Intelligent closed-loop linkage inspection method, model training method and system in centralized control system

Abstract

The invention provides an intelligent closed-loop linkage inspection method, a model training method and a model training system in a centralized control system, which comprehensively realize closed-loop inspection functions of fault discovery, fault location and fault alarm for the centralized control system by utilizing a deep learning network framework and combining an intelligent inspection robot inspection algorithm technology. In the whole closed-loop inspection process, the intelligent inspection robot is used for acquiring a large amount of redundant alarm information by utilizing a sensor and communication equipment, then the information sequence is put into the bidirectional long short-term memory network, and the output layer generates final output according to a hidden state. According to the structure, the Bi-LSTM network can effectively process time sequence data, capture the long-term dependency relationship in a data sequence, obtain equipment state information and position information, and finally feed back the information to the intelligent inspection robot, so that closed-loop inspection is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of substation inspection, and in particular to an intelligent closed-loop linkage inspection method and system in a centralized control system. Background Art

[0002] The scale of the company's substation equipment continues to expand, and the contradiction between the current operation and maintenance management model and the rapid growth of equipment has become increasingly prominent. With the rapid development of the company's power grid, the substantial growth of equipment scale, and the continuous upgrading of technology, the risks and pressures of safe operation of power grid equipment are increasing day by day. The existing substation operation and maintenance management model is difficult to adapt to the company's development strategy and power grid operation requirements. It is urgent to build a new generation of centralized control system to improve the degree of operation and maintenance informatization and meet the needs of comprehensive equipment monitoring. Using advanced communication technology and artificial intelligence technology, the centralized control system can realize functions such as "integrated monitoring, panoramic display, sequential control operation call, paperless operation, comprehensive error prevention, and unified alarm", remotely monitor and control unmanned substations, and ensure the safe and stable operation of power grid equipment. The core function of the centralized control system is equipment status perception and alarm, but equipment redundancy and data sampling redundancy may cause massive alarm information to appear instantly when a single device fails, making it difficult for operation and maintenance personnel to locate the source of the fault in time.

[0003] The problems existing in the prior art are:

[0004] (1) Traditional fault diagnosis methods rely entirely on manual judgment. When a centralized monitoring system is found to have a fault, the fault location is manually determined and the fault point is located step by step. This is time-consuming and labor-intensive, and there is a risk of misjudgment. It is impossible to accurately locate the root cause of the fault.

[0005] (2) There is no standardized fault management, the efficiency of fault location and processing is low and the process is chaotic. It is impossible to establish fault defect statistics, analysis and expert database, and there is a lack of intelligent analysis, filtering and automated processing.

[0006] (3) The existing technology not only requires high fault maintenance costs, but also has high complexity in manual positioning. Summary of the invention

[0007] The technical problem to be solved by the present invention is to provide an intelligent closed-loop linkage inspection method in a centralized control system for equipment rooms or substations where the substation equipment is relatively fixed.

[0008] The present invention solves the above technical problems through the following technical means:

[0009] The intelligent closed-loop linkage inspection method in the centralized control system includes the following steps:

[0010] The intelligent closed-loop linkage inspection method in the centralized control system includes the following steps:

[0011] Divide the space to be inspected into multiple grids; define a label for each grid; the label can reflect the coordinates of obstacles in the grid;

[0012] Randomly sample multiple points in the blank area of ​​the grid to generate potential path nodes and connect them. Then connect multiple sampling points in the grid to form local path segments. Multiple local path segments constitute a path network. Then, path planning is performed based on the path network.

[0013] The inspection robot moves along the optimal path and obtains the alarm signal in the inspection space. The alarm signals obtained by multiple inspection robots constitute an alarm data sequence;

[0014] The alarm data sequence is used as input of the trained alarm information model, and outputs device status and location information;

[0015] The inspection robot obtains the device status and location information output by the alarm information model and feeds it back to the host computer to complete the closed-loop inspection;

[0016] The alarm information model includes an input layer, a Bi-LSTM layer, a fully connected layer, a softmax layer, and a classification output layer;

[0017] Among them, the state activation function of the LSTM unit in the Bi-LSTM layer is tanh, and the gate activation function is sigmoid.

[0018] The method for optimizing the path network is specifically as follows:

[0019] A series of points are randomly sampled in the space to be inspected as potential path nodes through the PRM algorithm. The path nodes are randomly generated from the blank areas in the grid to ensure that the entire environment is covered. The local path planning algorithm A* algorithm is used to connect these sampling points in the grid to form local path segments of the path. Multiple local path segments constitute the path network of PRM, and path planning is performed based on the path network.

[0020] Among them, the A* algorithm calculates the priority of each randomly adopted path node through the following formula;

[0021] f(n)=g(n)+h(n)

[0022] Where: f(n) is the comprehensive priority of node n; g(n) is the cost of node n from the starting point; h(n) is the estimated cost of node n from the end point, and h(n) is the heuristic function of the A* algorithm;

[0023] The heuristic function h(n) is in the form of distance*multiple, that is:

[0024] h(n)=h1(n)*(1+(h2(n) / h3(n))*a)

[0025] Where h1(n) is the Euclidean distance from the current point to the target point, h2(n) is the Manhattan distance from the current point to the target point, h3(n) is the Manhattan distance from the starting point to the target point, and a is a variable multiple of the distance between the current node and the target node multiplied by;

[0026] During the operation of the A* algorithm, the node with the smallest f(n) value is selected from the priority queue each time as the next node to be traversed; finally, a priority node set is obtained, and the nodes in the priority set are used to construct a graph-like path network.

[0027] Furthermore, the grid division method is:

[0028] Multiple sets of sensors and cameras are deployed on site to collect real-time data on the operating status changes of each device in the inspection space; the processed operating status change data is added to the grid model to achieve the goal of real-time update of grid model information data;

[0029] Assume that the length of the intelligent inspection workspace is M, the width is N, and the grids are all squares with a side length of L. At this time, the calculation expression for the number of grids is:

[0030] R=(M / L)*(N / L)

[0031] Where R represents the number of small grids in the grid model; workspace grid map i It can be expressed as:

[0032] Map = {map i ,map i =0 / 1}

[0033] Use binary to indicate whether there are obstacles in the intelligent inspection workspace of the centralized control system, map i =0 indicates the free area for intelligent inspection; map i =1 indicates intelligent inspection of obstacle area.

