Thermal power plant environment monitoring and early warning method and system based on Internet of Things

Abstract

The invention provides a thermal power plant environment monitoring and early warning method and system based on the Internet of Things, and relates to the technical field of the Internet of Things, and the method comprises the steps: forming a preprocessing data stream for collected environment monitoring data through an edge calculation node; and carrying out real-time data stream processing by utilizing a distributed message queue, constructing a hierarchical distributed data storage structure, and carrying out homomorphic encryption and desensitization processing on data to form an encrypted data set. Based on an encrypted data set, a long-short term memory network, a random forest algorithm and a limit gradient boosting tree are adopted to construct a layered anomaly detection engine, environmental parameter spatial-temporal characteristics are extracted through a multi-head attention mechanism and residual connection, an early warning threshold value is iteratively calculated based on a Bayesian optimization algorithm, and an early warning level judgment standard is dynamically adjusted. And generating early warning information. And finally, through a three-dimensional visualization module and a knowledge inference engine based on a graph neural network and a deep reinforcement learning algorithm, visualization of early warning information and generation of a self-adaptive emergency processing strategy are realized respectively.

Description

Technical Field

[0001] The present invention relates to Internet of Things technology, and in particular to an Internet of Things-based thermal power plant environment monitoring and early warning method and system. Background Art

[0002] Traditional methods for monitoring the environment around thermal power plants typically rely on manual data collection and laboratory analysis, which can be inefficient, costly, and lacking real-time performance. With the development of the Internet of Things (IoT), IoT-based environmental monitoring systems are gaining popularity. By deploying various sensors, they can collect and transmit environmental parameters in real time. However, existing IoT-based environmental monitoring and early warning methods for thermal power plants still have some shortcomings.

[0003] First, the massive amount of sensor data places enormous pressure on data processing and transmission, resulting in high warning delays and difficulty meeting the demand for real-time warnings. Existing data processing methods typically use a centralized architecture, which is prone to single points of failure and computing bottlenecks.

[0004] Secondly, environmental monitoring data is multi-source, heterogeneous, and highly correlated in time and space. Existing early warning methods struggle to effectively extract the complex features and patterns within the data, resulting in low early warning accuracy. Traditional early warning models are typically based on simple threshold judgments and are unable to adapt to complex and changing environmental conditions.

[0005] Finally, existing early warning systems lack intelligent emergency response capabilities. Traditional early warning systems can only provide simple warning information, but cannot provide specific emergency response plans, making it difficult to effectively guide emergency response work. Summary of the Invention

[0006] The embodiments of the present invention provide a thermal power plant environment monitoring and early warning method and system based on the Internet of Things, which can solve the problems in the prior art.

[0007] According to a first aspect of the embodiments of the present invention,

[0008] Provides an IoT-based thermal power plant environmental monitoring and early warning method, including:

[0009] Environmental monitoring data is collected by environmental monitoring sensors installed in the thermal power plant. The environmental monitoring data is distributedly preprocessed by multiple edge computing nodes in a grid-based edge computing network. The distributed preprocessing includes data layer compression and dynamic scheduling of task priorities. Each edge computing node calculates task allocation based on a resource-aware load balancing strategy. The preprocessed data is transmitted to the data processing center via the Internet of Things gateway to generate a preprocessed data stream.

[0010] The pre-processed data stream is processed in real time through a distributed message queue, and a hierarchical distributed data storage structure based on data timeliness, acquisition frequency, and data relevance is constructed. The data storage structure adopts a master-multiple slave chain replication mechanism to perform multiple copies of data backup. The pre-processed data stream is desensitized based on homomorphic encryption for retrievable data, an asymmetric encryption algorithm is used to generate encrypted environmental monitoring data, and a multi-level index structure is established according to spatiotemporal characteristics to form an encrypted data set;

[0011] Based on the encrypted data set, a hierarchical anomaly detection engine is constructed using a long short-term memory network, a random forest algorithm, and an extreme gradient boosting tree. The hierarchical anomaly detection engine extracts the spatiotemporal characteristics of environmental parameters through a multi-head attention mechanism and residual connections, iteratively calculates the warning threshold based on the spatiotemporal characteristics of the environmental parameters through a Bayesian optimization algorithm, dynamically adjusts the warning level judgment standard according to the historical warning accuracy, and generates warning information including the warning level, warning area, and warning parameters; the warning information is converted into a multidimensional stereoscopic chart through a three-dimensional visualization module, and a knowledge reasoning engine is constructed based on a graph neural network and a deep reinforcement learning algorithm to generate an adaptive emergency response strategy.

[0012] In an optional embodiment,

[0013] Environmental monitoring data is collected by environmental monitoring sensors installed in the thermal power plant. The environmental monitoring data is distributed preprocessed by multiple edge computing nodes to form a grid edge computing network. The distributed preprocessing includes data layer compression and dynamic scheduling of task priorities. The steps of calculating task allocation based on resource-aware load balancing strategies of each edge computing node include:

[0014] Collect environmental monitoring data through dust concentration sensors, gas sensors, noise sensors, and temperature and humidity sensors installed in the thermal power plant. The environmental monitoring data includes device identification, timestamp, geographic location information, and measurement values;

[0015] The environmental monitoring data is input into multiple edge computing nodes according to the grid units divided by the physical space of the thermal power plant. The edge computing nodes build a grid edge computing network through Ethernet switches and point-to-point communication protocols. The distributed state synchronization protocol is used between the edge computing nodes for heartbeat maintenance;

[0016] Each edge computing node performs distributed preprocessing on the environmental monitoring data, performs wavelet transform and coefficient screening on the continuously changing environmental monitoring data to form first preprocessed data, performs differential encoding on the slowly changing environmental monitoring data to form second preprocessed data, and merges the first preprocessed data and the second preprocessed data to generate a preprocessed data packet, wherein the preprocessed data packet includes a compression type identifier and original data length information;

[0017] A load balancing strategy is used to allocate tasks. By obtaining the processor utilization, memory occupancy, and network bandwidth utilization of each edge computing node in real time, the pre-processed data packets are graded according to the urgency of the tasks based on real-time alarms, routine monitoring, and statistical analysis. The task priority is determined based on the data volume of the pre-processed data packets. The task allocation weight is dynamically adjusted according to the resource occupancy of each edge computing node, and the pre-processed data packets are allocated to the corresponding edge computing nodes. The edge computing nodes execute tasks in a time-slice rotation manner.

[0018] In an optional embodiment,

[0019] The pre-processed data stream is processed in real time through a distributed message queue, and a hierarchical distributed data storage structure based on data timeliness, acquisition frequency, and data relevance is constructed. The data storage structure uses a master-multiple slave chain replication mechanism to perform multiple copies of data backup. The pre-processed data stream is subjected to retrievable data desensitization processing based on homomorphic encryption, and an asymmetric encryption algorithm is used to generate encrypted environmental monitoring data. The steps include:

[0020] Processing the pre-processed data stream through a distributed message queue, wherein the distributed message queue uses an asynchronous batch method to send messages containing a pre-processed data stream identifier, a timestamp, data content, and a checksum, and each partition in the distributed message queue is provided with multiple copies and a chain replication mechanism is used to generate a message queue data stream;

[0021] The message queue data stream is divided into a hot data layer, a warm data layer and a cold data layer according to timeliness, the hot data layer uses an in-memory database to store data generated within two hours, the warm data layer uses a solid-state hard disk to store data from two hours to forty-eight hours, and the cold data layer uses a mechanical hard disk to store data exceeding forty-eight hours; the data is sharded and stored according to the data acquisition frequency, and the high-frequency acquisition data with a time interval of less than or equal to sixty seconds is sharded according to a ten-minute time interval, the medium-frequency acquisition data with a time interval of one minute to sixty minutes is sharded according to a thirty-minute time interval, and the low-frequency acquisition data with a time interval of more than sixty minutes is sharded according to a sixty-minute time interval; a multi-level index structure of data association degree is established, including a primary index, a secondary index and an associated index, wherein the primary index includes a timestamp and a device identifier, the secondary index includes a geographic location and a data type, and the associated index includes a device group and a monitoring indicator, so as to form hierarchical index data;

[0022] Based on the formation of hierarchical index data, a one-master-multiple-slave chain replication mechanism is used to perform data backup on the pre-processed data stream, wherein the master node receives a data write request and synchronizes the data to each slave node in a preset chain sequence, each slave node copies the data level by level and returns confirmation information, and the master node completes the data write after receiving confirmation information from all slave nodes, generating backup confirmation data;

[0023] According to the backup confirmation data, the hierarchically stored numerical data is encrypted using a homomorphic encryption algorithm, which supports addition operations in the encrypted state; the hierarchically stored non-numerical data is encrypted using an asymmetric encryption algorithm, whose private key is dispersedly stored in the authorized nodes to generate encrypted environmental monitoring data.

