Station area intelligent fusion terminal data processing system based on edge calculation

Abstract

The invention discloses a transformer area intelligent fusion terminal data processing system based on edge calculation, relates to the field of edge intelligence, and solves the problems of poor real-time performance and poor data exception processing accuracy of an existing intelligent fusion terminal data processing system. According to the method, communication between micro-services is efficiently managed through the container arrangement engine and the service grid, the message transmission delay is reduced, and resource allocation is further optimized through a resource scheduling algorithm based on reinforcement learning, so that system resources are used most reasonably; according to the method, anomaly detection is rapidly and accurately carried out through a space-time heterogeneous fusion network model, hierarchical processing of data and on-cloud training of the model are realized through a cloud-edge collaborative architecture, edge nodes participate in model training through a federated learning mechanism, the model performance is optimized, the transmission time of the data in the network is shortened, and the problem positioning speed is increased; the real-time performance and the reliability of an existing system are greatly improved.

Description

Technical Field

[0001] The present invention relates to the field of edge intelligence, and more specifically to an area intelligent fusion terminal data processing system based on edge computing. Background Art

[0002] With the acceleration of intelligent and digital transformation, the power industry faces huge challenges in improving operational efficiency and ensuring grid security. The intelligent fusion terminal data processing system based on edge computing has emerged. By performing data processing tasks at the edge of the network, it reduces dependence on central servers, reduces data transmission delays, and improves the response speed of the system. The application of this technology enables the power system to detect and recover faults more quickly, while providing strong support for the construction of smart grids.

[0003] The existing intelligent fusion terminal data processing systems in the substation area integrate advanced sensors, communication technologies and computing capabilities, and can realize real-time monitoring and data collection of power equipment. These systems are usually modular in design, which is easy to expand and upgrade. For example, some systems can analyze data through machine learning algorithms to predict and diagnose potential equipment failures. However, although these technologies have made significant progress in improving data processing capabilities, existing systems still face some challenges in terms of real-time performance and data anomaly processing.

[0004] Although edge computing has brought optimization to data processing for intelligent fusion terminals in substations, the existing system still has some shortcomings in terms of real-time performance and data anomaly processing. First, real-time performance requires the system to be able to quickly process and respond to large amounts of data, but the existing system may experience delays under high load conditions, affecting the timeliness of fault detection and response. Second, data anomaly processing is key to ensuring the stable operation of the power system. Existing systems may lack efficient anomaly detection algorithms, or the accuracy and robustness of anomaly detection may be insufficient when faced with complex and changing power grid environments.

[0005] Therefore, the present invention discloses a station area intelligent fusion terminal data processing system based on edge computing. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present invention discloses a data processing system for an intelligent fusion terminal in an area based on edge computing. The present invention can improve the real-time data processing and the accuracy of anomaly detection of the data processing system.

[0007] In order to achieve the above technical effects, the present invention adopts the following technical solutions:

[0008] A station area intelligent fusion terminal data processing system based on edge computing, comprising:

[0009] An edge acquisition module is used to collect and pre-process data from IoT devices; the edge acquisition module collects real-time data from the terminal devices in the substation area through the IoT communication network NB-IoT and the data transmission protocol MQTT; and performs preliminary filtering of the data through the sliding window averaging method; the terminal devices in the substation area include but are not limited to smart meters, environmental sensors and status monitors;

[0010] Dynamic scheduling and resource management module, used to optimize microservice performance and collaboration; the dynamic scheduling and resource management module realizes automatic management and efficient communication of microservices through the container orchestration engine Kubernetes and the service grid Istio, and dynamically adjusts the number of service instances according to the system load through the resource scheduling algorithm based on reinforcement learning, thereby optimizing resource utilization efficiency and task response speed;

[0011] Edge caching and preprocessing module, used for efficient caching and stream processing of data; the edge caching and preprocessing module dynamically caches data through a distributed caching system and a time series analysis and prediction model, and preprocesses data through a stream processing framework; the stream processing framework performs deep cleaning and structural processing of data through normalization, missing value filling and feature extraction;

[0012] Intelligent analysis and anomaly detection module, used to identify time series and image anomalies; the intelligent analysis and anomaly detection module realizes multi-dimensional anomaly detection and pattern recognition of data through a spatiotemporal heterogeneous fusion network model; the spatiotemporal heterogeneous fusion network model captures the correlation between the time series and spatial dimensions of data through a spatiotemporal convolutional network, and captures the forward and backward dependencies of abnormal data through a bidirectional long short-term memory network; the spatiotemporal heterogeneous fusion network model enhances the recognition ability of abnormal patterns in multimodal data through a heterogeneous feature attention mechanism;

[0013] The edge-cloud-edge collaboration module is used to coordinate data synchronization between the edge and the cloud, model updates, and the distribution and recovery of complex analysis tasks. The edge-cloud-edge collaboration module implements hierarchical data processing and model training on the cloud through a cloud-edge collaboration architecture, and allows edge nodes to participate in model training through a federated learning mechanism, thereby optimizing model performance while protecting privacy.

[0014] A security and privacy protection module is used to ensure the secure flow of data. The security and privacy protection module ensures the security of data transmission and the privacy of data in processing through the security protocol TLS / SSL and homomorphic encryption, and improves the system security level through differential privacy and zero-trust network strategies.

[0015] Among them, the output end of the edge acquisition module is connected to the input end of the edge cache and preprocessing module and the edge cloud edge collaboration module; the output end of the edge cache and preprocessing module and the edge cloud edge collaboration module are connected to the input end of the intelligent analysis and anomaly detection module; the security and privacy protection module and the dynamic scheduling and resource management module work throughout the process.

[0016] As a further technical solution of the present invention, the container orchestration engine Kubernetes ensures high availability and load balancing of services through controller mode and declarative update; the dynamic scheduling and resource management module includes a microservice management unit, a resource scheduling unit, a communication coordination unit and an intelligent optimization unit; the microservice management unit realizes the automatic deployment, expansion and management of microservices through the container orchestration engine Kubernetes; and provides intelligent routing, traffic management, policy execution and telemetry data collection between microservices through the service grid; the resource scheduling unit collects system resource usage through a resource monitoring tool; and dynamically adjusts the number of service instances through a resource scheduling algorithm based on reinforcement learning; the system resource usage includes at least CPU, memory and network bandwidth usage; the communication coordination unit realizes automatic discovery and communication between microservices through the service registration and discovery mechanism of the service grid Istio, and realizes at least a communication coordination strategy including A / B testing and canary deployment through the traffic segmentation and redirection function of the service grid Istio; the intelligent optimization unit predicts resource demand trends through an LSTM network, and automatically adjusts resource allocation according to real-time monitoring data and prediction results through a feedback control mechanism to optimize task response speed and resource utilization efficiency.

[0017] As a further technical solution of the present invention, the working steps of the resource scheduling algorithm based on reinforcement learning in the intelligent fusion terminal data processing system based on edge computing are as follows:

[0018] Step 1: Modeling the environment state: Modeling the resource state and task load in the edge computing environment through a multi-dimensional state representation formula. The resource scheduling algorithm based on reinforcement learning defines the state vector S t ={x1,x2,...,x i ,...,x n} represents the system state at time t, where x i Indicates the state of the i-th resource instance, the instance state includes CPU usage, memory usage, network bandwidth, number of service instances and task queue length parameters; the formula expression of the multi-dimensional state characterization formula is:

[0019] x i ={cpu it ,mem it ,bwit ,ins it ,q it} (1)

[0020]

[0021] In formula (1) and formula (2), cpu it ,mem it ,bw it Represents the usage status vectors of CPU, memory and network resources respectively; ins it is the system load; q it is the task queue length; for each resource u i is the utilization rate, c i is the resource load, t i is the resource capacity, l i is the task type;

[0022] Step 2: Resource scheduling strategy learning: Strategy T parameterized by deep neural network θ (a|S t ) Select the best action; strategy T θ (a|S t ) Output in a given state S t The probability of taking action a is calculated, and the network parameter θ is optimized by the gradient ascent method. The parameterized formula of the deep neural network is:

[0023]

[0024] In formula (3), θ represents the policy network parameter, which is used to determine the probability distribution of the network output; h(S t , a) The feature vector representing the state-action pair, which is used to represent the combined features of state and action in the policy network; D iI Indicates that in state S t The action taken under t , a) is in state S t The immediate reward after taking action a;

[0025] Step 3: Action execution and feedback: At each time step t, according to the current strategy T θ Select an action t And execute, action a t Including but not limited to increasing or decreasing the number of service instances and adjusting resource allocation; after executing the action, new feedback information is provided through the system dynamic feedback formula. The feedback information includes the new state S t+1 and return r t ; The formula expression of the system dynamic feedback formula is:

[0026]

[0027] In formula (4) and formula (5), γ represents the cost sensitivity parameter, c(a t ) is action a t cost; w1, w2 and w3 represent weight coefficients, which are used to balance the importance of different system performance indicators; S max Indicates the upper limit of resource capacity, used to standardize resource utilization and queue length; Δcpu it Indicates the change in CPU usage, which is used to measure the change in resource status; N(0, σ 2 ) is a noise term, which is used to represent the randomness and uncertainty of the environment;

[0028] Step 4: Long-term return and strategy evaluation: Evaluate the long-term return of the current strategy through the long-term return evaluation formula, that is, in strategy T θ Next from state S t The expected return sum at the beginning; the long-term return evaluation formula reflects the expectation of long-term return through the time difference learning update strategy; the formula expression of the long-term return evaluation formula is:

[0029] V = ∑log2(Q θ (S t , a) 2 -βE θ (S t )) 2 (6)

[0030] In formula (6), V is the strategy T θ long-term return; β is the discount factor, which is used to adjust the importance of future returns relative to immediate returns; Q θ To represent the state S t The expected return of taking action a, E θ (S t ) indicates that from state S t Start following strategy T θ expected return;

[0031] Step 5: Strategy iteration and optimization; Finally, based on the evaluation of long-term returns, the reinforcement learning-based resource scheduling algorithm continuously iterates and optimizes the strategy network through the strategy iteration optimization formula; the formula expression of the strategy iteration optimization formula is:

[0032]

[0033] In formula (7), T new is the updated parameter of the policy network; y is the expected target value of the return at multiple time steps, A θ (St , a t ) is the advantage function, representing action a t In state S t is the advantage of the policy network over the average policy; μ is the learning rate of the policy network, which is used to control the step size of parameter update; δ is the exploration rate, which is used to control the conservatism of policy update.