[0034] Furthermore, the path planning method is:

[0035] Through the constructed grid model, the objective function and variables of the intelligent inspection path planning task are determined and coded; any part of the code is randomly selected as the initialization population of the improved genetic algorithm, the feasible path from the starting point of the initialization population to the target node is determined, and a feasible path coding set is generated; the fitness of the individuals in the population is evaluated; fitness refers to the degree of advantage of the individual in the population, and its evaluation expression is:

[0036]

[0037] In the formula, F represents the fitness function; s represents the total number of individuals in the initialized population passing through the grid; D represents the sum of the straight-line distances between adjacent grids;

[0038] Through calculation, the fitness of the individuals in the initialized population is obtained. The larger the fitness value, the higher the possibility of being selected, and vice versa. The fitness is tested to determine whether it meets the optimal convergence conditions for the global planning of the intelligent inspection path of the centralized control system. If it does not meet the optimal convergence conditions, a new population is generated through selection, crossover, and mutation operations, and then the fitness of the individuals in the population is evaluated again, and the above steps are repeated. If the optimal convergence conditions are met, the individuals in the population are decoded and the optimal solution after decoding is output, which is the optimal path for intelligent inspection.

[0039] Furthermore, when the LSTM unit processes the alarm data sequence, the value of the alarm data sequence at each moment is used as the input of an LSTM unit. Assume that in the Bi-LSTM layer, the input of a certain LSTM unit is: the input sequence at time t is xt, the output value ht-1 of the LSTM unit corresponding to time t-1 and its memory unit state ct-1;

[0040] The output of the LSTM unit is: the output value ht of the LSTM unit corresponding to time t and the state ct of the memory unit at that time;

[0041] The input gate, output gate, and forget gate are used as activation functions to control the mutual influence between the data input into the LSTM unit. The input gate controls the influence of xt on ct, the forget gate controls the influence of ct-1 on ct, and the output gate controls the influence of ct on ht.

[0042] The calculation formula of the input gate is:

[0043]

[0044] The calculation formula of the output gate is:

[0045]

[0046] The calculation formula of the forget gate is:

[0047]

[0048] Among them, it represents the state of the input gate, ot represents the state of the output gate, ft represents the state of the forget gate, W is the weight matrix, b is the bias term, and the subscripts i, o, and f correspond to the input gate, output gate, and forget gate respectively; is the sigmoid activation function;

[0049] Therefore, the output of the LSTM unit corresponding to time t is as follows:

[0050] ht=ot⊙ψ(ct)

[0051] ct=ft⊙ct-1+it⊙ψ(Wisht-1+Wisxt+bis)

[0052] =ft⊙c t-1 +it⊙ψ(is)

[0053] Among them, ψ is the tanh activation function, Wis is the input state weight matrix, bis is the input state bias term, and ⊙ represents element-wise multiplication;

[0054] The fully connected layer takes the output of the LSTM layer as an input vector and outputs the probability distribution of the fault category, as follows:

[0055] y=Softmax(θd·x+b d )

[0056] Where x represents the output of LSTM, that is, the input of the fully connected layer, θd and bd are the weight matrix and bias matrix of the fully connected layer, y represents the final output of the fully connected layer, and Softmax is the activation function.

[0057] The present invention also provides a method for training an alarm information model, comprising the following steps:

[0058] The output of each unit in each layer is calculated by forward propagation, and the cost function loss is obtained. The formula is as follows:

[0059]

[0060] Among them, L(f(xi;θ),yi) is the cross loss function, N is the number of samples, xi is the sample value, f(xi;θ;bd) is the output result after the i-th sample is input into the five-layer network, θ and b are the weight matrix and bias matrix in the network, and yi is the expected output value of the corresponding sample;

[0061] Back propagation is used to calculate the deviation of the output of each unit in each layer. The error of the LSTM layer has two back propagation directions: one is to propagate to the upper layer; the other is to propagate back along time;

[0062] The gradient of each parameter in the loss function is calculated by the bias term and the parameter value is updated;

[0063] Continuously iterate and correct the parameter value to make the loss function smaller and smaller until it is less than the pre-set number of iterations;

[0064] Finally, the actual test data is combined for diagnosis and network update. The training process uses the AdamW algorithm to optimize the parameters of the Bi-LSTM layer. In AdamW, regularization becomes:

[0065] θ t =θ t-1 -γλθ t-1 Where gt is the gradient of the current step, θt-1 is the LSTM unit weight in the previous step, λ is the regularization coefficient, and r is the learning rate.

[0066] The present invention also provides an intelligent closed-loop linkage inspection system, comprising:

[0067] The grid division module of the space to be inspected divides the space to be inspected into multiple grids; defines a label for each grid; the label can reflect the coordinates of obstacles in the grid;

[0068] The path network construction module randomly samples multiple points in the blank area of ​​the grid to generate potential path nodes and connects them. Then, multiple sampling points are connected within the grid to form local path segments of the path. Multiple local path segments constitute a path network, and then path planning is performed based on the path network.

[0069] Inspection module: the inspection robot moves along the optimal path and obtains alarm signals in the inspection space. The alarm signals obtained by multiple inspection robots constitute an alarm data sequence.

[0070] An alarm information processing module, wherein the alarm data sequence is used as the input of a trained alarm information model and outputs device status and location information;

[0071] Feedback module: the inspection robot obtains the device status and location information output by the alarm information model and feeds it back to the host computer to complete the closed-loop inspection;

[0072] The alarm information model includes an input layer, a Bi-LSTM layer, a fully connected layer, a softmax layer, and a classification output layer;

[0073] Among them, the state activation function of the LSTM unit in the Bi-LSTM layer is tanh, and the gate activation function is sigmoid.

[0074] The method for optimizing the path network is specifically as follows:

[0075] A series of points are randomly sampled in the space to be inspected as potential path nodes through the PRM algorithm. The path nodes are randomly generated from the blank areas in the grid to ensure that the entire environment is covered. The local path planning algorithm A* algorithm is used to connect these sampling points in the grid to form local path segments of the path. Multiple local path segments constitute the path network of PRM, and path planning is performed based on the path network.

[0076] Among them, the A* algorithm calculates the priority of each randomly adopted path node through the following formula;

[0077] f(n)=g(n)+h(n)

[0078] Where: f(n) is the comprehensive priority of node n; g(n) is the cost of node n from the starting point; h(n) is the estimated cost of node n from the end point, and h(n) is the heuristic function of the A* algorithm;

[0079] The heuristic function h(n) is in the form of distance*multiple, that is:

[0080] h(n)=h1(n)*(1+(h2(n) / h3(n))*a)

[0081] Where h1(n) is the Euclidean distance from the current point to the target point, h2(n) is the Manhattan distance from the current point to the target point, h3(n) is the Manhattan distance from the starting point to the target point, and a is a variable multiple of the distance between the current node and the target node multiplied by;

[0082] During the operation of the A* algorithm, the node with the smallest f(n) value is selected from the priority queue each time as the next node to be traversed; finally, a priority node set is obtained, and the nodes in the priority set are used to construct a graph-like path network.

