Station environment security control method based on multi-module collaboration

Abstract

The present invention relates to a station environment safety control method based on multi-module collaboration. The method includes: multi-source sensors collect environmental data and pre-process it to obtain input data; split 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, and encrypt the intermediate data and send it to the back-end server; the back-end server decrypts the encrypted intermediate data, and inputs the decrypted intermediate data into the second neural network to obtain an intelligent control result. The segmentation of the recurrent neural network is based on the identification and replacement of the closed-loop structure that does not contain input nodes and output nodes in the network structure diagram, ensuring that the second neural network does not contain a coupled closed loop, thereby improving the consistency and accuracy of the control results. The present invention takes into account both data security and the stability of model reasoning, and is suitable for distributed monitoring and intelligent control of station environments.

Description

Technical Field

[0001] The present invention relates to the field of station environment control, and in particular to a station environment safety control method based on multi-module collaboration. Background Art

[0002] In the operation and management of modern smart power stations, numerous sensors are deployed throughout the facility to monitor key environmental parameters such as temperature and humidity, smoke concentration, gas leaks, and electrical partial discharges. To enable centralized analysis and unified control, these sensors typically transmit the collected data in real time to backend servers via wireless networks.

[0003] However, in existing technical architectures, sensors directly upload raw data to servers, which leads to a series of problems. First, due to the limited computing resources of sensors, the data encryption mechanisms used are often weak, resulting in security risks of data being monitored, tampered with, and counterfeited during transmission. Second, the large amount of raw data and high sampling frequency can easily increase pressure on wireless transmission bandwidth, increase system latency, and weaken the system's real-time performance and responsiveness. Finally, the centralized and in-depth analysis of all sensor data on the server also leads to the concentration of computing resources, heavy server load, and high energy consumption.

[0004] To address these issues, some technical solutions have attempted to introduce a recurrent neural network (RNN) structure, dividing the neural network into two stages: the front-end network, deployed at the edge, performs preliminary processing on raw sensor data and extracts intermediate features. The intermediate features are encrypted and sent to the back-end server, which decrypts them and generates the final intelligent control results through the back-end network. This "segmented" RNN processing model can effectively improve data transmission security, reduce transmission volume, and enhance overall system response efficiency. However, it is important to note that RNNs, as a network structure with strong temporal dependencies and complex computational paths, have numerous recurrent connections between nodes. If the network is directly split into two stages without properly analyzing and addressing the coupled closed-loop relationships within the structure, this can disrupt internal dependency paths, causing the back-end network to perform calculations without receiving the complete set of dependent inputs, resulting in distorted or incorrect output results. Summary of the Invention

[0005] The present disclosure provides a site environment security control method based on multi-module collaboration.

[0006] In a first aspect, the present disclosure provides a site environment security control method based on multi-module collaboration, comprising:

[0007] Multi-source sensors collect environmental data and pre-process the environmental data to obtain input data;

[0008] Split the recurrent neural network into a first neural network and a second neural network;

[0009] 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 backend server;

[0010] decrypting the encrypted intermediate data, and inputting the decrypted intermediate data into a second neural network to obtain an intelligent control result;

[0011] The second neural network includes preset parameters of a recurrent neural network, and the encrypted intermediate data is a lightweight encryption method.

[0012] The optional lightweight encryption method is the AES algorithm.

[0013] Optionally, the input data is an input data vector, which includes multi-dimensional time series features of each multi-source sensor in the current time window.

[0014] Optionally, the dividing the recurrent neural network into a first neural network and a second neural network further includes:

[0015] Obtaining a structural diagram of the recurrent neural network;

[0016] Finding a closed loop in the structure graph that does not contain an input node and an output node;

[0017] Replacing the closed loop with an integration node until no closed loop containing no input node and no output node can be found in the structure graph;

[0018] The recurrent neural network is divided into a first neural network and a second neural network, wherein the second neural network does not contain an integrated node.

[0019] Optionally, the step of finding a closed loop in the structure graph that does not include an input node and an output node further includes:

[0020] Starting from any non-input and non-output node, traverse all paths along the neural network without passing through any input or output nodes;

[0021] The path that returns to the starting node is a closed loop.

[0022] Optionally, after decrypting the encrypted intermediate data and inputting the decrypted intermediate data into the second neural network to obtain the intelligent control result, the method further includes:

[0023] generating an environmental control instruction according to the intelligent control result;

[0024] Sending the environmental control instruction to the execution module;

[0025] Receive feedback data returned by each execution module, and use the feedback data to optimize the first neural network or the second neural network.