[0034] As a further technical solution of the present invention, the distributed cache system includes a data cache module, a cache management module and a prediction model integration module; the data cache module includes a data sharding unit and a data serialization unit; the data sharding unit evenly distributes data requests to multiple nodes of the Redis cluster through hash allocation and consistent hashing algorithm; the data serialization unit compresses and serializes cached data through a serialization protocol PB; the cache management module ensures efficient use of cache space through a cache invalidation strategy, and ensures consistency and synchronous update of data in a distributed system through a publish / subscribe mode and a transaction mechanism; the cache invalidation strategy automatically eliminates infrequently used data through a least recently used mechanism; the cache management module includes a data security unit, a performance monitoring unit and a cache optimization unit; the data security unit encrypts the data in the cache through an elliptic encryption algorithm, and realizes data persistence and disaster recovery through regular snapshots and WAL pre-write logs; the performance monitoring unit monitors the performance indicators of cache operations in real time through a real-time monitoring tool; based on the monitoring data, the cache optimization unit dynamically adjusts the Redis configuration parameters through a dynamic cache strategy; the prediction model integration module provides intelligent support for the cache strategy through time series prediction and Akaike information criterion.

[0035] As a further technical solution of the present invention, the dynamic caching strategy first identifies the optimal configuration template under different working modes through cluster analysis, based on the cluster analysis results; the dynamic caching strategy dynamically adjusts the configuration parameters of the Redis cluster through a dynamic configuration adjustment mechanism, including but not limited to adjusting the maximum memory limit, expiration key strategy, number of replicas and sharding strategy; the dynamic configuration adjustment mechanism ensures the smooth progress of the adjustment operation without affecting service availability through the configuration update protocol RC; when the system recognizes that the load of a specific shard is too high, the dynamic caching strategy automatically triggers the data migration process through dynamic sharding and rebalancing methods, balances the data distribution between nodes, and ensures the efficiency of data migration and the consistency of data access through a consistent hashing algorithm.

[0036] As a further technical solution of the present invention, the time series analysis and prediction model includes an input layer, a data layer, a feature construction layer, a model training layer, a prediction generation layer, a prediction integration layer and an intelligent optimization layer; the working method of the time series analysis and prediction model in the intelligent fusion terminal data processing system based on edge computing is:

[0037] Step 1: Acquire the collected data of the edge collection module through the input layer; the input layer ensures the time consistency of data collection through the time protocol and timestamp synchronization mechanism;

[0038] Step 2: Clean and standardize the collected data through the data layer; the data layer performs signal denoising through adaptive Kalman filtering and processes missing data through spline interpolation;

[0039] Step 3, extracting time series features from the preprocessed data through the feature construction layer; the feature construction layer obtains time series features through wavelet transform and rolling statistics analysis, and removes redundant features through feature selection based on p value;

[0040] Step 4: Build and train a time series prediction model through the model training layer; the model training layer implements model construction and training through a machine learning framework, and trains model parameters through a batch gradient descent method

[0041] Step 5: Generate a prediction result through the prediction generation layer; the prediction generation layer calculates the prediction value through the forward propagation of the model, and provides a quantification of the uncertainty of the prediction through the confidence interval estimation;

[0042] Step 6: Fusion of multiple model prediction results through the prediction integration layer; the prediction integration layer improves prediction accuracy through integrated learning methods, weight adjustment and voting mechanisms;

[0043] Step 7: According to the deviation between the prediction result and the actual observation value, the model parameters and the prediction strategy are intelligently adjusted through the intelligent optimization layer; the intelligent optimization layer adjusts the model through an incremental learning method;

[0044] Step 8: Feedback the prediction performance to the model training layer through the intelligent optimization layer; the intelligent optimization layer continuously adjusts the model structure and parameters through a closed-loop feedback mechanism to adapt to changes in data distribution.

[0045] As a further technical solution of the present invention, the spatiotemporal heterogeneous fusion network model includes an input layer, a spatiotemporal convolution layer, a feature fusion layer, a bidirectional long short-term memory layer, an attention mechanism layer, a heterogeneous feature enhancement layer and an output layer; the working method steps of the spatiotemporal heterogeneous fusion network model in the intelligent fusion terminal data processing system based on edge computing are as follows:

[0046] S1, receiving raw data through the input layer, and the input layer eliminates dimensional differences between data by data resampling and standardization;

[0047] S2, extracting features through the spatiotemporal convolution layer; the spatiotemporal convolution layer extracts the time features of the time series and the spatial features of the image through a one-dimensional convolution layer and a two-dimensional convolution layer respectively; the one-dimensional convolution layer applies a one-dimensional convolution kernel to slide on the time axis, extracts features through convolution operation, and uses the filter size and step size parameters to control the granularity and coverage of feature extraction; the two-dimensional convolution layer slides on the image plane through a two-dimensional convolution kernel, and identifies different types of visual features through different filter groups, thereby enhancing the spatial understanding ability of the model;

[0048] S3, fusing the features extracted from the time series and image data through the feature fusion layer to form a unified spatiotemporal feature representation; the feature fusion layer ensures the effective integration of cross-modal information through feature cascading and feature mapping methods;

[0049] S4. Performing bidirectional analysis on time series data through the bidirectional long short-term memory layer to capture the forward and backward dependencies of the sequence and improve the recognition of abnormal event sequence patterns; the bidirectional long short-term memory layer captures the long-term dependencies in the sequence by combining forward and reverse LSTM units, and controls the information flow through a gating mechanism to enhance the ability to capture sequence dynamics;

[0050] S5. Dynamically adjust the importance of different features through the attention mechanism layer; the attention mechanism layer assigns weights to different features through a weighted sum strategy, and the weighted sum strategy improves the model's attention to abnormal patterns through dot product attention;

[0051] S6. Further refining and enhancing spatiotemporal features through the heterogeneous feature enhancement layer; the heterogeneous feature enhancement layer enhances the richness and robustness of feature representation through depth-separable convolution, residual connection and gated activation function;

[0052] S7, based on the features of the heterogeneous feature enhancement layer, generating anomaly probability and classification labels through the output layer;

[0053] The spatiotemporal convolution layer extracts information-rich spatiotemporal feature maps from the original data through a spatiotemporal feature extraction formula; the formula expression of the spatiotemporal feature extraction formula is:

[0054]

[0055] In formula (8), represents the spatiotemporal feature map of the output of layer I; , is a nonlinear activation function used to increase the expressive power of the model; τ is the slope of the linear regression statistical operation; k is the number of filters; is the bias term of layer I, used to adjust the convolution output; X I-1 Represents the output of the previous layer, which is used as the input feature of the current layer; Indicates that the Ith layer passes through k different spatiotemporal convolution kernels X on the previous layer I-1 The input is convolved to extract features of different scales and directions; * represents the convolution operator;

[0056] The bidirectional long short-term memory layer captures the front-end dependency of the time series data through a bidirectional recurrent network with a gating mechanism; the state update formula of the bidirectional long short-term memory layer is as follows:

[0057]

[0058] In formula (9), W is the hidden state of the forward and reverse LSTM units; ; is the sigmoid activation function, which is used to generate a value between 0 and 1 to control the degree of information passing; x is the bias term of the forward and reverse LSTM, which is used for bias adjustment; v is the weight matrix from the hidden state to the hidden state of the forward and reverse LSTM, which is used to transfer historical information; is the residual of linear regression statistics; c t is a comprehensive cell state that combines the forward and reverse hidden states and memory cell contents;

[0059] The attention mechanism layer and the heterogeneous feature enhancement layer selectively enhance the spatiotemporal features and the features of the bidirectional LSTM output through the heterogeneous feature attention allocation formula to ensure that the model focuses on the most relevant features; the formula expression of the heterogeneous feature attention allocation formula is:

[0060]

[0061] In formula (10), R is the final enhanced feature representation; e i is the attention score, which is used to measure the importance of the i-th feature; b a is the bias term of the attention layer, which is used to adjust the initial score of the feature combination; a i is the normalized attention weight, n is the number of feature vectors; W a is the weight matrix of the attention layer, which is used to learn the importance of different feature combinations; LeakyReLU represents the activation function, which is used to increase nonlinearity.

[0062] As a further technical solution of the present invention, the edge-cloud-edge collaboration module includes a data synchronization and distribution unit, a model update and optimization unit, and a federated learning management unit; the data synchronization and distribution unit performs data transmission and intelligent distribution through a two-way data transmission mechanism and a data layered processing strategy; the two-way data transmission mechanism ensures real-time synchronization of data between the edge side and the cloud side through the MQTT protocol and the gRPC service; the data layered processing strategy adopts an event-driven data distribution mechanism to automatically distribute data processing tasks between the edge side and the cloud side; the model update and optimization unit constructs a model training environment in the cloud side through a distributed computing framework; the distributed computing framework uses a stochastic gradient The model parameters are optimized by descending SGD; the model update and optimization unit realizes real-time model version management and seamless deployment update on the edge side through containerization and container orchestration system; the container orchestration system gradually promotes new models through grayscale release strategy; the federated learning management unit realizes efficient aggregation of model parameters through federated learning protocol and federated aggregation algorithm; the federated learning protocol enables multiple edge nodes to collaborate in training global models without leaking local data based on secure multi-party computing (SMPC) and differential privacy method; the federated aggregation algorithm aggregates model updates submitted by edge nodes in the cloud through federated averaging algorithm (FA), ensuring that the model reaches global optimality after multiple iterations.

[0063] As a further technical solution of the present invention, the working principle steps of the security and privacy protection module in the area intelligent fusion terminal data processing system based on edge computing are as follows:

[0064] Step 1: Security data reception and verification;

[0065] A secure channel is established at the early stage of data transmission through the TLS / SSL protocol; the TLS / SSL protocol verifies the identities of both parties in communication through a public key infrastructure and ensures the secure exchange of session keys through a handshake protocol;

[0066] Step 2: Data storage encryption;

[0067] The security protection of the stored data is achieved through a homomorphic encryption method, which allows direct calculation on the ciphertext through a polynomial ring;

[0068] Step 3: Preprocessing of privacy data;

[0069] Anonymize private data through differential privacy; the differential privacy balances privacy protection and data analysis accuracy by controlling the variance of noise through the Laplace equation;

[0070] Step 4: Access control and permission management;

[0071] Implement strict permission control on data access through role-based access control and attribute-based access control strategies;

[0072] Step 5: Dynamic security policy deployment;

[0073] Based on the current network status, dynamic security policy adjustment is performed through zero-trust network strategy; the zero-trust network strategy achieves fine control of network traffic through software-defined network SDN;

[0074] Step 6: Security audit and intrusion detection;

[0075] Continuously monitor the system security status and identify potential threats through a real-time security monitoring system and an intrusion detection system; the real-time security monitoring system identifies and responds to abnormal behaviors through machine learning models; the intrusion detection system tracks and records all key operations through log analysis and access pattern recognition methods;

[0076] Step 7: Data leakage prevention;

[0077] Monitor, identify and protect sensitive data through data loss prevention strategy DLP; the data loss prevention strategy DLP prevents unauthorized data transmission and leakage through access control and data encryption methods;

[0078] Step 8: Security incident response and recovery;

[0079] Ensure rapid response to security incidents and system recovery through security incident response mechanism and disaster recovery backup method.