[0083] Furthermore, the grid division method is:

[0084] Multiple sets of sensors and cameras are deployed on site to collect real-time data on the operating status changes of each device in the inspection space; the processed operating status change data is added to the grid model to achieve the goal of real-time update of grid model information data;

[0085] Assume that the length of the intelligent inspection workspace is M, the width is N, and the grids are all squares with a side length of L. At this time, the calculation expression for the number of grids is:

[0086] R=(M / L)*(N / L)

[0087] Where R represents the number of small grids in the grid model; the workspace grid mapi can be expressed as:

[0088] Map = {mapi, mapi = 0 / 1}

[0089] Binary system is used to indicate whether there are obstacles in the intelligent inspection workspace of the centralized control system. mapi=0 indicates the free area for intelligent inspection; mapi=1 indicates the obstacle area for intelligent inspection.

[0090] Furthermore, the path planning method is:

[0091] Through the constructed grid model, the objective function and variables of the intelligent inspection path planning task are determined and coded; any part of the code is randomly selected as the initialization population of the improved genetic algorithm, the feasible path from the starting point of the initialization population to the target node is determined, and a feasible path coding set is generated; the fitness of the individuals in the population is evaluated; fitness refers to the degree of advantage of the individual in the population, and its evaluation expression is:

[0092]

[0093] In the formula, F represents the fitness function; s represents the total number of individuals in the initialized population passing through the grid; D represents the sum of the straight-line distances between adjacent grids;

[0094] Through calculation, the fitness of the individuals in the initialized population is obtained. The larger the fitness value, the higher the possibility of being selected, and vice versa. The fitness is tested to determine whether it meets the optimal convergence conditions for the global planning of the intelligent inspection path of the centralized control system. If it does not meet the optimal convergence conditions, a new population is generated through selection, crossover, and mutation operations, and then the fitness of the individuals in the population is evaluated again, and the above steps are repeated. If the optimal convergence conditions are met, the individuals in the population are decoded and the optimal solution after decoding is output, which is the optimal path for intelligent inspection.

[0095] Furthermore, when the LSTM unit processes the alarm data sequence, the value of the alarm data sequence at each moment is used as the input of an LSTM unit. Assume that in the Bi-LSTM layer, the input of a certain LSTM unit is: the input sequence at time t is xt, the output value ht-1 of the LSTM unit corresponding to time t-1 and its memory unit state ct-1;

[0096] The output of the LSTM unit is: the output value ht of the LSTM unit corresponding to time t and the state ct of the memory unit at that time;

[0097] The input gate, output gate, and forget gate are used as activation functions to control the mutual influence between the data input into the LSTM unit. The input gate controls the influence of xt on ct, the forget gate controls the influence of ct-1 on ct, and the output gate controls the influence of ct on ht.

[0098] The calculation formula of the input gate is:

[0099]

[0100] The calculation formula of the output gate is:

[0101]

[0102] The calculation formula of the forget gate is:

[0103]

[0104] Among them, it represents the state of the input gate, ot represents the state of the output gate, ft represents the state of the forget gate, W is the weight matrix, b is the bias term, and the subscripts i, o, and f correspond to the input gate, output gate, and forget gate respectively; is the sigmoid activation function;

[0105] Therefore, the output of the LSTM unit corresponding to time t is as follows:

[0106] ht=ot⊙ψ(ct)

[0107] ct=ft⊙ct-1+it⊙ψ(Wisht-1+Wisxt+bis)

[0108] =ft⊙c t-1 +it⊙ψ(is)

[0109] Among them, ψ is the tanh activation function, Wis is the input state weight matrix, bis is the input state bias term, and ⊙ represents element-wise multiplication;

[0110] The fully connected layer takes the output of the LSTM layer as an input vector and outputs the probability distribution of the fault category, as follows:

[0111] y=Softmax(θd·x+bd)

[0112] Where x represents the output of LSTM, that is, the input of the fully connected layer, θd and bd are the weight matrix and bias matrix of the fully connected layer, y represents the final output of the fully connected layer, and Softmax is the activation function.

[0113] The present invention also provides a training system for an alarm information model, comprising:

[0114] The output of each unit in each layer is calculated by forward propagation, and the cost function loss is obtained. The formula is as follows:

[0115]

[0116] Among them, L(f(xi;θ),yi) is the cross loss function, N is the number of samples, xi is the sample value, f(xi;θ;bd) is the output result after the i-th sample is input into the five-layer network, θ and b are the weight matrix and bias matrix in the network, and yi is the expected output value of the corresponding sample;

[0117] Back propagation is used to calculate the deviation of the output of each unit in each layer. The error of the LSTM layer has two back propagation directions: one is to propagate to the upper layer; the other is to propagate back along time;

[0118] The gradient of each parameter in the loss function is calculated by the bias term and the parameter value is updated;

[0119] Continuously iterate and correct the parameter value to make the loss function smaller and smaller until it is less than the pre-set number of iterations;

[0120] Finally, the actual test data is combined for diagnosis and network update. The training process uses the AdamW algorithm to optimize the parameters of the Bi-LSTM layer. In AdamW, regularization becomes:

[0121] θ t =θ t-1 -γλθ t-1 Where gt is the gradient of the current step, θt-1 is the LSTM unit weight in the previous step, λ is the regularization coefficient, and r is the learning rate.

[0122] The effects of the present invention are:

[0123] The intelligent closed-loop linkage inspection system and method proposed in the present invention realizes fast, accurate and reliable closed-loop intelligent inspection of the centralized control system through the combination of an intelligent inspection robot and a Bi-LSTM network, thereby improving the efficiency of the entire centralized control system. The intelligent inspection robot uses sensors and communication equipment to obtain a large amount of redundant alarm information, while the Bi-LSTM network can effectively process time series data and capture long-term dependencies in data sequences, making the judgment of equipment status information and location information more accurate, and completing closed-loop intelligent inspection. BRIEF DESCRIPTION OF THE DRAWINGS

[0124] Figure 1 This is a flow chart of the intelligent closed-loop linkage inspection method in the centralized control system proposed in an embodiment of the present invention;

[0125] Figure 2 A schematic diagram of the inspection algorithm proposed in an embodiment of the present invention;

[0126] Figure 3 A diagram of a fault diagnosis network structure proposed in an embodiment of the present invention;

[0127] Figure 4 This is the Bi-LSTM network structure and LSTM unit diagram proposed in the present invention. DETAILED DESCRIPTION

[0128] In order to make the purpose, technical solution and advantages of the embodiment of the present embodiment clearer, the technical solution in the embodiment of the present embodiment will be clearly and completely described in combination with the embodiment of the present embodiment. Obviously, the described embodiment is a part of the embodiment of the present embodiment, not all of the embodiments. Based on the embodiment in the present embodiment, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present embodiment.

