A multi-space environment fire source positioning method and system based on Attention-LSTM

Abstract

The present invention discloses a multi-space fire source location method based on Attention-LSTM. This method belongs to the field of fire source location technology and includes the following steps: obtaining a multi-space, long-term measurement point temperature dataset using a temperature acquisition module; detecting outliers in the dataset using an outlier detection module; supplementing the dataset using a data completion module and normalizing the dataset; and inputting the data into a fire source location prediction module to obtain a fire location prediction value. The present invention uses an Attention-LSTM neural network to detect multi-space ambient temperatures, noting the importance of data at different times and thus intelligently and accurately identifying the fire source location. This enables intelligent urban fire protection, improves the efficiency of information transmission, enhances rescue capabilities, and reduces losses caused by fires.

Description

Technical Field

[0001] The present invention belongs to the field of fire safety technology, and specifically relates to a multi-space environment fire source positioning method based on Attention-LSTM. Background Art

[0002] Building fires can cause casualties and property damage. With the increasing diversification of functions within large buildings (hotels, office buildings, and apartment buildings), their interiors are divided into multiple independent spaces, each operating independently of the others, creating a multi-space environment. Due to this complex internal environment, a serious fire can spread to other spaces, causing incalculable damage. Therefore, multi-space building environments present significant challenges for firefighting design and fire rescue efforts.

[0003] During the design phase, fire scenarios are common in multi-space building environments, making it difficult to properly delineate fire and smoke zones. Before conducting firefighting and rescue operations at a multi-space building fire scene, it's necessary to assess the on-site personnel situation and quickly locate the fire source and develop a corresponding rescue plan. However, due to the complex structure of multi-space environments, fire commanders struggle to quickly locate the fire source and develop effective firefighting and rescue measures. Therefore, rapid and accurate fire location technology is a critical component of firefighting design and rescue in multi-space environments, not only providing targets for firefighting measures but also improving firefighting and rescue efficiency.

[0004] Currently, commonly used fire source location methods are generally based on image theory. For example, a Chinese patent (patent number: 201510569144.3) uses infrared imaging technology to locate the fire source using pixels around the highest temperature point. Another Chinese patent (patent number: 202310004375.4) uses an image acquisition module to capture indoor wall images during fire source location. However, image location methods are susceptible to obstruction by objects, resulting in large positioning errors. Installing cameras in living environments may raise privacy concerns. Therefore, the use of environmental monitoring sensors (such as temperature and smoke concentration) can be considered for fire source location.

[0005] With the development of computational fluid dynamics, fire simulation software (PyroSim) can realistically simulate typical fire scenarios within buildings. By inputting the fire location, intensity, and environmental conditions, the spread of a fire can be simulated. This model has been validated by fire experiments both domestically and internationally. However, numerical simulations must be based on known fire scenario conditions. Many conditions in real-world fires are unknown. Therefore, predicting fires within buildings based on the results of typical fire scenario research is crucial for the design of fire protection systems.

[0006] In recent years, artificial neural networks have demonstrated remarkable results in object recognition and feature prediction. Neural networks consist of layers of nodes, including an input layer, one or more hidden layers, and an output layer. Each node is interconnected and has associated weights and thresholds. When a node's output value exceeds the specified threshold, the node is activated and data is sent to the next layer. Neural network models continuously adjust the weights of each layer according to specific rules until the weights converge, marking the end of training. Trained neural network models can be used to solve a variety of real-life and scientific problems.

[0007] A Chinese patent (patent number: 202310004375.4) uses an artificial neural network to identify wall images after a fire and locate the fire source, enabling intelligent fire source location identification during fire investigations. However, the method used in this patent indiscriminately focuses on the overall input data, ignoring the varying importance of each input data. The prediction results obtained after applying the same weight to all data are less accurate. Furthermore, this patent is primarily intended for post-incident investigations of fire accidents.

[0008] Traditional image-based fire detection methods are susceptible to external interference, leading to false detections and inability to make timely and accurate judgments on fires of varying manifestations. Single intelligent algorithms ignore the correlation and importance of data, and can only make predictions based on fixed calculation patterns, resulting in low accuracy.

[0009] The real-time changes of the temperature field in the area near the fire source in the fire, using the changes in temperature data as the input value of the artificial neural network prediction, will be mentioned in the on-site fire location.