[0026] Optionally, the preset parameters of the recurrent neural network include: node weights, biases, edge parameters between networks, and the activation function used.

[0027] Optionally, an order interval of the unidirectional graph is determined; based on the order interval, the external nodes are used as nodes of the second neural network.

[0028] Optionally, the order interval of the unidirectional graph is determined based on a depth-first search, wherein the first value of the order interval is the order information of the node turning gray, and the second value of the order interval is the order information of the node turning black.

[0029] Optionally, the recurrent neural network is a long short-term memory network or a gated recurrent unit.

[0030] The beneficial effects of the present disclosure are that, compared with the prior art, the present disclosure has the following advantages:

[0031] 1) By deploying a first neural network on the sensor side and encrypting the intermediate data it outputs to the backend server, this method avoids the security risks associated with transmitting raw sensor data over the network, effectively improving the confidentiality and tamper resistance of data transmission. Furthermore, by dividing the recurrent neural network into two sub-networks, some computational tasks are completed at the edge, reducing the load on the backend server and improving the overall system's responsiveness and real-time performance. This makes it suitable for large-scale, multi-point distributed station environments.

[0032] 2) The present invention proposes a method for identifying and processing closed loops in a neural network structure diagram. By replacing closed loops that do not contain input nodes and output nodes with integrated nodes and ensuring that such closed loops are not contained in the second neural network, the problem of disordered reasoning logic after model segmentation is avoided, and the consistency of the results before and after segmented execution is ensured.

[0033] 3) The present invention supports the use of feedback data for online optimization of neural networks, enabling the first and second neural networks to continuously adjust parameters based on their actual performance during operation, thereby enhancing the system's adaptability and control accuracy.

[0034] 4) The present invention creatively introduces the order interval to determine the second neural network to the maximum extent possible, thereby ensuring the maximum security performance of encryption. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0036] Figure 1 A schematic diagram of a station environment security control system based on multi-module collaboration provided by an embodiment of the present disclosure;

[0037] Figure 2 A schematic diagram of a station environment security control system based on multi-module collaboration provided by an embodiment of the present disclosure;

[0038] Figure 3A 、 3B A schematic diagram of a set of recurrent neural network error segmentation provided by an embodiment of the present disclosure;

[0039] Figure 4 A schematic diagram of correct segmentation of a recurrent neural network provided by an embodiment of the present disclosure;

[0040] Figure 5 A schematic diagram of sequential interval-based segmentation of a first neural network and a second neural network according to an embodiment of the present disclosure.

[0041] The above drawings illustrate specific embodiments of the present disclosure, which will be described in more detail below. These drawings and textual descriptions are not intended to limit the scope of the present disclosure in any way, but rather to illustrate the concepts of the present disclosure to those skilled in the art by reference to specific embodiments. DETAILED DESCRIPTION

[0042] The present disclosure is further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present disclosure and are not intended to limit the scope of protection of the present disclosure.

[0043] Figure 1 A schematic diagram of a station environment security control method based on multi-module collaboration provided by an embodiment of the present disclosure. Figure 1 , each step is now discussed in detail in conjunction with this embodiment.

[0044] S100 , multi-source sensors collect environmental data, and pre-process the environmental data to obtain input data.

[0045] In this embodiment, the multi-source sensor in step S100 includes a temperature and humidity sensor, a smoke sensing module, Gas concentration detection modules, partial discharge monitoring modules and passive high-voltage node temperature sensors are deployed in different functional areas within the station.

[0046] Each sensor collects environmental data at a fixed period, and the sampling period is set to 5 seconds. Specifically:

[0047] The temperature and humidity sensor collects ambient temperature and humidity every 5 seconds. For example, the first collected value is 24.7°C and the relative humidity is 63.2%. The smoke sensor module returns a unit concentration value. For example, the first value 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.

[0048] After the raw data is collected, it is pre-processed uniformly by the pre-processing module. This includes timestamp alignment, unit conversion (e.g., converting temperature to degrees Celsius and concentration to ppm), sliding window filtering (with a window size of 3), and normalization to map each channel data to the [0, 1] range.

[0049] For example, the original temperature data sequence is [24.7, 24.6, 24.9], which is transformed into [0.4, 0.2, 0.8] after normalization. The concentration sequence is [99.84, 99.79, 99.81], which is transformed into [0.75, 0.25, 0.5] after normalization.