[0024] In an optional embodiment,

[0025] Based on the encrypted data set, the steps of constructing a hierarchical anomaly detection engine using a long short-term memory network, a random forest algorithm, and an extreme gradient boosting tree include:

[0026] Performing multi-scale feature extraction on the encrypted data set, constructing a multi-dimensional feature representation by combining time domain features and frequency domain features through multi-scale decomposition based on wavelet transform, and generating multi-scale feature data;

[0027] Based on the multi-scale feature data, an attention-enhanced encrypted domain long short-term memory network model is constructed. The encrypted domain long short-term memory network model includes a time attention mechanism and a feature attention mechanism. The time attention mechanism calculates the importance weight of the time domain features, and the feature attention mechanism performs correlation analysis on the frequency domain features. An encrypted domain attention score matrix is constructed based on the obtained time domain weights and frequency domain correlations. The multi-scale feature data is weighted reconstructed using the attention score matrix. Sequence modeling is performed based on the reconstructed feature data. An adversarial training module for sequence prediction error is constructed, and a time series anomaly detection result is generated based on the prediction error.

[0028] The multi-scale feature data is optimized using a random forest algorithm in a ciphertext space, a decision tree splitting criterion is constructed based on homomorphic encryption, and an initial decision tree set is generated through approximate calculation. The information gain value of the encrypted feature is calculated for each tree in the initial decision tree set, and a feature importance evaluation index is determined based on the information gain value, and the feature sampling probability is dynamically adjusted; the initial decision tree is reconstructed according to the adjusted sampling probability, the node redundancy is calculated based on the reconstructed decision tree set, and the nodes whose redundancy exceeds a preset threshold are deleted to complete the pruning operation; and the optimized random forest algorithm is used to generate spatial anomaly detection results.

[0029] In an optional embodiment,

[0030] The method further comprises:

[0031] In the encrypted domain, a second-order expansion approximation is performed on the homomorphically encrypted data through Taylor expansion, a ciphertext domain loss function is constructed based on the approximation result, the first-order and second-order gradient information of the ciphertext domain loss function with respect to the encrypted feature is calculated, the first-order and second-order gradient information is used to construct a splitting gain evaluation criterion for the encrypted feature, the optimal splitting feature and splitting point are selected based on the splitting gain evaluation criterion, entropy value calculation is performed on the encrypted feature in the multi-scale feature data, the feature binning boundary is determined according to the entropy value, the feature binning boundary result is combined with the splitting gain evaluation result to generate an extreme gradient boosting tree, and a pattern anomaly detection result is output by iteratively accumulating the prediction results of the extreme gradient boosting tree;

[0032] The information entropy of the time series anomaly detection results, spatial anomaly detection results and pattern anomaly detection results is calculated to obtain the uncertainty level of each detection result. The mutual information between any two detection results is calculated based on the probability distribution of each detection result. Each detection result is randomly perturbed by Monte Carlo sampling in the encrypted domain to obtain the prediction confidence. The adaptive fusion weight of each detection result is calculated according to the uncertainty level, the prediction confidence and the historical performance coefficient of each detection result. The historical performance coefficient is dynamically updated through a sliding time window. The adaptive fusion weight is used to weight each anomaly detection result to obtain an initial fusion result. The initial fusion result is multi-layered with the mutual information and the historical anomaly score based on the time series attenuation factor. The historical anomaly score is attenuated and accumulated considering the time series correlation to generate the final anomaly detection result.

[0033] In an optional embodiment,

[0034] The hierarchical anomaly detection engine extracts the spatiotemporal features of environmental parameters through a multi-head attention mechanism and residual connections, iteratively calculates the warning threshold based on the spatiotemporal features of the environmental parameters through a Bayesian optimization algorithm, and dynamically adjusts the warning level judgment standard according to the historical warning accuracy. The steps of generating warning information including the warning level, warning area, and warning parameters include:

[0035] The environmental parameter sequence is feature extracted through a multi-head attention mechanism. The product of the query matrix, key matrix and value matrix is calculated for each attention head of the environmental parameter sequence. The output of each attention head is residually connected with the original environmental parameter sequence to obtain the spatiotemporal features of the environmental parameters.

[0036] An adaptive sampling strategy is constructed based on the spatiotemporal characteristics of the environmental parameters. The sampling density function is dynamically adjusted based on the fluctuation change rate of the environmental parameters. The spatiotemporal characteristics of the environmental parameters are dynamically sampled to obtain a sampling data set. A Gaussian process regression model with multiple kernel functions of different scales is constructed using the sampling data set. The weight coefficients of the kernel functions are adaptively adjusted according to the characteristics of the environmental parameters. A particle swarm is initialized based on the prediction results of the Gaussian process regression model. A particle swarm velocity term with a dynamic exploration factor is introduced into a Bayesian optimization process. The warning threshold is iteratively optimized using a multi-objective optimization function that includes data coverage, parameter fluctuation amplitude, time series correlation, and spatial clustering degree.

[0037] A deep Q network is constructed to dynamically adjust the warning level. The historical warning accuracy sequence, the spatiotemporal characteristics of the environmental parameters, and the current warning threshold set are used as state space inputs. The threshold adjustment amount, the merging and splitting operations of the warning level, and the time window size adjustment are used as action space outputs. A comprehensive reward function is constructed based on the warning time delay rate, false alarm rate, missed alarm rate, and warning level change frequency. A dual network architecture consisting of a target network and an evaluation network is used to optimize the warning level determination strategy.

[0038] Multimodal warning information is generated based on the warning threshold and the output value of the deep Q network, wherein the multimodal warning information includes a warning level, a warning area, warning parameters and a confidence assessment result, wherein the confidence assessment result is calculated based on the ratio of the output value of the deep Q network to the maximum Q value.

[0039] In an optional embodiment,

[0040] The steps of converting the warning information into a multi-dimensional stereoscopic chart through a three-dimensional visualization module, and constructing a knowledge reasoning engine based on a graph neural network and a deep reinforcement learning algorithm to generate an adaptive emergency response strategy include:

[0041] The warning information is constructed into a three-layer rendering structure. The bottom geographic information layer uses quadtree grid division for detailed hierarchical management. The middle warning information layer uses adaptive grid subdivision to spatially discretize environmental parameters and constructs the parameter distribution field through bilinear interpolation. The upper dynamic annotation layer uses frustum clipping and annotation avoidance algorithms to overlay information. Combined with a color band mapping scheme based on perceptual uniformity and a local variance adaptive transparency adjustment mechanism, 3D visualization data of the warning information is generated.

[0042] Based on the three-dimensional visualization data, a warning knowledge graph is constructed, with warning events, environmental parameters, spatial locations, and emergency measures as node types. Spatial proximity relationships are calculated using R-tree indexes, and semantic similarity is calculated using word vectors to obtain event associations. The knowledge graph is modeled using a heterogeneous graph neural network, which includes a type feature conversion layer to process different node features, a message passing layer combined with an attention mechanism to adjust edge weights, and a feature aggregation layer and a position encoding module with jump connections to form a knowledge graph representation vector.

[0043] The knowledge graph representation vector and the environmental parameter time series features in the three-dimensional visualization data are spliced together to construct a state space, and a hierarchical architecture of a deep reinforcement learning algorithm is used for decision-making, wherein the meta-strategy layer determines the sub-goal sequence based on long-term reward prediction, and the execution strategy layer generates specific emergency instructions based on short-term rewards; a composite reward function is constructed based on the node association, edge weight and spatial position information in the early warning knowledge graph, and the composite reward function includes parameter improvement, impact area range, resource scheduling cost, response delay and knowledge consistency weight terms. An experience replay module based on the attention mechanism is introduced to store the historical data of the composite reward function, and the relevance of the action sequence is analyzed through a recursive neural network, and an emergency processing strategy is output based on an asynchronous advantage strategy network.

[0044] According to a second aspect of the embodiments of the present invention,

[0045] Provide an IoT-based thermal power plant environmental monitoring and early warning system, including:

[0046] The first unit is used to collect environmental monitoring data through environmental monitoring sensors installed in the thermal power plant. The environmental monitoring data is distributedly preprocessed by a grid edge computing network constructed by multiple edge computing nodes. The distributed preprocessing includes data layer compression and dynamic scheduling of task priorities. Each edge computing node calculates task allocation based on a resource-aware load balancing strategy. The preprocessed data is transmitted to the data processing center via the Internet of Things gateway to generate a preprocessed data stream;

[0047] The second unit is used to process the pre-processed data stream in real time through a distributed message queue, build a hierarchical distributed data storage structure based on data timeliness, acquisition frequency and data relevance, and use a master-multiple slave chain replication mechanism to perform multiple copies of data backup. The pre-processed data stream is desensitized based on homomorphic encryption for retrievable data, an asymmetric encryption algorithm is used to generate encrypted environmental monitoring data, and a multi-level index structure is established according to spatiotemporal characteristics to form an encrypted data set.

[0048] The third unit is used to construct a hierarchical anomaly detection engine based on the encrypted data set, using a long short-term memory network, a random forest algorithm, and an extreme gradient boosting tree. The hierarchical anomaly detection engine extracts the spatiotemporal characteristics of environmental parameters through a multi-head attention mechanism and residual connections, iteratively calculates the warning threshold based on the spatiotemporal characteristics of the environmental parameters through a Bayesian optimization algorithm, dynamically adjusts the warning level judgment standard according to the historical warning accuracy, and generates warning information including warning level, warning area, and warning parameters; converts the warning information into a multidimensional stereoscopic chart through a three-dimensional visualization module, and constructs a knowledge reasoning engine based on a graph neural network and a deep reinforcement learning algorithm to generate an adaptive emergency response strategy.