[0080] Positive beneficial effects:

[0081] The present invention realizes the automatic expansion and contraction of microservices through the container orchestration engine Kubernetes, ensuring that the number of service instances is dynamically adjusted according to the real-time system load. Combined with the service grid Istio, the system can efficiently manage the communication between microservices and reduce the message transmission delay. In addition, the resource scheduling algorithm based on reinforcement learning further optimizes resource allocation, so that system resources are used in the most reasonable way and the task response speed is improved. This intelligent resource management and scheduling mechanism effectively avoids the processing delays that occur during the peak period of data traffic and ensures the smooth operation of the system.

[0082] The intelligent analysis and anomaly detection module significantly improves the accuracy and robustness of anomaly detection by adopting a spatiotemporal heterogeneous fusion network model. This model combines spatiotemporal convolutional networks and bidirectional long short-term memory networks to capture dynamic patterns in time series data and distribution characteristics of spatial data, thereby more accurately identifying anomalies. In addition, the heterogeneous feature attention mechanism enhances the model's ability to recognize abnormal patterns in multimodal data, enabling the system to accurately detect potential anomalies in complex power grid environments. The application of this deep learning and pattern recognition technology provides the system with powerful data analysis capabilities, ensuring that anomaly detection can be performed quickly and accurately when faced with changes in power grid status.

[0083] The edge-cloud-edge collaboration module realizes hierarchical data processing and cloud-based model training through the cloud-edge collaboration architecture, and uses the federated learning mechanism to allow edge nodes to participate in model training and optimize model performance. This architecture not only improves the accuracy of the model, but also enables rapid local data analysis and processing through edge computing capabilities when an anomaly is detected, reducing the data transmission time in the network and speeding up problem location. In addition, the rich contextual information provided by the intelligent analysis and anomaly detection module provides operation and maintenance personnel with a basis for quickly locating the source of the problem, thereby improving processing efficiency and system stability. This rapid response mechanism greatly shortens the time for fault detection and recovery and enhances the reliability of the power grid. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative work, among which:

[0085] Figure 1 This is a framework diagram of a data processing system for intelligent fusion terminals in a substation area based on edge computing according to the present invention;

[0086] Figure 2 A working step diagram of the resource scheduling algorithm based on reinforcement learning of the present invention;

[0087] Figure 3 It is a working principle diagram of the information management database module of the present invention;

[0088] Figure 4 It is a framework diagram of the spatiotemporal heterogeneous fusion network model of the present invention. DETAILED DESCRIPTION

[0089] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0090] like Figure 1-Figure 4 As shown, a data processing system for intelligent fusion terminals in a substation based on edge computing includes: an edge acquisition module for collecting and preprocessing data from IoT devices; the edge acquisition module collects real-time data from the substation terminal devices through the IoT communication network NB-IoT and the data transmission protocol MQTT; and performs preliminary filtering on the data through the sliding window averaging method; the substation terminal devices include but are not limited to smart meters, environmental sensors and status monitors;

[0091] Dynamic scheduling and resource management module, used to optimize microservice performance and collaboration; the dynamic scheduling and resource management module realizes automatic management and efficient communication of microservices through the container orchestration engine Kubernetes and the service grid Istio, and dynamically adjusts the number of service instances according to the system load through the resource scheduling algorithm based on reinforcement learning, thereby optimizing resource utilization efficiency and task response speed;

[0092] Edge caching and preprocessing module, used for efficient caching and stream processing of data; the edge caching and preprocessing module dynamically caches data through a distributed caching system and a time series analysis and prediction model, and preprocesses data through a stream processing framework; the stream processing framework performs deep cleaning and structural processing of data through normalization, missing value filling and feature extraction;

[0093] Intelligent analysis and anomaly detection module, used to identify time series and image anomalies; the intelligent analysis and anomaly detection module realizes multi-dimensional anomaly detection and pattern recognition of data through a spatiotemporal heterogeneous fusion network model; the spatiotemporal heterogeneous fusion network model captures the correlation between the time series and spatial dimensions of data through a spatiotemporal convolutional network, and captures the forward and backward dependencies of abnormal data through a bidirectional long short-term memory network; the spatiotemporal heterogeneous fusion network model enhances the recognition ability of abnormal patterns in multimodal data through a heterogeneous feature attention mechanism;

[0094] The edge-cloud-edge collaboration module is used to coordinate data synchronization between the edge and the cloud, model updates, and the distribution and recovery of complex analysis tasks. The edge-cloud-edge collaboration module implements hierarchical data processing and model training on the cloud through a cloud-edge collaboration architecture, and allows edge nodes to participate in model training through a federated learning mechanism, thereby optimizing model performance while protecting privacy.

[0095] A security and privacy protection module is used to ensure the secure flow of data. The security and privacy protection module ensures the security of data transmission and the privacy of data in processing through the security protocol TLS / SSL and homomorphic encryption, and improves the system security level through differential privacy and zero-trust network strategies.

[0096] Among them, the output end of the edge acquisition module is connected to the input end of the edge cache and preprocessing module and the edge cloud edge collaboration module; the output end of the edge cache and preprocessing module and the edge cloud edge collaboration module are connected to the input end of the intelligent analysis and anomaly detection module; the security and privacy protection module and the dynamic scheduling and resource management module work throughout the process.

[0097] In a specific embodiment, when the system starts, the intelligent data acquisition module (SDC) is initialized, and the connection with the terminal devices such as smart meters and environmental sensors in the substation is configured, and a secure communication link is established using LPWAN technology (such as LoRaWAN) and MQTT protocol. The terminal device periodically sends real-time data (electrical parameters, environmental status, etc.) to the edge computing node. After the edge side receives the data, it immediately performs preliminary cleaning of the data through lightweight preprocessing logic (such as simple filtering rules, sliding average denoising) to screen out obvious errors or outliers. The dynamic microservice orchestration module (DMO) uses AI algorithms to automatically adjust the number of microservice instances and optimize resource allocation according to the current system load. Kubernetes is responsible for container scheduling, and Istio manages communication and traffic control between services. According to data processing requirements, DMO intelligently allocates data streams to corresponding processing services to ensure efficient use of edge computing resources and reduce processing delays. The intelligent data caching and preprocessing module (IDCP) uses Redis Cluster to dynamically cache hot data according to prediction models (such as ARIMA) to reduce the need for repeated access to the cloud. The received data is stream processed through Apache Kafka, and operations such as feature extraction (PCA dimensionality reduction) are performed to prepare for subsequent analysis. The adaptive multimodal anomaly detection system (AMED) analyzes both time series data (LSTM analyzes electrical parameter fluctuations) and image data (CNN identifies device status) to achieve comprehensive anomaly detection. According to the detection results and real-time feedback, the online learning mechanism in AMED automatically adjusts the model parameters to improve the accuracy and adaptability of anomaly identification. The edge cloud collaborative intelligent analysis and decision service (ECADS) shares the processing tasks, with real-time data processing and simple analysis completed at the edge, and complex model training and deep analysis tasks performed in the cloud. During the model update process, ECADS uses the federated learning mechanism to ensure that the data does not leave the edge node, only sharing the model update parameters to protect data privacy. The security and privacy protection module (SPP) ensures that all data transmission is encrypted through TLS / SSL to protect the security of data during transmission. The security and privacy protection module (SPP) uses homomorphic encryption technology to protect data content during data processing, and the differential privacy mechanism further increases the privacy protection of data output to ensure that user privacy is not leaked. Finally, the data after comprehensive analysis and processing is converted into operation and maintenance decision suggestions, which are presented to operation and maintenance personnel through decision support and visualization services to achieve fault warning, resource optimization configuration, etc. In addition, during the operation of the system, the processing results and performance data of each module are continuously collected and fed back to the decision and optimization system to form a closed loop, continuously iterate the optimization algorithm and resource allocation strategy, and improve the overall system performance.

[0098] In the above embodiment, the container orchestration engine Kubernetes ensures high availability and load balancing of services through controller mode and declarative update; the dynamic scheduling and resource management module includes a microservice management unit, a resource scheduling unit, a communication coordination unit and an intelligent optimization unit; the microservice management unit realizes the automatic deployment, expansion and management of microservices through the container orchestration engine Kubernetes; and provides intelligent routing, traffic management, policy execution and telemetry data collection between microservices through the service grid; the resource scheduling unit collects system resource usage through resource monitoring tools; and dynamically adjusts the number of service instances through a resource scheduling algorithm based on reinforcement learning; the system resource usage includes at least CPU, memory and network bandwidth usage; the communication coordination unit realizes automatic discovery and communication between microservices through the service registration and discovery mechanism of the service grid Istio, and realizes at least communication coordination strategies including A / B testing and canary deployment through the traffic segmentation and redirection function of the service grid Istio; the intelligent optimization unit predicts resource demand trends through the LSTM network, and automatically adjusts resource allocation according to real-time monitoring data and prediction results through a feedback control mechanism to optimize task response speed and resource utilization efficiency.

[0099] In a specific embodiment, the combined use of the container orchestration engine Kubernetes and the service mesh Istio provides powerful support for the microservice architecture. Specifically, Kubernetes is an open source container orchestration system for automating the deployment, expansion and management of containerized applications. It manages the container groups (Pods) on the nodes through control plane components (such as API servers, schedulers, controller managers, etc.). Istio is an open source service mesh that provides traffic management, security, observability and other functions for the microservice architecture. It configures and manages agents (such as Envoy) in the data plane through a control plane (including components such as Pilot, Mixer, Citadel, etc.). Specifically, the working method of the container orchestration engine is: first, the microservices are packaged into containers, and each container runs a microservice instance. Next, Kubernetes is used to deploy the container, and the required Pods, services (Services), deployments (Deployments), etc. are declared by defining YAML files. Third, the Istio control plane and data plane agent are deployed in the Kubernetes cluster to provide traffic management functions such as intelligent routing and fault recovery for microservices. Then, using reinforcement learning algorithms, the number of microservice instances and service resource allocation are dynamically adjusted based on real-time monitoring data to respond to changes in system load. Finally, the Istio control plane performs intelligent routing through the data plane proxy based on traffic management policies to ensure efficient communication between services.

