Robust fire early warning model based on multi-source time series data fusion

Abstract

The invention belongs to the technical field of fire early warning, traditional fire early warning mainly utilizes a sensor to carry out state detection and identification, although the sensor technology has noticeable progress in the aspect of fire early warning, many challenges still exist in the aspects of further shortening fire early warning time, reducing false alarm rate and missing report rate and the like, and aiming at the problems, the fire early warning method is provided. The invention provides a robust fire early warning model based on multi-source time series data fusion, which mainly depends on three types of sensors, namely a temperature sensor, a CO concentration sensor and a smoke concentration sensor, and utilizes time series data collected by the temperature sensor, the CO concentration sensor and the smoke concentration sensor to carry out fire early warning on the fire. According to the method, parameters of 12 dimensions are extracted from the parameters to serve as input of the neural network for multi-source data fusion, the situation of abnormal data collection is considered at the same time, and compared with other existing methods, the method has the advantages that the false alarm rate and the missing report rate of fire early warning are greatly reduced, and the fire early warning time is obviously shortened.

Description

Technical Field

[0001] The present invention relates to the field of fire warning technology, and in particular to a robust fire warning model based on multi-source time series data fusion. Background Art

[0002] At present, the mainstream methods for detecting fires include vision-based and sensor-based research. Compared with the use of visual methods to identify the occurrence of fires, sensor-based fire detection is more suitable for most environments. Using sensors for fire detection has the advantages of wide application range, low cost, simple maintenance, and fast response. At present, traditional fire sensors based on temperature, smoke and gas concentration are widely used for fire detection in indoor places in my country, among which smoke sensors account for the vast majority. However, this type of sensor can only detect a single environmental information and is easily interfered by other factors. In order to reduce the interference of environmental factors on sensor information and provide more timely and accurate alarms, we propose a robust fire warning model based on multi-source time series data fusion.

[0003] The mainstream smoke sensors on the market currently make fire judgments by monitoring the smoke concentration in a single area at a single moment. In order to improve the performance of the fire alarm system, the BP neural network is trained by collecting temperature, CO concentration, smoke concentration, etc. as input data to reduce the errors caused by interference from a single environmental factor. Although this method can significantly improve the accuracy of the alarm compared to the single sensor method, the existing method does not take into account the temporal evolution of sensor data and possible acquisition anomalies. If the actual value is abnormal, it will also cause a great deviation in the fire warning, making the false alarm and missed alarm rate of the alarm system still high. Therefore, it is urgent to develop a highly accurate and robust fire warning system based on the deficiencies of existing research. Summary of the invention

[0004] Based on the technical problems that the existing fire alarm system in the background technology has high false alarm rate, low accuracy and poor robustness, the present invention proposes a robust fire warning model based on multi-source time series data fusion.

[0005] The present invention proposes a robust fire warning model based on multi-source time series data fusion, including

[0006] Data exception processing module:

[0007] The fire warning system contains three types of sensors in hardware: temperature sensor, CO concentration sensor and smoke concentration sensor. The time series data collected at a certain time point is used Indicates, where l represents the sensor type and k represents the time series;

[0008] First, the collected time series data is used for anomaly detection and judgment. Is it within the confidence interval?

[0009]

[0010] k 1 , k 2 is the interval value set according to the actual situation. Beyond this range, it means The data is abnormal and needs to be processed. According to the data missing experiment, the autoregressive model has a good effect on the prediction of unknown data. So the autoregressive formula is used to predict the abnormal value, that is:

[0011]

[0012] In the autoregressive formula, ε(t) is the random error value. The autocorrelation coefficient represents the linear relationship between the observed value at the current moment and the observed value at the past p moments. After predicting the data at the abnormal moment, the predicted value Replace the original abnormal data. The above is the abnormal data processing module of the entire model;

[0013] Neural Network Module:

[0014] Obtain time series data of factors with obvious changes in the fire process from known data sets, including temperature, carbon monoxide concentration, and smoke concentration, and extract characteristic values ​​that can accurately reflect the data change process;

[0015] The input data is represents the value of sensor data at time l, L represents the number of different types of sensors, and K is the time label sequence of the data set;

[0016] Information extraction is performed on it. The entire information extraction content includes: moving average, benchmark error, predicted benchmark error, and cyclic trend index. Moving average is one of the well-known technical indicators, which is used to predict future data in time series analysis. Moving average is used to enhance the judgment of data changes. The formula of moving average is:

[0017]

[0018] u l (k) represents the moving average of sensor data l at time k;

[0019] Moving average is a key element in extracting time series changes. On this basis, the benchmark error and the benchmark error in the prediction stage can be calculated. The benchmark error is calculated by the difference between the current value of the sensor and the benchmark value of the environment. This can directly reflect the degree of deviation between the current data and the environmental benchmark. The benchmark value of the environment is replaced by the moving average. The formula for the moving benchmark error is expressed as:

[0020]

