistency of power grid monitoring data. Denoising and smoothing methods effectively eliminate high-frequency noise and abrupt changes, improving signal stability and ensuring the model's sensitivity and accuracy to real-time changes in the power grid. Missing value imputation avoids gaps or misleading information caused by missing data, improving data integrity and enabling large models to be trained and inferred based on complete datasets, thereby improving prediction accuracy and decision reliability. This process makes the automatic monitoring and handling of power grid alarm signals more efficient and accurate, better supporting intelligent power grid management.

[0110] 2. Intelligent alarm signal inference based on large models

[0111] By integrating multiple advanced technologies, intelligent signal inference based on a large model has been achieved. This model can efficiently and accurately process power grid alarm signals and continuously optimize inference results in dynamically changing power grid environments. Figure 3 As shown;

[0112] 2.1 Input Data Preparation and Embedded Representation

[0113] A dataset matrix S was obtained through data acquisition and preprocessing. total Its dimension is 7×d, where d is the dimension of each signal characteristic (e.g., current, voltage, temperature, etc.). Next, we will discuss S. total Signal embedding processing is performed to prepare suitable inputs for the Transformer, the core component of the large model.

[0114] For the input signal Stotal, feature embedding is first performed to ensure that the signal can be effectively processed by the large model. Let each signal s... t ∈R d Let be the signal feature vector at time step t, representing the feature space of a single signal, and d be the dimension of the feature. Then the embedding process is as follows:

[0115]

[0116] Here, `Embed(.)` represents the feature embedding layer, and its output is a vector of dimension `h`, where `h` is the dimension of the model's hidden layers. The entire input signal matrix `Stotal` is transformed into an embedding matrix:

[0117]

[0118] The entire dataset `Stotal` is a T×d matrix, where T is the number of time steps, d is the feature dimension of each signal, and h is the dimension of the embedded signal. Through the embedding process, each signal is transformed from `Rd` to `Rh`, thus obtaining `R`. T×hThe data matrix is ​​used as input for subsequent large-scale model processing.

[0119] 2.2 Temporal Modeling of Large Models Based on Transformer

[0120] The application of Transformer in time series modeling, with its powerful self-attention mechanism, provides an efficient solution for modeling long-range dependencies in alarm signal processing. In this invention, the Transformer model is used to perform global dependency modeling on power grid alarm signal sequences, capturing potential time series patterns in the signals.

[0121] The core idea of ​​self-attention mechanisms is to achieve information exchange and aggregation between signals by calculating the similarity (correlation) between signals at different positions in the input sequence. Assume the input signal sequence X∈R. T×h To calculate the similarity between signals, the embedding matrix input X is first mapped to a query, key, and value through a linear transformation:

[0122] Q = XW Q K = XW K V = XW V

[0123] Where WQ, WK, WV∈R h×hk It is a learnable weight matrix, h k It refers to the dimensions of query, key, and value.

[0124] Next, we calculate the relevance between query Q and key K, and obtain the similarity matrix S∈R through dot product. T×T :

[0125]

[0126] To enhance numerical stability and improve model performance, the similarity matrix is ​​normalized to ensure that the sum of the weights is 1.

[0127]

[0128] These weight matrices α represent the correlation between each query signal (position) and all key signals (positions) (corresponding to α). ij The larger the weight, the stronger the dependency between signals. ij are elements in the similarity matrix S.

[0129] The calculated attention weights α are applied to the value matrix V, and the final output representation is obtained by weighted summation:

[0130] A=αV

[0131] Where A∈R T×hk This is the value matrix weighted by attention weights. This is the output of the self-attention mechanism, representing the weighted representation of the signal at each position after considering all other signals. The final output of the self-attention mechanism is a weighted representation of each signal, which incorporates the dependencies between positions in the input sequence. In this way, each signal, based on its own features, can combine the influence of other signals to form a context-aware representation. This output is passed to subsequent model layers for further processing or prediction.

[0132] A key characteristic of the Transformer model is its independence from recursion; therefore, when processing time-series data, positional encoding is needed to inject positional information into the sequence. Positional encoding P∈R T×h The position encoding is used to represent the position of each signal in the sequence, and the position encoding is performed using the following formula:

[0133]

[0134] Where i is the time step index in the sequence, j is the feature dimension index, and h is the embedding dimension. This positional encoding uses sine and cosine functions to encode the positional information of each time step, ensuring that positional information of different dimensions has different scales.

