Station environment safety regulation and control method based on multi-module cooperation

Abstract

The invention relates to a station environment safety regulation and control method based on multi-module cooperation. The method comprises the steps that a multi-source sensor collects environmental data and preprocesses the environmental data to obtain input data; segmenting the recurrent neural network into a first neural network and a second neural network; inputting the input data into a first neural network to obtain intermediate data, encrypting the intermediate data, and sending the encrypted intermediate data to a back-end server; and the back-end server decrypts the encrypted intermediate data and inputs the decrypted intermediate data into the second neural network to obtain an intelligent regulation and control result. The segmentation of the recurrent neural network is based on identification and replacement processing of a closed-loop structure which does not contain an input node and an output node in a network structure chart, and it is ensured that the second neural network does not contain a coupled closed loop, so that the consistency and accuracy of regulation and control results are improved. The method gives consideration to both data security and model reasoning stability, and is suitable for distributed monitoring and intelligent management and control of a station environment.

Description

Technical Field

[0001] The present invention relates to the field of substation environment regulation, and particularly to a substation environment safety regulation method based on multi-module collaboration. Background Art

[0002] In the operation and management of modern intelligent substations, a large number of sensors are deployed in various areas of the substation to monitor key environmental parameters such as temperature and humidity, smoke concentration, gas leakage, and partial electrical discharge. To achieve centralized analysis and unified regulation, these sensors usually transmit the collected data to the backend server in real time through a wireless network.

[0003] However, in the existing technical architecture, the direct upload of raw data from sensors to the server will cause a series of problems. First, due to the limited computing resources of sensors, the data encryption mechanism used is often weak, resulting in security risks such as data being eavesdropped, tampered with, and forged during transmission. Second, the large amount of raw data and high sampling frequency are likely to cause an increase in wireless transmission bandwidth pressure, increase system latency, and weaken the real-time performance and response ability of the system. Finally, the unified in-depth analysis of all sensor data on the server side will also cause problems such as concentrated computing resources, heavy server load, and high energy consumption.

[0004] To address the above problems, some technical solutions attempt to introduce a recurrent neural network (RNN) structure, which divides the neural network into two stages: the front-stage network is deployed at the edge side to perform preliminary processing on the raw sensor data and extract intermediate features; the intermediate features are encrypted and sent to the backend server; the backend server then decrypts them and generates the final intelligent regulation result through the back-stage network. This "segmented" RNN processing mode can effectively improve the security of data transmission, reduce the transmission volume, and enhance the overall response efficiency of the system. However, it should be noted that as a network structure with strong temporal dependence and complex computational paths, there are a large number of cyclic connections between the nodes of RNN. If the coupled closed-loop relationship in the network structure is not reasonably analyzed and processed during the segmentation process and it is directly split into two segments, it may interrupt the internal dependence path, causing the back-stage network to perform calculations without obtaining the complete dependent input, resulting in distorted or incorrect output results. Summary of the Invention

[0005] The present disclosure provides a substation environment safety regulation method based on multi-module collaboration.

[0006] In a first aspect, the present disclosure provides a substation environment safety regulation method based on multi-module collaboration, including: Multi-source sensors collect environmental data, and obtain input data after preprocessing the environmental data; Divide the recurrent neural network into a first neural network and a second neural network; Input the input data into the first neural network to obtain intermediate data, encrypt the intermediate data, and send the encrypted intermediate data to the backend server; Decrypt the encrypted intermediate data, and input the decrypted intermediate data into the second neural network to obtain an intelligent regulation result; Wherein, the second neural network includes preset parameters of a recurrent neural network, and the encrypted intermediate data is in a lightweight encryption method.

[0007] Optionally, the lightweight encryption method is the AES algorithm.

[0008] Optionally, the input data is an input data vector, including multi-dimensional time series features of each multi-source sensor within the current time window.

[0009] Optionally, splitting the recurrent neural network into the first neural network and the second neural network further includes: Obtain the structure diagram of the recurrent neural network; Find the closed loop in the structure diagram that does not include input nodes and output nodes; Replace the closed loop with an integration node until no closed loop that does not include input nodes and output nodes can be found in the structure diagram; Split the recurrent neural network into the first neural network and the second neural network, wherein the second neural network does not contain integration nodes.

[0010] Optionally, finding the closed loop in the structure diagram that does not include input nodes and output nodes further includes: Starting from any non-input and non-output node, traverse all paths along the neural network without passing through input and output nodes; The path that can return to the starting node is a closed loop.

[0011] Optionally, after decrypting the encrypted intermediate data, inputting the decrypted intermediate data into the second neural network to obtain an intelligent regulation result, it further includes: Generate an environment regulation instruction according to the intelligent regulation result; Send the environment regulation instruction to the execution module; Receive the feedback data returned by each execution module, and use the feedback data to optimize the first neural network or the second neural network.

[0012] Optionally, the preset parameters of the recurrent neural network include: weights of nodes, biases, edge connection parameters between networks, and activation functions used.