[0100] In specific implementation, Kubernetes' automatic deployment, expansion, and management functions reduce manual intervention and improve operation and maintenance efficiency. Istio service mesh provides intelligent traffic management, optimizes communication paths between services, and reduces latency. In addition, Istio's load balancing function can intelligently distribute traffic based on the health and load of service instances. Secondly, Istio provides encryption and authorization mechanisms for communication between services, enhancing the security of the system.

[0101] In the above embodiment, the working steps of the resource scheduling algorithm based on reinforcement learning in the intelligent fusion terminal data processing system based on edge computing are as follows:

[0102] Step 1: Modeling the environment state: Modeling the resource state and task load in the edge computing environment through a multi-dimensional state representation formula. The resource scheduling algorithm based on reinforcement learning defines the state vector S t ={x1,x2,...,x i ,...,x n} represents the system state at time t, where x i Indicates the state of the i-th resource instance, the instance state includes CPU usage, memory usage, network bandwidth, number of service instances and task queue length parameters; the formula expression of the multi-dimensional state characterization formula is:

[0103] x i ={cpu it ,mem it ,bw it ,ins it ,q it} (1)

[0104]

[0105] In formula (1) and formula (2), cpu it ,mem it ,bw it Represents the usage status vectors of CPU, memory and network resources respectively; ins it is the system load; q it is the task queue length; for each resource u i is the utilization rate, c i is the resource load, t i is the resource capacity, l i is the task type;

[0106] Step 2: Resource scheduling strategy learning: Strategy T parameterized by deep neural network θ (a|S t) Select the best action; strategy T θ (a|S t ) Output in a given state S t The probability of taking action a is calculated, and the network parameter θ is optimized by the gradient ascent method. The parameterized formula of the deep neural network is:

[0107]

[0108] In formula (3), θ represents the policy network parameter, which is used to determine the probability distribution of the network output; h(S t , a) The feature vector representing the state-action pair, which is used to represent the combined features of state and action in the policy network; D iI Indicates that in state S t The action taken under t , a) is in state S t The immediate reward after taking action a;

[0109] Step 3: Action execution and feedback: At each time step t, according to the current strategy T θ Select an action t And execute, action a t Including but not limited to increasing or decreasing the number of service instances and adjusting resource allocation; after executing the action, new feedback information is provided through the system dynamic feedback formula. The feedback information includes the new state S t+1 and return r t ; The formula expression of the system dynamic feedback formula is:

[0110]

[0111] In formula (4) and formula (5), γ represents the cost sensitivity parameter, c(a t ) is action a t cost; w1, w2 and w3 represent weight coefficients, which are used to balance the importance of different system performance indicators; S max Indicates the upper limit of resource capacity, used to standardize resource utilization and queue length; Δcpu it Indicates the change in CPU usage, which is used to measure the change in resource status; N(0, σ 2 ) is a noise term, which is used to represent the randomness and uncertainty of the environment;

[0112] Step 4: Long-term return and strategy evaluation: Evaluate the long-term return of the current strategy through the long-term return evaluation formula, that is, in strategy T θ Next from state S tThe expected return sum at the beginning; the long-term return evaluation formula reflects the expectation of long-term return through the time difference learning update strategy; the formula expression of the long-term return evaluation formula is:

[0113] V = ∑log2(Q θ (S t , a) 2 -βE θ (S t )) 2 (6)

[0114] In formula (6), V is the strategy T θ long-term return; β is the discount factor, which is used to adjust the importance of future returns relative to immediate returns; Q θ To represent the state S t The expected return of taking action a, E θ (S t ) indicates that from state S t Start following strategy T θ expected return;

[0115] Step 5: Strategy iteration and optimization; Finally, based on the evaluation of long-term returns, the reinforcement learning-based resource scheduling algorithm continuously iterates and optimizes the strategy network through the strategy iteration optimization formula; the formula expression of the strategy iteration optimization formula is:

[0116]

[0117] In formula (7), T new is the updated parameter of the policy network; y is the expected target value of the return at multiple time steps, A θ (S t , a t ) is the advantage function, representing action a t In state S t is the advantage of the policy network over the average policy; μ is the learning rate of the policy network, which is used to control the step size of parameter update; δ is the exploration rate, which is used to control the conservatism of policy update.

[0118] In a specific embodiment, the resource scheduling algorithm based on reinforcement learning automatically adjusts the number of service instances through a continuous trial and error learning process to optimize resource utilization and improve task response speed. The technical principles mainly include: State space definition: The algorithm first defines a state space, which includes indicators such as the current system load (CPU, memory usage), service request volume, response time, etc. These state information reflects the current resource utilization and task processing capabilities of the system. Action space definition: The action space defines possible resource scheduling actions, such as increasing or decreasing the number of instances of a service. The algorithm selects an action to execute based on the current state to try to improve system performance. Reward mechanism design: The algorithm measures the effect of the selected action by designing a reward function. The reward function can be designed based on multiple dimensions such as resource utilization, task response time, service stability, etc., in order to encourage the algorithm to find an action combination that can improve the overall performance of the system. Reinforcement learning model: Using reinforcement learning algorithms such as Q-Learning, Deep Q-Network (DQN) or Proximal Policy Optimization (PPO), through interaction with the environment (i.e., executing actions and receiving rewards), gradually learn the optimal scheduling strategy. The algorithm balances exploration (trying new actions) and exploitation (selecting the best known actions) to learn efficiently.

[0119] The specific implementation process is as follows: First, set the initial microservice configuration and system status, including the number of service instances, resource allocation, etc. Then, periodically collect system status data, including but not limited to CPU usage, memory usage, task queue length, etc. Based on the current status, the algorithm selects an action to execute, such as adjusting the number of instances of a service. Next, adjust the number of microservice instances through Kubernetes, and Istio optimizes the communication efficiency between services. Calculate and give rewards based on changes in system performance after executing the action (such as shortened response time and improved resource utilization). Based on reward feedback, update the reinforcement learning model and optimize the decision-making strategy in the hope of performing better in subsequent decisions.

[0120] Specifically, the design experiment 1 aims to evaluate the performance of the resource scheduling algorithm based on reinforcement learning in the intelligent fusion terminal data processing system of edge computing through practical application tests; the experimental environment is:

[0121] The edge computing node configuration includes multi-core CPU, appropriate memory, high-speed storage, and necessary network communication facilities. Based on the Kubernetes container orchestration engine and Istio service grid, the deployment and management of microservices are realized, providing highly flexible and dynamic resource allocation capabilities. The deep Q network (DQN) is used as a reinforcement learning model to realize automated decision-making for resource scheduling. The experimental process is as follows:

[0122] Configure a standard test environment, deploy a set of representative microservice applications, and set the initial resource allocation state. Next, record the baseline data of the system operation, including CPU utilization, memory usage, task response time, etc. Then, connect the resource scheduling algorithm based on reinforcement learning to the system and start interacting with the environment. Next, the algorithm dynamically adjusts the number of microservice instances according to the current system status, and records the changes in system status after each adjustment. Finally, through multiple rounds of scheduling adjustments, collect data to evaluate key performance indicators such as resource utilization efficiency and task response time. Experimental indicators include: resource utilization efficiency: CPU, memory utilization; task response time: average response delay of service requests; stability: system failure rate, service interruption time; algorithm convergence speed: the time required for the algorithm to reach a better strategy. The experimental test data is shown in Table 1:

[0123] Table 1 Experimental data record table

[0124]

[0125] Through the dynamic adjustment of the algorithm, it can be seen that the CPU usage rate tends to a more reasonable range after adjustment, indicating that the algorithm can effectively avoid over-allocation or under-allocation of resources. With the iteration of the algorithm, the average response time shows a downward trend, indicating that the algorithm has successfully improved the service response speed and optimized the user experience. The increase in cumulative rewards reflects that the algorithm continuously optimizes decisions during the learning process and gradually finds a better resource allocation strategy. The fixed decision cycle ensures the real-time nature of the algorithm and shows that the algorithm can effectively respond and adjust strategies within a limited time.

[0126] In addition, Experiment 2 is designed to evaluate the performance of the algorithm, and the indicators examined include resource utilization efficiency, task response time, system stability, etc. The experimental environment simulates an edge computing environment that contains multiple microservices and applications. The workload is to generate various types of workloads, including data-intensive, compute-intensive, and hybrid tasks. The algorithm includes the implementation of a resource scheduling algorithm based on reinforcement learning (RL) and a traditional polling scheduling algorithm (Round Robin, RR) for comparison. The test indicators include: average response time (milliseconds), resource utilization (average value of CPU and memory utilization), system throughput (number of tasks processed per second), and scheduling delay (the time it takes for the algorithm to make a scheduling decision, in milliseconds); the experimental data records are shown in Table 2:

[0127] Table 2 Performance comparison of algorithms

[0128]

[0129] As can be seen from the data table 2, the resource scheduling algorithm based on reinforcement learning (RL) is superior to the traditional polling scheduling algorithm (RR) in terms of average response time, resource utilization, and system throughput. Although the scheduling delay of the RL algorithm is slightly higher than that of RR, this increase is acceptable considering the improvement in overall performance. Especially when processing compute-intensive and mixed workloads, the RL algorithm shows more significant advantages. These results show that the resource scheduling algorithm based on reinforcement learning can more effectively manage resources in the edge computing environment and improve the task processing speed while maintaining high resource utilization and system stability. Therefore, for edge computing application scenarios that require fast response and high resource efficiency, the use of reinforcement learning resource scheduling algorithm is a more appropriate choice.

[0130] In the above embodiment, the distributed cache system includes a data cache module, a cache management module and a prediction model integration module; the data cache module includes a data sharding unit and a data serialization unit; the data sharding unit evenly distributes data requests to multiple nodes of the Redis cluster through hash allocation and consistent hashing algorithm; the data serialization unit compresses and serializes cached data through the serialization protocol PB; the cache management module ensures efficient use of cache space through a cache invalidation strategy, and ensures consistency and synchronous update of data in a distributed system through a publish / subscribe mode and a transaction mechanism; the cache invalidation strategy automatically eliminates infrequently used data through a least recently used mechanism; the cache management module includes a data security unit, a performance monitoring unit and a cache optimization unit; the data security unit encrypts the data in the cache through an elliptic encryption algorithm, and realizes data persistence and disaster recovery through regular snapshots and WAL pre-write logs; the performance monitoring unit monitors the performance indicators of cache operations in real time through a real-time monitoring tool; based on the monitoring data, the cache optimization unit dynamically adjusts the Redis configuration parameters through a dynamic cache strategy; the prediction model integration module provides intelligent support for the cache strategy through time series prediction and Akaike information criterion.