[0010] Therefore, the present invention proposes an intelligent and efficient fire location method and system that integrates multiple intelligent algorithms for firefighting design and rescue. By assigning different weights to input data based on their importance, this method achieves high-precision predictions in multiple scenarios. This method effectively combines fire numerical simulation technology with artificial neural networks, facilitating rapid and accurate fire location in multiple spatial environments. Furthermore, the fire location system can be integrated with building FAS systems to enable rapid fire rescue in multiple spatial environments, minimizing loss of life and property. Summary of the Invention

[0011] To solve the problems existing in the prior art, the present invention provides a multi-space fire source positioning method and system based on Attention-LSTM, which uses an environmental detector to measure the ambient temperature, locates the fire source through the Attention-LSTM model, and realizes intelligent fire source location identification. It solves the problems mentioned in the above background technology that data collection is easily affected by the external environment and the low accuracy caused by indiscriminate calculation of all input data.

[0012] To achieve the above objectives, the present invention provides the following technical solution: a multi-space fire source positioning method based on Attention-LSTM, comprising the following steps:

[0013] Step S1: The temperature acquisition module collects data from each temperature measurement point arranged in multiple spaces, and the collected data from each temperature measurement point is used as a multi-space long-term measurement point temperature data set;

[0014] Step S2, performing outlier detection on the temperature measurement point data in the measurement point temperature data set by an outlier detection module to obtain an outlier value of the temperature data set;

[0015] Step S3, using the data completion module to supplement the abnormal values ​​of the temperature data set in step 2 above, to obtain a complete temperature data set and perform normalization processing;

[0016] In step S4, the normalized temperature data set is input into the fire source location prediction module, and the fire source location prediction value is obtained through back propagation.

[0017] Preferably, the outlier detection module in step S2 uses the isolation forest algorithm. The isolation forest algorithm is to segment out abnormal samples by building a "tree" or itree, relying on random features. The outlier detection module has a trained itree, and the itree training steps include:

[0018] S21, use the fire simulation software PyroSim to simulate fire data and construct a fire dataset, including training set, validation set, and test set;

[0019] S22, randomly select a time point from the fire data set and randomly select a data points x at that time point as subsamples;

[0020] S23, randomly generating cutting points in succession within the current sub-sample data range, setting the height of a single tree, and completing the generation of a single tree;

[0021] S24, repeating steps S22 and S23 for the remaining data points at the same time point in the fire dataset to complete the generation of multiple trees;

[0022] S25, repeating steps S22, S23, and S24 for the data points at each time point in the fire dataset to complete the training and save the parameters of the isolated forest;

[0023] S26, calculate the outlier score s by comprehensively calculating the results of each tree for each sample x through each isolated tree. If the outlier score is close to 1, the data point is an outlier; if the outlier score is close to 0.5, it is not an outlier; if the outlier score of each point is around 0.5, the sample does not have an outlier. The specific formula is as follows:

[0024]

[0025] Where h(x) is the height of x in each tree, c(Ψ) is the average path length for a given sample number Ψ, which is used to standardize the path length h(x) of sample x, and E is the mathematical expectation.

[0026] Preferably, the data completion module in step S3 adopts a time similarity completion method, and the specific formula is as follows:

[0027]

[0028] In the formula, N represents the total number of non-abnormal data at a certain moment, x s,t represents the non-abnormal data at the t-th measurement point at the s-th second, x` s,t Represents abnormal data, x t For abnormal data x` s,t Corrected data. Add time constraint:

[0029] s'-d1≤s≤s'+d2

[0030] Where d1 and d2 are set to 1 and 1 respectively.

[0031] Preferably, the specific steps of simulating fire data using the fire simulation software PyroSim in step S21 are as follows:

[0032] S211: Build a fire combustion model and set parameters, including room geometry, fire source location, fire source heat release rate, and fire source area;

[0033] S212, set t temperature measuring points and record the temperatures at the measuring points;

[0034] S213, setting simulation time according to the fire development state;

[0035] S213, obtain the simulated measurement point temperature data set ﹛x s,t ﹜,x s,t Represents the temperature data x at the tth measuring point at the sth second;

[0036] S214: Feature extraction is performed on the measurement point data according to the number of rooms n, and the temperature measurement point data is packaged into a data set according to the extracted feature dimensions, where the vertical direction represents different times, the horizontal direction represents different measurement points, and multiple measurement points are represented as one room:

[0037]

[0038] In step S3, the data is normalized after completion.