[0013] Optionally, determine the order interval of the directed graph; based on the order interval, use the external nodes as the nodes of the second neural network.

[0014] Optionally, the order interval of the unidirectional graph is determined based on depth-first search, where the first value of the order interval is the order information when the node becomes gray, and the second value of the order interval is the order information when the node becomes black.

[0015] Optionally, the recurrent neural network is a long short-term memory network or a gated recurrent unit.

[0016] The beneficial effects of the present disclosure are as follows. Compared with the prior art, the present disclosure has the following advantages: 1) By deploying the first neural network on the sensor side and encrypting and transmitting the intermediate data output by it to the backend server, the present invention avoids the security risks brought by the transmission of raw sensor data on the network, effectively improving the confidentiality and tamper resistance of data transmission. At the same time, by dividing the recurrent neural network into two sub-networks before and after, part of the computing tasks are completed at the edge, reducing the load on the backend server and improving the response efficiency and real-time performance of the overall system, which is applicable to large-scale and multi-point distributed station environment application scenarios.

[0017] 2) The present invention proposes a method for identifying and processing closed loops in the neural network structure diagram. By replacing the closed loops that do not contain input nodes and output nodes with integrated nodes and ensuring that the second neural network does not contain such closed loops, the problem of disordered inference logic after model segmentation is avoided, ensuring the consistency of the results before and after segmented execution.

[0018] 3) The present invention supports using feedback data for online optimization of the neural network, enabling the first neural network and the second neural network to continuously adjust parameters according to the actual performance during operation, thereby enhancing the adaptive ability and regulation accuracy of the system.

[0019] 4) The present invention creatively introduces the order interval to maximize the determination of the second neural network, ensuring the maximization of the security performance of encryption. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The accompanying drawings herein are incorporated into the specification and form a part of the specification, showing embodiments consistent with the present disclosure and, together with the specification, are used to explain the principles of the present disclosure.

[0021] Figure 1 A schematic diagram of station environment safety regulation based on multi-module collaboration provided by an embodiment of the present disclosure; Figure 2 A schematic diagram of station environment safety regulation based on multi-module collaboration provided by an embodiment of the present disclosure; Figure 3A 、 3B A set of schematic diagrams of incorrect segmentation of recurrent neural networks provided by an embodiment of the present disclosure; Figure 4 Schematic diagram of correct segmentation of a recurrent neural network provided by an embodiment of the present disclosure; Figure 5 Schematic diagram of segmentation of a first neural network and a second neural network based on an order interval provided by an embodiment of the present disclosure.

[0022] Through the above-mentioned drawings, specific embodiments of the present disclosure have been shown, and more detailed descriptions will be given in the following text. These drawings and text descriptions are not intended to limit the scope of the concept of the present disclosure in any way, but to illustrate the concept of the present disclosure to those skilled in the art by referring to specific embodiments. Detailed implementation manners

[0023] The present disclosure will be further described below with reference to the drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present disclosure and cannot be used to limit the protection scope of the present disclosure.

[0024] Figure 1 Schematic diagram of a station environment safety regulation method based on multi-module collaboration provided by an embodiment of the present disclosure. Refer to Figure 1 , and specific discussions will be carried out for each step in combination with this embodiment.

[0025] S100. Multisource sensors collect environmental data, and input data is obtained after preprocessing the environmental data.

[0026] In this embodiment, the multisource sensors in step S100 include a temperature and humidity sensor, a smoke sensing module, a gas concentration detection module, a partial discharge monitoring module, and a passive high-voltage node temperature sensor, which are respectively deployed in different functional areas within the station.

[0027] Each sensor collects environmental data at a fixed period, and the sampling period is set to 5 seconds. Specifically: The temperature and humidity sensor collects the environmental temperature and humidity every 5 seconds. For example, the collected values at the first time point are a temperature of 24.7 °C and a relative humidity of 63.2%; the smoke sensing module returns a unit concentration value. For example, the value at the first time point is 12 ppm; The gas detection module returns the volume fraction percentage. For example, the value at the first time point is 99.84%; the insulation terminal temperature measured by the passive high-voltage node temperature sensor is 58.1 °C; the partial discharge monitoring module returns a characteristic discharge amplitude value, such as 37 dBμV at the corresponding time point.

[0028] After the above-mentioned raw data is collected, it is uniformly preprocessed by the preprocessing module. Specifically, it includes timestamp alignment, unit conversion (such as unifying the temperature to Celsius and the concentration to ppm), sliding window filtering (window size is 3), and normalization processing to map the data of each channel to the interval [0, 1].

[0029] For example: The original temperature data sequence is [24.7, 24.6, 24.9], and after normalization, it is transformed into [0.4, 0.2, 0.8]; The concentration sequence is [99.84, 99.79, 99.81], and after normalization, it is transformed into [0.75, 0.25, 0.5].