[0050] After preprocessing, all channel data is combined into an input data vector, which serves as the input data to the first neural network in the subsequent step S200. This input data vector contains multidimensional time series features from each sensor within several recent time windows. It has strong expressive power and structural consistency, making it suitable for processing in deep neural network models.

[0051] S200: Split the recurrent neural network into a first neural network and a second neural network.

[0052] Specifically, the recurrent neural network can be a long short-term memory network (LSTM) or a gated recurrent unit (GRU).

[0053] In this embodiment, a recurrent neural network is used to process multi-sensor environmental data. The network is composed of multiple neurons, each of which may have sequential connections, feedback connections, or recursive connections, forming a complex recurrent network. To achieve functional separation between the front-end and back-end, the recurrent neural network is divided into two independent sub-networks: the first neural network and the second neural network.

[0054] First, the neural network's structural information is obtained, in the form of a diagram consisting of nodes and connections. Nodes represent neurons, and connections represent information flow paths. According to pre-set rules, the input and output nodes are labeled, corresponding to the sensor data input and the final intelligent control output of the station, respectively.

[0055] Then, using a structured traversal method, we identify all paths that start from non-input nodes and return to the starting point without passing through any input or output nodes. These paths are defined as "closed loops consisting solely of hidden layers." For example, if we identify several circular paths like "Node A → Node B → Node C → Node A" during structured traversal, these are typical closed loops.

[0056] After identification, each closed-loop region is replaced with a logical "integration node." The replacement process involves: all connections that originally entered the closed loop are now connected to the integration node; all connections that originally exited the closed loop are now output from the integration node. This replacement ensures that the closed loop's external input and output interfaces remain consistent while eliminating internal coupling dependencies.

[0057] Finally, the recurrent neural network is split into two parts, where the second neural network does not contain integration nodes.

[0058] S300: Input the input data into a first neural network to obtain intermediate data, encrypt the intermediate data, and send the encrypted intermediate data to a backend server.

[0059] It is understandable that the encrypted intermediate data is a lightweight encryption method. Optionally, the lightweight encryption method is an AES algorithm.

[0060] In this embodiment, the input data in step S300 is a multi-dimensional time series vector after pre-processing in step S100, which contains normalized feature data of multiple sensors at multiple consecutive time points. For example, the input data consists of temperature, humidity, The input matrix consists of six channels: 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.

[0061] The input data is fed into the first neural network deployed on the edge device for processing. This first neural network is an LSTM network structure consisting of two layers of LSTM units, each with 32 neurons, and outputs a one-dimensional vector of length 64 as intermediate data.

[0062] For example, in a certain round of reasoning, the input is the following matrix:

[0063] [0.41, 0.39, 0.43, 0.45, 0.42],# Temperature channel

[0064] [0.63, 0.65, 0.66, 0.64, 0.62], # humidity channel

[0065] [0.95, 0.94, 0.93, 0.94, 0.96],# Concentration channel

[0066] [0.12, 0.10, 0.11, 0.13, 0.12],# Smoke channel

[0067] [0.48, 0.47, 0.46, 0.49, 0.50],# High pressure temperature channel

[0068] [0.37, 0.36, 0.35, 0.38, 0.39],# Partial discharge channel

[0069] After inputting this matrix, the first neural network outputs an intermediate data vector, for example, a one-dimensional floating-point vector of length 64: [0.052, 0.137, 0.228, ..., 0.099]. To ensure data confidentiality during transmission, this intermediate data is encrypted using the AES-128 symmetric encryption algorithm. The key is pre-negotiated and regularly updated between the edge device and the backend server. The encrypted ciphertext data is sent to the backend server via a lightweight security protocol such as MQTT or HTTPS. This process typically completes within 500ms, ensuring that the control response latency meets the real-time requirements of station-level monitoring.

[0070] This encrypted transmission mechanism for intermediate data, on the one hand, significantly reduces the risk of original data being transmitted on the public network, and on the other hand, reduces the overall data volume, making network transmission more efficient and secure.

[0071] S400: decrypting the encrypted intermediate data, and inputting the decrypted intermediate data into a second neural network to obtain an intelligent control result.

[0072] The second neural network includes preset parameters of a recurrent neural network, which include node weights, biases, edge parameters between networks, and activation functions.

[0073] In this embodiment, the decryption operation 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 a preset symmetric key to restore the original intermediate data vector.