[0049] According to a third aspect of the embodiments of the present invention,

[0050] An electronic device is provided, comprising:

[0051] processor;

[0052] a memory for storing processor-executable instructions;

[0053] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0054] According to a fourth aspect of the embodiments of the present invention,

[0055] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0056] The present invention adopts a distributed preprocessing method of edge computing, including data layered compression and dynamic scheduling of task priorities, and performs task allocation based on a resource-aware load balancing strategy, which can effectively reduce data transmission volume and processing delay and improve preprocessing efficiency.

[0057] This invention adopts a hierarchical distributed data storage structure and a chain replication mechanism to perform multiple data copies to improve data reliability. At the same time, homomorphic encryption technology is used to desensitize the data and an asymmetric encryption algorithm is used to encrypt the data to ensure data security.

[0058] This invention uses a layered anomaly detection engine, combining multiple algorithms to extract spatiotemporal characteristics of environmental parameters. It then iteratively calculates warning thresholds using a Bayesian optimization algorithm, dynamically adjusting the criteria for determining warning levels to improve warning accuracy. Furthermore, a knowledge inference engine is constructed using graph neural networks and deep reinforcement learning algorithms to generate adaptive emergency response strategies, improving efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1This is a flow chart of a method for monitoring and early warning of a thermal power plant environment based on the Internet of Things according to an embodiment of the present invention;

[0060] Figure 2 The figure is a schematic structural diagram of an Internet of Things-based thermal power plant environment monitoring and early warning system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings 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 making creative efforts shall fall within the scope of protection of the present invention.

[0062] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0063] Figure 1 FIG. 1 is a flow chart of a method for monitoring and warning the environment of a thermal power plant based on the Internet of Things according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0064] S1. Environmental monitoring data is collected by environmental monitoring sensors installed within the thermal power plant. This data is distributedly preprocessed by multiple edge computing nodes in a gridded edge computing network. This distributed preprocessing includes data layer compression and dynamic scheduling of task priorities. Each edge computing node allocates computational tasks based on a resource-aware load balancing strategy. The preprocessed data is transmitted to the data processing center via an IoT gateway, generating a preprocessed data stream.

[0065] S2. Process the preprocessed data stream in real time through a distributed message queue, construct a hierarchical distributed data storage structure based on data timeliness, acquisition frequency, and data relevance, and use a master-multiple slave chain replication mechanism to back up multiple copies of data. Perform homomorphic encryption-based searchable data desensitization on the preprocessed data stream, generate encrypted environmental monitoring data using an asymmetric encryption algorithm, and establish a multi-level index structure based on spatiotemporal characteristics to form an encrypted data set.

[0066] S3. Based on the encrypted data set, a hierarchical anomaly detection engine is constructed using a long short-term memory network, a random forest algorithm, and an extreme gradient boosting tree. The hierarchical anomaly detection engine extracts the spatiotemporal characteristics of environmental parameters through a multi-head attention mechanism and residual connections, iteratively calculates the warning threshold based on the spatiotemporal characteristics of the environmental parameters through a Bayesian optimization algorithm, dynamically adjusts the warning level judgment standard according to the historical warning accuracy, and generates warning information including warning level, warning area, and warning parameters; the warning information is converted into a multidimensional stereogram through a three-dimensional visualization module, and a knowledge reasoning engine is constructed based on a graph neural network and a deep reinforcement learning algorithm to generate an adaptive emergency response strategy.

[0067] In an optional embodiment,

[0068] Environmental monitoring data is collected by environmental monitoring sensors installed in the thermal power plant. The environmental monitoring data is distributed preprocessed by multiple edge computing nodes to form a grid edge computing network. The distributed preprocessing includes data layer compression and dynamic scheduling of task priorities. The steps of calculating task allocation based on resource-aware load balancing strategies of each edge computing node include:

[0069] Collect environmental monitoring data through dust concentration sensors, gas sensors, noise sensors, and temperature and humidity sensors installed in the thermal power plant. The environmental monitoring data includes device identification, timestamp, geographic location information, and measurement values;

[0070] The environmental monitoring data is input into multiple edge computing nodes according to the grid units divided by the physical space of the thermal power plant. The edge computing nodes build a grid edge computing network through Ethernet switches and point-to-point communication protocols. The distributed state synchronization protocol is used between the edge computing nodes for heartbeat maintenance;

[0071] Each edge computing node performs distributed preprocessing on the environmental monitoring data, performs wavelet transform and coefficient screening on the continuously changing environmental monitoring data to form first preprocessed data, performs differential encoding on the slowly changing environmental monitoring data to form second preprocessed data, and merges the first preprocessed data and the second preprocessed data to generate a preprocessed data packet, wherein the preprocessed data packet includes a compression type identifier and original data length information;

[0072] A load balancing strategy is used to allocate tasks. By obtaining the processor utilization, memory occupancy, and network bandwidth utilization of each edge computing node in real time, the pre-processed data packets are graded according to the urgency of the tasks based on real-time alarms, routine monitoring, and statistical analysis. The task priority is determined based on the data volume of the pre-processed data packets. The task allocation weight is dynamically adjusted according to the resource occupancy of each edge computing node, and the pre-processed data packets are allocated to the corresponding edge computing nodes. The edge computing nodes execute tasks in a time-slice rotation manner.

[0073] For example, first, various types of environmental monitoring sensors such as dust concentration sensors, gas sensors, noise sensors, and temperature and humidity sensors are deployed in the thermal power plant. The data collected by these sensors include device identification, timestamp, geographic location information, and measurement values. For example, the dust concentration value "150mg / m 3 ", temperature value "30℃".

[0074] The collected environmental monitoring data is then fed into the corresponding edge computing nodes according to the grid cells that define the physical space of the thermal power plant. For example, data from Area A is transmitted to Edge Node 1, data from Area B is transmitted to Edge Node 2, and so on. These edge computing nodes form a grid-like edge computing network using Ethernet switches and point-to-point communication protocols. Distributed state synchronization protocols, such as Raft or Paxos, are used for heartbeat maintenance to ensure communication and state synchronization between nodes.

[0075] Next, each edge computing node performs distributed preprocessing on the received environmental monitoring data. For slowly changing environmental monitoring data such as temperature and humidity, differential encoding is used for compression. For example, if the current temperature is 30°C and the previous temperature was 29°C, the 1°C difference is recorded. For rapidly changing data such as dust concentration and noise, wavelet transform and coefficient screening are used for compression. For example, after performing a wavelet transform on dust concentration data, only important coefficients are retained, while coefficients with less impact on the data are discarded, thus achieving data compression. The two preprocessed data are combined to generate a preprocessed data packet, which contains the compression type identifier and the original data length information.

[0076] The system then uses a resource-aware load balancing strategy to allocate tasks. First, it obtains real-time resource information for each edge computing node, including processor utilization, memory utilization, and network bandwidth utilization. For example, edge node 1 has a CPU utilization of 70%, a memory utilization of 50%, and a network bandwidth utilization of 30%. It then prioritizes preprocessed data packets based on their urgency, categorizing them as real-time alarms, routine monitoring, and statistical analysis. The priority of these tasks is determined based on the size of the preprocessed data packets. For example, real-time alarms have the highest priority, followed by routine monitoring, and finally statistical analysis. Tasks with larger data volumes also have a higher priority. Task allocation weights are dynamically adjusted based on the resource utilization of each edge computing node, assigning preprocessed data packets to the appropriate edge computing node. For example, if edge node 1 has a lower resource utilization, the task assigned to it will receive a higher weight. Finally, the edge computing nodes execute the assigned tasks using a time-slicing rotation.

[0077] The present invention can effectively reduce data transmission volume and processing time and improve data processing efficiency through data layer compression and dynamic scheduling of task priorities; through distributed preprocessing and load balancing strategies, it can effectively reduce data processing delays and enhance the real-time performance of the system, thereby responding to abnormal situations in thermal power plants more quickly; the gridded edge computing network and distributed state synchronization protocol can improve the fault tolerance and reliability of the system, and even if some nodes fail, it will not affect the normal operation of the entire system.