[0129] In this embodiment, the intelligent closed-loop linkage inspection method in the centralized control system uses a deep learning network framework and combines it with the intelligent inspection robot inspection algorithm technology to comprehensively realize the closed-loop inspection functions of fault discovery, fault location and fault alarm for the centralized control system (space to be inspected). The entire closed-loop inspection process will use intelligent inspection robots to use sensors and communication equipment to obtain a large amount of redundant alarm information, and then put these information sequences into a bidirectional long short-term memory network. The output layer of the bidirectional long short-term memory network generates the final output based on the hidden state. This structure enables the Bi-LSTM network to effectively process time series data and capture long-term dependencies in the data sequence to obtain equipment status information and location information, which are finally fed back to the intelligent inspection robot to achieve closed-loop inspection. The entire closed-loop inspection process is like this. Figure 1 Specifically, it is carried out in the following three parts:

[0130] 1. Grid division of the space to be inspected and robot path planning

[0131] Step 1: Use multiple groups of intelligent inspection robots to collect alarm information data from multiple locations in the centralized control system;

[0132] Step 1.1: Determine the key parts and parameters for alarm information data collection, including but not limited to main equipment, key sensors and environmental parameters.

[0133] Step 1.2: Design the sensor configuration of the intelligent inspection robot to ensure that each robot can monitor the status and environmental parameters of key parts in real time.

[0134] In this example, a voltage divider voltage sensor will be installed on the intelligent inspection robot. The voltage divider voltage sensor is a sensor based on the voltage division principle, usually composed of a voltage divider and a signal processing circuit. There is a linear relationship between its output voltage and the measured voltage. Its output voltage V out With the measured voltage V in The relationship between can be expressed by the following formula:

[0135] V out＝ V in *(R2 / (R1+R2))

[0136] R1 and R2 are two series resistors in the circuit.

[0137] Step 1.3: Develop the robot's data collection and transmission system to ensure that the robot can accurately and timely collect alarm information and transmit it to the centralized control system through wireless communication.

[0138] ZigBee wireless communication technology is adopted. The ZigBee wireless communication module is mainly used to realize the wireless transmission and reception of data by the inspection staff. The module consists of two parts: the ZigBee terminal node module and the ZigBee coordinator module. Both use the CC2530 chip of Chipcon Company. The chip contains a high-performance RF transceiver and an enhanced 8051MCU, online programmable program memory Flash, 8KB data storage RAM and other modules. The chip supports 8-channel 12-bit ADC, 2 USART interfaces and 21 GPIO interfaces. The chip also has a short sleep wake-up conversion time, which can meet the requirements of ultra-low power consumption system.

[0139] The ZigBee terminal node uses the serial port to communicate with STM32, configure CC2530 and complete data transmission, perform ZigBee networking, and collect all messages in the ZigBee network and send them to the host computer through the serial port.

[0140] Step 1.4: Figure 2 As shown in the figure, an inspection plan and route algorithm are formulated, and multiple groups of intelligent inspection robots are assigned to various parts of the centralized control system for regular inspections to ensure comprehensive coverage and timely collection of alarm information.

[0141] In this embodiment, multiple sets of sensors and cameras are deployed on site to collect the data of the operation status change of the centralized control system in real time. On this basis, the uncertain data in the system operation status change data is technically processed, and the processed data is added to the grid model to achieve the goal of real-time update of the grid model information data. The length of the intelligent inspection workspace is set to M, the width is N, and the grids are all squares with a side length of L. At this time, the calculation expression of the number of grids is:

[0142] R=(M / L)*(N / L)

[0143] In the formula, R represents the number of small grids in the grid model. Workspace grid map i It can be expressed as:

[0144] Map = {map i ,map i =0 / 1}

[0145] Use binary to indicate whether there are obstacles in the intelligent inspection workspace of the centralized control system, mapi =0 indicates the free area for intelligent inspection; map i =1 indicates that there are obstacles in the intelligent inspection area and intelligent inspection cannot be performed directly. The constructed grid model can divide the intelligent inspection workspace of the centralized control system into multiple grids. The number of grids is determined according to the size of the intelligent inspection workspace. Through the specific feature information of each grid, it can fully and multi-dimensionally describe whether there are obstacles in the intelligent inspection workspace of the centralized control system and the changes in the operating status.

[0146] The PRM (Probabilistic Roadmap) algorithm randomly samples a series of points in the environment as potential path nodes through node sampling. These nodes are usually randomly generated from blank areas in the grid model to ensure that the entire environment is covered, and the path connection is performed: these sampled points are connected in the grid model using a local path planning algorithm (such as the A* algorithm) to form local segments of the path. These local path segments constitute the graph structure of the PRM, and finally the path optimization is performed: the generated path network is optimized, usually by removing redundant path segments or merging approximate paths to improve the quality and feasibility of the path.

[0147] For points on the randomly sampled map, the local path planner A* algorithm is used. The A* algorithm calculates the priority of each node through the following function.

[0148] f(n)=g(n)+h(n)

[0149] Where: f(n) is the comprehensive priority of node n. When this embodiment selects the next node to be traversed, this embodiment will always select the node with the highest comprehensive priority (smallest value). g(n) is the cost of node n from the starting point. h(n) is the estimated cost of node n from the end point, which is the heuristic function of the A* algorithm.

[0150] Sometimes, the heuristic function may not be accurate enough, causing the A* algorithm to search for suboptimal solutions or search too large a space. In this case, you can consider dynamically adjusting the heuristic function to provide a more accurate estimate. You can adjust the weight of the heuristic function based on the actual cost during the search to make it closer to the actual situation.