[0074] The decrypted intermediate data is a one-dimensional vector of length 64. This intermediate data is input into the second neural network deployed on the backend server to make the final intelligent control decision. In one actual implementation scenario, this 64-dimensional vector is mapped to a classification probability vector of length 3, corresponding to the three system states: "normal operation," "warning state," and "immediate control": [Normal operation: 0.12, Warning state: 0.18, Immediate control: 0.70]. Based on this, the system generates a control result of "immediate control" and triggers subsequent execution strategies, such as activating the dehumidification module, closing doors and windows, and activating audible and visual alarms.

[0075] Furthermore, this embodiment discloses a station environment safety control method based on multi-module collaboration, wherein the recurrent neural network is divided into a first neural network and a second neural network, and further includes: obtaining a structural diagram of the recurrent neural network; finding a closed loop that does not contain an input node and an output node in the structural diagram; replacing the closed loop with an integrated node until the closed loop that does not contain an input node and an output node can no longer be found in the structural diagram; and dividing the recurrent neural network into a first neural network and a second neural network, wherein the second neural network does not contain an integrated node.

[0076] First, a structural diagram of the recurrent neural network to be divided is obtained, which is used to describe the connection relationship between all neuron nodes in the network, including the directed connections between input nodes, output nodes and intermediate hidden nodes.

[0077] Next, we identify all closed-loop structures within the structure graph that do not contain input or output nodes—that is, feedback loops consisting solely of hidden layer nodes. These closed-loop structures introduce temporal dependencies and coupled logic into the network execution process. If removed without further processing, these closed-loops can lead to mismatches in the reasoning logic of upstream and downstream subnetworks.

[0078] To this end, each closed-loop structure is logically replaced with an integration node, that is, the closed loop is replaced with a functionally equivalent node, and all connections originally entering the closed loop are uniformly redirected to the integration node, and the connections originally output by the closed loop are uniformly issued by the integration node.

[0079] The above replacement operation is repeated until there is no closed loop in the structure graph that does not contain input and output nodes, and finally a structure graph without coupled closed loops is formed.

[0080] First, Figure 3AFor example, assuming the operator op is an AND operation and x2 is 0, whether x1 or x3 is 0 or 1 will not affect the result y = 0. This means that when x2 = 1, the operator must wait for data from x1 or x3 to arrive before executing the op operation. However, when x2 = 0, the op operation can be executed directly without waiting for x1 or x3. Therefore, timing is a significant issue in real-world operations.

[0081] by Figure 3B For example, if a recurrent neural network is split into the first and second recurrent neural networks using u0, the computation order of nodes B and C is assumed to be consistent: that is, node 32 is affected by input node B during the first iteration. However, if the recurrent neural network is considered as a whole (unsplit), the computation order of nodes B and C may be inconsistent. For example, in the iteration C->31->32, node 32 may not be affected by node B until several iterations later (because multiple nodes precede node B and have their computation timings calculated after node B). Clearly, the timing of the effects of B and C on node 32 changes when the recurrent neural network is split compared to when the recurrent neural network is considered as a whole, leading to inconsistent results between the two methods. Therefore, this method of segmentation is incorrect.

[0082] In order to solve this problem, Figure 4 For example, loops can be eliminated one by one. G0 first eliminates the loop formed by 4->5->6->7->8, replacing it with integrated node 13, and inheriting the directed edge connection relationship between this loop and other external nodes, forming G1. Next, G1 eliminates the loop formed by 13->9->10, replacing it with integrated node 14, and inheriting the directed edge connection relationship between this loop and other external nodes, forming G2. Finally, based on G2, all integrated nodes are used as the left side of the split, so that the second neural network on the right has no integrated nodes, and the above problem does not exist.

[0083] Furthermore, this embodiment discloses a station environment safety control method based on multi-module collaboration, and the method of finding a closed loop in the structure diagram that does not contain input nodes and output nodes also includes: starting from any non-input and non-output node, traversing all paths along the neural network without passing through the input and output nodes; the path that can return to the starting node is a closed loop.

[0084] Starting from any hidden layer node that is neither an input node nor an output node, all possible paths are traversed forward along the directed connections of the neural network without passing through any input or output nodes. If a path eventually returns to the starting node, the path is considered to form a closed loop.