[0135] After position encoding, the input signal X is added to the position code P to form an input with timing information:

[0136] X input =X+P

[0137] The input is then passed to the Transformer's multi-layer encoder to ensure that the model can understand the temporal order of the signals.

[0138] By adding location encoding to signal features, the model can simultaneously process both content and temporal information of the signal, ensuring that each signal not only understands its own characteristics but also obtains global contextual information through its correlation with other signals. The self-attention mechanism further enhances the model's temporal modeling capabilities, effectively capturing long-range dependencies in alarm signals.

[0139] 2.3 Graph Reasoning and Equipment Association Modeling

[0140] To effectively model the topological relationships of devices in a power grid and capture the complex relationships such as signal propagation and fault propagation between devices, Graph Neural Networks (GNNs) and their improved algorithms are introduced to improve the accuracy and robustness of alarm signal inference. While traditional GNN methods can propagate information through adjacency relationships between nodes, they often face problems such as insufficient information propagation and inadequate node feature updates in practical applications. Therefore, based on the core ideas of Graph Convolutional Networks (GCNs) and Graph Attention Networks (GATs), an enhanced graph inference method is proposed.

[0141] Based on the physical or logical connections of power grid equipment, a graph structure is constructed between the equipment. Let G = (V, E) be the equipment graph of the power grid, where V represents the set of equipment nodes and E is the set of edges between equipment (representing the connection or signal transmission relationship between equipment). Each equipment vi ∈ V has a feature vector, hi ∈ Rd, which represents the status of the equipment (such as equipment voltage, current, temperature, etc.) and related information of alarm signals.

[0142] To enhance the information transmission capability between devices, an adjacency matrix A∈R is defined. V|×|V| And normalization is applied:

[0143]

[0144] in, It is the adjacency matrix after adding self-connections. I is the degree matrix, and I is the identity matrix. This effectively ensures the self-information transmission capability of each node in graph reasoning.

[0145] Graph Convolutional Networks (GCNs) are used for information transfer and node feature updates in the device graph. GCNs update the feature representation of each node by weighted summation of the features of its neighbors. Let H = [h1, h2, ..., hN] be the feature matrix of all nodes in the graph. To avoid the gradient vanishing problem in multi-layer networks, residual connections are added to each layer of the graph convolutional network. The GCN update rule can be expressed as:

[0146] H (l+1) =σ(A norm H (l) W (l) +H (l) )

[0147] Among them, W (1) σ is the learnable weight matrix of layer 1, and σ is a non-linear activation function (such as ReLU). This formula represents the learning weight matrix through the adjacency matrix A. normThe features of each node are weighted and aggregated, and then transformed using learnable weights to update the node representation. This allows information in the network to remain stable across multiple layers, thereby enhancing the efficiency of information flow.

[0148] To further enhance the flexibility of information transmission between devices, a Graph Attention Network (GAT) based on a global attention mechanism is introduced, enabling each device node to consider the impact of the entire power grid structure during updates. The weights of the GAT based on the global attention mechanism are calculated adaptively using β.

[0149]

[0150] Where 'a' is the attention weight of the learning process, || represents the concatenation operation, and h... i and h j Let N(i) be the feature vectors of nodes i and j, respectively, and N(i) be the set of neighbors of node i. This attention mechanism dynamically adjusts the information propagation weights based on the relative importance of each node. The updated features of node i are:

[0151]

[0152] 2.4 Dynamic Knowledge Injection and Incremental Learning

[0153] In power grid monitoring systems, the status and fault modes of equipment change continuously over time. Therefore, to maintain the accuracy and adaptability of the inference model, a dynamic knowledge injection mechanism must be introduced, and incremental learning must be used to gradually adapt to new alarm signals and equipment status changes. This invention employs incremental learning and meta-learning methods, enabling the model to learn and update new knowledge and fault modes in real time without losing its ability to process historical data.