[0021] Indicates the l sensor data at time k with u l The difference between (k) and (k) is used as an input to the neural network;

[0022] In order to better predict the occurrence of fire in advance, the model judges by predicting the difference between the next stage data and the set benchmark. The predicted benchmark error uses the exponential smoothing formula, namely:

[0023]

[0024] in and The two constants have no direct relationship and are used to weight the historical value and the current value. It represents the benchmark error of sensor data at the prediction stage at time k. Then the model extracts the cyclic trend of the sensor's previous benchmark error, current benchmark error, and next predicted benchmark error. By using the feature that sgn(x) only returns -1, 0, and 1, it returns the cyclic trend indicators of different sensors:

[0025]

[0026] The cyclic trend index is the sum of the previous moment's benchmark error, the current moment's benchmark error, and the next moment's benchmark error under the sgn(x) function, which is specifically expressed as:

[0027]

[0028] It represents the cyclic trend index of sensor data l at time k, which is a comprehensive reflection of the trend change degree of the same sensor at the previous moment, the current moment and the next moment;

[0029] Since the method adopted by the model includes BP neural network, in the BP neural network, the model uses the S-type function as the activation function. In order to prevent the smaller values ​​in the output data from being swallowed by the larger data, the data after abnormal processing needs to be And the extracted information and Normalization is performed, and the specific normalization formula is as follows:

[0030]

[0031] in is the input vector, Indicates the normalized value of the parameter input, y min Represents the minimum value of the input vector, y max Represents the maximum value of the input vector;

[0032] The model will The normalized data is used as the input of the neural network, and the fire probability y is used as the output of the neural network. In view of the problems of low accuracy, high false alarm rate, poor robustness, etc. faced by traditional fire warning methods, the present invention proposes a robust fire warning model based on multi-source time series data fusion. The model uses the time series characteristics of sensor data to train a model with high accuracy, low false alarm rate and low missed alarm rate through a BP neural network. At the same time, the abnormal data collected by the sensor is estimated and replaced, which effectively improves the robustness of the fire warning. By making full use of the time series evolution law of multi-sensor data and the preprocessing of collected abnormal conditions, a fire warning model with high robustness, high accuracy and low false alarm rate is developed.

[0033] Preferably, the model considers the input of the neural network as: They respectively represent the i-th group of temperature, CO concentration, smoke concentration, temperature benchmark error, CO concentration benchmark error, smoke concentration benchmark error, temperature prediction benchmark error, CO concentration prediction benchmark error, smoke concentration prediction benchmark error, temperature cycle trend index, CO concentration cycle trend index, and smoke concentration cycle trend index, which are conducive to the formation of the model.

[0034] Preferably, the algorithm flow of the robust fire warning model based on multi-source time series data fusion includes initializing the network, taking the input data as the input vector of the BP neural network linear model, each group of input data, i.e. each row, corresponds to a fire probability label, and determining the network structure as a three-layer structure, namely, an input layer, a hidden layer, and an output layer, to facilitate subsequent operations.

[0035] Preferably, the following is the information of each layer:

[0036] Expected output Y i =[y i ] represents the fire probability of the i-th group of expected output;

[0037] The input to the hidden layer is:

[0038] The output of the hidden layer is:

[0039] The input to the output layer is:

[0040] The output of the output layer is:

[0041] The weights of the input layer and the hidden layer are w ij , the threshold is θ j ;

[0042] The weights v of the hidden layer and the output layer jt , the threshold is γ t ;

[0043] Randomly select a group of samples as input and target samples: the kth group of input And the corresponding expected output Y i =[y], select the error function e, set the calculation accuracy and number of learning times.

[0044] Preferably, forward propagation: passing the input data through the network, calculating the output of each neuron, and calculating the weighted input and the output of the activation function in each layer;

[0045] Calculate the input of each unit in the hidden layer:

[0046]

[0047] Output of the hidden layer:

[0048]

[0049] Input of each neuron in the output layer:

[0050]

[0051] Output of each neuron in the output layer:

[0052]

[0053] Preferably, calculating the error: comparing the output of the neural network with the actual target, and calculating the error using an error function;

[0054] Calculate the unit error between the actual output layer and the target output layer:

[0055] d k =(y k -c)c(1-c) (13)

[0056] Calculate the error of each unit in the middle layer:

[0057]

[0058] Preferably, back propagate: starting from the output layer, calculate the error gradient of each layer:

[0059] Use the chain rule to calculate the contribution of each parameter to the error, propagate the error gradient back to each layer, and update each parameter to minimize the error:

[0060] Correct the weights and thresholds from the output layer to the hidden layer:

[0061] v jt (N+1)=v jt (N)+αd t b j γ (15)

[0062] γ j (N+1)=γ j (N)+αd t (16)

[0063] Correct the weights and thresholds from the hidden layer to the input layer:

[0064]

[0065] θ j (N+1)=θ j (N)+α[(1-η)e j +ηe j ] (18)

[0066] Among them, α is the learning rate, 0<α<1.