[0085] Taking a specific recurrent neural network as an example, assume its structure diagram contains several nodes, where nodes 1 through 4 are input nodes, nodes 35 through 40 are output nodes, and the remaining nodes are hidden layer nodes. This step begins by selecting a non-input and non-output hidden layer node as the starting point, for example, node 10. Starting from node 10, the path is traversed forward along the directed connections in the neural network, restricting the traversal process to not passing through input nodes (i.e., nodes 1 through 4) or output nodes (i.e., nodes 35 through 40). During the path search process, if a complete path such as: node 10 → node 13 → node 17 → node 10 exists, it is considered a closed loop. After finding any closed loop, it can be replaced with an integration node. The search for closed loops is repeated until no closed loops are found.

[0086] Furthermore, the present embodiment discloses a method for controlling the environmental safety of a station based on multi-module collaboration, which, after decrypting the encrypted intermediate data and inputting the decrypted intermediate data into a second neural network to obtain an intelligent control result, also includes: generating an environmental control instruction according to the intelligent control result; sending the environmental control 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.

[0087] Furthermore, in this embodiment, the phrase "after decrypting the encrypted intermediate data and inputting the decrypted intermediate data into the second neural network to obtain the intelligent control result" also includes the following operations: First, the system generates corresponding environmental control instructions based on the output of the second neural network. For example, if the output result is a "humidity exceeds the limit + temperature is low" status label, the system automatically generates the following combined instructions: close the intelligent doors and windows, activate the far-infrared heating and dehumidification module, and set the fan speed control to low-speed ventilation mode. This control instruction is encapsulated using a standard protocol and distributed to each execution module via a wired or wireless control channel.

[0088] The execution module then initiates the corresponding action according to the command content and transmits feedback data back to the system after the action is completed. For example, the far-infrared heating module returns the actual operating power curve, the fan module returns the current speed value, and the door and window module returns the physical opening and closing status.

[0089] The system integrates and analyzes the feedback data returned by each execution module and compares it with the control target. If there is an obvious deviation, such as "insufficient humidity reduction after execution" or "delayed fan speed response", the system will automatically trigger the model update mechanism.

[0090] This update mechanism selectively optimizes parameter weights in either the first or second neural network. For example, if the deviation primarily arises from the difference in response between prediction and execution, the output layer weights of the second neural network are adjusted inversely based on the feedback error. If the deviation stems from distortion in sensor data during the front-end extraction phase, the input mapping structure of the first neural network is optimized.

[0091] In principle, the sequence interval [tb,te] should be described as the time interval of each node in the one-way graph: Figure 4 Taking G2 as an example, we want node 14's tb to be strictly greater than node 1 and 2's te, and node 14's te to be strictly less than node 12's tb. Since nodes 11 and 14 have no temporal relationship, their timing intervals typically need to be set to overlap. That is, node 11's tb must be strictly less than node 14's te, and node 14's tb must be strictly less than node 11's te. However, this determination is actually quite difficult. We use a sequence interval instead of a temporal interval, with the difference that the definitions of te and tb are modified.

[0092] Figure 5 This is a diagram showing the splitting of the first and second neural networks based on sequence intervals. Each pair of brackets represents a node. The left bracket represents the first value of the sequence interval, and the right bracket represents the second value of the sequence interval, arranged in ascending order. Square brackets indicate integrated nodes, represented by capital letters (e.g., E, G, I); small brackets indicate non-integrated nodes, represented by lowercase letters (e.g., a, b, c, d, f, h, j, k, l).

[0093] Among them, the external node is defined as a non-integrated node with no integrated node in the periphery. Figure 5 In the example, only a, b, c, and d are external nodes. As can be understood, since f is surrounded by E and h is surrounded by G, f and h are not external nodes.

[0094] It can be understood that a unidirectional graph is formed by replacing the closed loops with integrated nodes until no closed loops containing no input nodes and no output nodes can be found in the structure diagram. Simply determining the second neural network is the key to determining how to separate the first and second neural networks.

[0095] To summarize, the following steps are: first, determine the order interval of the unidirectional graph; then, based on this order interval, use external nodes as nodes of the second neural network; correspondingly, use non-external nodes (including integrated nodes) as nodes of the first neural network. The order interval of the unidirectional graph is determined using a depth-first search, where the first value tb of the order interval is the order information for nodes that turn gray, and the second value te of the order interval is the order information for nodes that turn black.

[0096] It can be understood that white, gray and black are the marks of the nodes after the 0th visit node, the 1st visit node and the second visit node in the depth-first search.

[0097] Figure 2 A schematic diagram of a station environment safety control system based on multi-module collaboration provided in an embodiment of the present disclosure.

[0098] The multi-source sensor is used to collect environmental data, and the pre-processing module is used to pre-process the environmental data to obtain input data.