[0151] This patent uses the form of dynamically adjusting the heuristic function to optimize the A* algorithm. The specific implementation is: the heuristic function adopts the form of distance*multiple, that is:

[0152] h(n)=h1(n)*(1+(h2(n) / h3(n))*a)

[0153] Among them, h1(n) is the Euclidean distance from the current point to the target point, h2(n) is the Manhattan distance from the current point to the target point, h3(n) is the Manhattan distance from the starting point to the target point, and a is a variable multiple of the distance between the current node and the target node, which makes it a good adjustment for the heuristic function.

[0154] In this part h1, the Euclidean distance is used because it is necessary to find the minimum distance limit, and the Manhattan distance is used in the multiple because it takes less time to calculate without changing the properties.

[0155] During the operation of the A* algorithm, the node with the smallest f(n) value (highest priority) is selected from the priority queue each time as the next node to be traversed.

[0156] In addition, the A* algorithm uses two sets to represent the nodes to be traversed and the nodes that have been traversed, which are usually called open_set and close_set. The following is the specific algorithm process:

[0157] * Initialize open_set and close_set;

[0158] *Add the starting point to open_set and set the priority to 0 (highest priority);

[0159] *If open_set is not empty, select the node n with the highest priority from open_set:

[0160] *If node n is the end point, then:

[0161] *Start from the end point and gradually trace the parent node until it reaches the starting point;

[0162] *Return the found result path and the algorithm ends;

[0163] * If node n is not an endpoint, then:

[0164] * Delete node n from open_set and add it to close_set;

[0165] *Traverse all neighboring nodes of node n:

[0166] *If the neighboring node m is in close_set, then:

[0167] * Skip and select the next adjacent node

[0168] *If the neighboring node m is not in open_set either, then:

[0169] *Set the parent of node m to node n

[0170] * Calculate the priority of node m

[0171] *Add node m to open_set

[0172] Connections are created between these points to build a graph-like path network. This enables the robot to plan paths in complex, incompletely known environments and to flexibly cope with dynamic obstacles.

[0173] In general, in the construction of the grid model based on the PRM algorithm, in the initial stage, the grid model can help determine the location of static obstacles and perform preliminary path planning. The PRM algorithm further optimizes the path on this basis and uses its global path search capability to find a better path solution. PRM can consider the information in the grid model, such as obstacle boundaries and passability, during the random sampling and connection process to avoid paths crossing obstacles or passing through inaccessible areas. When the position of obstacles in the environment changes, the grid model can be quickly updated to provide new obstacle information. The PRM algorithm can adapt to new environmental conditions by resampling new path nodes and reconnecting the path.

[0174] The PRM algorithm can generate candidate paths, but may need further optimization to meet multiple objectives, such as shortest path, least collision, minimum energy consumption, etc.

[0175] The improved genetic algorithm can globally optimize the path generated by PRM through the evolutionary optimization process. For example, the population evolution and fitness function evaluation of the genetic algorithm are used to select the optimal path and adjust the node positions in the path to improve the path quality and execution efficiency. Therefore, after completing the construction of the grid model based on the PRM algorithm, it can be known whether there are obstacles in the intelligent inspection workspace and the operating status changes of the centralized control system. On this basis, the inspection path is planned using the improved genetic algorithm.

[0176] Through the constructed grid model of the centralized control system, the objective function and variables of the intelligent inspection path planning task are determined and encoded. Randomly select any part from the code as the initialization population of the improved genetic algorithm, determine the feasible path from the starting point of the initialization population to the target node, and generate a feasible path encoding set. Evaluate the fitness of the individuals in the population. Fitness refers to the degree of advantage of an individual in the population, and its evaluation expression is:

[0177]

[0178] Where F represents the fitness function; s represents the total number of individuals in the initialized population passing through the grid; and D represents the sum of the straight-line distances between adjacent grids.

[0179] Through calculation, the fitness of the individuals in the initialized population is obtained. The larger the fitness value, the higher the possibility of being selected, and vice versa. The fitness is tested to determine whether it meets the optimal convergence conditions for the global planning of the intelligent inspection path of the centralized control system. If it does not meet the optimal convergence conditions, a new population is generated through selection, crossover, and mutation operations, and then the fitness of the individuals in the population is evaluated again, and the above steps are repeated. If the optimal convergence conditions are met, the individuals in the population are decoded and the optimal solution after decoding is output, which is the optimal path for the intelligent inspection of the centralized control system, achieving the goal of inspection path planning.

[0180] 2. Construction and training of alarm information model

[0181] like Figure 3 As shown in the figure, the alarm information model is a five-layer network built based on the Bi-LSTM network, including: input layer, Bi-LSTM layer, fully connected layer, softmax layer, and classification output layer.

[0182] The state activation function of the LSTM unit in the Bi-LSTM layer is tanh, the gate activation function is sigmoid, and the Bi-LSTM layer is composed of a reverse combination of LSTM units with 100 hidden units;

[0183] When the LSTM unit processes the above alarm data sequence, the value of the sequence at each moment is used as the input of an LSTM unit. Assume that in the Bi-LSTM layer, the input of a LSTM unit is: the input sequence at time t is xt, the output value of the LSTM unit corresponding to time t-1 is ht-1, and its memory unit state ct-1;

[0184] The output of the LSTM unit is: the output value ht of the LSTM unit corresponding to time t and the state ct of the memory unit at that time;

[0185] The output of the LSTM unit corresponding to each moment is not only subject to the input sequence value at this moment, but also depends on the impact of historical data on the data at that moment. The structured representation of the impact of different data on the final output results in three gates: input gate, output gate, and forget gate.

[0186] The three gates are equivalent to using three activation functions to control the mutual influence between the data input into the LSTM unit. The three gates are the components of the LSTM unit, and multiple LSTM units constitute the LSTM layer. The reverse combination of two layers of LSTM constitutes the Bi-LSTM layer, and the Bi-LSTM layer is only one layer in the deep learning framework. Among them, the input gate controls the influence of xt on ct, the forget gate controls the influence of ct-1 on ct, and the output gate controls the influence of ct on ht.