[0154] First, a knowledge graph K = {(h,r,t)} related to power grid equipment is constructed, where each triple represents a relationship between devices (such as "fault mode" or "device connection relationship"). To learn and utilize this knowledge more effectively, an improved version, called TransE++, is proposed based on traditional knowledge graph embedding methods (such as TransE). By introducing adaptive regularization and temporal attention mechanisms, the model can effectively handle temporal changes and complex relationships between devices during inference.

[0155] In TransE++, the triples (h, r, t) in the knowledge graph are embedded as vectors h, r, t ∈ R. d And use the following optimization objectives:

[0156]

[0157] Here, K represents all triples in the knowledge graph, h′ represents negative samples (i.e., triples that do not conform to a certain failure mode), and γ is a hyperparameter that controls the distance between positive and negative samples. To address the temporal dependency problem, a temporal attention mechanism is introduced. The updated knowledge graph embedding is composed of temporal information z. t Weighted average yields:

[0158] h t+1 =h t +α·z t

[0159] Where α is the adaptive weight, z t This indicates the timing features added at time t, which can dynamically adjust the model's response to new devices or new failure modes.

[0160] When new knowledge (such as new equipment failure types, equipment status changes, etc.) is added, the model needs to update the existing knowledge graph. Specifically, given an existing knowledge graph K... t And the current model M, and the newly added knowledge graph K t+1 The model is dynamically updated using the following formula:

[0161] h t+1 =h t +Δh t+1

[0162] Where ht represents the embedding representation of the existing knowledge graph, Δh t+1 Incremental knowledge vectors are optimized through incremental learning, aiming to minimize conflicts between newly added knowledge and existing knowledge while preserving the model's understanding of older knowledge. Incremental learning methods use meta-learning techniques to learn how to update the existing model when new knowledge is added.

[0163] To avoid global retraining when new knowledge is introduced, a Model-Agnostic Meta-Learning (MAML) strategy is employed. In MAML, the model needs to quickly adapt to new tasks or new data. Within this framework, given a new task T... t+1 (For example, new fault types), the goal is for the model to be able to adjust its parameters to suit the task in minimal steps.

[0164] For each task T t+1 The goal of updating the model parameters θ is to minimize the task loss function L(θ):

[0165]

[0166] Where α is the learning rate. Based on the current task T t+1 The gradient. Through meta-learning, the model can learn from the experience of previous tasks how to adapt to new knowledge inputs more quickly, so that each time new knowledge is added, the model can quickly update its parameters θ′ to adapt to new device failure modes and state changes.

[0167] To further enhance the effectiveness of incremental learning, dynamic adjustments are made by optimizing the knowledge graph embedding and updating strategy. Specifically, incremental knowledge K... t+1 The triples (h′, r′, t′) in the knowledge graph are embedded into the existing knowledge graph, and the embedding of the knowledge graph is fine-tuned using a loss function-based update mechanism.

[0168] Given the embedding representation Ht of the original knowledge graph Kt and the embeddings h′, r′, t′ corresponding to the new knowledge triples (h′, r′, t′), we define the loss function as:

[0169]

[0170] Here, / / . / / represents the Euclidean distance. The goal of minimizing the loss function is to make the embedding of the new triple as close as possible to its corresponding tail entity t′, while maintaining the stability and consistency of the existing knowledge graph. In this way, the system can gradually optimize the embedding representation of the knowledge graph during incremental learning, ensuring that each newly added knowledge is effectively absorbed.

[0171] By introducing incremental learning and meta-learning techniques, this invention enables the model to adapt to dynamically changing power grid monitoring environments, effectively handling new equipment fault types and equipment state changes. In this process, the incremental update and dynamic adjustment mechanism of the knowledge graph embedding ensures that the model can quickly adapt to new tasks or fault modes without losing its ability to process old knowledge, thereby maintaining the accuracy and efficiency of inference.