[0030] All the channel data after preprocessing are combined into an input data vector, which is used as the input data input to the first neural network in the subsequent step S200. This input data vector contains the multi-dimensional time series features of each sensor in the recent several time windows, has strong expression ability and structural consistency, and is suitable for processing by the deep neural network model.

[0031] S200. Split the recurrent neural network into a first neural network and a second neural network.

[0032] Specifically, the recurrent neural network can be a long short-term memory network (LSTM) or a gated recurrent unit (GRU).

[0033] In this embodiment, the recurrent neural network is used to process multi-sensor environment data. The network is composed of multiple neurons, and there may be time series connections, feedback connections or recursive connections between each neuron, forming a complex recurrent network. To achieve the functional separation between the front end and the back end, it is necessary to divide this recurrent neural network into two independent sub-networks, namely the first neural network and the second neural network.

[0034] First, obtain the structure information of the neural network, in the form of a structure diagram composed of nodes and connection relationships. Nodes represent neurons, and connections represent information flow paths. According to the preset rules, mark the input nodes and output nodes among them, corresponding to the sensor data input and the final intelligent control output of the substation respectively.

[0035] Subsequently, through the structure traversal method, identify all paths that start from non-input nodes and can return to the starting point without passing through input or output nodes. Such paths are defined as "closed loops composed only of hidden layers". For example, in the structure traversal, several circular paths such as "node A → node B → node C → node A" are identified, which are typical closed loops.

[0036] After recognition is completed, each closed-loop area is replaced by a logically "integrated node". The replacement process is as follows: all the connections that originally entered the closed loop are uniformly connected to this integrated node; all the connections that originally output from the inside of the closed loop to the outside are uniformly output by this integrated node. This replacement operation ensures that the input and output interfaces of the closed loop to the outside remain consistent, and at the same time eliminates the coupling and dependency relationships inside the structure.

[0037] Finally, the recurrent neural network is divided into two parts, where the second neural network does not contain an integrated node.

[0038] S300. Input the input data into the first neural network to obtain intermediate data, encrypt the intermediate data, and send the encrypted intermediate data to the backend server.

[0039] Understandably, the encrypted intermediate data is a lightweight encryption method. Optionally, the lightweight encryption method is the AES algorithm.

[0040] In this embodiment, the input data in step S300 is the multi-dimensional time series vector after preprocessing in step S100, which contains the normalized feature data of multiple sensors at multiple consecutive time points. For example, this input data consists of six channels: temperature, humidity, concentration, smoke concentration, high-voltage node temperature, and partial discharge intensity. Each channel contains sampling values of 5 time steps, and finally forms a 6×5 input matrix.

[0041] This input data is sent to the first neural network deployed in the edge device for processing. The first neural network is a segment of LSTM network structure, which contains 2 layers of LSTM units in total, with 32 neurons in each layer, and outputs a one-dimensional vector with a length of 64 as the intermediate data.

[0042] For example, in a certain round of inference, the input is the following matrix: [0.41, 0.39, 0.43, 0.45, 0.42],# Temperature channel [0.63, 0.65, 0.66, 0.64, 0.62],# Humidity channel [0.95, 0.94, 0.93, 0.94, 0.96],# Concentration channel [0.12, 0.10, 0.11, 0.13, 0.12],# Smoke channel [0.48, 0.47, 0.46, 0.49, 0.50],# High-voltage temperature channel [0.37, 0.36, 0.35, 0.38, 0.39],# Partial discharge channel After inputting this matrix, the first neural network outputs an intermediate data vector, for example: [0.052, 0.137, 0.228, ..., 0.099], a one-dimensional floating-point vector with a length of 64. To ensure the confidentiality of data during transmission, this intermediate data is encrypted using the AES-128 symmetric encryption algorithm. The key is pre-negotiated between the edge device and the backend server and updated regularly. The encrypted ciphertext data is sent to the backend server through lightweight security protocols such as MQTT or HTTPS. This process generally completes within 500 ms, ensuring that the regulation response latency meets the real-time requirements of substation-level monitoring.

[0043] This encrypted transmission mechanism for intermediate data significantly reduces the risk of raw data transmission over the public network on the one hand, and also reduces the overall data volume on the other hand, making network transmission more efficient and secure.

[0044] S400. Decrypt the encrypted intermediate data and input the decrypted intermediate data into the second neural network to obtain an intelligent regulation result.

[0045] Among them, the second neural network includes preset parameters of a recurrent neural network. The preset parameters of the recurrent neural network include: the weights of the nodes, biases, edge parameters between networks, and activation functions used.

[0046] In this embodiment, the decryption operation described in step S400 is performed by the backend server. After receiving the encrypted intermediate data transmitted from the edge device, the backend server performs AES-128 decryption using the preset symmetric key to restore the original intermediate data vector.