[0099] In this embodiment, the multi-source sensor includes a temperature and humidity sensor, a smoke sensing module, Gas concentration detection modules, partial discharge monitoring modules and passive high-voltage node temperature sensors are deployed in different functional areas within the station.

[0100] Each sensor collects environmental data at a fixed period, and the sampling period is set to 5 seconds. Specifically:

[0101] The temperature and humidity sensor collects ambient temperature and humidity every 5 seconds. For example, the first collected value is 24.7°C and the relative humidity is 63.2%. The smoke sensor module returns a unit concentration value. For example, the first value 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.

[0102] After the raw data is collected, it is pre-processed uniformly by the pre-processing module. This includes timestamp alignment, unit conversion (e.g., converting temperature to degrees Celsius and concentration to ppm), sliding window filtering (with a window size of 3), and normalization to map each channel data to the [0, 1] range.

[0103] For example, the original temperature data sequence is [24.7, 24.6, 24.9], which is transformed into [0.4, 0.2, 0.8] after normalization. The concentration sequence is [99.84, 99.79, 99.81], which is transformed into [0.75, 0.25, 0.5] after normalization.

[0104] After preprocessing, all channel data is combined into an input data vector, which serves as the input data for the first neural network in the subsequent step S200. This input data vector contains the multidimensional time series features of each multi-source sensor within the current time window. It has strong expressive power and structural consistency, making it suitable for processing in deep neural network models.

[0105] The neural network management module is used to divide the recurrent neural network into a first neural network and a second neural network.

[0106] In this embodiment, a recurrent neural network is used to process multi-sensor environmental data. The network is composed of multiple neurons, each of which may have sequential connections, feedback connections, or recursive connections, forming a complex recurrent network. To achieve functional separation between the front-end and back-end, the recurrent neural network is divided into two independent sub-networks: the first neural network and the second neural network.

[0107] First, the neural network's structural information is obtained, in the form of a diagram consisting of nodes and connections. Nodes represent neurons, and connections represent information flow paths. According to pre-set rules, the input and output nodes are labeled, corresponding to the sensor data input and the final intelligent control output of the station, respectively.

[0108] Then, using a structured traversal method, we identify all paths that start from non-input nodes and return to the starting point without passing through any input or output nodes. These paths are defined as "closed loops consisting solely of hidden layers." For example, if we identify several circular paths like "Node A → Node B → Node C → Node A" during structured traversal, these are typical closed loops.

[0109] After identification, each closed-loop region is replaced with a logical "integration node." The replacement process involves: all connections that originally entered the closed loop are now connected to the integration node; all connections that originally exited the closed loop are now output from the integration node. This replacement ensures that the closed loop's external input and output interfaces remain consistent while eliminating internal coupling dependencies.

[0110] Finally, the recurrent neural network is split into two parts, where the second neural network does not contain integration nodes.

[0111] The pre-processing module is further used 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.

[0112] In this embodiment, for example, the input data consists of temperature, humidity, The input matrix consists of six channels: 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.

[0113] The input data is fed into the first neural network deployed on the edge device for processing. This first neural network is an LSTM network structure consisting of two layers of LSTM units, each with 32 neurons, and outputs a one-dimensional vector of length 64 as intermediate data.

[0114] For example, in a certain round of reasoning, the input is the following matrix:

[0115] [0.41, 0.39, 0.43, 0.45, 0.42],# Temperature channel

[0116] [0.63, 0.65, 0.66, 0.64, 0.62], # humidity channel

[0117] [0.95, 0.94, 0.93, 0.94, 0.96],# Concentration channel

[0118] [0.12, 0.10, 0.11, 0.13, 0.12],# Smoke channel

[0119] [0.48, 0.47, 0.46, 0.49, 0.50],# High pressure temperature channel

[0120] [0.37, 0.36, 0.35, 0.38, 0.39],# Partial discharge channel

[0121] After inputting this matrix, the first neural network outputs an intermediate data vector, for example, a one-dimensional floating-point vector of length 64: [0.052, 0.137, 0.228, ..., 0.099]. To ensure data confidentiality during transmission, this intermediate data is encrypted using the AES-128 symmetric encryption algorithm. The key is pre-negotiated and regularly updated between the edge device and the backend server. The encrypted ciphertext data is sent to the backend server via a lightweight security protocol such as MQTT or HTTPS. This process typically completes within 500ms, ensuring that the control response latency meets the real-time requirements of station-level monitoring.