[0039] Preferably, the fire source location prediction module in step S4 is a trained long short-term memory network (Attention-LSTM) model combined with an attention mechanism, and the training steps of the attention mechanism-long short-term memory network (Attention-LSTM) model include:

[0040] S41, build a long short-term memory network, set the input feature dimension, time step, dropout, hidden layer state dimension, number of stacking layers, activation function, and set the loss function;

[0041] S42, long short-term memory network weight initialization;

[0042] S43, inputting the normalized data set to train the long short-term memory network and continuously adjusting the weights of the long short-term memory network;

[0043] S44, setting a similarity function and an attention mechanism function, and inputting the long short-term memory network hidden layer state for calculation;

[0044] S45, output to the fully connected layer and output the result;

[0045] S46, complete the training and save the parameters of the attention mechanism-long short-term memory network.

[0046] Preferably, the specific steps of constructing the long short-term memory network in step S41 include: setting the input feature dimension, i.e., the number of temperature measurement points and time, according to the extracted feature dimension, setting the time step according to the data time range, setting the hidden layer state dimension according to the input feature dimension, setting the dropout initial value, the number of stacking layers, the activation function, and the loss function.

[0047] Preferably, the specific steps of training the long short-term memory network in step S43 include:

[0048] S431, input the normalized data training set, and obtain the error of the output layer, i.e., the loss value, through forward propagation;

[0049] S432, reversely solve the weights of each layer based on the output layer error to complete the weight update;

[0050] S433, input the validation set and obtain the loss value;

[0051] S434, repeat step S431, step S432 and step S433 until the loss value of the verification data set no longer decreases or reaches the number of data training times.

[0052] Preferably, the specific steps of adding the attention mechanism in step S44 include:

[0053] S441, setting a similarity function Score function to calculate the similarity score between the final hidden layer output of the LSTM network in step S43 and the hidden layer output at each time point;

[0054] S442, set the SoftMax function to calculate the attention weight at each time point;

[0055] Preferably, the final result output step of step S45 is to multiply the weight obtained in step S442 by the hidden layer output of the corresponding LSTM at different time points, and input it into the fully connected layer to obtain the final prediction result.

[0056] To achieve the above objectives, the present invention further provides the following technical solution: a fire source location system based on Attention-LSTM, the fire source location system specifically comprising:

[0057] Temperature acquisition module, used to collect temperature values ​​at different times at each location;

[0058] An outlier detection module, used to detect and remove outliers in temperature data;

[0059] The data completion module is used to complete the deleted abnormal values ​​of the temperature data and normalize the overall temperature data;

[0060] The fire source location prediction module is used to take the normalized data as input and obtain the fire location prediction value through forward propagation and back propagation.

[0061] The beneficial effects of the present invention are:

[0062] (1) The present invention can automatically identify the location of the fire source based on the on-site environmental data (temperature), thus avoiding the situation where personnel investigation at the fire scene in multiple space environments is inefficient and dangerous;

[0063] (2) The present invention combines the Attention mechanism with LSTM to consider the importance of fire-affected features at different locations in the environment at different times, calculate more weights for important information, and ultimately achieve accurate prediction of the fire source location;

[0064] (3) The present invention avoids errors that are easily caused by camera obstruction in multi-space environments in most image recognition processes, and at the same time avoids privacy issues that may arise from installing cameras, without affecting people's daily lives;

[0065] (4) The fire source location system of the present invention can be trained over a long period of time. After large-scale application, the prediction system can be trained with more data, thereby continuously improving the prediction accuracy. The present invention has a data collection function and can provide certain basic data for my country's fire data system, thereby improving the level of information technology in the field of fire prevention and control. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 This is a system module diagram of the present invention.

[0067] Figure 2 Schematic diagram of the arrangement of temperature measurement points according to an embodiment of the present invention.

[0068] Figure 3 This is the outlier detection module of the present invention.

[0069] Figure 4 Schematic diagram of the attention mechanism structure of the present invention.

[0070] Figure 5 Schematic diagram of the neural network structure.