[0172] 2.5 Anomaly Detection and Fault Reasoning

[0173] In the input data preparation stage, alarm signals from power grid equipment are first collected and preprocessed, converting the original signal matrix `Stotal` into a feature embedding matrix `X` to provide suitable input for subsequent processing of the large model. Next, a Transformer model is used for temporal modeling, employing a self-attention mechanism to capture long-range dependencies in the alarm signals. By calculating the similarity between signals and using weighted summation, the Transformer can provide a context-aware representation of each signal, enabling the model to effectively capture temporal patterns. Furthermore, to model the complex topological relationships between power grid equipment, a Graph Neural Network (GNN) is used. Graph convolution and graph attention mechanisms are employed to update the features of equipment nodes, improving the accuracy and robustness of information propagation between equipment. Finally, by introducing dynamic knowledge injection and incremental learning techniques, the model can adapt to new alarm signals and equipment state changes in real time, rapidly updating the knowledge graph and inference capabilities to ensure the real-time performance and accuracy of the inference results. Through the combination of these technologies, this invention achieves efficient and accurate intelligent alarm signal inference, continuously optimizing inference performance in dynamically changing power grid environments. Assume y t ∈{0, 1, ..., C} represents the alarm signal s t The classification labels are C, where C is the number of fault categories. The large model ultimately outputs a probability distribution:

[0174]

[0175] Among them W cls 'b' is the weight of the classification layer, and 'b' is the bias term. This represents the enhanced feature vector resulting from fusing the features of the alarm signal with relevant knowledge extracted from the knowledge graph (e.g., equipment failure modes, operating procedures, etc.). Finally, the predicted probability for each category is calculated using the softmax function. Based on the inference results, the system can generate processing suggestions for each alarm signal, automatically performing equipment adjustments or issuing new alarms.

[0176] 3. Event association and feedback learning mechanism based on large model

[0177] By dynamically correlating alarm signals with equipment status, the system calculates the correlation degree and determines signal priorities, optimizing the signal handling sequence. Subsequently, an automated alarm collection and execution mechanism identifies signals in real time and generates corresponding handling plans, ensuring efficient response to equipment changes. Through feedback optimization and continuous learning, the system continuously adjusts its processing strategies, improving the accuracy and response speed of alarm handling, and adapting to complex changes in power grid operation, such as... Figure 4 As shown.

[0178] 3.1 Signal Event Correlation and Adaptive Handling Optimization Method

[0179] The event correlation method based on a large model enables precise correlation and priority determination of alarm signals from power grid equipment, and automatically optimizes and generates targeted handling solutions based on the real-time operating status and historical data of the equipment. Relying on the reasoning capabilities of the large model, after an alarm signal occurs, the system first uses the model to dynamically correlate the signal event with the equipment state. Given the real-time state vector x(t) of the power grid equipment, historical data dh, and alarm signal type y(t), the system quantifies the degree of correlation between the alarm signal and the equipment's operating status by calculating the correlation degree A.

[0180]

[0181] Among them, f i It is the characteristic correlation function for each specific signal, w i Here, is the importance weight of each feature, and N is the number of devices involved. This correlation degree A(t) measures the degree of match between the alarm signal and the current state of the device. In this way, the system can accurately identify which alarm signals are highly correlated with the abnormal operating state of the device.

[0182] Based on correlation, the system needs to assign a priority P(t) to each signal, which determines the urgency of signal processing. Assume... To optimize the signal priority, the system uses the following objective function to achieve adaptive optimization:

[0183]

[0184] Where C(t) is the handling cost, R(t) is the expected risk of equipment recovery, and λ1, λ2, and λ3 are the corresponding weighting coefficients. This optimization process can dynamically adjust the priority of alarm signals according to the current load and status of the equipment, prioritizing the processing of high-risk and highly correlated signals.

[0185] 3.2 Automated Alarm Collection and Execution Mechanism

[0186] Once the signal priority is determined, the system monitors and automatically identifies all alarm signals in real time through the automated alarm signal collection module. Assuming y(t) represents the alarm signal event at each time point t, the system uses a large-scale model inference engine to instantly determine the type and urgency of the alarm signal and generate a matching handling plan s(t):

[0187]

[0188] Among them, M e The execution module is responsible for generating the corresponding equipment handling plan. This process combines the reasoning results with the current state of the equipment through an integrated optimization algorithm to generate precise handling instructions.

[0189] During execution, the system can adjust its handling steps based on real-time feedback and equipment changes. For example, assuming the equipment's state changes during processing, the equipment's state vector x′(t) may need to be updated with the corresponding handling plan.

[0190]

[0191] Through this real-time dynamic adjustment, the system can flexibly respond to different operating conditions based on equipment feedback, ensuring the efficiency and safety of the response measures.