[0131] In a specific embodiment, the distributed cache system stores and manages data by working together with multiple nodes. It uses a consistent hashing algorithm or a distributed hash table (DHT) to distribute data to ensure high availability and scalability of the data. The specific working method of the distributed cache system is as follows: First, the system receives real-time data from smart meters, sensors and other devices. The data cache module quickly stores the received data in the memory database. The cache management module manages the life cycle and storage space of the data according to the preset strategy. Check the cache status regularly, clean up expired data, and ensure cache efficiency. Then, use historical data to train the prediction model, and deploy the model to the prediction model integration module. Then, the prediction model integration module performs real-time analysis and prediction of the data in the cache. Then, the prediction results are used for power grid status monitoring, load balancing, preventive maintenance, etc. According to the deviation between the prediction results and the actual data, adjust the prediction model parameters to optimize the prediction accuracy.

[0132] In the above embodiment, the dynamic cache strategy first identifies the optimal configuration template under different working modes through cluster analysis, based on the cluster analysis results; the dynamic cache strategy dynamically adjusts the configuration parameters of the Redis cluster through a dynamic configuration adjustment mechanism, including but not limited to adjusting the maximum memory limit, expiration key strategy, number of replicas and sharding strategy; the dynamic configuration adjustment mechanism ensures the smoothness of the adjustment operation without affecting service availability through the configuration update protocol RC; when the system recognizes that the load of a specific shard is too high, the dynamic cache strategy automatically triggers the data migration process through dynamic sharding and rebalancing methods, balances the data distribution between nodes, and ensures the efficiency of data migration and the consistency of data access through a consistent hashing algorithm.

[0133] In a specific embodiment, the dynamic caching strategy continuously monitors the data processing requirements and performance indicators of edge devices, such as CPU usage, memory usage, network latency, etc. By using time series prediction and Akaike Information Criterion, future data processing requirements and resource usage are predicted. Based on the monitoring data and prediction results, an algorithm is used to determine how to adjust the configuration parameters of Redis, such as increasing the cache size, adjusting the expiration time, etc. Then, the new configuration parameters are dynamically applied to the Redis cache system to optimize its performance. The performance changes after adjustment are monitored and these data are fed back into the prediction model to continuously optimize the caching strategy.

[0134] In specific implementation, by dynamically adjusting the cache configuration, resource usage can be optimized according to actual needs to avoid resource waste or shortage. In addition, the optimized cache strategy can reduce the delay of data access and processing and improve the response speed of the system. Secondly, the dynamic cache strategy enables the system to better adapt to the changing workload and data processing requirements. At the same time, by intelligently adjusting the cache parameters, the system can more effectively respond to emergencies or peak loads and maintain stable operation.

[0135] In the above embodiment, the time series analysis and prediction model includes an input layer, a data layer, a feature construction layer, a model training layer, a prediction generation layer, a prediction integration layer and an intelligent optimization layer; the working method of the time series analysis and prediction model in the intelligent fusion terminal data processing system based on edge computing is:

[0136] Step 1: Acquire the collected data of the edge collection module through the input layer; the input layer ensures the time consistency of data collection through the time protocol and timestamp synchronization mechanism;

[0137] Step 2: Clean and standardize the collected data through the data layer; the data layer performs signal denoising through adaptive Kalman filtering and processes missing data through spline interpolation;

[0138] Step 3, extracting time series features from the preprocessed data through the feature construction layer; the feature construction layer obtains time series features through wavelet transform and rolling statistics analysis, and removes redundant features through feature selection based on p value;

[0139] Step 4: Build and train a time series prediction model through the model training layer; the model training layer implements model construction and training through a machine learning framework, and trains model parameters through a batch gradient descent method

[0140] Step 5: Generate a prediction result through the prediction generation layer; the prediction generation layer calculates the prediction value through the forward propagation of the model, and provides a quantification of the uncertainty of the prediction through the confidence interval estimation;

[0141] Step 6: Fusion of multiple model prediction results through the prediction integration layer; the prediction integration layer improves prediction accuracy through integrated learning methods, weight adjustment and voting mechanisms;

[0142] Step 7: According to the deviation between the prediction result and the actual observation value, the model parameters and the prediction strategy are intelligently adjusted through the intelligent optimization layer; the intelligent optimization layer adjusts the model through an incremental learning method;

[0143] Step 8: Feedback the prediction performance to the model training layer through the intelligent optimization layer; the intelligent optimization layer continuously adjusts the model structure and parameters through a closed-loop feedback mechanism to adapt to changes in data distribution.

[0144] In a specific embodiment, the working method of the time series analysis prediction model is as follows: first, real-time data such as grid load, temperature, humidity, etc. are collected from terminal devices and sensors. Next, the data is preliminarily cleaned through the stream processing framework, including normalization, missing value filling, and feature extraction. Third, the prediction model is trained using historical time series data. Common algorithms include ARIMA, LSTM networks, etc. Fourth, the trained model is deployed on the edge device to perform predictive analysis on the real-time data stream. Then, the data caching strategy is dynamically adjusted according to the prediction results, and key data that may affect future decisions is cached first. Finally, the predicted data is further structured for use by upper-level applications and decision support systems.

[0145] In specific implementation, by predicting future data trends, data can be cached and preprocessed more intelligently, reducing unnecessary computing and storage overhead. In addition, accurate time series prediction provides a scientific basis for real-time decision-making, improving the accuracy and timeliness of decisions. Secondly, by performing data preprocessing and prediction on edge devices, the need to transmit raw data to the central server is reduced, thereby reducing communication costs. At the same time, the real-time prediction capability of edge devices enables the system to respond quickly to external changes, improving the response speed and reliability of the entire system. Finally, the prediction model can help the system allocate computing and storage resources more reasonably, ensuring that critical tasks are supported by sufficient resources.

[0146] In the above embodiment, the spatiotemporal heterogeneous fusion network model includes an input layer, a spatiotemporal convolution layer, a feature fusion layer, a bidirectional long short-term memory layer, an attention mechanism layer, a heterogeneous feature enhancement layer and an output layer; the working method steps of the spatiotemporal heterogeneous fusion network model in the intelligent fusion terminal data processing system based on edge computing are as follows:

[0147] S1, receiving raw data through the input layer, and the input layer eliminates dimensional differences between data by data resampling and standardization;

[0148] S2, extracting features through the spatiotemporal convolution layer; the spatiotemporal convolution layer extracts the time features of the time series and the spatial features of the image through a one-dimensional convolution layer and a two-dimensional convolution layer respectively; the one-dimensional convolution layer applies a one-dimensional convolution kernel to slide on the time axis, extracts features through convolution operation, and uses the filter size and step size parameters to control the granularity and coverage of feature extraction; the two-dimensional convolution layer slides on the image plane through a two-dimensional convolution kernel, and identifies different types of visual features through different filter groups, thereby enhancing the spatial understanding ability of the model;

[0149] S3, fusing the features extracted from the time series and image data through the feature fusion layer to form a unified spatiotemporal feature representation; the feature fusion layer ensures the effective integration of cross-modal information through feature cascading and feature mapping methods;

[0150] S4. Performing bidirectional analysis on time series data through the bidirectional long short-term memory layer to capture the forward and backward dependencies of the sequence and improve the recognition of abnormal event sequence patterns; the bidirectional long short-term memory layer captures the long-term dependencies in the sequence by combining forward and reverse LSTM units, and controls the information flow through a gating mechanism to enhance the ability to capture sequence dynamics;

[0151] S5. Dynamically adjust the importance of different features through the attention mechanism layer; the attention mechanism layer assigns weights to different features through a weighted sum strategy, and the weighted sum strategy improves the model's attention to abnormal patterns through dot product attention;

[0152] S6. Further refine and enhance the spatiotemporal features through the heterogeneous feature enhancement layer; the heterogeneous feature enhancement layer enhances the richness and robustness of feature representation through depthwise separable convolution, residual connection and gated activation function; S7. Based on the features of the heterogeneous feature enhancement layer, generate anomaly probability and classification labels through the output layer;

[0153] The spatiotemporal convolution layer extracts information-rich spatiotemporal feature maps from the original data through a spatiotemporal feature extraction formula; the formula expression of the spatiotemporal feature extraction formula is:

[0154]

[0155] In formula (8), represents the spatiotemporal feature map of the output of layer I; m is a nonlinear activation function used to increase the expressive power of the model; τ is the slope of the linear regression statistical operation; k is the number of filters; is the bias term of layer I, used to adjust the convolution output; X I-1 Represents the output of the previous layer, which is used as the input feature of the current layer; Indicates that the Ith layer passes through k different spatiotemporal convolution kernels X on the previous layer I-1 The input is convolved to extract features of different scales and directions; * represents the convolution operator;

[0156] The bidirectional long short-term memory layer captures the front-end dependency of the time series data through a bidirectional recurrent network with a gating mechanism; the state update formula of the bidirectional long short-term memory layer is as follows:

[0157]

[0158] In formula (9), W is the hidden state of the forward and reverse LSTM units; l is the sigmoid activation function, which is used to generate a value between 0 and 1 to control the degree of information passing; x is the bias term of the forward and reverse LSTM, which is used for bias adjustment; v is the weight matrix from the hidden state to the hidden state of the forward and reverse LSTM, which is used to transfer historical information; is the residual of linear regression statistics; c t is a comprehensive cell state that combines the forward and reverse hidden states and memory cell contents;

[0159] The attention mechanism layer and the heterogeneous feature enhancement layer selectively enhance the spatiotemporal features and the features of the bidirectional LSTM output through the heterogeneous feature attention allocation formula to ensure that the model focuses on the most relevant features; the formula expression of the heterogeneous feature attention allocation formula is:

[0160]

[0161] In formula (10), R is the final enhanced feature representation; e i is the attention score, which is used to measure the importance of the i-th feature; b a is the bias term of the attention layer, which is used to adjust the initial score of the feature combination; a i is the normalized attention weight, n is the number of feature vectors; W a is the weight matrix of the attention layer, which is used to learn the importance of different feature combinations; LeakyReLU represents the activation function, which is used to increase nonlinearity.