[0078] In an optional embodiment,

[0079] The pre-processed data stream is processed in real time through a distributed message queue, and a hierarchical distributed data storage structure based on data timeliness, acquisition frequency, and data relevance is constructed. The data storage structure uses a master-multiple slave chain replication mechanism to perform multiple copies of data backup. The pre-processed data stream is subjected to retrievable data desensitization processing based on homomorphic encryption, and an asymmetric encryption algorithm is used to generate encrypted environmental monitoring data. The steps include:

[0080] Processing the pre-processed data stream through a distributed message queue, wherein the distributed message queue uses an asynchronous batch method to send messages containing a pre-processed data stream identifier, a timestamp, data content, and a checksum, and each partition in the distributed message queue is provided with multiple copies and a chain replication mechanism is used to generate a message queue data stream;

[0081] The message queue data stream is divided into a hot data layer, a warm data layer and a cold data layer according to timeliness, the hot data layer uses an in-memory database to store data generated within two hours, the warm data layer uses a solid-state hard disk to store data from two hours to forty-eight hours, and the cold data layer uses a mechanical hard disk to store data exceeding forty-eight hours; the data is sharded and stored according to the data acquisition frequency, and the high-frequency acquisition data with a time interval of less than or equal to sixty seconds is sharded according to a ten-minute time interval, the medium-frequency acquisition data with a time interval of one minute to sixty minutes is sharded according to a thirty-minute time interval, and the low-frequency acquisition data with a time interval of more than sixty minutes is sharded according to a sixty-minute time interval; a multi-level index structure of data association degree is established, including a primary index, a secondary index and an associated index, wherein the primary index includes a timestamp and a device identifier, the secondary index includes a geographic location and a data type, and the associated index includes a device group and a monitoring indicator, so as to form hierarchical index data;

[0082] Based on the formation of hierarchical index data, a one-master-multiple-slave chain replication mechanism is used to perform data backup on the pre-processed data stream, wherein the master node receives a data write request and synchronizes the data to each slave node in a preset chain sequence, each slave node copies the data level by level and returns confirmation information, and the master node completes the data write after receiving confirmation information from all slave nodes, generating backup confirmation data;

[0083] According to the backup confirmation data, the hierarchically stored numerical data is encrypted using a homomorphic encryption algorithm, which supports addition operations in the encrypted state; the hierarchically stored non-numerical data is encrypted using an asymmetric encryption algorithm, whose private key is dispersedly stored in the authorized nodes to generate encrypted environmental monitoring data.

[0084] For example, preprocessed data streams are transmitted in real time via a distributed message queue. This distributed message queue uses Kafka, configured with multiple partitions to increase throughput. Three replicas are set for each partition, using a chained replication mechanism to ensure data reliability. Messages are sent asynchronously in batches, each containing the following: a preprocessed data stream identifier using a universally unique identifier (UID), a millisecond-accurate timestamp based on a unified timestamp, data content in JavaScript Object Notation (JavaScript Object Notation), and a checksum based on the CRC32 algorithm.

[0085] Next, the message queue data stream is stored in layers based on data timeliness, collection frequency, and data relevance. Timeliness division: hot data within two hours is stored in the Redis cluster; warm data between two hours and forty-eight hours is stored in the SSD database Cassandra; cold data over forty-eight hours is stored in the distributed file system. Collection frequency division: high-frequency data, such as temperature, is collected every 30 seconds and stored in shards at ten-minute time intervals; medium-frequency data, such as PM2.5, is collected every ten minutes and stored in shards at thirty-minute time intervals; low-frequency data, such as soil moisture, is collected every hour and stored in shards at sixty-minute time intervals. Data relevance division: a multi-level index structure is constructed. The primary index contains timestamps and device identifiers and is stored in Elasticsearch; the secondary index contains geographic location and data type and is stored in MongoDB; the association index contains device groups and monitoring indicators and is stored in the Neo4j graph database.

[0086] Data backup is then performed using a master-slave chain replication mechanism. Taking Cassandra as an example, three data replica nodes are set up: one is the master node, and the other two are slave nodes. The master node receives a data write request and synchronizes the data to the first slave node. The first slave node replicates the data and synchronizes it to the second slave node. The second slave node replicates the data and returns a confirmation message to the first slave node, which then returns a confirmation message to the master node. After receiving confirmation messages from all slave nodes, the master node completes the data write operation and generates backup confirmation data.

[0087] Finally, the hierarchically stored data undergoes retrievable data desensitization. Numerical data is encrypted using the Paillier homomorphic encryption algorithm, which allows addition operations to be performed while encrypted, such as summing encrypted temperature data. Non-numeric data is encrypted using an asymmetric encryption algorithm, with the private key distributed across multiple authorized nodes. For example, the Shamir secret sharing algorithm is used to split the private key into three parts, each stored on three authorized nodes. Decryption requires the private keys of at least two authorized nodes. The encrypted data is then stored in the corresponding data storage tier.

[0088] The present invention utilizes distributed message queues and chain replication mechanisms to ensure real-time data transmission and multiple copy backup, thereby improving data reliability and system fault tolerance. A hierarchical storage structure based on data timeliness, acquisition frequency, and data relevance, as well as a multi-level indexing mechanism, enables rapid data retrieval and efficient storage, reducing storage costs. Homomorphic encryption and asymmetric encryption algorithms are used to desensitize data, ensuring data confidentiality and security and preventing data leakage and tampering.

[0089] In an optional embodiment,

[0090] Based on the encrypted data set, the steps of constructing a hierarchical anomaly detection engine using a long short-term memory network, a random forest algorithm, and an extreme gradient boosting tree include:

[0091] Performing multi-scale feature extraction on the encrypted data set, constructing a multi-dimensional feature representation by combining time domain features and frequency domain features through multi-scale decomposition based on wavelet transform, and generating multi-scale feature data;

[0092] Based on the multi-scale feature data, an attention-enhanced encrypted domain long short-term memory network model is constructed. The encrypted domain long short-term memory network model includes a time attention mechanism and a feature attention mechanism. The time attention mechanism calculates the importance weight of the time domain features, and the feature attention mechanism performs correlation analysis on the frequency domain features. An encrypted domain attention score matrix is constructed based on the obtained time domain weights and frequency domain correlations. The multi-scale feature data is weighted reconstructed using the attention score matrix. Sequence modeling is performed based on the reconstructed feature data. An adversarial training module for sequence prediction error is constructed, and a time series anomaly detection result is generated based on the prediction error.

[0093] The multi-scale feature data is optimized using a random forest algorithm in a ciphertext space, a decision tree splitting criterion is constructed based on homomorphic encryption, and an initial decision tree set is generated through approximate calculation. The information gain value of the encrypted feature is calculated for each tree in the initial decision tree set, and a feature importance evaluation index is determined based on the information gain value, and the feature sampling probability is dynamically adjusted; the initial decision tree is reconstructed according to the adjusted sampling probability, the node redundancy is calculated based on the reconstructed decision tree set, and the nodes whose redundancy exceeds a preset threshold are deleted to complete the pruning operation; and the optimized random forest algorithm is used to generate spatial anomaly detection results.

[0094] For example, multi-scale feature extraction is first performed on the encrypted dataset. Using wavelet transform, the original encrypted data is decomposed into data sequences of different scales. Time domain features (such as mean, variance, and peak) and frequency domain features (such as spectral energy and dominant frequency) are extracted for each scale. The extracted time and frequency domain features are combined to form a multidimensional feature vector, which serves as input for the subsequent model.

[0095] For example, consider a segment of encrypted time series data [10, 12, 15, 14, 16, 18, 20, 19, 17, 15]. After wavelet transform, we can obtain two scaled data sequences: a low-frequency sequence [13, 15, 18, 18, 16] and a high-frequency sequence [-3, 2, -1, 2, 2, 1, -1, -2, -2]. For the low-frequency sequence, we can extract a mean of 16 and a variance of 4; for the high-frequency sequence, we can extract an energy of 26. These features are combined to form a multidimensional feature vector [16, 4, 26]. The same process is performed on all data in the dataset to generate multi-scale feature data.

[0096] Next, we construct an attention-enhanced encrypted domain long short-term memory (LSTM) network model. This model incorporates both temporal and feature-attention mechanisms. The temporal attention mechanism assigns different weights to features at different time steps based on their importance. The feature-attention mechanism analyzes the correlation between different frequency-domain features and assigns different weights to different features based on this correlation. The temporal weights and feature correlations are combined to construct an attention score matrix. The attention score matrix is used to perform weighted reconstruction of multi-scale feature data. Sequence modeling is performed based on the reconstructed feature data, and the LSTM network model is trained to predict future data values. The difference between the predicted value and the true value is used as the prediction error. An adversarial training module is constructed to enable the model to distinguish prediction errors generated by normal data from those generated by anomalous data. The magnitude of the prediction error determines whether a temporal anomaly exists.

[0097] For example, a temporal attention mechanism might assign higher weight to the most recent time step, while a feature attention mechanism might assign higher weight to energy features. Assume the attention score matrix is [[0.8, 0.2, 0.5], [0.7, 0.3, 0.6], [0.9, 0.1, 0.7]]. Multiply this matrix by the multi-scale feature data to obtain reconstructed feature data. A long short-term memory network model is trained based on this reconstructed feature data. If the prediction error exceeds a preset threshold, a time series anomaly is determined.

[0098] The multi-scale feature data is then optimized using a random forest algorithm in ciphertext space. Homomorphic encryption technology is used to perform calculations without decrypting the data. A decision tree splitting criterion based on homomorphic encryption is constructed, and an initial set of decision trees is generated through approximate calculations. The information gain of the encrypted features is calculated for each tree in the set. The importance of the features is assessed based on the information gain, and the sampling probability of the features is dynamically adjusted. The decision tree is reconstructed based on the adjusted sampling probability. The redundancy of the nodes in the reconstructed decision tree set is calculated, and nodes with redundancy exceeding a preset threshold are removed to complete the pruning operation. The optimized random forest algorithm is used to generate spatial anomaly detection results.