[0192] 3.3 Feedback Optimization and Continuous Learning Mechanism

[0193] After each alarm signal is handled, the system adjusts its decisions based on the results to enhance the model's adaptability; a feedback reward R(t) is set as an indicator of successful handling (such as repair success rate, equipment recovery time, etc.).

[0194]

[0195] Where δ(k) is the feedback reward of device k at time point k, typically the recovery time or repair success rate, and γ is a discount factor. The system continuously learns from each feedback, updates and optimizes its strategy, and adjusts the handling scheme more precisely when the next alarm signal occurs. Under this mechanism, the system can not only continuously optimize the alarm signal processing flow based on historical feedback, but also adaptively adjust the processing strategy according to new device states and alarm types, achieving dynamic self-adjustment and optimization. The gradual iteration of feedback learning enables the system to gradually adapt to the constantly changing fault modes and equipment configurations in the power grid, improving the accuracy and response speed of alarm signal processing. By combining large model inference capabilities, event correlation algorithms, adaptive handling optimization, automated execution, and continuous feedback mechanisms, this application provides an efficient alarm signal processing method capable of real-time automated identification and response to power grid equipment faults. Through continuous optimization and updating of decision-making strategies, the system can continuously improve processing efficiency and reduce fault risks during power grid operation.

[0196] This application effectively solves these problems by introducing innovative signal data acquisition and preprocessing technologies and a large-model-based intelligent alarm signal inference method. In terms of signal data acquisition and preprocessing, the system acquires real-time operating data of power grid equipment from multiple core data sources such as SCADA and PMU through multi-source data acquisition technology, including key parameters such as voltage, current, and frequency. Noise in the signals is removed by moving average filtering and exponential smoothing, and missing values ​​are filled using KNN, ensuring the stability and integrity of the data and providing high-quality input data for subsequent inference, thus improving the accuracy and stability of the processing. In terms of intelligent inference, this invention adopts a large-model Transformer structure, capturing long-range dependencies between signals through a self-attention mechanism. This enables global dependency modeling of power grid alarm signals, ensuring that each signal is accurately interpreted within its context. Simultaneously, the combination of Graph Neural Networks (GNN) and Graph Attention Networks (GAT) allows the system to model the topological relationships of power grid equipment more flexibly and accurately. To avoid the gradient vanishing problem in deep networks, residual connections are added to each layer of the graph convolutional network, ensuring stable information transmission and further enhancing the model's deep learning capabilities. By employing incremental learning and meta-learning techniques, the system can adapt to new equipment fault modes and state changes in real time, continuously optimizing reasoning and handling strategies, thus improving the intelligent management level of power grid equipment. This solution significantly enhances the automation, intelligence, and adaptability of power grid alarm signal processing. Combining these advanced technologies, this invention significantly improves the automation level and real-time response capability of power grid alarm signal processing, effectively addressing power grid equipment faults and emergencies, and ensuring the efficient, stable, and safe operation of the power grid in complex environments.

Claims

1. A method for alarm signal inference and adaptive handling optimization based on a large model, characterized in that, Includes the following steps: (1) Signal data acquisition and preprocessing, including data acquisition, data removal and cleaning, and missing value filling; (2) Intelligent alarm signal reasoning based on large model, including input data preparation and embedding representation, large model time series modeling based on Transformer, graph reasoning and device association modeling, dynamic knowledge injection and incremental learning, and anomaly detection and fault reasoning. (3) Event correlation and feedback learning mechanism based on large model, including signal event correlation and adaptive handling optimization method, automated alarm collection and execution mechanism, and feedback optimization and continuous learning mechanism.