[0071] Figure 6 Flowchart for model data training. DETAILED DESCRIPTION

[0072] The present invention is described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are merely some embodiments of the present invention, not all embodiments. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are intended to fall within the scope of protection of the present invention.

[0073] See also Figure 1-5 ,The present invention provides a technical solution: a multi-space fire source positioning system based on Attention-LSTM, such as Figure 1 As shown, the fire source positioning system specifically includes: temperature acquisition module, outlier detection module, data completion module, and fire source location prediction module.

[0074] Temperature acquisition module, used to collect temperature values ​​at different locations at different times;

[0075] An outlier detection module, used to detect and remove outliers in temperature data;

[0076] The data completion module is used to complete the deleted abnormal values ​​of the temperature data and normalize the overall temperature data;

[0077] The fire source location prediction module is used to take the normalized data as input and obtain the fire location prediction value through forward propagation and back propagation.

[0078] A multi-space fire source positioning method based on Attention-LSTM, characterized by comprising the following steps:

[0079] S1, using the temperature measurement points arranged in multiple spaces as temperature acquisition modules, obtains a multi-space long-term measurement point temperature data set, such as Figure 2 shown.

[0080] S2, perform outlier detection on the temperature data of the measuring points through the outlier detection module to obtain the outliers of the temperature data set.

[0081] The outlier detection module in step S2 uses the isolation forest algorithm. The isolation forest algorithm is to segment out abnormal samples by building a "tree" or itree and relying on random features. Figure 3 As shown, the outlier detection module has a trained iTree. The training steps of iTree include:

[0082] S21, use the fire simulation software PyroSim to simulate fire data and construct fire data training set, validation set, and test set;

[0083] Use PyroSim to build a fire combustion model and set parameters, including room geometry, fire source location, fire source heat release rate and fire source area, such as Figure 2 As shown, set t temperature measurement points, record the temperature of the measurement points, set the simulation time, and obtain the temperature data set of the measurement points after simulation. s,t ﹜,x s,t Represents the temperature data x at the s-second of the t-th measurement point. Feature extraction is performed on the measurement point data according to the number of rooms n. The temperature measurement point data is packaged into a dataset according to the extracted feature dimensions, where the vertical direction represents different times, the horizontal direction represents different measurement points, and multiple measurement points are represented as one room:

[0084]

[0085] S22, randomly select a time point from the fire data set and randomly select a data points x at that time point as subsamples;

[0086] S23, randomly generating cutting points in succession within the current sub-sample data range, setting the height of a single tree, and completing the generation of a single tree;

[0087] S24, repeating steps S22 and S23 for the remaining data points at the same time point in the fire dataset to complete the generation of multiple trees;

[0088] S25, repeating steps S22, S23, and S24 for the data points at each time point in the fire dataset to complete the training and save the parameters of the isolated forest;

[0089] S26, calculate the outlier score s by comprehensively calculating the results of each tree for each sample x through each isolated tree. If the outlier score is close to 1, the data point is an outlier; if the outlier score is close to 0.5, it is not an outlier; if the outlier score of each point is around 0.5, the sample does not have an outlier. The specific formula is as follows:

[0090]

[0091] Where h(x) is the height of x in each tree, c(Ψ) is the average path length for a given sample number Ψ, which is used to standardize the path length h(x) of sample x, and E is the mathematical expectation.

[0092] S3, using the data completion module to supplement the above abnormal values, obtain a complete temperature data set and perform normalization processing.

[0093] The data completion module in step S3 adopts the time similarity completion method, and the specific formula is as follows:

[0094]

[0095] In the formula, N represents the total number of non-abnormal data at a certain moment, x s,t represents the non-abnormal data at the t-th measurement point at the s-th second, x` s,t Represents abnormal data, x t For abnormal data x` s,t Corrected data. Add time constraint:

[0096] s'-d1≤s≤s'+d2

[0097] Where d1 and d2 are set to 1 and 1 respectively.

[0098] The data normalization formula in step S3 is as follows, which can improve the convergence speed and model accuracy of the fire source location prediction model.

[0099]

[0100] Where x norm is the normalized data, x max is the maximum value of the original data set.

[0101] S4, input the normalized data set into the fire source location prediction module, and obtain the fire source location prediction value through back propagation.