[0047] The decrypted intermediate data is a one-dimensional vector with a length of 64. This intermediate data is input into the second neural network deployed on the backend server to complete the final intelligent regulation decision. In a certain actual execution scenario, this 64-dimensional vector is mapped to a classification probability vector with a length of 3, corresponding to three system states: "normal operation", "warning state", and "immediate regulation": [normal operation: 0.12, warning state: 0.18, immediate regulation: 0.70]. Based on this, the system generates a regulation result of "immediate regulation" and triggers subsequent execution strategies, such as starting the dehumidification module, closing doors and windows, and emitting sound and light alarms.

[0048] Further, a substation environment safety regulation method based on multi-module collaboration disclosed in this embodiment, where the recurrent neural network is segmented into a first neural network and a second neural network, further includes: obtaining a structure diagram of the recurrent neural network; finding closed loops in the structure diagram that do not include input nodes and output nodes; replacing the closed loops with integrated nodes until no closed loops that do not include input nodes and output nodes can be found in the structure diagram; and segmenting the recurrent neural network into a first neural network and a second neural network, where the second neural network does not contain integrated nodes.

[0049] First, obtain a structure diagram of the recurrent neural network to be partitioned. This structure diagram is used to describe the connection relationships between all neuron nodes in the network, including the directed connections between input nodes, output nodes, and intermediate hidden nodes.

[0050] Subsequently, identify all closed-loop structures in the structure diagram that do not include input nodes and output nodes, that is, feedback loops composed only of hidden layer nodes. These closed loops introduce time dependencies and coupling logics during the execution of the network. If directly cut without processing, it will cause the mismatch of the inference logics of the upstream and downstream sub-networks.

[0051] Therefore, each closed-loop structure is logically replaced with an integrated node, that is, a closed loop is replaced with a functionally equivalent node, and all connections that originally entered the closed loop are uniformly redirected to this integrated node, and all connections that were originally output by the closed loop are uniformly sent out by this integrated node.

[0052] Repeat the above replacement operation until no closed loops that do not include input and output nodes exist in the structure diagram, and finally form a structure diagram without coupled closed loops.

[0053] First, take Figure 3A as an example. Assume that the operator op is "AND operation", and at the same time assume that x2 is 0. Then, regardless of whether x1 and x3 are 0 or 1, it will not affect the result of y = 0. This means that when x2 = 1, it must wait for the data of x1 or x3 to arrive before performing the op operation; while when x2 = 0, it can directly perform the op operation unilaterally without waiting for x1 or x3. Therefore, in actual operations, timing is a meaningful issue.

[0054] Take Figure 3BFor example, if a recurrent neural network is divided into a first recurrent neural network and a second recurrent neural network through u0: then the computational order information of node B and node C can only be defaulted to be the same: that is, at the first iteration of node 32, it is affected by the input node B. However, when this recurrent neural network is regarded as a whole (before splitting), the computational order information of node B and node C may be inconsistent. For example: in the iteration of C->31->32, it may be until several iterations later (because there are multiple nodes before node B, and these nodes have a timing of operations, and the timing of B is after this), that node 32 is affected by node B. Obviously, in the case of splitting the recurrent neural network and regarding the recurrent neural network as a whole, the timing of B and C acting on node 32 has changed, and further this will lead to inconsistent results calculated by the two methods. That is, this splitting method is incorrect.

[0055] To solve this problem, taking Figure 4 as an example, the loops can be eliminated one by one. G0 first eliminates the loop formed by 4->5->6->7->8, takes the integrated node 13 as a replacement, and inherits the connection relationship of the directed edges of this loop with other external nodes to form G1. Immediately afterwards, G1 eliminates the loop formed by 13->9->10, takes the integrated node 14 as a replacement, and inherits the connection relationship of the directed edges of this loop with other external nodes to form G2. Finally, on the basis of G2, all integrated nodes are used as the left side of the split, so that there are no integrated nodes in the second neural network on the right side, and the above problem does not exist.

[0056] Furthermore, for a station environment safety regulation method based on multi-module collaboration disclosed in this embodiment, finding the closed loop in the structure diagram that does not include input nodes and output nodes further includes: starting from any non-input and non-output node, without passing through input and output nodes, traversing all paths along the neural network; the path that can return to the starting node is a closed loop.

[0057] Starting from any hidden layer node that is neither an input node nor an output node as the starting point, without passing through any input node or output node, traverse all possible paths forward along the directed connection relationship of the neural network. If in a certain path, the traversal finally returns to the starting node, it is determined that this path forms a closed loop.

[0058] Taking a specific recurrent neural network as an example, assume that its structure diagram contains several nodes, where nodes 1 to 4 are input nodes, nodes 35 to 40 are output nodes, and the remaining nodes are hidden layer nodes. In this step, first select any one of the hidden layer nodes that are neither input nor output as the starting point, for example, select node 10. Then, starting from node 10, traverse the path forward in sequence along the directed connections in the neural network, and it is specified that the input nodes (i.e., nodes 1 to 4) or output nodes (i.e., nodes 35 to 40) are not allowed to be passed through during the traversal process. During the path search process, if there is a complete path in the form of: node 10 → node 13 → node 17 → node 10, it means that this path forms a closed loop. After finding any closed loop, the closed loop can be replaced with an integrated node. Repeat the process of finding closed loops until no closed loop can be found.