2. The alarm signal reasoning and adaptive handling optimization method based on a large model according to claim 1, characterized in that, The data collection in step (1) includes: (1.1) Real-time telemetry and teleindication data are collected through the SCADA system and set as S s ={s1, s2, ..., s n }, where s i S represents the device status and telemetry data at a certain moment. s As the core data source, it connects with field devices and uses communication protocols to transmit data to the monitoring center in real time; (1.2) Obtain high-precision, time-synchronized voltage / current phasors, frequency, and rate of change of frequency via PMU; set as S p ={v i i i f i , Δf i }, where v i and i i f represents the voltage and current phasors, respectively. i For frequency, Δf i For the frequency change rate, the protection device and fault recorder provide the system with detailed fault event records, protection action timing, and fault waveforms, set as S. Pr = {ej, tj, wj}, where e j Indicates the type of fault event, t j For the fault timestamp, w j This is fault waveform data; data from the equipment condition monitoring system, set as S. Eq ={o k ,pd k , t k v k }, where o k This represents oil chromatography data, pd k For partial discharge, t k v represents the equipment temperature. k For mechanical vibration data, the energy management system is set to S. E ={Tg,Ps,Mn}, where Tg represents the power grid topology, Ps is the dispatch plan, Mn is the network model information, and the maintenance management system is set to S. C ={hm,pm,dm,rm}, where hm is the equipment maintenance history, pm is the maintenance plan, dm is the defect record, and rm is the test report; (1.3) The meteorological information system provides environmental data and is set as follows: Where θ is temperature. For humidity, v w For wind speed, r w For rainfall, 1 w Lightning information is aggregated and processed to form a multi-dimensional signal dataset, Statal.

3. The alarm signal reasoning and adaptive handling optimization method based on a large model according to claim 1, characterized in that, Step (2), input data preparation and embedding representation, includes obtaining a dataset matrix S through data acquisition and preprocessing. total Its dimension is 7×d; For the input signal Stotal, we first perform feature embedding on it, assuming each signal s t ∈R d Let be the signal feature vector at time step t, representing the feature space of a single signal, and d be the dimension of the feature. Then the embedding process is as follows: Where Embed(·) represents the feature embedding layer, and the output is a vector of dimension h, where h is the hidden layer dimension of the model. The entire input signal matrix Sttotal is transformed into an embedding matrix: The entire dataset Stotal is a T×d matrix, where T is the number of time steps, d is the feature dimension of each signal, and h is the dimension of the embedded signal.

4. The alarm signal reasoning and adaptive handling optimization method based on a large model according to claim 1, characterized in that, Step (2) of large-scale temporal modeling based on Transformer includes: The Transformer model is used to perform global dependency modeling on the power grid alarm signal sequence to capture potential temporal patterns in the signal. Assume the input signal sequence X∈R T×h The embedded matrix input X is mapped to a query, key, and value through a linear transformation: Q=XW Q ,K=XW K ,V=XW V Where WQ, WK, WV∈R h×hk It is a learnable weight matrix, h k It refers to the dimensions of query, key, and value; The relevance between query Q and key K is calculated, and the similarity matrix S∈R is obtained by dot product. T×T : The similarity matrix is ​​normalized to ensure that the sum of the weights is 1. Where the weight matrix α represents the correlation between each query signal and all key signals, S ij These are the elements in the similarity matrix S; The calculated attention weights α are applied to the value matrix V, and the final output representation is obtained by weighted summation: A=αV Where A∈R T×hk The positional encoding P∈R is obtained by weighting the value matrix with attention weights. T×h The position encoding is used to represent the position of each signal in the sequence, and the position encoding is performed using the following formula: Where i is the time step index in the sequence, j is the feature dimension index, and h is the embedding dimension. This positional encoding uses sine and cosine functions to encode the positional information of each time step, ensuring that positional information of different dimensions has different scales. After position encoding, the input signal X is added to the position code P to form an input with timing information: X input =X+P The input is then passed to the Transformer's multi-layer encoder to ensure that the model can understand the temporal order of the signals.

5. The alarm signal reasoning and adaptive handling optimization method based on a large model according to claim 1, characterized in that, Step (2) of graph reasoning and device association modeling includes: Let G = (V, E) be the equipment graph of the power grid, where V represents the set of equipment nodes, E is the set of edges between equipment, and each equipment vi ∈ V has an eigenvector, hi ∈ R. d It represents the state of the device, and the adjacency matrix A∈R is defined. |V|×|V| And normalization is applied: in, It is the adjacency matrix after adding self-connections. Here, H is the degree matrix, and I is the identity matrix. Based on the Graph Convolutional Network (GCN), information transfer and node feature updates are performed on the device graph. Let H = [h1, h2, ..., hN] be the feature matrix of all nodes in the graph. The GCN update rule is expressed as: H (l+1) =σ(A norm H (l) W (l) +H (l) ) Among them, W (1) σ is the learnable weight matrix of the first layer, and σ is the non-linear activation function; A graph attention network (GAT) based on a global attention mechanism is introduced. The weights of the graph attention network based on the global attention mechanism are calculated in an adaptive manner, β: The update characteristics of node i are: Where 'a' is the attention weight of the learning process, || represents the concatenation operation, and h... i and h j Let i and j be the feature vectors of node i and node j respectively, and N(i) be the set of neighbors of node i.