[0099] For example, suppose there are two features, Feature A and Feature B. By calculating the information gain, it is found that Feature A has a higher information gain than Feature B. Therefore, the sampling probability is adjusted to increase the sampling probability of Feature A. The decision tree is reconstructed based on the adjusted sampling probability. If the redundancy of a node exceeds a preset threshold, the node is deleted.

[0100] Finally, the temporal anomaly detection results generated by the long short-term memory network and the spatial anomaly detection results generated by the random forest algorithm are fused to obtain the final anomaly detection results.

[0101] The multi-scale feature extraction and attention mechanism of the present invention can capture data features more comprehensively and improve the accuracy of anomaly detection; the use of homomorphic encryption technology allows calculations to be performed without decrypting the data, thus protecting the privacy and security of the data; the adversarial training module and the optimization strategy of the random forest algorithm enhance the robustness of the model, enabling it to adapt to different data distributions and noise interference.

[0102] In an optional embodiment,

[0103] The method further comprises:

[0104] In the encrypted domain, a second-order expansion approximation is performed on the homomorphically encrypted data through Taylor expansion, a ciphertext domain loss function is constructed based on the approximation result, the first-order and second-order gradient information of the ciphertext domain loss function with respect to the encrypted feature is calculated, the first-order and second-order gradient information is used to construct a splitting gain evaluation criterion for the encrypted feature, the optimal splitting feature and splitting point are selected based on the splitting gain evaluation criterion, entropy value calculation is performed on the encrypted feature in the multi-scale feature data, the feature binning boundary is determined according to the entropy value, the feature binning boundary result is combined with the splitting gain evaluation result to generate an extreme gradient boosting tree, and a pattern anomaly detection result is output by iteratively accumulating the prediction results of the extreme gradient boosting tree;

[0105] The information entropy of the time series anomaly detection results, spatial anomaly detection results and pattern anomaly detection results is calculated to obtain the uncertainty level of each detection result. The mutual information between any two detection results is calculated based on the probability distribution of each detection result. Each detection result is randomly perturbed by Monte Carlo sampling in the encrypted domain to obtain the prediction confidence. The adaptive fusion weight of each detection result is calculated according to the uncertainty level, the prediction confidence and the historical performance coefficient of each detection result. The historical performance coefficient is dynamically updated through a sliding time window. The adaptive fusion weight is used to weight each anomaly detection result to obtain an initial fusion result. The initial fusion result is multi-layered with the mutual information and the historical anomaly score based on the time series attenuation factor. The historical anomaly score is attenuated and accumulated considering the time series correlation to generate the final anomaly detection result.

[0106] For example, data preprocessing is performed first. Multi-source time series data, such as sensor readings and network traffic, is collected. The data is cleaned, for example, by removing noise and filling missing values. The cleaned data is then normalized, for example, using min-max normalization or Z-score normalization, to convert the data to a uniform scale.

[0107] Next, we perform multi-scale feature extraction. We perform multi-scale decomposition on the preprocessed time series data, using techniques such as wavelet transform and empirical mode decomposition, to extract features at different time scales. For example, we can extract features at multiple scales, such as the original time series, short-term trends, and long-term trends. Consider a set of time series data: [10, 12, 15, 14, 16, 18, 20, 19, 22, 25]. By performing a wavelet transform on this data, we can obtain the original sequence, the high-frequency component, and the low-frequency component.

[0108] The extracted multi-scale features are then encrypted. Homomorphic encryption is used to encrypt the multi-scale feature data to protect data privacy. For example, the Paillier homomorphic encryption algorithm can be used. Assuming that the extracted feature value is 10, after using Paillier encryption, the ciphertext C is obtained.

[0109] An extreme gradient boosting tree model is constructed in the encrypted domain. In the encrypted domain, a second-order expansion of the encrypted data is approximated using Taylor expansion. A ciphertext domain loss function is constructed based on the approximation. The first- and second-order gradient information of the ciphertext domain loss function with respect to the encrypted features is calculated. The first- and second-order gradient information is used to construct a splitting gain evaluation criterion for the encrypted features. The optimal splitting features and splitting points are selected based on the splitting gain evaluation criterion. Entropy values are calculated for the encrypted features in the multi-scale feature data. Feature binning boundaries are determined based on the entropy values. The feature binning boundary results are combined with the splitting gain evaluation results to generate an extreme gradient boosting tree. Assuming the entropy value of a feature is 0.8, it is divided into three bins according to the preset binning strategy. The pattern anomaly detection results are output by iteratively accumulating the prediction results of the extreme gradient boosting tree.

[0110] Perform both temporal and spatial anomaly detection. Perform temporal analysis on raw time series data, such as sliding window statistics and time series decomposition, to detect anomalies in the temporal dimension. Perform spatial analysis on multidimensional time series data, such as cluster analysis and outlier detection, to detect anomalies in the spatial dimension. If the value of time series data at a certain point in time suddenly increases significantly, a temporal anomaly is considered to exist at that point in time. If the data from a particular sensor differs significantly from that from other sensors, a spatial anomaly is considered to exist in that sensor's data.

[0111] Multi-source detection results are fused. Information entropy is calculated for the temporal, spatial, and pattern anomaly detection results to determine the uncertainty level of each detection result. The mutual information between any two detection results is calculated based on the probability distribution of each detection result. Each detection result is randomly perturbed in the encrypted domain using Monte Carlo sampling to obtain the prediction confidence. An adaptive fusion weight is calculated for each detection result based on the uncertainty level, prediction confidence, and the historical performance coefficient of each detection result. The historical performance coefficient is dynamically updated using a sliding time window. The anomaly detection results are weighted using the adaptive fusion weight to obtain an initial fusion result. The initial fusion result is then multi-layered with the mutual information and historical anomaly scores based on a temporal decay factor. The historical anomaly scores are accumulated and decayed to account for temporal correlation. The final anomaly detection result is generated. For example, if the temporal anomaly detection result has a weight of 0.4, the spatial anomaly detection result has a weight of 0.3, and the pattern anomaly detection result has a weight of 0.3, the final anomaly detection result is the weighted average of these three.

[0112] The present invention uses homomorphic encryption technology to encrypt data, realizing anomaly detection while protecting data privacy, effectively avoiding the risk of data leakage, and is suitable for scenarios with high data security requirements. It integrates multi-scale features, temporal information and spatial information, and can more comprehensively capture abnormal patterns in the data, improve detection accuracy, reduce false alarm rate and missed alarm rate, and dynamically adjust the fusion weight according to the uncertainty level, prediction confidence and historical performance coefficient of each detection result, realizing adaptive fusion and improving the reliability and robustness of the fusion result.

[0113] In an optional embodiment,

[0114] The hierarchical anomaly detection engine extracts the spatiotemporal features of environmental parameters through a multi-head attention mechanism and residual connections, iteratively calculates the warning threshold based on the spatiotemporal features of the environmental parameters through a Bayesian optimization algorithm, and dynamically adjusts the warning level judgment standard according to the historical warning accuracy. The steps of generating warning information including the warning level, warning area, and warning parameters include:

[0115] The environmental parameter sequence is feature extracted through a multi-head attention mechanism. The product of the query matrix, key matrix and value matrix is calculated for each attention head of the environmental parameter sequence. The output of each attention head is residually connected with the original environmental parameter sequence to obtain the spatiotemporal features of the environmental parameters.

[0116] An adaptive sampling strategy is constructed based on the spatiotemporal characteristics of the environmental parameters. The sampling density function is dynamically adjusted based on the fluctuation change rate of the environmental parameters. The spatiotemporal characteristics of the environmental parameters are dynamically sampled to obtain a sampling data set. A Gaussian process regression model with multiple kernel functions of different scales is constructed using the sampling data set. The weight coefficients of the kernel functions are adaptively adjusted according to the characteristics of the environmental parameters. A particle swarm is initialized based on the prediction results of the Gaussian process regression model. A particle swarm velocity term with a dynamic exploration factor is introduced into a Bayesian optimization process. The warning threshold is iteratively optimized using a multi-objective optimization function that includes data coverage, parameter fluctuation amplitude, time series correlation, and spatial clustering degree.

[0117] A deep Q network is constructed to dynamically adjust the warning level. The historical warning accuracy sequence, the spatiotemporal characteristics of the environmental parameters, and the current warning threshold set are used as state space inputs. The threshold adjustment amount, the merging and splitting operations of the warning level, and the time window size adjustment are used as action space outputs. A comprehensive reward function is constructed based on the warning time delay rate, false alarm rate, missed alarm rate, and warning level change frequency. A dual network architecture consisting of a target network and an evaluation network is used to optimize the warning level determination strategy.

[0118] Multimodal warning information is generated based on the warning threshold and the output value of the deep Q network, wherein the multimodal warning information includes a warning level, a warning area, warning parameters and a confidence assessment result, wherein the confidence assessment result is calculated based on the ratio of the output value of the deep Q network to the maximum Q value.

[0119] For example, first, obtain an environmental parameter sequence. Assume that the obtained environmental parameter sequence includes multiple environmental parameters such as temperature, humidity, and wind speed. Each parameter is recorded in a time series format, for example, data is collected every hour. Taking temperature as an example, the temperature data sequence for a certain area over a week is [25, 26, 27, 28, 29, 30, 28].