[0059] Further, for a station environment safety regulation method based on multi-module collaboration disclosed in this embodiment, after decrypting the encrypted intermediate data and inputting the decrypted intermediate data into the second neural network to obtain the intelligent regulation result, it further includes: generating an environment regulation instruction according to the intelligent regulation result; sending the environment regulation instruction to the execution module; receiving the feedback data returned by each execution module, and using the feedback data to optimize the first neural network or the second neural network.

[0060] Further, in this embodiment, after "decrypting the encrypted intermediate data and inputting the decrypted intermediate data into the second neural network to obtain the intelligent regulation result", the following operations are further included: First, the system generates a corresponding environment regulation instruction according to the output result of the second neural network. For example, when the output result is a status label of "humidity exceeding the limit + temperature being too low", the system automatically generates the following combined instruction: close the intelligent doors and windows, turn on the far-infrared heating and dehumidification module, and set the fan speed to the low-speed ventilation mode. The regulation instruction is encapsulated in a standard protocol and sent to each execution module through a wired or wireless control channel.

[0061] Subsequently, the execution module starts the corresponding action process according to the instruction content, and after the action is completed, the feedback data is sent back to the system. For example, the far-infrared heating module sends back the actual operating power curve, the fan module sends back the current rotation speed value, and the doors and windows module sends back the physical opening and closing state.

[0062] The system integrates and analyzes the feedback data returned by each execution module and compares it with the regulation target. If there are obvious deviations, such as "the humidity decrease amplitude after execution is insufficient" or "the fan speed response is lagging", the system will automatically trigger the model update mechanism.

[0063] The update mechanism can selectively optimize the parameter weights in the first neural network or the second neural network. For example, if the deviation mainly appears in the response difference between prediction and execution, the output layer weights of the second neural network are adjusted backward by the feedback error; if the deviation comes from the distortion of sensor data in the front-end extraction stage, the input mapping structure of the first neural network is optimized.

[0064] The order interval [tb, te] should in principle be described as the timing intervals of the respective nodes in the unidirectional graph: taking Figure 4 G2 as an example, we hope that the tb of node 14 must be strictly greater than the te of nodes 1 and 2; and the te of node 14 must be strictly less than the tb of node 12. Since there is no timing relationship between node 11 and node 14, its timing interval usually needs to be set to be coincident, that is, the tb of node 11 must be strictly less than the te of node 14; and the tb of node 14 must be strictly less than the te of node 11. However, this determination method is actually quite difficult. We choose the order interval to replace the timing interval, and the difference is that the definitions of te and tb have changed.

[0065] Figure 5 is a schematic diagram of dividing the first neural network and the second neural network based on the order interval. Among them, each pair of parentheses represents a node, the left parenthesis represents the first value of the order interval, the right parenthesis represents the second value of the order interval, and they are arranged in ascending order. The square brackets represent the integrated nodes, which are represented by capital letters (for example: E, G, I); the small parentheses represent the non-integrated nodes, which are represented by lowercase letters (for example: a, b, c, d, f, h, j, k, l).

[0066] Among them, the external nodes are defined as non-integrated nodes without integrated nodes on the periphery. In Figure 5 only a, b, c, d are external nodes. It can be understood that since f is surrounded by E and h is surrounded by G, f and h do not belong to the external nodes.

[0067] It can be understood that the unidirectional graph is formed by replacing the closed loop with an integrated node until no closed loop that does not contain input nodes and output nodes can be found in the structure diagram. As long as the second neural network is determined, then how to divide the first neural network and the second neural network is determined.

[0068] In summary, it can be generalized as: first determine the order interval of the unidirectional graph; then, based on the order interval, use the external nodes as the nodes of the second neural network; correspondingly, use the non-external nodes (including the integrated nodes) as the nodes of the first neural network. The order interval of the unidirectional graph is determined based on depth-first search. Among them, the first value tb of the order interval is the order information when the node becomes gray, and the second value te of the order interval is the order information when the node becomes black.

[0069] It is understandable that white, gray, and black are the markings of the nodes after the 0th visit, the 1st visit, and the second visit of the nodes in the depth-first search.

[0070] Figure 2 FIG. 5 is a schematic diagram of a substation environmental safety regulation and control system based on multi-module collaboration provided by an embodiment of the present disclosure.

[0071] A multi-source sensor is used to collect environmental data by the multi-source sensor, and a preprocessing module is used to obtain input data after preprocessing the environmental data.

[0072] In this embodiment, the multi-source sensor includes a temperature and humidity sensor, a smoke sensing module, a gas concentration detection module, a partial discharge monitoring module, and a passive high-voltage node temperature sensor, which are respectively deployed in different functional areas in the substation.