[0102] The fire source location prediction module in step S4 is a trained long short-term memory network (Attention-LSTM) model combined with the attention mechanism, such as Figure 5 As shown in the figure, the training steps of the Attention-LSTM model include:

[0103] S41, build a long short-term memory network, set the input feature dimension, time step, dropout, hidden layer state dimension, number of stacking layers, activation function, and loss function;

[0104] Set the input feature dimension, i.e., the number of temperature measurement points and time, according to the extracted feature dimension, set the time step according to the data time range, and set the hidden layer state dimension according to the input feature dimension;

[0105] The dropout is initially set to 0.1 to prevent data overfitting;

[0106] The number of stacking layers is set to 2 to process multi-dimensional and multi-time data;

[0107] The activation function is selected from sigmoid function and tanh function.

[0108] The sigmoid function is a smooth step function that can convert any value to (0, 1), where 0 means it cannot be passed (discarded) and 1 means it can be passed (retained). Different data changes in emphasis over time. The sigmoid function can activate the forget gate of the LSTM model to forget unnecessary information. The formula is as follows:

[0109]

[0110] The output range of the tanh function is [-1, 1]. The input gate of the LSTM model is activated to process the input information. The formula is as follows:

[0111]

[0112] The loss function uses the cross entropy function:

[0113]

[0114] Where y is the predicted probability distribution, is the true probability distribution.

[0115] S42, long short-term memory network weight initialization;

[0116] S43, inputting the normalized data set to train the long short-term memory network and continuously adjusting the weights of the long short-term memory network;

[0117] The specific steps of training the long short-term memory network in step S43 include:

[0118] S431, input the normalized data training set, and obtain the error of the output layer, i.e., the loss value, through forward propagation;

[0119] S432, reversely solve the weights of each layer based on the output layer error to complete the weight update;

[0120] S433, input the validation set and calculate the loss value;

[0121] S434, repeat step S431, step S432 and step S433 until the loss value of the verification data set no longer decreases or reaches the number of data training times.

[0122] S44, build attention mechanism, such as Figure 4 As shown, similarity function and attention mechanism function are set, and the state value of the hidden layer of the long short-term memory network is input for calculation;

[0123] Set the similarity function Score function to calculate the final hidden layer output h of the LSTM network in step S43 i And the hidden layer output h at each time point k The similarity score is as follows:

[0124]

[0125] Set the SoftMax function to calculate the attention weight at each time point,

[0126]

[0127] S45, output to the fully connected layer and output the result;

[0128] The final result output step of step S45 is to combine the weight obtained in step S442 with the corresponding LSTM hidden layer output h at different time points. i The multiplication and summation are used to obtain the output vector p of the attention layer, which is input into the fully connected layer to obtain the final prediction result vector y. The final result vector y represents the probability of a fire in each room.

[0129] S46, complete the training, save the attention mechanism - long short-term memory network parameters, the final model training process is as follows Figure 6 shown.