[0120] Next, the spatiotemporal features of the environmental parameters are extracted. A multi-head attention mechanism is used to extract features from the acquired environmental parameter sequence. The environmental parameter sequence is fed into multiple attention heads, each focusing on a different aspect of the sequence. Each attention head calculates the product of the query matrix, key matrix, and value matrix to generate the output of that attention head. The outputs of all attention heads are concatenated and residual-connected with the original environmental parameter sequence to obtain environmental parameter features that contain spatiotemporal information. For example, after processing the temperature sequence through the multi-head attention mechanism, features such as temperature trend and periodicity are obtained.

[0121] Next, the warning threshold is iteratively calculated based on the spatiotemporal characteristics of the environmental parameters. An adaptive sampling strategy is constructed based on these spatiotemporal characteristics. For example, when environmental parameters fluctuate significantly, the sampling density is increased; when fluctuations are minor, the sampling density is reduced. Dynamic sampling of the spatiotemporal characteristics of the environmental parameters is performed to generate a sampled dataset. A Gaussian process regression model is constructed using the sampled dataset. This model uses multiple Gaussian kernel functions of different scales, with the kernel weight coefficients adaptively adjusted based on the characteristics of the environmental parameters. For example, when environmental parameters fluctuate dramatically, the model assigns greater weight to the short-term kernel function. The prediction results of the Gaussian process regression model are used to initialize the particle swarm algorithm. A particle swarm velocity term with a dynamic exploration factor is introduced into the Bayesian optimization process. Using data coverage, parameter fluctuation amplitude, time series correlation, and spatial clustering as multi-objective optimization functions, the warning threshold is iteratively optimized to obtain the optimal warning threshold. For example, the optimization algorithm determines the temperature warning threshold to be 32 degrees Celsius.

[0122] Subsequently, the warning level is dynamically adjusted. A deep Q-network is constructed, taking the historical warning accuracy sequence, the spatiotemporal characteristics of environmental parameters, and the current warning threshold set as state-space inputs. The threshold adjustment, the merging and splitting of warning levels, and the adjustment of the time window size are output as action-space outputs. A comprehensive reward function is constructed based on the warning delay rate, false alarm rate, missed alarm rate, and the frequency of warning level changes. A dual-network architecture, consisting of a target network and an evaluation network, is used to optimize the warning level determination strategy. For example, if the temperature exceeds 32 degrees Celsius for three consecutive days, the warning level is raised from low to medium.

[0123] Finally, a multimodal warning message is generated. Based on the warning threshold and the output of the deep Q-network, a multimodal warning message is generated, including the warning level, warning area, warning parameters, and confidence assessment results. The confidence assessment result is calculated as the ratio of the deep Q-network output value to the maximum Q value. For example, a warning message might be generated: "The temperature in area A exceeds the warning threshold of 32 degrees Celsius, the warning level is medium, the warning parameter is temperature, and the confidence level is 90%."

[0124] The present invention extracts the spatiotemporal characteristics of environmental parameters through a multi-head attention mechanism, and combines the Bayesian optimization algorithm and deep Q network to dynamically adjust the warning threshold and warning level, which can more accurately identify abnormal situations, reduce false alarm and missed alarm rates, and improve warning accuracy; the adaptive sampling strategy and the mechanism of adaptive adjustment of kernel function weights enable the model to be dynamically adjusted according to the characteristics of environmental parameters, enhancing the adaptability of the warning method to different environmental conditions; the generated multimodal warning information includes warning level, warning area, warning parameters and confidence assessment results, providing more comprehensive information to help users better understand and respond to abnormal situations.

[0125] In an optional embodiment,

[0126] The steps of converting the warning information into a multi-dimensional stereoscopic chart through a three-dimensional visualization module, and constructing a knowledge reasoning engine based on a graph neural network and a deep reinforcement learning algorithm to generate an adaptive emergency response strategy include:

[0127] The warning information is constructed into a three-layer rendering structure. The bottom geographic information layer uses quadtree grid division for detailed hierarchical management. The middle warning information layer uses adaptive grid subdivision to spatially discretize environmental parameters and constructs the parameter distribution field through bilinear interpolation. The upper dynamic annotation layer uses frustum clipping and annotation avoidance algorithms to overlay information. Combined with a color band mapping scheme based on perceptual uniformity and a local variance adaptive transparency adjustment mechanism, 3D visualization data of the warning information is generated.

[0128] Based on the three-dimensional visualization data, a warning knowledge graph is constructed, with warning events, environmental parameters, spatial locations, and emergency measures as node types. Spatial proximity relationships are calculated using R-tree indexes, and semantic similarity is calculated using word vectors to obtain event associations. The knowledge graph is modeled using a heterogeneous graph neural network, which includes a type feature conversion layer to process different node features, a message passing layer combined with an attention mechanism to adjust edge weights, and a feature aggregation layer and a position encoding module with jump connections to form a knowledge graph representation vector.

[0129] The knowledge graph representation vector and the environmental parameter time series features in the three-dimensional visualization data are spliced together to construct a state space, and a hierarchical architecture of a deep reinforcement learning algorithm is used for decision-making, wherein the meta-strategy layer determines the sub-goal sequence based on long-term reward prediction, and the execution strategy layer generates specific emergency instructions based on short-term rewards; a composite reward function is constructed based on the node association, edge weight and spatial position information in the early warning knowledge graph, and the composite reward function includes parameter improvement, impact area range, resource scheduling cost, response delay and knowledge consistency weight terms. An experience replay module based on the attention mechanism is introduced to store the historical data of the composite reward function, and the relevance of the action sequence is analyzed through a recursive neural network, and an emergency processing strategy is output based on an asynchronous advantage strategy network.

[0130] For example, a 3D visualization of warning information is first constructed. Warning information is organized into a three-level rendering structure: the bottom layer is the geographic information layer, which uses quadtree gridding to manage the detailed level of geographic information. For example, for urban areas, the grid can be divided down to the street level, while for suburban areas, it can be divided into larger areas. The middle layer is the warning information layer, which uses adaptive grid subdivision to spatially discretize environmental parameters such as temperature and pollutant concentration, and uses bilinear interpolation to construct the parameter distribution field. For example, if the temperatures at the four corners of a region are 25°C, 28°C, 26°C, and 27°C, bilinear interpolation can be used to calculate the temperature value at any point in the region. The top layer is the dynamic annotation layer, which uses frustum clipping and annotation avoidance algorithms to overlay information and prevent occlusion. A color band mapping scheme based on perceptual uniformity and a local variance-adaptive transparency adjustment mechanism are combined. For example, in areas with large variance, transparency is increased to display more details, thus generating 3D visualization data for warning information.

[0131] Next, a warning knowledge graph is constructed based on the 3D visualization data. Warning events, environmental parameters, spatial locations, and emergency measures are used as node types. For example, "fire" is an event node, "temperature" is a parameter node, "city center" is a location node, and "evacuate" is a measure node. Spatial proximity is calculated using an R-tree index. For example, two events are considered spatially proximal if their locations are less than 1 kilometer apart. Event associations are calculated using word embeddings to measure semantic similarity. For example, "fire" and "thick smoke" have high semantic similarity. The knowledge graph is modeled using a heterogeneous graph neural network. This network includes a type feature conversion layer to process different node features. For example, event nodes can be characterized by event type and occurrence time, while parameter nodes can be characterized by parameter value and measurement time. A message passing layer with an attention mechanism adjusts edge weights. For example, if two events are highly temporally correlated, the edge weight between them will be higher. A feature aggregation layer with skip connections and a position encoding module form a knowledge graph representation vector, which contains information about all nodes and edges in the graph.

[0132] Next, emergency strategies are generated. The state space is constructed by concatenating the knowledge graph representation vector with the temporal features of environmental parameters from the 3D visualization data. For example, parameters such as current temperature and wind speed are concatenated with the knowledge graph representation vector. A hierarchical architecture based on a deep reinforcement learning algorithm is employed for decision-making. The meta-strategy layer determines the sub-goal sequence based on long-term reward predictions, for example, controlling the fire first, then evacuating the crowd. The execution strategy layer generates specific emergency instructions based on short-term rewards, for example, dispatching a fire truck to a designated location. A composite reward function is constructed based on node associations, edge weights, and spatial location information in the early warning knowledge graph. This function includes parameters such as the improvement degree (e.g., the degree of temperature reduction), the impact area (e.g., the size of the area affected by the fire), the resource scheduling cost (e.g., the cost of dispatching a fire truck), the response latency (e.g., the time from alarm to action), and a knowledge consistency weight (e.g., whether the action taken is consistent with existing knowledge in the knowledge graph). An experience replay module based on an attention mechanism is introduced to store historical data for the composite reward function, for example, recording past successful emergency strategies and their corresponding rewards. Recurrent neural networks analyze the relevance of action sequences, for example, analyzing the effectiveness of controlling a fire before evacuating people. Asynchronous advantage strategy networks output emergency response strategies, such as specific instructions like how many fire trucks to dispatch to which locations.