[0073] Each sensor collects environmental data at a fixed period, and the sampling period is set to 5 seconds. Specifically: The temperature and humidity sensor collects the environmental temperature and humidity every 5 seconds. For example, the collected value at the first time point is a temperature of 24.7 °C and a relative humidity of 63.2%; the smoke sensing module returns a unit concentration value. For example, the value at the first time point is 12 ppm; The gas detection module returns the volume fraction percentage. For example, the value at the first time point is 99.84%; the insulation terminal temperature measured by the passive high-voltage node temperature sensor is 58.1 °C; the partial discharge monitoring module returns a characteristic discharge amplitude value, such as 37 dBμV at the corresponding time point.

[0074] After the above raw data is collected, it is uniformly preprocessed by the preprocessing module. Specifically, it includes timestamp alignment, unit conversion (such as unifying the temperature to degrees Celsius and the concentration to ppm), sliding window filtering (window size is 3), and normalization processing to map the data of each channel to the [0, 1] interval.

[0075] For example: the original temperature data sequence is [24.7, 24.6, 24.9], and after normalization, it is transformed into [0.4, 0.2, 0.8]; the concentration sequence is [99.84, 99.79, 99.81], and after normalization, it is transformed into [0.75, 0.25, 0.5].

[0076] All the channel data after preprocessing are combined into an input data vector, which is used as the input data input to the first neural network in the subsequent step S200. This input data vector contains the multi-dimensional time series features of each multi-source sensor within the current time window, has strong expression ability and structural consistency, and is suitable for processing by the deep neural network model.

[0077] A neural network management module for splitting a recurrent neural network into a first neural network and a second neural network.

[0078] In this embodiment, the recurrent neural network is used to process multi-sensor environmental data. The network consists of multiple neurons, and there may be temporal connections, feedback connections, or recursive connections between each neuron, forming a complex recurrent network. To achieve the functional separation between the front end and the back end, the recurrent neural network needs to be divided into two independent sub-networks, namely the first neural network and the second neural network.

[0079] First, obtain the structural information of the neural network, in the form of a structure diagram composed of nodes and connection relationships. Nodes represent neurons, and connections represent information flow paths. According to preset rules, mark the input nodes and output nodes among them, corresponding to the sensor data input and the final intelligent control output of the substation respectively.

[0080] Subsequently, identify all paths that start from non-input nodes and can return to the starting point without passing through input or output nodes through a structure traversal method. Such paths are defined as "closed loops composed only of hidden layers". For example, several circular paths such as "node A → node B → node C → node A" are identified during the structure traversal, which are typical closed loops.

[0081] After the identification is completed, each closed-loop area is replaced with a logically "integrated node". The replacement process is as follows: all connections that originally entered the closed loop are uniformly connected to this integrated node; all connections that originally output from the inside of the closed loop to the outside are uniformly output by this integrated node. This replacement operation ensures that the input-output interfaces of the closed loop to the outside remain consistent, and at the same time eliminates the coupling dependency relationships inside the structure.

[0082] Finally, the recurrent neural network is split into two parts, where the second neural network does not contain integrated nodes.

[0083] The preprocessing module is further configured to input the input data into the first neural network, obtain intermediate data, encrypt the intermediate data, and send the encrypted intermediate data to the back-end server.

[0084] In this embodiment, for example, the input data consists of six channels: temperature, humidity, concentration, smoke concentration, high-voltage node temperature, and partial discharge intensity. Each channel contains sampling values of 5 time steps, and finally forms a 6×5 input matrix.

[0085] The input data is sent to the first neural network deployed in the edge device for processing. The first neural network is a segment of LSTM network structure, which contains 2 layers of LSTM units, with 32 neurons in each layer, and outputs a one-dimensional vector with a length of 64 as the intermediate data.

[0086] For example, in a certain round of inference, the input is the following matrix: [0.41, 0.39, 0.43, 0.45, 0.42], # Temperature channel [0.63, 0.65, 0.66, 0.64, 0.62], # Humidity channel [0.95, 0.94, 0.93, 0.94, 0.96], # Concentration channel [0.12, 0.10, 0.11, 0.13, 0.12], # Smoke channel [0.48, 0.47, 0.46, 0.49, 0.50], # High-voltage temperature channel [0.37, 0.36, 0.35, 0.38, 0.39], # Partial discharge channel After inputting this matrix, the first neural network outputs an intermediate data vector, for example: [0.052, 0.137, 0.228,..., 0.099], a one-dimensional floating-point vector with a length of 64. To ensure the confidentiality of the data during transmission, this intermediate data is encrypted by the AES-128 symmetric encryption algorithm. The key is pre-negotiated between the edge device and the backend server and updated regularly. The encrypted ciphertext data is sent to the backend server through lightweight security protocols such as MQTT or HTTPS. This process is generally completed within 500 ms to ensure that the regulation response delay meets the real-time requirements of substation-level monitoring.