[0187] The calculation formula of the input gate is:

[0188]

[0189] The calculation formula of the output gate is:

[0190]

[0191] The calculation formula of the forget gate is:

[0192]

[0193] Among them, it represents the state of the input gate, ot represents the state of the output gate, ft represents the state of the forget gate, W is the weight matrix, b is the bias term, and the subscripts i, o, and f correspond to the input gate, output gate, and forget gate respectively; is the sigmoid activation function;

[0194] Therefore, the output of the LSTM unit corresponding to time t is as follows:

[0195] ht=ot⊙ψ(ct) (14)

[0196] ct=ft⊙ct-1+it⊙ψ(Wisht-1+Wisxt+bis) (15)

[0197] =ft⊙c t-1 +it⊙ψ(is)

[0198] Among them, ψ is the tanh activation function, Wis is the input state weight matrix, bis is the input state bias term, and ⊙ represents element-wise multiplication;

[0199] The fully connected layer takes the output of the LSTM layer as an input vector and outputs the probability distribution of the fault category, as follows:

[0200] y=Softmax(θd·x+b d ) (16)

[0201] Where x represents the output of LSTM, that is, the input of the fully connected layer, θd and bd are the weight matrix and bias matrix of the fully connected layer, y represents the final output of the fully connected layer, and Softmax is the activation function.

[0202] During the training process of the alarm information model, the AdamW algorithm is used to optimize the cost function loss, and the optimized loss function is used to make the prediction results of the alarm information model closer to the actual situation.

[0203] For massive alarm signal sequences, which have complex time dynamics and frequency domain characteristics, most of the previous optimizers use the Adam optimizer. In the training of different types of alarm signals collected in the centralized control system, the L2 regularization (weight decay) of the parameters is not ideal, which will lead to parameter bias problems, resulting in model distortion, affecting the minimization of the loss function, and affecting the subsequent prediction of the alarm information location and alarm information type. Therefore, the optimizer is replaced with the AdamW optimizer. The AdamW optimizer introduces a more effective weight decay mechanism that can better control the size of the alarm signal model parameters, thereby reducing the risk of overfitting. The regularization parameter makes the alarm signal prediction model smoother by penalizing large weight values, reduces the interference of noise on the model, and thus improves the generalization ability. Therefore, by using the AdamW optimizer to adjust the parameters of the model, the prediction model can better fit the loss function loss on the training data and converge to the optimal solution.

[0204] The loss function measures the difference between the model's predicted output and the actual alarm signal type, location, and time. In supervised learning, the model adjusts parameters by minimizing the loss function so that the model's prediction results are as close to the true value as possible. The gradient information of the loss function is used to update the model parameters in optimizer optimization. By calculating the partial derivative of the loss function with respect to each parameter, it is possible to determine in which direction the parameters should be adjusted to minimize the loss function. The loss function directly reflects the performance of the model on the training data. Lower loss values ​​generally mean that the model can predict the target variable more accurately, while high losses indicate that the model's prediction effect is poor or needs further optimization. During the training process, the loss function is also used to evaluate the current state of the model. By monitoring the changes in the loss function, it can be determined whether the model is gradually converging and whether the learning rate or other training parameters need to be adjusted.

[0205] By using the AdamW optimizer to adjust the model parameters, the model can better fit the loss function loss on the training data and converge to the optimal solution

[0206] Specifically, the output of each unit in each layer is calculated using forward propagation to obtain the cost function loss, as shown below:

[0207]

[0208] Among them, L(f(xi;θ),yi) is the cross loss function, N is the number of samples, xi is the sample value, f(xi;θ;bd) is the output result after the i-th sample is input into the five-layer network, θ and b are the weight matrix and bias matrix in the network, and yi is the expected output value of the corresponding sample;

[0209] Back propagation is used to calculate the deviation of the output of each unit in each layer. The error of the LSTM layer has two back propagation directions: one is to propagate to the upper layer; the other is to propagate back along time;

[0210] The gradient of each parameter is calculated by the bias term and the parameter value is updated;

[0211] Continuously iterate and correct the parameter value to make the loss function smaller and smaller until it is less than the pre-set number of iterations;

[0212] Finally, diagnosis and network update are performed based on actual test data.

[0213] The training process uses the AdamW algorithm; AdamW is an improved version of Adam that solves the bias problem of Adam in parameter optimization by introducing weight decay. In AdamW, regularization becomes:

[0214] θ t =θ t-1 -γλθ t-1 Among them, gt is the gradient of the current step, θt-1 is the model weight in the previous step, λ is the regularization coefficient, and r is the learning rate. Therefore, by attenuating the weight, it can also be ahead of Adam in terms of convergence speed.

[0215] 3. Robot inspection and feedback

[0216] The detection model is obtained based on the trained alarm information model to accurately obtain the actual equipment status and location information and complete the closed-loop inspection.

[0217] Step 3.1: According to the model, the alarm information collected by the robot is input as input information, the device information and location information are obtained, and then fed back to the robot;

[0218] Step 3.2: The intelligent inspection robot repairs and notifies the alarmed equipment based on the equipment information and location information;

[0219] Step 3.3: Complete the intelligent closed-loop inspection of the entire centralized control system.

[0220] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. 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 embodiments of the present invention.

Claims

1. An intelligent closed-loop linkage inspection method in a centralized control system, characterized in that: The following steps are involved: Divide the space to be inspected into multiple grids; define a label for each grid; the label can reflect the coordinates of obstacles in the grid; Randomly sample multiple points in the blank area of ​​the grid to generate potential path nodes and connect them. Then connect multiple sampling points in the grid to form local path segments. Multiple local path segments constitute a path network. Then, path planning is performed based on the path network. The inspection robot moves along the optimal path and obtains alarm signals in the inspection space. The alarm signals obtained by multiple inspection robots constitute an alarm data sequence. The alarm data sequence is used as input of the trained alarm information model, and outputs device status and location information; The inspection robot obtains the device status and location information output by the alarm information model and feeds it back to the host computer to complete the closed-loop inspection; The alarm information model includes an input layer, a Bi-LSTM layer, a fully connected layer, a softmax layer, and a classification output layer; Among them, the state activation function of the LSTM unit in the Bi-LSTM layer is tanh, and the gate activation function is sigmoid. The method for optimizing the path network is specifically as follows: A series of points are randomly sampled in the space to be inspected as potential path nodes through the PRM algorithm. The path nodes are randomly generated from the blank areas in the grid to ensure that the entire environment is covered. The local path planning algorithm A* algorithm is used to connect these sampling points in the grid to form local path segments of the path. Multiple local path segments constitute the path network of PRM, and path planning is performed based on the path network. Among them, the A* algorithm calculates the priority of each randomly adopted path node through the following formula; f(n)=g(n)+h(n) Where: f(n) is the comprehensive priority of node n; g(n) is the cost of node n from the starting point; h(n) is the estimated cost of node n from the end point, and h(n) is the heuristic function of the A* algorithm; The heuristic function h(n) is in the form of distance*multiple, that is: h(n)=h1(n)*(1+(h2(n) / h3(n))*a) Where h1(n) is the Euclidean distance from the current point to the target point, h2(n) is the Manhattan distance from the current point to the target point, h3(n) is the Manhattan distance from the starting point to the target point, and a is a variable multiple of the distance between the current node and the target node multiplied by; During the operation of the A* algorithm, the node with the smallest f(n) value is selected from the priority queue each time as the next node to be traversed; finally, a priority node set is obtained, and the nodes in the priority set are used to construct a graph-like path network.