[0133] Through three-dimensional visualization technology, the present invention transforms the originally abstract warning information into intuitive and easy-to-understand three-dimensional charts, making it easier for decision makers to quickly understand and grasp the spatiotemporal distribution characteristics of the warning information and improve situational awareness. Based on graph neural networks and deep reinforcement learning algorithms, a knowledge reasoning engine is constructed, which can automatically learn the correlation between warning events and the optimal strategies for emergency measures, generate adaptive emergency response plans, and improve emergency response efficiency. Through the optimization of the composite reward function, it can effectively reduce resource scheduling costs, shorten response delays, and ensure the consistency of emergency strategies with existing knowledge, thereby improving the overall effectiveness of emergency response.

[0134] Figure 2 FIG. 1 is a schematic diagram of the structure of a thermal power plant environment monitoring and early warning system based on the Internet of Things according to an embodiment of the present invention. Figure 2 As shown, the system includes:

[0135] The first unit is used to collect environmental monitoring data through environmental monitoring sensors installed in the thermal power plant. The environmental monitoring data is distributedly preprocessed by a grid edge computing network constructed by multiple edge computing nodes. The distributed preprocessing includes data layer compression and dynamic scheduling of task priorities. Each edge computing node calculates task allocation based on a resource-aware load balancing strategy. The preprocessed data is transmitted to the data processing center via the Internet of Things gateway to generate a preprocessed data stream;

[0136] The second unit is used to process the pre-processed data stream in real time through a distributed message queue, build a hierarchical distributed data storage structure based on data timeliness, acquisition frequency and data relevance, and use a master-multiple slave chain replication mechanism to perform multiple copies of data backup. The pre-processed data stream is desensitized based on homomorphic encryption for retrievable data, an asymmetric encryption algorithm is used to generate encrypted environmental monitoring data, and a multi-level index structure is established according to spatiotemporal characteristics to form an encrypted data set.

[0137] The third unit is used to construct a hierarchical anomaly detection engine based on the encrypted data set, using a long short-term memory network, a random forest algorithm, and an extreme gradient boosting tree. The hierarchical anomaly detection engine extracts the spatiotemporal characteristics of environmental parameters through a multi-head attention mechanism and residual connections, iteratively calculates the warning threshold based on the spatiotemporal characteristics of the environmental parameters through a Bayesian optimization algorithm, dynamically adjusts the warning level judgment standard according to the historical warning accuracy, and generates warning information including warning level, warning area, and warning parameters; converts the warning information into a multidimensional stereoscopic chart through a three-dimensional visualization module, and constructs a knowledge reasoning engine based on a graph neural network and a deep reinforcement learning algorithm to generate an adaptive emergency response strategy.

[0138] According to a third aspect of the embodiments of the present invention,

[0139] An electronic device is provided, comprising:

[0140] Processor;

[0141] a memory for storing processor-executable instructions;

[0142] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0143] According to a fourth aspect of the embodiments of the present invention,

[0144] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0145] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0146] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. The environmental monitoring and early warning method for thermal power plants based on the Internet of Things is characterized by: include: Environmental monitoring data is collected by environmental monitoring sensors installed in the thermal power plant. The environmental monitoring data is distributed pre-processed by multiple edge computing nodes to form a grid edge computing network. The distributed pre-processing includes data layer compression and dynamic scheduling of task priorities. Each edge computing node calculates task allocation based on a resource-aware load balancing strategy. The preprocessed data is transmitted to the data processing center through the IoT gateway to generate a preprocessed data stream; The pre-processed data stream is processed in real time through a distributed message queue, and a hierarchical distributed data storage structure based on data timeliness, acquisition frequency, and data relevance is constructed. The data storage structure adopts a master-multiple slave chain replication mechanism to perform multiple copies of data backup. The pre-processed data stream is desensitized based on homomorphic encryption for retrievable data, an asymmetric encryption algorithm is used to generate encrypted environmental monitoring data, and a multi-level index structure is established according to spatiotemporal characteristics to form an encrypted data set; Based on the encrypted data set, a hierarchical anomaly detection engine is constructed using a long short-term memory network, a random forest algorithm, and an extreme gradient boosting tree. The hierarchical anomaly detection engine extracts the spatiotemporal characteristics of environmental parameters through a multi-head attention mechanism and residual connections, iteratively calculates the warning threshold based on the spatiotemporal characteristics of the environmental parameters through a Bayesian optimization algorithm, dynamically adjusts the warning level judgment standard according to the historical warning accuracy, and generates warning information including the warning level, warning area, and warning parameters; the warning information is converted into a multidimensional stereoscopic chart through a three-dimensional visualization module, and a knowledge reasoning engine is constructed based on a graph neural network and a deep reinforcement learning algorithm to generate an adaptive emergency response strategy.

2. The method according to claim 1, characterized in that Environmental monitoring data is collected by environmental monitoring sensors installed in the thermal power plant. The environmental monitoring data is distributed preprocessed by multiple edge computing nodes to form a grid edge computing network. The distributed preprocessing includes data layer compression and dynamic scheduling of task priorities. The steps of calculating task allocation based on resource-aware load balancing strategies of each edge computing node include: Collect environmental monitoring data through dust concentration sensors, gas sensors, noise sensors, and temperature and humidity sensors installed in the thermal power plant. The environmental monitoring data includes device identification, timestamp, geographic location information, and measurement values; The environmental monitoring data is input into multiple edge computing nodes according to the grid units divided by the physical space of the thermal power plant. The edge computing nodes build a grid edge computing network through Ethernet switches and point-to-point communication protocols. The distributed state synchronization protocol is used between the edge computing nodes for heartbeat maintenance; Each edge computing node performs distributed preprocessing on the environmental monitoring data, performs wavelet transform and coefficient screening on the continuously changing environmental monitoring data to form first preprocessed data, performs differential encoding on the slowly changing environmental monitoring data to form second preprocessed data, and merges the first preprocessed data and the second preprocessed data to generate a preprocessed data packet, wherein the preprocessed data packet includes a compression type identifier and original data length information; A load balancing strategy is used to allocate tasks. By obtaining the processor utilization, memory occupancy, and network bandwidth utilization of each edge computing node in real time, the pre-processed data packets are graded according to the urgency of the tasks based on real-time alarms, routine monitoring, and statistical analysis. The task priority is determined based on the data volume of the pre-processed data packets. The task allocation weight is dynamically adjusted according to the resource occupancy of each edge computing node, and the pre-processed data packets are allocated to the corresponding edge computing nodes. The edge computing nodes execute tasks in a time-slice rotation manner.

3. The method according to claim 1, characterized in that The pre-processed data stream is processed in real time through a distributed message queue, and a hierarchical distributed data storage structure based on data timeliness, acquisition frequency, and data relevance is constructed. The data storage structure uses a master-multiple slave chain replication mechanism to perform multiple copies of data backup. The pre-processed data stream is subjected to retrievable data desensitization processing based on homomorphic encryption, and an asymmetric encryption algorithm is used to generate encrypted environmental monitoring data. The steps include: Processing the pre-processed data stream through a distributed message queue, wherein the distributed message queue uses an asynchronous batch method to send messages containing a pre-processed data stream identifier, a timestamp, data content, and a checksum, and each partition in the distributed message queue is provided with multiple copies and a chain replication mechanism is used to generate a message queue data stream; The message queue data stream is divided into a hot data layer, a warm data layer and a cold data layer according to timeliness, the hot data layer uses an in-memory database to store data generated within two hours, the warm data layer uses a solid-state hard disk to store data from two hours to forty-eight hours, and the cold data layer uses a mechanical hard disk to store data exceeding forty-eight hours; the data is sharded and stored according to the data acquisition frequency, and the high-frequency acquisition data with a time interval of less than or equal to sixty seconds is sharded according to a ten-minute time interval, the medium-frequency acquisition data with a time interval of one minute to sixty minutes is sharded according to a thirty-minute time interval, and the low-frequency acquisition data with a time interval of more than sixty minutes is sharded according to a sixty-minute time interval; a multi-level index structure of data association degree is established, including a primary index, a secondary index and an associated index, wherein the primary index includes a timestamp and a device identifier, the secondary index includes a geographic location and a data type, and the associated index includes a device group and a monitoring indicator, so as to form hierarchical index data; Based on the formation of hierarchical index data, a one-master-multiple-slave chain replication mechanism is used to perform data backup on the pre-processed data stream, wherein the master node receives a data write request and synchronizes the data to each slave node in a preset chain sequence, each slave node copies the data level by level and returns confirmation information, and the master node completes the data write after receiving confirmation information from all slave nodes, generating backup confirmation data; According to the backup confirmation data, the hierarchically stored numerical data is encrypted using a homomorphic encryption algorithm, which supports addition operations in the encrypted state; the hierarchically stored non-numerical data is encrypted using an asymmetric encryption algorithm, whose private key is dispersedly stored in the authorized nodes to generate encrypted environmental monitoring data.