[0087] This intermediate data encryption and transmission mechanism, on the one hand, greatly reduces the risk of raw data transmission in the public network, and on the other hand, also reduces the overall data volume, making the network transmission more efficient and secure.

[0088] The backend server is used to decrypt the encrypted intermediate data and input the decrypted intermediate data into the second neural network to obtain the intelligent regulation result.

[0089] In this embodiment, the decryption operation described in step S400 is executed by the backend server. After receiving the encrypted intermediate data transmitted from the edge device, the backend server performs AES-128 decryption with the preset symmetric key to restore the original intermediate data vector.

[0090] The decrypted intermediate data is a one-dimensional vector with a length of 64. This intermediate data is input into the second neural network deployed on the backend server to complete the final intelligent regulation decision. In a certain actual execution scenario, this 64-dimensional vector is mapped to a classification probability vector with a length of 3, corresponding to three system states: "normal operation", "warning state", and "immediate regulation": [normal operation: 0.12, warning state: 0.18, immediate regulation: 0.70]. Based on this, the system generates a regulation result of "immediate regulation" and triggers subsequent execution strategies, such as starting the dehumidification module, closing the doors and windows, and emitting a sound and light alarm.

[0091] Furthermore, for a substation environment safety regulation system based on multi-module collaboration disclosed in this embodiment, where the recurrent neural network is divided into a first neural network and a second neural network, it further includes: obtaining the structure diagram of the recurrent neural network; finding the closed loops in the structure diagram that do not contain input nodes and output nodes; replacing the closed loops with integration nodes until no closed loops that do not contain input nodes and output nodes can be found in the structure diagram; dividing the recurrent neural network into a first neural network and a second neural network, where the second neural network does not contain integration nodes.

[0092] First, obtain the structure diagram of the recurrent neural network to be divided. This structure diagram is used to describe the connection relationships between all neuron nodes in the network, including the directed connections between input nodes, output nodes, and intermediate hidden nodes.

[0093] Subsequently, identify all closed-loop structures in the structure diagram that do not contain input nodes and output nodes, that is, feedback loops composed only of hidden layer nodes. These closed loops introduce time dependencies and coupling logics during the network execution process. If directly cut without processing, it will lead to a mismatch in the inference logics of the upstream and downstream sub-networks.

[0094] To this end, each closed-loop structure is logically replaced with an integration node, that is, a functionally equivalent node is used to replace the closed loop, and all the connections that originally entered the closed loop are uniformly redirected to this integration node, and the connections that were originally output by the closed loop are uniformly emitted by this integration node.

[0095] Repeat the above replacement operation until there are no closed loops that do not contain input and output nodes in the structure diagram, and finally form a structure diagram without coupled closed loops.

[0096] Furthermore, for a substation environment safety regulation system based on multi-module collaboration disclosed in this embodiment, when finding the closed loops in the structure diagram that do not contain input nodes and output nodes, it further includes: starting from any non-input and non-output node, traversing all paths along the neural network without passing through input and output nodes; a path that can return to the starting node is a closed loop.

[0097] Starting from any hidden layer node that is neither an input node nor an output node as the starting point, without passing through any input node or output node, traverse all possible paths forward along the directed connection relationship of the neural network. If in a certain path, the traversal finally returns to the starting node, it is determined that this path forms a closed loop.

[0098] Taking a specific recurrent neural network as an example, assume that its structure diagram contains several nodes, where nodes 1 to 4 are input nodes, nodes 35 to 40 are output nodes, and the remaining nodes are hidden layer nodes. In this step, first select any one of the hidden layer nodes that are neither input nor output as the starting point, for example, select node 10. Then starting from node 10, traverse the path forward in sequence along the directed connection in the neural network, and it is specified that the traversal process shall not pass through input nodes (i.e., nodes 1 to 4) or output nodes (i.e., nodes 35 to 40). During the path search process, if there is a complete path in the form of: node 10 → node 13 → node 17 → node 10, it indicates that this path forms a closed loop. After finding any closed loop, the closed loop can be replaced with an integrated node. Repeat finding closed loops until no closed loop can be found.

[0099] Furthermore, for a substation environment safety regulation and control system based on multi-module collaboration disclosed in this embodiment, after decrypting the encrypted intermediate data and inputting the decrypted intermediate data into the second neural network to obtain the intelligent regulation and control result, it further includes: generating an environment regulation instruction according to the intelligent regulation and control result; sending the environment regulation instruction to the execution module; receiving the feedback data returned by each execution module, and using the feedback data to optimize the first neural network or the second neural network.

[0100] Furthermore, in this embodiment, the following operations are further included after "decrypting the encrypted intermediate data and inputting the decrypted intermediate data into the second neural network to obtain the intelligent regulation and control result": First, the system generates a corresponding environment regulation instruction according to the output result of the second neural network. For example, when the output result is the status label of "humidity exceeding the limit + temperature being low", the system automatically generates the following combined instruction: close the intelligent doors and windows, turn on the far-infrared heating and dehumidification module, and set the fan speed regulation to the low-speed ventilation mode. The regulation instruction is encapsulated by a standard protocol and sent to each execution module through a wired or wireless control channel.