[0130] Example Function and Effect

[0131] The method of the present invention can predict the location of a fire source based on fire scene environmental data, avoiding errors caused by image recognition. The present invention not only predicts the location of a fire source in multiple rooms of any spatial structure, but also detects outliers, alerting users to issues with on-site environmental monitoring equipment. While the present invention has been described in detail with reference to the aforementioned embodiments, it will be apparent to those skilled in the art that any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A multi-space fire source positioning method based on Attention-LSTM, characterized in that: The steps include: Step S1: The temperature acquisition module collects data from each temperature measurement point arranged in multiple spaces, and the collected data from each temperature measurement point is used as a multi-space long-term measurement point temperature data set; Step S2, performing outlier detection on the temperature measurement point data in the measurement point temperature data set by an outlier detection module to obtain outliers of the temperature data set; the outlier detection module in step S2 uses an isolation forest algorithm; the isolation forest algorithm is to segment out abnormal samples by establishing a "tree" (iTree) based on random features; the outlier detection module has a trained iTree; In step S3, the abnormal values ​​of the temperature data set in step S2 are supplemented by a data completion module to obtain a complete temperature data set and perform normalization processing. The data completion module in step S3 adopts a time similarity completion method, and the specific formula is as follows: In the formula, N represents the total number of non-abnormal data at a certain moment, x s,t represents the non-abnormal data at the t-th measurement point at the s-th second, x` s, t represents abnormal data, x t For abnormal data x` s, tCorrected data; add time constraint: s'-d1≤s≤s'+d2 Where d1, d2, are set to 1, 1 respectively; After the data is completed in step S3, normalization is performed; Step S4: input the normalized temperature data set into the fire source location prediction module, and obtain the fire source location prediction value through back propagation; The fire source location prediction module in step S4 is a trained Attention-LSTM model that combines an attention mechanism with a long short-term memory network. The training steps of the Attention-LSTM model include: S41, build a long short-term memory network, set the input feature dimension, time step, dropout, hidden layer state dimension, number of stacking layers, activation function, and loss function; S42, long short-term memory network weight initialization; S43, inputting the normalized data set to train the long short-term memory network and continuously adjusting the weights of the long short-term memory network; S44, setting a similarity function and an attention mechanism function, and inputting the long short-term memory network hidden layer state value for calculation; S45, output to the fully connected layer and output the result; S46, complete the training and save the parameters of the attention mechanism-long short-term memory network; The specific steps of adding the attention mechanism in step S44 include: S441, setting a similarity function Score function to calculate the similarity score between the final hidden layer output of the LSTM network in step S43 and the hidden layer output at each time point; S442, set the SoftMax function to calculate the attention weight at each time point; The final result output step of step S45 is to multiply the weight obtained in step S442 by the hidden layer output of the corresponding LSTM at different time points, and input it into the fully connected layer to obtain the final prediction result.

2. The multi-space fire source positioning method based on Attention-LSTM according to claim 1 is characterized in that: The training steps of itree include: S21, using fire simulation software to simulate fire data and construct fire data training set, validation set, and test set; S22, randomly select a time point from the fire data set and randomly select a data points x at that time point as subsamples; S23, randomly generating cutting points in succession within the current sub-sample data range, setting the height of a single tree, and completing the generation of a single tree; S24, repeating steps S22 and S23 for the remaining data points at the same time point in the fire dataset to complete the generation of multiple trees; S25, repeating steps S22, S23, and S24 for the data points at each time point in the fire dataset to complete the training and save the parameters of the isolated forest; S26, calculate the outlier score s by comprehensively calculating the results of each tree for each sample x through each isolated tree; if the outlier score is close to 1, the data point is an outlier; if the outlier score is close to 0.5, it is not an outlier; if the outlier score of each point is around 0.5, the sample does not have an outlier; the specific formula is as follows: Where h(x) is the height of x in each tree, c(Ψ) is the average path length for a given sample number Ψ, which is used to standardize the path length h(x) of sample x, and E is the mathematical expectation.

3. The multi-space fire source positioning method based on Attention-LSTM according to claim 2 is characterized in that: The specific steps of using the fire simulation software PyroSim to simulate fire data in step S21 are as follows: S211: Build a fire combustion model and set parameters, including room geometry, fire source location, fire source heat release rate, and fire source area; S212, set t temperature measuring points and record the temperatures at the measuring points; S213, setting simulation time; S214, obtain the simulated measurement point temperature data set ﹛x s,t ﹜,x s,t Represents the temperature data x at the tth measuring point at the sth second; S215 , extracting features from the measurement point data according to the number of rooms n, and packaging the temperature measurement point data into a data set according to the extracted feature dimensions, where the vertical direction represents different times, the horizontal direction represents different measurement points, and multiple measurement points are represented as one room.

4. The multi-space fire source positioning method based on Attention-LSTM according to claim 1 is characterized in that: The specific steps of constructing the long short-term memory network in step S41 include: setting the input feature dimension, i.e., the number of temperature measurement points and time, according to the extracted feature dimension, setting the time step according to the data time range, setting the hidden layer state dimension according to the input feature dimension, setting the dropout initial value, the number of stacking layers, the activation function, and the loss function.

5. The multi-space fire source positioning method based on Attention-LSTM according to claim 1 is characterized in that: The specific steps of step S43 of training the long short-term memory network include: S431, input the normalized data training set, and obtain the error of the output layer, i.e., the loss value, through forward propagation; S432, reversely solve the weights of each layer based on the output layer error to complete the weight update; S433, input the validation set and obtain the loss value; S434, repeat step S431, step S432 and step S433 until the loss value of the verification data set no longer decreases or reaches the number of data training times.

Images