[0162] In a specific embodiment, the spatiotemporal convolutional network combines the spatial convolution and temporal convolution characteristics of the traditional convolutional neural network, and can simultaneously capture the spatial features (such as the pixel layout in the image) and temporal features (such as the temporal changes of the time series data) in the data. By applying different filters on the time axis and the spatial axis, the STCNN can extract the temporal dependency of the time series data and the spatial context information of the image data. The bidirectional long short-term memory network (BiLSTM) can capture not only the forward dependency of the data sequence (future data depends on past data) but also the backward dependency (past data may be affected by future data) by processing the forward sequence and backward sequence of the data at the same time, which is particularly effective for identifying anomalies in complex time series. The bidirectional characteristics of BiLSTM increase the comprehensiveness and accuracy of the model's recognition of abnormal patterns in sequence data. When processing multimodal data, data of different modalities often contribute to anomaly detection to different degrees. The heterogeneous feature attention mechanism automatically assigns different weights to the model by learning the importance of different features, emphasizing the features that are more valuable for anomaly detection and ignoring or weakening insignificant features. This mechanism significantly enhances the model's ability to recognize abnormal patterns in multimodal data.

[0163] Specifically, Experiment 3 is designed to evaluate the anomaly detection capability of the spatiotemporal heterogeneous fusion network model when processing time series and image data. The data set includes: Time series data: Select the power load data of a certain substation within one year, including normal and abnormal load patterns. Image data: Thermal images of key equipment in the substation, including normal working conditions and abnormal heating conditions. The model structure is as follows:

[0164] Spatiotemporal convolutional network (STCNN) layer: extracts temporal dependencies in time series data and spatial features of image data. Bidirectional long short-term memory network (BiLSTM) layer: captures bidirectional temporal dependencies in time series data. Heterogeneous feature attention mechanism layer: dynamically adjusts the importance of different modal features and focuses on abnormal patterns. Fusion and classification layer: combines spatiotemporal features for anomaly detection and classification.

[0165] The experimental process is as follows: First, clean the data, standardize it, and split it into training set / test set. Then, use Python programming language, TensorFlow or PyTorch framework to implement the model. Evaluate the model performance through indicators such as accuracy, recall rate, and F1 score. Compare the performance with the isolation forest and unimodal network models. The experimental test data is shown in Table 3:

[0166] Table 3 Model experiment data record table

[0167]

[0168]

[0169] As can be seen from the data table 3, the spatiotemporal heterogeneous fusion network model performs well in intelligent analysis and anomaly detection tasks by integrating STCNN, BiLSTM and heterogeneous feature attention mechanism. Especially when processing multimodal data, it can more accurately identify abnormal situations, which is of great significance for improving the monitoring and early warning capabilities of the intelligent fusion terminal in the substation area. The experimental results verify the positive role of the model in improving the detection accuracy, recall rate and comprehensive performance index F1 score, providing strong technical support for anomaly detection applications in power systems and other fields.

[0170] In addition, Experiment 4 is designed to compare the performance of the spatiotemporal heterogeneous fusion network model (STHFN) and the traditional support vector machine (SVM) in the anomaly detection task of time series and image data. The experimental environment includes: processor: Intel Corei7-9700K@3.60GHz, memory: 32GB DDR4 RAM, storage: 1TB SSD; operating system: Ubuntu 20.04LTS, programming language: Python 3.8, dependent libraries: TensorFlow2.4, Keras 2.4, Scikit-learn 0.23; the experimental data sets are data sets 1 to 10, each data set contains time series and image data, and the data volume ranges from 10,000 to 50,000. The specific experimental methods are:

[0171] Data preprocessing: Normalization: Normalize all data to the interval [0,1]. Data partitioning: Divide the dataset into training set (70%), validation set (15%) and test set (15%).

[0172] Model construction: STHFN model: Use spatiotemporal convolutional network (ST-ConvNet) to extract temporal and spatial features. Adopt bidirectional long short-term memory network (BiLSTM) to capture the temporal dependency of abnormal data. Introduce heterogeneous feature attention mechanism to enhance the recognition ability of abnormal patterns. SVM model: Use traditional support vector machine for anomaly detection, without deep learning method.

[0173] Training and evaluation: Training: Use the training set to train the STHFN and SVM models. Validation: Use the validation set to adjust the model parameters. Testing: Use the test set to evaluate the model performance and record the accuracy, recall, F1 score, and running time. The experimental data records are shown in Table 4:

[0174] Table 4 Model comparison experimental data table

[0175]

[0176] From the experimental data, it can be seen that the spatiotemporal heterogeneous fusion network model (STHFN) outperforms the traditional support vector machine (SVM) in terms of anomaly detection accuracy, recall rate and F1 score. Although the running time of STHFN is slightly higher than that of SVM, this increase is acceptable considering its significant improvement in anomaly detection performance. Especially when dealing with complex data sets, STHFN shows more significant advantages.

[0177] In the above embodiment, the edge-cloud-edge collaboration module includes a data synchronization and distribution unit, a model update and optimization unit, and a federated learning management unit; the data synchronization and distribution unit performs data transmission and intelligent distribution through a two-way data transmission mechanism and a data hierarchical processing strategy; the two-way data transmission mechanism ensures real-time synchronization of data between the edge side and the cloud through the MQTT protocol and the gRPC service; the data hierarchical processing strategy adopts an event-driven data distribution mechanism to automatically distribute data processing tasks between the edge side and the cloud; the model update and optimization unit constructs a model training environment in the cloud through a distributed computing framework; the distributed computing framework uses a stochastic gradient descent S GD optimizes model parameters; the model update and optimization unit realizes real-time model version management and seamless deployment and update on the edge side through containerization and container orchestration system; the container orchestration system gradually promotes new models through grayscale release strategy; the federated learning management unit realizes efficient aggregation of model parameters through federated learning protocol and federated aggregation algorithm; the federated learning protocol enables multiple edge nodes to collaborate in training global models without leaking local data based on secure multi-party computing (SMPC) and differential privacy methods; the federated aggregation algorithm aggregates model updates submitted by edge nodes in the cloud through the federated averaging algorithm (FA), ensuring that the model reaches the global optimum after multiple iterations.

[0178] In a specific embodiment, the edge cloud-edge collaboration module uses message queues (such as RabbitMQ, Kafka) and APIGateway technology to achieve real-time synchronization of data between edge nodes and the cloud. Through intelligent scheduling algorithms (such as priority-based task queues or load-based dynamic allocation), the cloud-edge collaboration module intelligently distributes complex analysis tasks to the cloud or appropriate edge nodes for processing according to the resource status and task requirements of the edge computing nodes, and recovers the completed task results. At the same time, a hierarchical model processing strategy is adopted to leave simple analysis tasks with high real-time requirements to edge nodes for processing, while complex model training tasks with large computing resource requirements and rich training data sets are deployed in the cloud. This hierarchical processing ensures the effective use of resources and rapid response of tasks. The cloud uses large-scale computing resources to train and optimize the model, and accelerates the learning process of the model through distributed training frameworks (such as TensorFlow Distributed, PyTorch Distributed). Federated learning is a distributed machine learning paradigm that allows edge nodes to collaboratively train global models without directly sharing local data. Each edge node trains a local copy of the model on the data it holds, and only uploads updates to the model parameters (gradients or model weights) to the cloud server, which integrates these updates through aggregation algorithms (such as FedAvg) to form a global model, and then distributes the updated model back to the edge nodes. This process protects data privacy while achieving continuous optimization of model performance.

[0179] Specifically, the edge-cloud-edge collaboration module first collects data from terminal devices through edge nodes, uploads necessary data to the cloud after preliminary processing, and synchronizes the latest model parameters on the cloud. Next, the cloud-edge collaboration module decides whether to perform the task on the cloud or edge node based on the task type and resource conditions. Complex analysis tasks are assigned to the cloud or edge nodes with sufficient resources, while real-time processing needs are directly responded to by edge nodes. The cloud trains the model based on the collected data, and uses the federated learning mechanism to regularly collect model parameter updates from edge nodes, summarize and update the global model, and then distribute it to edge nodes to update their local models. The optimized model is deployed to the edge node through the cloud-edge collaboration module to ensure that the edge always runs the latest model version, improving processing efficiency and accuracy.

[0180] In specific implementation, edge computing reduces the time it takes for data to travel back and forth to the cloud, allowing urgent and real-time tasks to be responded to quickly locally, improving user experience. In addition, by intelligently distributing tasks, cloud-edge collaboration maximizes resource utilization, avoids unnecessary data transmission, and reduces network bandwidth requirements and cloud service costs. Secondly, the federated learning mechanism ensures that user data does not leave the local device, complies with data privacy protection regulations, and enhances user trust. At the same time, the continuous training and optimization of the model, through the federated learning mechanism, utilizes the rich data resources at the edge, making the model more accurate and more adaptable.

[0181] In the above embodiment, the working principle steps of the security and privacy protection module in the area intelligent fusion terminal data processing system based on edge computing are as follows:

[0182] Step 1: Secure data reception and verification: Establish a secure channel at the beginning of data transmission through the TLS / SSL protocol; the TLS / SSL protocol verifies the identities of both parties in communication through the public key infrastructure and ensures the secure exchange of session keys through the handshake protocol;

[0183] Step 2: Data storage encryption: Implement security protection of stored data through a homomorphic encryption method, where the homomorphic encryption method allows direct calculation on the ciphertext through a polynomial ring;

[0184] Step 3: Preprocessing of private data: Anonymizing the private data through differential privacy; the differential privacy balances the privacy protection and the accuracy of data analysis by controlling the variance of noise through the Laplace equation;

[0185] Step 4: Access control and permission management: Implement strict permission control on data access through role-based access control and attribute-based access control strategies;

[0186] Step 5: Dynamic security policy deployment: Based on the current network status, dynamic security policy adjustment is performed through zero-trust network policy; the zero-trust network policy achieves fine control of network traffic through software-defined network SDN;

[0187] Step 6: Security audit and intrusion detection: Continuously monitor the system security status and identify potential threats through real-time security monitoring system and intrusion detection system; the real-time security monitoring system identifies and responds to abnormal behaviors through machine learning models; the intrusion detection system tracks and records all key operations through log analysis and access pattern recognition methods;

[0188] Step 7: Data leakage prevention: Monitor, identify and protect sensitive data through data loss prevention strategy DLP; the data loss prevention strategy DLP prevents unauthorized data transmission and leakage through access control and data encryption methods;

[0189] Step 8: Security incident response and recovery: Ensure rapid response to security incidents and system recovery through security incident response mechanisms and disaster recovery backup methods.