4. The method according to claim 1, wherein Based on the encrypted data set, the steps of constructing a hierarchical anomaly detection engine using a long short-term memory network, a random forest algorithm, and an extreme gradient boosting tree include: Performing multi-scale feature extraction on the encrypted data set, constructing a multi-dimensional feature representation by combining time domain features and frequency domain features through multi-scale decomposition based on wavelet transform, and generating multi-scale feature data; Based on the multi-scale feature data, an attention-enhanced encrypted domain long short-term memory network model is constructed. The encrypted domain long short-term memory network model includes a time attention mechanism and a feature attention mechanism. The time attention mechanism calculates the importance weight of the time domain features, and the feature attention mechanism performs correlation analysis on the frequency domain features. An encrypted domain attention score matrix is constructed based on the obtained time domain weights and frequency domain correlations. The multi-scale feature data is weighted reconstructed using the attention score matrix. Sequence modeling is performed based on the reconstructed feature data. An adversarial training module for sequence prediction error is constructed, and a time series anomaly detection result is generated based on the prediction error. The multi-scale feature data is optimized using a random forest algorithm in a ciphertext space, a decision tree splitting criterion is constructed based on homomorphic encryption, and an initial decision tree set is generated through approximate calculation. The information gain value of the encrypted feature is calculated for each tree in the initial decision tree set, and a feature importance evaluation index is determined based on the information gain value, and the feature sampling probability is dynamically adjusted; the initial decision tree is reconstructed according to the adjusted sampling probability, the node redundancy is calculated based on the reconstructed decision tree set, and the nodes whose redundancy exceeds a preset threshold are deleted to complete the pruning operation; and the optimized random forest algorithm is used to generate spatial anomaly detection results.

5. The method according to claim 4, characterized in that The method further comprises: In the encrypted domain, a second-order expansion approximation is performed on the homomorphically encrypted data through Taylor expansion, a ciphertext domain loss function is constructed based on the approximation result, the first-order and second-order gradient information of the ciphertext domain loss function with respect to the encrypted feature is calculated, the first-order and second-order gradient information is used to construct a splitting gain evaluation criterion for the encrypted feature, the optimal splitting feature and splitting point are selected based on the splitting gain evaluation criterion, entropy value calculation is performed on the encrypted feature in the multi-scale feature data, the feature binning boundary is determined according to the entropy value, the feature binning boundary result is combined with the splitting gain evaluation result to generate an extreme gradient boosting tree, and a pattern anomaly detection result is output by iteratively accumulating the prediction results of the extreme gradient boosting tree; The information entropy of the time series anomaly detection results, spatial anomaly detection results and pattern anomaly detection results is calculated to obtain the uncertainty level of each detection result. The mutual information between any two detection results is calculated based on the probability distribution of each detection result. Each detection result is randomly perturbed by Monte Carlo sampling in the encrypted domain to obtain the prediction confidence. The adaptive fusion weight of each detection result is calculated according to the uncertainty level, the prediction confidence and the historical performance coefficient of each detection result. The historical performance coefficient is dynamically updated through a sliding time window. The adaptive fusion weight is used to weight each anomaly detection result to obtain an initial fusion result. The initial fusion result is multi-layered with the mutual information and the historical anomaly score based on the time series attenuation factor. The historical anomaly score is attenuated and accumulated considering the time series correlation to generate the final anomaly detection result.

6. The method according to claim 1, characterized in that The hierarchical anomaly detection engine extracts the spatiotemporal features of environmental parameters through a multi-head attention mechanism and residual connections, iteratively calculates the warning threshold based on the spatiotemporal features of the environmental parameters through a Bayesian optimization algorithm, and dynamically adjusts the warning level judgment standard according to the historical warning accuracy. The steps of generating warning information including the warning level, warning area, and warning parameters include: The environmental parameter sequence is feature extracted through a multi-head attention mechanism. The product of the query matrix, key matrix and value matrix is calculated for each attention head of the environmental parameter sequence. The output of each attention head is residually connected with the original environmental parameter sequence to obtain the spatiotemporal features of the environmental parameters. An adaptive sampling strategy is constructed based on the spatiotemporal characteristics of the environmental parameters. The sampling density function is dynamically adjusted based on the fluctuation change rate of the environmental parameters. The spatiotemporal characteristics of the environmental parameters are dynamically sampled to obtain a sampling data set. A Gaussian process regression model with multiple kernel functions of different scales is constructed using the sampling data set. The weight coefficients of the kernel functions are adaptively adjusted according to the characteristics of the environmental parameters. A particle swarm is initialized based on the prediction results of the Gaussian process regression model. A particle swarm velocity term with a dynamic exploration factor is introduced into a Bayesian optimization process. The warning threshold is iteratively optimized using a multi-objective optimization function that includes data coverage, parameter fluctuation amplitude, time series correlation, and spatial clustering degree. A deep Q network is constructed to dynamically adjust the warning level. The historical warning accuracy sequence, the spatiotemporal characteristics of the environmental parameters, and the current warning threshold set are used as state space inputs. The threshold adjustment amount, the merging and splitting operations of the warning level, and the time window size adjustment are used as action space outputs. A comprehensive reward function is constructed based on the warning time delay rate, false alarm rate, missed alarm rate, and warning level change frequency. A dual network architecture consisting of a target network and an evaluation network is used to optimize the warning level determination strategy. Multimodal warning information is generated based on the warning threshold and the output value of the deep Q network, wherein the multimodal warning information includes a warning level, a warning area, warning parameters and a confidence assessment result, wherein the confidence assessment result is calculated based on the ratio of the output value of the deep Q network to the maximum Q value.

7. The method according to claim 1, characterized in that The steps of converting the warning information into a multi-dimensional stereoscopic chart through a three-dimensional visualization module, and constructing a knowledge reasoning engine based on a graph neural network and a deep reinforcement learning algorithm to generate an adaptive emergency response strategy include: The warning information is constructed into a three-layer rendering structure. The bottom geographic information layer uses quadtree grid division for detailed hierarchical management. The middle warning information layer uses adaptive grid subdivision to spatially discretize environmental parameters and constructs the parameter distribution field through bilinear interpolation. The upper dynamic annotation layer uses frustum clipping and annotation avoidance algorithms to overlay information. Combined with a color band mapping scheme based on perceptual uniformity and a local variance adaptive transparency adjustment mechanism, 3D visualization data of the warning information is generated. Based on the three-dimensional visualization data, a warning knowledge graph is constructed, with warning events, environmental parameters, spatial locations, and emergency measures as node types. Spatial proximity relationships are calculated using R-tree indexes, and semantic similarity is calculated using word vectors to obtain event associations. The knowledge graph is modeled using a heterogeneous graph neural network, which includes a type feature conversion layer to process different node features, a message passing layer combined with an attention mechanism to adjust edge weights, and a feature aggregation layer and a position encoding module with jump connections to form a knowledge graph representation vector. The knowledge graph representation vector and the environmental parameter time series features in the three-dimensional visualization data are spliced together to construct a state space, and a hierarchical architecture of a deep reinforcement learning algorithm is used for decision-making, wherein the meta-strategy layer determines the sub-goal sequence based on long-term reward prediction, and the execution strategy layer generates specific emergency instructions based on short-term rewards; a composite reward function is constructed based on the node association, edge weight and spatial position information in the early warning knowledge graph, and the composite reward function includes parameter improvement, impact area range, resource scheduling cost, response delay and knowledge consistency weight terms. An experience replay module based on the attention mechanism is introduced to store the historical data of the composite reward function, and the relevance of the action sequence is analyzed through a recursive neural network, and an emergency processing strategy is output based on an asynchronous advantage strategy network.

8. An Internet of Things-based thermal power plant environmental monitoring and early warning system, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to collect environmental monitoring data through environmental monitoring sensors installed in the thermal power plant. The environmental monitoring data is distributed pre-processed by a grid edge computing network constructed by multiple edge computing nodes. The distributed pre-processing includes data layer compression and dynamic scheduling of task priorities. Each edge computing node calculates task allocation based on a resource-aware load balancing strategy; The preprocessed data is transmitted to the data processing center through the IoT gateway to generate a preprocessed data stream; The second unit is used to process the pre-processed data stream in real time through a distributed message queue, build a hierarchical distributed data storage structure based on data timeliness, acquisition frequency and data relevance, and use a master-multiple slave chain replication mechanism to perform multiple copies of data backup. The pre-processed data stream is desensitized based on homomorphic encryption for retrievable data, an asymmetric encryption algorithm is used to generate encrypted environmental monitoring data, and a multi-level index structure is established according to spatiotemporal characteristics to form an encrypted data set. The third unit is used to construct a hierarchical anomaly detection engine based on the encrypted data set, using a long short-term memory network, a random forest algorithm, and an extreme gradient boosting tree. The hierarchical anomaly detection engine extracts the spatiotemporal characteristics of environmental parameters through a multi-head attention mechanism and residual connections, iteratively calculates the warning threshold based on the spatiotemporal characteristics of the environmental parameters through a Bayesian optimization algorithm, dynamically adjusts the warning level judgment standard according to the historical warning accuracy, and generates warning information including warning level, warning area, and warning parameters; converts the warning information into a multidimensional stereoscopic chart through a three-dimensional visualization module, and constructs a knowledge reasoning engine based on a graph neural network and a deep reinforcement learning algorithm to generate an adaptive emergency response strategy.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images