2. The intelligent closed-loop linkage inspection method in the centralized control system according to claim 1 is characterized in that: The grid division method is: Multiple sets of sensors and cameras are deployed on site to collect real-time data on the operating status changes of each device in the inspection space; the processed operating status change data is added to the grid model to achieve the goal of real-time update of grid model information data; Assume that the length of the intelligent inspection workspace is M, the width is N, and the grids are all squares with a side length of L. At this time, the calculation expression for the number of grids is: R=(M / L)*(N / L) Where R represents the number of small grids in the grid model; workspace grid map i It can be expressed as: Map={map i ,map i =0 / 1} Use binary to indicate whether there are obstacles in the intelligent inspection workspace of the centralized control system, map i =0 indicates the free area for intelligent inspection; map i =1 indicates intelligent inspection of obstacle area.

3. The intelligent closed-loop linkage inspection method in the centralized control system according to claim 1 is characterized in that: The path planning method is: Through the constructed grid model, the objective function and variables of the intelligent inspection path planning task are determined and encoded; any part of the code is randomly selected as the initialization population of the improved genetic algorithm, the feasible path from the starting point of the initialization population to the target node is determined, and a feasible path encoding set is generated; the fitness of the individuals in the population is evaluated; fitness refers to the degree of advantage of an individual in the population, and its evaluation expression is: In the formula, F represents the fitness function; s represents the total number of individuals in the initialized population passing through the grid; D represents the sum of the straight-line distances between adjacent grids; Through calculation, the fitness of the individuals in the initialized population is obtained. The larger the fitness value, the higher the possibility of being selected, and vice versa. The fitness is tested to determine whether it meets the optimal convergence conditions of the global planning of the intelligent inspection path of the centralized control system. If it does not meet the optimal convergence conditions, a new population is generated through selection, crossover, and mutation operations, and then the fitness of the individuals in the population is evaluated again, and the above steps are repeated. If the optimal convergence conditions are met, the individuals in the population are decoded and the optimal solution after decoding is output, which is the optimal path for intelligent inspection.

4. The intelligent closed-loop linkage inspection method in the centralized control system according to claim 1 is characterized in that: When the LSTM unit processes the alarm data sequence, the value of the alarm data sequence at each moment is used as the input of an LSTM unit. Assume that in the Bi-LSTM layer, the input of a certain LSTM unit is: the input sequence at time t is xt, the output value ht-1 of the LSTM unit corresponding to time t-1 and its memory unit state ct-1; The output of the LSTM unit is: the output value ht of the LSTM unit corresponding to time t and the state ct of the memory unit at that time; The input gate, output gate, and forget gate are used as activation functions to control the mutual influence between the data input into the LSTM unit. The input gate controls the influence of xt on ct, the forget gate controls the influence of ct-1 on ct, and the output gate controls the influence of ct on ht. The calculation formula of the input gate is: The calculation formula of the output gate is: The calculation formula of the forget gate is: Among them, it represents the state of the input gate, ot represents the state of the output gate, ft represents the state of the forget gate, W is the weight matrix, b is the bias term, and the subscripts i, o, and f correspond to the input gate, output gate, and forget gate respectively; is the sigmoid activation function; Therefore, the output of the LSTM unit corresponding to time t is as follows: ht=ot⊙ψ(ct) ct=ft⊙ct-1+it⊙ψ(Wisht-1+Wisxt+bis) =ft⊙c t-1 +it⊙ψ(is) Among them, ψ is the tanh activation function, Wis is the input state weight matrix, bis is the input state bias term, and ⊙ represents element-wise multiplication; The fully connected layer takes the output of the LSTM layer as an input vector and outputs the probability distribution of the fault category, as follows: y=Softmax(θd·x+b d ) Where x represents the output of LSTM, that is, the input of the fully connected layer, θd and bd are the weight matrix and bias matrix of the fully connected layer, y represents the final output of the fully connected layer, and Softmax is the activation function.

5. A method for training an alarm information model, characterized in that: The following steps are involved: The output of each unit in each layer is calculated by forward propagation, and the cost function loss is obtained. The formula is as follows: Among them, L(f(xi;θ),yi) is the cross loss function, N is the number of samples, xi is the sample value, f(xi;θ;bd) is the output result after the i-th sample is input into the five-layer network, θ and b are the weight matrix and bias matrix in the network, and yi is the expected output value of the corresponding sample; Back propagation is used to calculate the deviation of the output of each unit in each layer. The error of the LSTM layer has two back propagation directions: one is to propagate to the upper layer; the other is to propagate back along time; The gradient of each parameter in the loss function is calculated by the bias term and the parameter value is updated; Continuously iterate and correct the parameter value to make the loss function smaller and smaller until it is less than the pre-set number of iterations; Finally, the actual test data is combined for diagnosis and network update. The training process uses the AdamW algorithm to optimize the parameters of the Bi-LSTM layer. In AdamW, regularization becomes: θ t =θ t-1 -γλθ t-1 Where gt is the gradient of the current step, θt-1 is the LSTM unit weight in the previous step, λ is the regularization coefficient, and r is the learning rate.