[0190] In a specific embodiment, the security and privacy protection module uses the TLS / SSL protocol to encrypt data transmission to ensure that the data is not eavesdropped or tampered with when it is transmitted in the network. TLS / SSL establishes a secure connection through a handshake protocol, uses public key encryption to exchange keys, and then uses symmetric key encryption for data transmission, providing end-to-end security protection. Homomorphic encryption technology allows direct operations on encrypted data, such as addition, multiplication, etc., without decryption first. This means that data can be processed in an encrypted state, and the true content of the data remains confidential even in an untrusted environment, greatly enhancing privacy protection in the data processing stage. Differential privacy protects personal privacy by adding random noise to the data query results to ensure that the addition or removal of any individual data does not significantly change the query results. This method makes it difficult for attackers to infer information about specific individuals by analyzing query results, and is suitable for statistical analysis and data publishing scenarios. The zero-trust security model assumes that there are threats both inside and outside the network, and requires strict authentication and authorization for all access requests, regardless of their source. Through fine-grained access control, continuous monitoring, and the principle of least privilege, the zero-trust strategy ensures that only verified and authorized entities can access system resources, enhancing overall security.

[0191] In the intelligent fusion terminal data processing system based on edge computing, the TLS / SSL protocol is used to encrypt the data before the data leaves the terminal or is transmitted between the cloud and the edge, ensuring that the data cannot be interpreted even if it is intercepted during transmission. Next, homomorphic encryption technology is applied to encrypt sensitive data stored in edge nodes or the cloud, and then computing tasks such as data aggregation or model training are performed directly on the encrypted data without decryption. Before data analysis, a differential privacy algorithm is used to add random noise to the output data to ensure that the analysis results do not leak personal information while maintaining the effectiveness of the analysis. At the same time, a zero-trust network strategy is implemented to continuously authenticate, authorize, and monitor all network access requests to ensure that only verified users and devices can access system resources and any abnormal behavior can be identified and responded to in a timely manner.

[0192] In the specific implementation, encryption technology and protocols are used to effectively prevent data from being illegally accessed or tampered with during transmission and processing, thus protecting the integrity of the data. In addition, the application of differential privacy technology ensures that the release of data analysis results will not leak personal privacy information and maintain user trust. Secondly, the implementation of the zero-trust strategy strengthens the overall defense mechanism of the system and reduces the risk of internal threats and external attacks.

[0193] Although the specific embodiments of the present invention are described above, it should be understood by those skilled in the art that these specific embodiments are only illustrative, and those skilled in the art may omit, replace, and change the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, merging the above method steps so as to perform substantially the same functions in substantially the same manner to achieve substantially the same results is within the scope of the present invention. Therefore, the scope of the present invention is limited only by the appended claims.

Claims

1. A station area intelligent fusion terminal data processing system based on edge computing, characterized by: include: Edge acquisition module, used to collect and pre-process data from IoT devices; The edge acquisition module collects real-time data from the terminal devices in the substation area through the Internet of Things communication network NB-IoT and the data transmission protocol MQTT; and preliminarily filters the data through the sliding window averaging method; the terminal devices in the substation area include but are not limited to smart meters, environmental sensors and status monitors; Dynamic scheduling and resource management module, used to optimize microservice performance and collaboration; the dynamic scheduling and resource management module realizes automatic management and efficient communication of microservices through the container orchestration engine Kubernetes and the service grid Istio, and dynamically adjusts the number of service instances according to the system load through the resource scheduling algorithm based on reinforcement learning, thereby optimizing resource utilization efficiency and task response speed; Edge caching and preprocessing module, used for efficient caching and stream processing of data; the edge caching and preprocessing module dynamically caches data through a distributed caching system and a time series analysis and prediction model, and preprocesses data through a stream processing framework; the stream processing framework performs deep cleaning and structural processing of data through normalization, missing value filling and feature extraction; Intelligent analysis and anomaly detection module, used to identify time series and image anomalies; the intelligent analysis and anomaly detection module realizes multi-dimensional anomaly detection and pattern recognition of data through a spatiotemporal heterogeneous fusion network model; the spatiotemporal heterogeneous fusion network model captures the correlation between the time series and spatial dimensions of data through a spatiotemporal convolutional network, and captures the forward and backward dependencies of abnormal data through a bidirectional long short-term memory network; the spatiotemporal heterogeneous fusion network model enhances the recognition ability of abnormal patterns in multimodal data through a heterogeneous feature attention mechanism; The edge-cloud-edge collaboration module is used to coordinate data synchronization between the edge and the cloud, model updates, and the distribution and recovery of complex analysis tasks. The edge-cloud-edge collaboration module implements hierarchical data processing and model training on the cloud through a cloud-edge collaboration architecture, and allows edge nodes to participate in model training through a federated learning mechanism, thereby optimizing model performance while protecting privacy. A security and privacy protection module is used to ensure the secure flow of data. The security and privacy protection module ensures the security of data transmission and the privacy of data in processing through the security protocol TLS / SSL and homomorphic encryption, and improves the system security level through differential privacy and zero-trust network strategies. Among them, the output end of the edge acquisition module is connected to the input end of the edge cache and preprocessing module and the edge cloud edge collaboration module; the output end of the edge cache and preprocessing module and the edge cloud edge collaboration module are connected to the input end of the intelligent analysis and anomaly detection module; the security and privacy protection module and the dynamic scheduling and resource management module work throughout the process.

2. According to the edge computing-based intelligent fusion terminal data processing system of claim 1, it is characterized by: The container orchestration engine Kubernetes ensures high availability and load balancing of services through controller mode and declarative updates; the dynamic scheduling and resource management module includes a microservice management unit, a resource scheduling unit, a communication coordination unit and an intelligent optimization unit; The microservice management unit implements the automated deployment, expansion and management of microservices through the container orchestration engine Kubernetes; and provides intelligent routing, traffic management, policy execution and telemetry data collection between microservices through the service grid; the resource scheduling unit collects system resource usage through resource monitoring tools; and dynamically adjusts the number of service instances through a resource scheduling algorithm based on reinforcement learning; the system resource usage includes at least CPU, memory and network bandwidth usage; the communication coordination unit implements automatic discovery and communication between microservices through the service registration and discovery mechanism of the service grid Istio, and implements communication coordination strategies including at least A / B testing and canary deployment through the traffic segmentation and redirection functions of the service grid Istio; The intelligent optimization unit predicts resource demand trends through the LSTM network, and automatically adjusts resource allocation according to real-time monitoring data and prediction results through a feedback control mechanism to optimize task response speed and resource utilization efficiency.

3. According to the edge computing-based intelligent fusion terminal data processing system of claim 1, it is characterized by: The working steps of the resource scheduling algorithm based on reinforcement learning in the intelligent fusion terminal data processing system based on edge computing are as follows: Step 1: Environmental state modeling; The resource state and task load in the edge computing environment are modeled by a multi-dimensional state representation formula. The resource scheduling algorithm based on reinforcement learning defines the state vector S t ={x1,x2,...,x i ,...,x n } represents the system state at time t, where x i Indicates the state of the i-th resource instance, the instance state includes CPU usage, memory usage, network bandwidth, number of service instances and task queue length parameters; the formula expression of the multi-dimensional state characterization formula is: x i ={cpu it ,mem it ,bw it ,ins it ,q it } (1) In formula (1) and formula (2), cpu it ,mem it ,bw it Represents the usage status vectors of CPU, memory and network resources respectively; ins it is the system load; q it is the task queue length; for each resource u i is the utilization rate, c i is the resource load, t i is the resource capacity, l i is the task type; Step 2: Learning resource scheduling strategies; The policy T parameterized by a deep neural network θ (a|S t ) Select the best action; Strategy T θ (a|S t ) Output in a given state S t The probability of taking action a is calculated, and the network parameter θ is optimized by the gradient ascent method. The parameterized formula of the deep neural network is: In formula (3), θ represents the policy network parameter, which is used to determine the probability distribution of the network output; h(S t , a) The feature vector representing the state-action pair is used to represent the combined features of state and action in the policy network; D iI Indicates that in state S t The action taken under t , a) is in state S t The immediate reward after taking action a; Step 3: Action execution and feedback; At each time step t, according to the current strategy T θ Select an action t And execute, action a t Including but not limited to increasing or decreasing the number of service instances and adjusting resource allocation; after executing the action, new feedback information is provided through the system dynamic feedback formula. The feedback information includes the new state S t+1 and return r t ; The formula expression of the system dynamic feedback formula is: In formula (4) and formula (5), γ represents the cost sensitivity parameter, c(a t ) is action a t cost; w1, w2 and w3 represent weight coefficients, which are used to balance the importance of different system performance indicators; S max Indicates the upper limit of resource capacity, used to standardize resource utilization and queue length; Δcpu it Indicates the change in CPU usage, which is used to measure the change in resource status; N(0, σ 2 ) is a noise term, which is used to represent the randomness and uncertainty of the environment; Step 4: Long-term returns and strategy evaluation; The long-term return evaluation formula is used to evaluate the long-term return of the current strategy, that is, in the strategy T θ Next from state S t The expected return sum at the beginning; the long-term return evaluation formula reflects the expectation of long-term return through the time difference learning update strategy; the formula expression of the long-term return evaluation formula is: V=∑log2(Q θ (S t ,a) 2 -βE θ (S t )) 2 (6) In formula (6), V is the strategy T θ long-term return; β is the discount factor, which is used to adjust the importance of future returns relative to immediate returns; Q θ To represent the state S t The expected return of taking action a, E θ (S t ) indicates that from state S t Start following strategy T θ expected return; Step 5: Strategy iteration and optimization; Finally, based on the evaluation of long-term returns, the reinforcement learning-based resource scheduling algorithm continuously iterates and optimizes the policy network through the policy iteration optimization formula; the formula expression of the policy iteration optimization formula is: In formula (7), T new is the updated parameter of the policy network; y is the expected target value of the return at multiple time steps, A θ (S t , a t ) is the advantage function, representing action a t In state S t is the advantage of the policy network over the average policy; μ is the learning rate of the policy network, which is used to control the step size of parameter update; δ is the exploration rate, which is used to control the conservatism of policy update.