[0122] This encrypted transmission mechanism for intermediate data, on the one hand, significantly reduces the risk of original data being transmitted on the public network, and on the other hand, reduces the overall data volume, making network transmission more efficient and secure.

[0123] The back-end server is used to decrypt the encrypted intermediate data and input the decrypted intermediate data into the second neural network to obtain an intelligent control result.

[0124] In this embodiment, the decryption operation 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 a preset symmetric key to restore the original intermediate data vector.

[0125] The decrypted intermediate data is a one-dimensional vector of length 64. This intermediate data is input into the second neural network deployed on the backend server to make the final intelligent control decision. In one actual implementation scenario, this 64-dimensional vector is mapped to a classification probability vector of length 3, corresponding to the three system states: "normal operation," "warning state," and "immediate control": [Normal operation: 0.12, Warning state: 0.18, Immediate control: 0.70]. Based on this, the system generates a control result of "immediate control" and triggers subsequent execution strategies, such as activating the dehumidification module, closing doors and windows, and activating audible and visual alarms.

[0126] Furthermore, this embodiment discloses a station environment safety control system based on multi-module collaboration, wherein the segmentation of the recurrent neural network into a first neural network and a second neural network also includes: obtaining a structural diagram of the recurrent neural network; finding a closed loop that does not contain an input node and an output node in the structural diagram; replacing the closed loop with an integrated node until the closed loop that does not contain an input node and an output node can no longer be found in the structural diagram; and segmenting the recurrent neural network into a first neural network and a second neural network, wherein the second neural network does not contain an integrated node.

[0127] First, a structural diagram of the recurrent neural network to be divided is obtained, which is used to describe the connection relationship between all neuron nodes in the network, including the directed connections between input nodes, output nodes and intermediate hidden nodes.

[0128] Next, we identify all closed-loop structures within the structure graph that do not contain input or output nodes—that is, feedback loops consisting solely of hidden layer nodes. These closed-loop structures introduce temporal dependencies and coupled logic into the network execution process. If removed without further processing, these closed-loops can lead to mismatches in the reasoning logic of upstream and downstream subnetworks.

[0129] To this end, each closed-loop structure is logically replaced with an integration node, that is, the closed loop is replaced with a functionally equivalent node, and all connections originally entering the closed loop are uniformly redirected to the integration node, and the connections originally output by the closed loop are uniformly issued by the integration node.

[0130] The above replacement operation is repeated until there is no closed loop in the structure graph that does not contain input and output nodes, and finally a structure graph without coupled closed loops is formed.

[0131] Furthermore, the present embodiment discloses a station environment safety control system based on multi-module collaboration, and the method of finding a closed loop in the structure diagram that does not contain input nodes and output nodes also includes: starting from any non-input and non-output node, traversing all paths along the neural network without passing through the input and output nodes; the path that can return to the starting node is a closed loop.

[0132] Starting from any hidden layer node that is neither an input node nor an output node, all possible paths are traversed forward along the directed connections of the neural network without passing through any input or output nodes. If a path eventually returns to the starting node, the path is considered to form a closed loop.

[0133] Taking a specific recurrent neural network as an example, assume its structure diagram contains several nodes, where nodes 1 through 4 are input nodes, nodes 35 through 40 are output nodes, and the remaining nodes are hidden layer nodes. This step begins by selecting a non-input and non-output hidden layer node as the starting point, for example, node 10. Starting from node 10, the path is traversed forward along the directed connections in the neural network, restricting the traversal process to not passing through input nodes (i.e., nodes 1 through 4) or output nodes (i.e., nodes 35 through 40). During the path search process, if a complete path such as: node 10 → node 13 → node 17 → node 10 exists, it is considered a closed loop. After finding any closed loop, it can be replaced with an integration node. The search for closed loops is repeated until no closed loops are found.

[0134] Furthermore, the present embodiment discloses a station environment safety control system based on multi-module collaboration, which, after decrypting the encrypted intermediate data and inputting the decrypted intermediate data into the second neural network to obtain the intelligent control result, also includes: generating an environmental control instruction according to the intelligent control result; sending the environmental control instruction to the 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.

[0135] Furthermore, in this embodiment, the phrase "after decrypting the encrypted intermediate data and inputting the decrypted intermediate data into the second neural network to obtain the intelligent control result" also includes the following operations: First, the system generates corresponding environmental control instructions based on the output of the second neural network. For example, if the output result is a "humidity exceeds the limit + temperature is low" status label, the system automatically generates the following combined instructions: close the intelligent doors and windows, activate the far-infrared heating and dehumidification module, and set the fan speed control to low-speed ventilation mode. This control instruction is encapsulated using a standard protocol and distributed to each execution module via a wired or wireless control channel.