6. Intelligent closed-loop linkage inspection system, characterized in that: include: A grid division module for the space to be inspected divides the space to be inspected into multiple grids; Define labels for each grid; The label can reflect the coordinates of the obstacle in the grid; The path network construction module randomly samples multiple points in the blank area of ​​the grid to generate potential path nodes and connects them. Then, multiple sampling points are connected within the grid to form local path segments of the path. Multiple local path segments constitute a path network, and then path planning is performed based on the path network. Inspection module: the inspection robot moves along the optimal path and obtains alarm signals in the inspection space. The alarm signals obtained by multiple inspection robots constitute an alarm data sequence. An alarm information processing module, wherein the alarm data sequence is used as the input of a trained alarm information model and outputs device status and location information; Feedback module: the inspection robot obtains the device status and location information output by the alarm information model and feeds it back to the host computer to complete the closed-loop inspection; The alarm information model includes an input layer, a Bi-LSTM layer, a fully connected layer, a softmax layer, and a classification output layer; Among them, the state activation function of the LSTM unit in the Bi-LSTM layer is tanh, and the gate activation function is sigmoid. The method for optimizing the path network is specifically as follows: A series of points are randomly sampled in the space to be inspected as potential path nodes through the PRM algorithm. The path nodes are randomly generated from the blank areas in the grid to ensure that the entire environment is covered. The local path planning algorithm A* algorithm is used to connect these sampling points in the grid to form local path segments of the path. Multiple local path segments constitute the path network of PRM, and path planning is performed based on the path network. Among them, the A* algorithm calculates the priority of each randomly adopted path node through the following formula; f(n)=g(n)+h(n) Where: f(n) is the comprehensive priority of node n; g(n) is the cost of node n from the starting point; h(n) is the estimated cost of node n from the end point, and h(n) is the heuristic function of the A* algorithm; The heuristic function h(n) is in the form of distance*multiple, that is: h(n)=h1(n)*(1+(h2(n) / h3(n))*a) Where h1(n) is the Euclidean distance from the current point to the target point, h2(n) is the Manhattan distance from the current point to the target point, h3(n) is the Manhattan distance from the starting point to the target point, and a is a variable multiple of the distance between the current node and the target node multiplied by; During the operation of the A* algorithm, the node with the smallest f(n) value is selected from the priority queue each time as the next node to be traversed; finally, a priority node set is obtained, and the nodes in the priority set are used to construct a graph-like path network.

7. The intelligent closed-loop linkage inspection system according to claim 6 is characterized in that: The grid division method is: Multiple sets of sensors and cameras are deployed on site to collect real-time data on the operating status changes of each device in the inspection space; the processed operating status change data is added to the grid model to achieve the goal of real-time update of grid model information data; Assume that the length of the intelligent inspection workspace is M, the width is N, and the grids are all squares with a side length of L. At this time, the calculation expression for the number of grids is: R=(M / L)*(N / L) Where R represents the number of small grids in the grid model; the workspace grid mapi can be expressed as: Map = {mapi, mapi = 0 / 1} Binary system is used to indicate whether there are obstacles in the intelligent inspection workspace of the centralized control system. mapi=0 indicates the free area for intelligent inspection; mapi=1 indicates the obstacle area for intelligent inspection.

8. The intelligent closed-loop linkage inspection system according to claim 6 is characterized in that: The path planning method is: Through the constructed grid model, the objective function and variables of the intelligent inspection path planning task are determined and encoded; any part of the code is randomly selected as the initialization population of the improved genetic algorithm, the feasible path from the starting point of the initialization population to the target node is determined, and a feasible path encoding set is generated; the fitness of the individuals in the population is evaluated; fitness refers to the degree of advantage of an individual in the population, and its evaluation expression is: In the formula, F represents the fitness function; s represents the total number of individuals in the initialized population passing through the grid; D represents the sum of the straight-line distances between adjacent grids; Through calculation, the fitness of the individuals in the initialized population is obtained. The larger the fitness value, the higher the possibility of being selected, and vice versa. The fitness is tested to determine whether it meets the optimal convergence conditions of the global planning of the intelligent inspection path of the centralized control system. If it does not meet the optimal convergence conditions, a new population is generated through selection, crossover, and mutation operations, and then the fitness of the individuals in the population is evaluated again, and the above steps are repeated. If the optimal convergence conditions are met, the individuals in the population are decoded and the optimal solution after decoding is output, which is the optimal path for intelligent inspection.

9. The intelligent closed-loop linkage inspection system according to any one of claims 6 to 8, characterized in that: When the LSTM unit processes the alarm data sequence, the value of the alarm data sequence at each moment is used as the input of an LSTM unit. Assume that in the Bi-LSTM layer, the input of a certain LSTM unit is: the input sequence at time t is xt, the output value ht-1 of the LSTM unit corresponding to time t-1 and its memory unit state ct-1; The output of the LSTM unit is: the output value ht of the LSTM unit corresponding to time t and the state ct of the memory unit at that time; The input gate, output gate, and forget gate are used as activation functions to control the mutual influence between the data input into the LSTM unit. The input gate controls the influence of xt on ct, the forget gate controls the influence of ct-1 on ct, and the output gate controls the influence of ct on ht. The calculation formula of the input gate is: The calculation formula of the output gate is: The calculation formula of the forget gate is: Among them, it represents the state of the input gate, ot represents the state of the output gate, ft represents the state of the forget gate, W is the weight matrix, b is the bias term, and the subscripts i, o, and f correspond to the input gate, output gate, and forget gate respectively; is the sigmoid activation function; Therefore, the output of the LSTM unit corresponding to time t is as follows: ht=ot⊙ψ(ct) ct=ft⊙ct-1+it⊙ψ(Wisht-1+Wisxt+bis) =ft⊙c t-1 +it⊙ψ(is) Among them, ψ is the tanh activation function, Wis is the input state weight matrix, bis is the input state bias term, and ⊙ represents element-wise multiplication; The fully connected layer takes the output of the LSTM layer as an input vector and outputs the probability distribution of the fault category, as follows: y=Softmax(θd·x+b d ) Where x represents the output of LSTM, that is, the input of the fully connected layer, θd and bd are the weight matrix and bias matrix of the fully connected layer, y represents the final output of the fully connected layer, and Softmax is the activation function.

10. A training system for an alarm information model, characterized in that: include: The output of each unit in each layer is calculated by forward propagation, and the cost function loss is obtained. The formula is as follows: Among them, L(f(xi;θ),yi) is the cross loss function, N is the number of samples, xi is the sample value, f(xi;θ;bd) is the output result after the i-th sample is input into the five-layer network, θ and b are the weight matrix and bias matrix in the network, and yi is the expected output value of the corresponding sample; Back propagation is used to calculate the deviation of the output of each unit in each layer. The error of the LSTM layer has two back propagation directions: one is to propagate to the upper layer; the other is to propagate back along time; The gradient of each parameter in the loss function is calculated by the bias term and the parameter value is updated; Continuously iterate and correct the parameter value to make the loss function smaller and smaller until it is less than the pre-set number of iterations; Finally, combined with the actual test data, diagnosis and network update are performed. The training process uses the AdamW algorithm to optimize the parameters of the Bi-LSTM layer. In AdamW, regularization becomes: θ t =θ t-1 -γλθ t-1 , Where gt is the gradient of the current step, θt-1 is the LSTM unit weight in the previous step, λ is the regularization coefficient, and r is the learning rate.

Images