4. According to the edge computing-based intelligent fusion terminal data processing system of claim 1, it is characterized by: The distributed cache system includes a data cache module, a cache management module and a prediction model integration module; the data cache module includes a data sharding unit and a data serialization unit; the data sharding unit evenly distributes data requests to multiple nodes of the Redis cluster through hash distribution and consistent hashing algorithm; The data serialization unit compresses and serializes the cache data through the serialization protocol PB; The cache management module ensures efficient use of cache space through cache invalidation strategy, and ensures consistency and synchronous update of data in the distributed system through publish / subscribe mode and transaction mechanism; the cache invalidation strategy automatically eliminates infrequently used data through the least recently used mechanism; the cache management module includes a data security unit, a performance monitoring unit and a cache optimization unit; The data security unit encrypts the data in the cache through an elliptic encryption algorithm, and realizes data persistence and disaster recovery through periodic snapshots and WAL write-ahead logs; the performance monitoring unit monitors the performance indicators of cache operations in real time through real-time monitoring tools; based on the monitoring data, the cache optimization unit dynamically adjusts the Redis configuration parameters through dynamic cache strategies; the prediction model integration module provides intelligent support for cache strategies through time series prediction and Akaike information criterion.

5. According to the edge computing-based intelligent fusion terminal data processing system of claim 4, it is characterized by: The dynamic cache strategy first identifies the optimal configuration template under different working modes through cluster analysis, based on the cluster analysis results; the dynamic cache strategy dynamically adjusts the configuration parameters of the Redis cluster through a dynamic configuration adjustment mechanism, including but not limited to adjusting the maximum memory limit, expiration key strategy, number of replicas and sharding strategy; the dynamic configuration adjustment mechanism ensures the smoothness of the adjustment operation without affecting service availability through the configuration update protocol RC; when the system identifies that the load of a specific shard is too high, the dynamic cache strategy automatically triggers the data migration process through dynamic sharding and rebalancing methods, balances the data distribution between nodes, and ensures the efficiency of data migration and the consistency of data access through the consistent hashing algorithm.

6. According to the edge computing-based intelligent fusion terminal data processing system of claim 1, it is characterized by: The time series analysis and prediction model includes an input layer, a data layer, a feature construction layer, a model training layer, a prediction generation layer, a prediction integration layer and an intelligent optimization layer; the working method of the time series analysis and prediction model in the station area intelligent fusion terminal data processing system based on edge computing is: Step 1: Acquire the collected data of the edge collection module through the input layer; the input layer ensures the time consistency of data collection through the time protocol and timestamp synchronization mechanism; Step 2: Clean and standardize the collected data through the data layer; the data layer performs signal denoising through adaptive Kalman filtering and processes missing data through spline interpolation; Step 3: extracting time series features from the preprocessed data through the feature construction layer; The feature construction layer obtains time series features through wavelet transform and rolling statistics analysis, and removes redundant features through feature selection based on p-value; Step 4: Build and train a time series prediction model through the model training layer; the model training layer implements model construction and training through a machine learning framework, and trains model parameters through a batch gradient descent method Step 5: Generate a prediction result through the prediction generation layer; the prediction generation layer calculates the prediction value through the forward propagation of the model, and provides a quantification of the uncertainty of the prediction through the confidence interval estimation; Step 6: Fusion of multiple model prediction results through the prediction integration layer; the prediction integration layer improves prediction accuracy through integrated learning methods, weight adjustment and voting mechanisms; Step 7: According to the deviation between the prediction result and the actual observation value, the model parameters and the prediction strategy are intelligently adjusted through the intelligent optimization layer; the intelligent optimization layer adjusts the model through an incremental learning method; Step 8: Feedback the prediction performance to the model training layer through the intelligent optimization layer; the intelligent optimization layer continuously adjusts the model structure and parameters through a closed-loop feedback mechanism to adapt to changes in data distribution.

7. The area intelligent fusion terminal data processing system based on edge computing according to claim 1 is characterized in that: The spatiotemporal heterogeneous fusion network model includes an input layer, a spatiotemporal convolution layer, a feature fusion layer, a bidirectional long short-term memory layer, an attention mechanism layer, a heterogeneous feature enhancement layer and an output layer; the working method steps of the spatiotemporal heterogeneous fusion network model in the intelligent fusion terminal data processing system based on edge computing are as follows: S1, receiving raw data through the input layer, and the input layer eliminates dimensional differences between data by data resampling and standardization; S2, extracting features through the spatiotemporal convolution layer; the spatiotemporal convolution layer extracts the time features of the time series and the spatial features of the image through a one-dimensional convolution layer and a two-dimensional convolution layer respectively; The one-dimensional convolution layer applies a one-dimensional convolution kernel to slide on the time axis, extracts features through convolution operations, and uses filter size and step size parameters to control the granularity and coverage of feature extraction; the two-dimensional convolution layer slides on the image plane through a two-dimensional convolution kernel, and identifies different types of visual features through different filter groups, thereby enhancing the spatial understanding ability of the model; S3, fusing the features extracted from the time series and image data through the feature fusion layer to form a unified spatiotemporal feature representation; The feature fusion layer ensures the effective integration of cross-modal information through feature cascading and feature mapping methods; S4. Performing bidirectional analysis on time series data through the bidirectional long short-term memory layer to capture the forward and backward dependencies of the sequence and improve the recognition of abnormal event sequence patterns; the bidirectional long short-term memory layer captures the long-term dependencies in the sequence by combining forward and reverse LSTM units, and controls the information flow through a gating mechanism to enhance the ability to capture sequence dynamics; S5, dynamically adjusting the importance of different features through the attention mechanism layer; The attention mechanism layer assigns weights to different features through a weighted sum strategy, and the weighted sum strategy improves the model's attention to abnormal patterns through dot product attention; S6. Further refining and enhancing spatiotemporal features through the heterogeneous feature enhancement layer; the heterogeneous feature enhancement layer enhances the richness and robustness of feature representation through depth-separable convolution, residual connection and gated activation function; S7, based on the features of the heterogeneous feature enhancement layer, generating anomaly probability and classification labels through the output layer; The spatiotemporal convolution layer extracts information-rich spatiotemporal feature maps from the original data through a spatiotemporal feature extraction formula; the formula expression of the spatiotemporal feature extraction formula is: In formula (8), represents the spatiotemporal feature map of the output of layer I; m is a nonlinear activation function used to increase the expressive power of the model; τ is the slope of the linear regression statistical operation; k is the number of filters; is the bias term of layer I, used to adjust the convolution output; X I-1 Represents the output of the previous layer, which is used as the input feature of the current layer; Indicates that the Ith layer passes through k different spatiotemporal convolution kernels X on the previous layer I-1 The input is convolved to extract features of different scales and directions; * represents the convolution operator; The bidirectional long short-term memory layer captures the front-end dependency of the time series data through a bidirectional recurrent network with a gating mechanism; the state update formula of the bidirectional long short-term memory layer is as follows: In formula (9), W is the hidden state of the forward and reverse LSTM units; l is the sigmoid activation function, which is used to generate a value between 0 and 1 to control the degree of information passing; x is the bias term of the forward and reverse LSTM, which is used for bias adjustment; v is the weight matrix from the hidden state to the hidden state of the forward and reverse LSTM, which is used to transfer historical information; is the residual of linear regression statistics; c t is a comprehensive cell state that combines the forward and reverse hidden states and memory cell contents; The attention mechanism layer and the heterogeneous feature enhancement layer selectively enhance the spatiotemporal features and the features of the bidirectional LSTM output through the heterogeneous feature attention allocation formula to ensure that the model focuses on the most relevant features; the formula expression of the heterogeneous feature attention allocation formula is: In formula (10), R is the final enhanced feature representation; e i is the attention score, which is used to measure the importance of the i-th feature; b a is the bias term of the attention layer, which is used to adjust the initial score of the feature combination; a i is the normalized attention weight, n is the number of feature vectors; W a is the weight matrix of the attention layer, which is used to learn the importance of different feature combinations; LeakyReLU represents the activation function, which is used to increase nonlinearity.

8. The area intelligent fusion terminal data processing system based on edge computing according to claim 1 is characterized by: The edge-cloud-edge collaboration module includes a data synchronization and distribution unit, a model update and optimization unit, and a federated learning management unit; the data synchronization and distribution unit performs data transmission and intelligent distribution through a two-way data transmission mechanism and a data layered processing strategy; the two-way data transmission mechanism ensures real-time synchronization of data between the edge and the cloud through the MQTT protocol and the gRPC service; the data layered processing strategy uses an event-driven data distribution mechanism to automatically distribute data processing tasks between the edge and the cloud; the model update and optimization unit builds a model training environment in the cloud through a distributed computing framework; the distributed computing framework optimizes model parameters through stochastic gradient descent SGD; the model update and optimization unit realizes real-time model version management and seamless deployment and update on the edge through containerization and container orchestration system; the container orchestration system gradually promotes new models through a grayscale release strategy; the federated learning management unit realizes efficient aggregation of model parameters through a federated learning protocol and a federated aggregation algorithm; The federated learning protocol enables multiple edge nodes to collaboratively train a global model without leaking local data based on secure multi-party computing (SMPC) and differential privacy methods. The federated aggregation algorithm aggregates model updates submitted by edge nodes in the cloud through the federated averaging algorithm FA, ensuring that the model reaches the global optimum after multiple iterations.

9. The area intelligent fusion terminal data processing system based on edge computing according to claim 1 is characterized in that: The working principle and steps of the security and privacy protection module in the area intelligent fusion terminal data processing system based on edge computing are as follows: Step 1: Security data reception and verification; A secure channel is established at the early stage of data transmission through the TLS / SSL protocol; the TLS / SSL protocol verifies the identities of both parties in communication through a public key infrastructure and ensures the secure exchange of session keys through a handshake protocol; Step 2: Data storage encryption; The security protection of the stored data is achieved through a homomorphic encryption method, which allows direct calculation on the ciphertext through a polynomial ring; Step 3: Preprocessing of privacy data; Anonymize private data through differential privacy; the differential privacy balances privacy protection and data analysis accuracy by controlling the variance of noise through the Laplace equation; Step 4: Access control and permission management; Implement strict permission control on data access through role-based access control and attribute-based access control strategies; Step 5: Dynamic security policy deployment; Based on the current network status, dynamic security policy adjustment is performed through zero-trust network strategy; the zero-trust network strategy achieves fine control of network traffic through software-defined network SDN; Step 6: Security audit and intrusion detection; Continuously monitor the system security status and identify potential threats through real-time security monitoring systems and intrusion detection systems; The real-time security monitoring system identifies and responds to abnormal behavior through machine learning models; The intrusion detection system tracks and records all key operations through log analysis and access pattern recognition methods; Step 7: Data leakage prevention; Monitor, identify and protect sensitive data through data loss prevention strategy DLP; The data loss prevention strategy DLP prevents unauthorized data transmission and leakage through access control and data encryption methods; Step 8: Security incident response and recovery; Ensure rapid response to security incidents and system recovery through security incident response mechanism and disaster recovery backup method.

Images