[0136] The execution module then initiates the corresponding action according to the command content and transmits feedback data back to the system after the action is completed. For example, the far-infrared heating module returns the actual operating power curve, the fan module returns the current speed value, and the door and window module returns the physical opening and closing status.

[0137] The system integrates and analyzes the feedback data returned by each execution module and compares it with the control target. If there is an obvious deviation, such as "insufficient humidity reduction after execution" or "delayed fan speed response", the system will automatically trigger the model update mechanism.

[0138] This update mechanism selectively optimizes parameter weights in either the first or second neural network. For example, if the deviation primarily arises from the difference in response between prediction and execution, the output layer weights of the second neural network are adjusted inversely based on the feedback error. If the deviation stems from distortion in sensor data during the front-end extraction phase, the input mapping structure of the first neural network is optimized.

[0139] According to an embodiment of the present disclosure, an electronic device is provided. The electronic device may include: a processor, a communications interface, a memory, and a communications bus. The processor, the communications interface, and the memory communicate with each other via the communications bus. The processor may invoke logic instructions in the memory to execute a configuration software-based soft authorization implementation method.

[0140] In addition, the logical instructions in the above-mentioned memory can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present disclosure is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present disclosure. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0141] On the other hand, the present disclosure further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which is implemented when the computer program is executed by a processor to execute the configuration software soft authorization implementation method provided by the above methods.

[0142] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0143] 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, or of course, by hardware. Based on this understanding, the essence of the above technical solution or the part that contributes to the existing technology can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling 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 certain parts of the embodiments.

[0144] 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 aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present disclosure.

Claims

1. A site environment security control method based on multi-module collaboration, comprising: Multi-source sensors collect environmental data and pre-process the environmental data to obtain input data; Split 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 backend server; decrypting the encrypted intermediate data, and inputting the decrypted intermediate data into a second neural network to obtain an intelligent control result; The second neural network includes preset parameters of a recurrent neural network; the encrypted intermediate data is encrypted in a lightweight manner; The method of dividing the recurrent neural network into a first neural network and a second neural network further includes: Obtaining a structural diagram of the recurrent neural network; Finding a closed loop in the structure graph that does not contain an input node and an output node; Replacing the closed loop with an integration node until no closed loop containing no input node and no output node can be found in the structure diagram; The recurrent neural network is divided into a first neural network and a second neural network, wherein the second neural network does not contain an integrated node.

2. A station environment safety control method based on multi-module collaboration according to claim 1, characterized in that: The lightweight encryption method is the AES algorithm.

3. A station environment safety control method based on multi-module collaboration according to claim 1, characterized in that: The input data is an input data vector, which contains the multi-dimensional time series features of each multi-source sensor in the current time window.

4. A station environment safety control method based on multi-module collaboration according to claim 1, characterized in that: The step of finding a closed loop in the structure graph that does not include an input node and an output node further includes: Starting from any non-input and non-output node, traverse all paths along the neural network without passing through any input or output nodes; The path that returns to the starting node is a closed loop.

5. The method for controlling station environment security based on multi-module collaboration according to claim 1 is characterized in that: After decrypting the encrypted intermediate data and inputting the decrypted intermediate data into the second neural network to obtain the intelligent control result, the method further includes: generating an environmental control instruction according to the intelligent control result; Sending the environmental control instruction to the execution module; Receive feedback data returned by each execution module, and use the feedback data to optimize the first neural network or the second neural network.

6. The method for controlling station environment safety based on multi-module collaboration according to claim 1 is characterized in that: The preset parameters of the recurrent neural network include: node weights, biases, edge parameters between networks, and the activation function used.

7. The method for controlling station environment security based on multi-module collaboration according to claim 1 is characterized in that: Determine the order interval of the unidirectional graph; and based on the order interval, use the external nodes as nodes of the second neural network.

8. A station environment safety control method based on multi-module collaboration according to claim 7, characterized in that: The order interval of the unidirectional graph is determined based on a depth-first search, wherein the first value of the order interval is the order information of the node turning gray, and the second value of the order interval is the order information of the node turning black.

9. The method for controlling station environment safety based on multi-module collaboration according to claim 1 is characterized in that: Recurrent neural networks are long short-term memory networks or gated recurrent units.

Images