6. The alarm signal reasoning and adaptive handling optimization method based on a large model according to claim 1, characterized in that, Step (2), dynamic knowledge injection and incremental learning, includes: Construct a knowledge graph K = {(h, r, t)} related to power grid equipment. In TransE++, embed the triples (h, r, t) in the knowledge graph as vectors h, r, t ∈ R. d And use the following optimization objectives: Where K represents all triples in the knowledge graph, h′ represents negative samples, and γ is a hyperparameter. A temporal attention mechanism is introduced, and the updated knowledge graph embedding is composed of temporal information z. t Weighted average yields: h t+1 =h t +α·z t Where α is the adaptive weight, z t This indicates the temporal feature added at time t. Employing a model-independent meta-learning strategy, in MAML, given a new task T... t+1 For each task T t+1 The goal of updating the model parameters θ is to minimize the task loss function L(θ): Where α is the learning rate. Based on the current task T t+1 The gradient; Incremental knowledge K t+1 The triples (h′, r′, t′) in the original knowledge graph Kt are embedded into the existing knowledge graph. Given the embedding representation Ht of the original knowledge graph Kt and the embeddings h′, r′, t′ corresponding to the new knowledge triples (h′, r′, t′), the loss function is defined as: Here, || represents the Euclidean distance, and the goal of minimizing the loss function is to make the embedding of the new triplet as close as possible to its corresponding tail entity t′.

7. The alarm signal reasoning and adaptive handling optimization method based on a large model according to claim 1, characterized in that, The optimization method for signal event correlation and adaptive handling in step (3) includes: The model dynamically correlates signal events with equipment states. Given the real-time state vector x(t) of the power grid equipment and historical data d, h In addition to the alarm signal type y(t), the system quantifies the correlation between the alarm signal and the equipment operating status by calculating the correlation degree A: Among them, f i It is the characteristic correlation function for each specific signal, w i Here, N is the importance weight of each feature, and N is the number of devices involved. The system assigns a priority P(t) to each signal. Assume... To optimize the signal priority, the system uses the following objective function to achieve adaptive optimization: Where C(t) is the disposal cost, R(t) is the expected risk of equipment recovery, and λ1, λ2, and λ3 are the corresponding weighting coefficients.

8. The alarm signal reasoning and adaptive handling optimization method based on a large model according to claim 1, characterized in that, The automated alarm collection and execution mechanism in step (3) includes: Assuming y(t) represents an alarm signal event at each time point t, the system uses a large model inference engine to instantly determine the type and urgency of the alarm signal and generate a matching handling plan s(t): Among them, M e The execution module is responsible for generating corresponding equipment handling plans; during execution, the system can adjust the handling steps based on real-time feedback and equipment changes.

9. The alarm signal reasoning and adaptive handling optimization method based on a large model according to claim 1, characterized in that, The feedback optimization and continuous learning mechanism in step (3) includes: After each alarm signal is processed, the system adjusts its decision based on the results, setting the feedback reward R(t) as the indicator of whether the processing was successful. Where δ(k) is the feedback reward of device k at time point k, which is usually the recovery time of the device or the success rate of the repair, and γ is the discount factor.

10. A large-model-based alarm signal inference and adaptive handling optimization system, characterized in that, include: The data processing module is used for signal data acquisition and preprocessing, including data acquisition, data removal and cleaning, and missing value filling. The signal reasoning module is used for intelligent alarm signal reasoning based on large models, including input data preparation and embedding representation, large model time series modeling based on Transformer, graph reasoning and device association modeling, dynamic knowledge injection and incremental learning, and anomaly detection and fault reasoning. The feedback learning module is used for event correlation and feedback learning mechanisms based on large models, including signal event correlation and adaptive handling optimization methods, automated alarm collection and execution mechanisms, and feedback optimization and continuous learning mechanisms.

Images