[0101] Subsequently, the execution module starts the corresponding action process according to the instruction content, and after the action is completed, the feedback data is transmitted back to the system. For example, the far-infrared heating module transmits the actual operating power curve, the fan module transmits the current rotation speed value, and the doors and windows module transmits the physical opening and closing state.

[0102] The system integrates and analyzes the feedback data returned by each execution module, and compares it with the regulation target. If there are obvious deviations, such as "the humidity drop after execution is insufficient" or "the response of the fan speed is lagging", etc., the system will automatically trigger the model update mechanism.

[0103] This update mechanism can selectively optimize the parameter weights in the first neural network or the second neural network. For example, if the deviation mainly appears in the response difference between prediction and execution, the output layer weights of the second neural network are adjusted backward through the feedback error; if the deviation comes from the distortion of sensor data in the front-end extraction stage, the input mapping structure of the first neural network is optimized.

[0104] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device, which may include: a processor, a communications interface, a memory, and a communication bus. Among them, the processor, the communications interface, and the memory communicate with each other through the communication bus. The processor can call the logical instructions in the memory to execute the implementation method based on the soft authorization of the configuration software.

[0105] In addition, when the logical instructions in the above-mentioned memory are implemented in the form of software functional units and sold or used as an independent product, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present disclosure, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present disclosure. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs that can store program codes.

[0106] On the other hand, the present disclosure also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it is implemented to execute the implementation method based on the soft authorization of the configuration software provided by the above-mentioned various methods. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without creative effort.

[0107] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the above technical solution, in essence, or the part that contributes to the prior art can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0108] It should be understood that the above embodiments are only used to illustrate the technical solutions of the present disclosure, rather than to limit them; although the present disclosure has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure.

Claims

1. A station environment safety regulation method based on multi-module collaboration, comprising: Collecting environmental data by multi-source sensors, and obtaining input data after preprocessing the environmental data; Dividing a recurrent neural network into a first neural network and a second neural network; Inputting the input data into the first neural network to obtain intermediate data, encrypting the intermediate data, and sending the encrypted intermediate data to a backend server; Decrypting the encrypted intermediate data, and inputting the decrypted intermediate data into the second neural network to obtain an intelligent regulation result; Wherein, the second neural network includes preset parameters of the recurrent neural network; the encrypted intermediate data is a lightweight encryption method.

2. The method for regulating the safety of the substation environment based on multi-module collaboration according to claim 1, wherein The lightweight encryption method is the AES algorithm.

3. The method for regulating the safety of the substation environment based on multi-module collaboration according to claim 1, wherein The input data is an input data vector, including multi-dimensional time series features of each multi-source sensor within the current time window.

4. A method for regulating the safety of a substation environment based on multi-module collaboration according to claim 1, characterized in that The dividing the recurrent neural network into a first neural network and a second neural network further includes: Obtaining the structure diagram of the recurrent neural network; Finding a closed loop in the structure diagram that does not include input nodes and output nodes; Replacing the closed loop with an integrated node until no closed loop that does not include input nodes and output nodes can be found in the structure diagram; Dividing the recurrent neural network into a first neural network and a second neural network, wherein the second neural network does not contain integrated nodes.

5. The method for regulating and controlling the safety of the substation environment based on multi-module collaboration according to claim 4, wherein The finding a closed loop in the structure diagram that does not include input nodes and output nodes further includes: Starting from any non-input and non-output node, traversing all paths along the neural network without passing through input and output nodes; A path that can return to the starting node is a closed loop.

6. The station environment safety regulation method based on multi-module collaboration according to claim 4, wherein After the decrypting the encrypted intermediate data, and inputting the decrypted intermediate data into the second neural network to obtain an intelligent regulation result, it further includes: Generating an environmental regulation instruction according to the intelligent regulation result; Sending the environmental regulation instruction to an execution module; Receiving feedback data returned by each execution module, and using the feedback data to optimize the first neural network or the second neural network.

7. A method for regulating the safety of the substation environment based on multi-module collaboration according to claim 1, characterized in that The preset parameters of the recurrent neural network include: weights of nodes, biases, edge parameters between networks, and activation functions used.

8. The method for regulating the safety of the substation environment based on multi-module collaboration according to claim 4, wherein, Determining the order interval of a directed graph; based on the order interval, taking external nodes as nodes of the second neural network.

9. The method for regulating the safety of the substation environment based on multi-module collaboration according to claim 8, characterized in that, The order interval of the directed graph is determined based on depth-first search, wherein the first value of the order interval is the order information when a node becomes gray, and the second value of the order interval is the order information when a node becomes black.

10. A method for regulating the safety of a substation environment based on multi-module collaboration according to claim 1, characterized in that, The recurrent neural network is a long short-term memory network or a gated recurrent unit.

Images