[0067] Preferably, update iteration: determine the number of training iterations according to the loss of training data, and repeat forward propagation and back propagation until the number of learning times reaches a preset value.

[0068] Preferably, the model uses python to train and simulate neural networks, selects the public data set of the National Institute of Standards and Technology (NIST) as the training sample, sets the learning rate to 0.001, sets the training batch to 32, and cycles 1000 times to obtain the final model.

[0069] The beneficial effects of the present invention are as follows: in response to the problems of low accuracy, high false alarm rate, poor robustness, etc. faced by traditional fire warning methods, the present invention proposes a robust fire warning model based on multi-source time series data fusion, which utilizes the time series characteristics of sensor data to train a model with high accuracy, low false alarm rate, and low missed alarm rate through a BP neural network, and at the same time estimates and replaces the abnormal data collected by the sensor, effectively improving the robustness of the fire warning, and developing a fire warning model with high robustness, high accuracy, and low false alarm rate by making full use of the time series evolution laws of multi-sensor data and preprocessing of collected abnormal situations. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 A model flow chart of a robust fire warning model based on multi-source time series data fusion proposed by the present invention;

[0071] Figure 2 This is an algorithm flow chart of a robust fire warning model based on multi-source time series data fusion proposed by the present invention. DETAILED DESCRIPTION

[0072] The present invention will be further explained below in conjunction with specific embodiments.

[0073] Example

[0074] refer to Figure 1-2 In this embodiment, a robust fire warning model based on multi-source time series data fusion is proposed, including

[0075] Data exception processing module:

[0076] The fire warning system contains three types of sensors in hardware: temperature sensor, CO concentration sensor and smoke concentration sensor. The time series data collected at a certain time point is used Indicates, where l represents the sensor type and k represents the time series;

[0077] First, the collected time series data is used for anomaly detection and judgment. Is it within the confidence interval?

[0078]

[0079] k 1 , k 2 is the interval value set according to the actual situation. Beyond this range, it means The data is abnormal and needs to be processed. According to the data missing experiment, the autoregressive model has a good effect on the prediction of unknown data. So the autoregressive formula is used to predict the abnormal value, that is:

[0080]

[0081] In the autoregressive formula, ε(t) is the random error value. The autocorrelation coefficient represents the linear relationship between the observed value at the current moment and the observed value at the past p moments. After predicting the data at the abnormal moment, the predicted value Replace the original abnormal data. The above is the abnormal data processing module of the entire model;

[0082] Neural Network Module:

[0083] Obtain time series data of factors with obvious changes in the fire process from known data sets, including temperature, carbon monoxide concentration, and smoke concentration, and extract characteristic values ​​that can accurately reflect the data change process;

[0084] The input data is represents the value of sensor data at time l, L represents the number of different types of sensors, and K is the time label sequence of the data set;

[0085] Information extraction is performed on it. The entire information extraction content includes: moving average, benchmark error, predicted benchmark error, and comprehensive change index. Moving average is one of the well-known technical indicators, which is used to predict future data in time series analysis. Moving average is used to enhance the judgment of data changes. The formula of moving average is:

[0086]

[0087] u l (k) represents the moving average of sensor data l at time k;

[0088] Moving average is a key element in extracting time series changes. On this basis, the benchmark error and the benchmark error in the prediction stage can be calculated. The benchmark error is calculated by the difference between the current value of the sensor and the benchmark value of the environment. This can directly reflect the degree of deviation between the current data and the environmental benchmark. The benchmark value of the environment is replaced by the moving average. The benchmark error formula is expressed as:

[0089]

[0090] Indicates the l sensor data at time k with u l The difference between (k) and (k) is used as an input to the neural network;

[0091] In order to better predict the occurrence of fire in advance, the model judges by predicting the difference between the next stage data and the set benchmark. The predicted benchmark error uses the exponential smoothing formula, namely:

[0092]

[0093] in and The two constants have no direct relationship and are used to weight the historical value and the current value. It represents the moving reference error of the l sensor data in the prediction phase at time k. Then the model extracts the cyclic trend of the sensor's previous phase reference error, reference error, and reference error in the prediction phase. By using the feature that sgn(x) only returns -1, 0, and 1, it returns the cyclic trend indicators of different sensors:

[0094]

[0095] The cyclic trend index is the sum of the previous stage benchmark error, benchmark error and moving benchmark error in the prediction stage under the sgn(x) function. The cyclic trend index is expressed as:

[0096]

[0097] It represents the comprehensive moving reference error of the l sensor data at time k, which is a comprehensive reflection of the degree of change of the same sensor at the current time;

[0098] Since the method adopted by the model includes BP neural network, in the BP neural network, the model uses the S-type function as the activation function. In order to prevent the smaller values ​​in the output data from being swallowed by the larger data, the data after abnormal processing needs to be And the extracted information and Normalization is performed, and the specific normalization formula is as follows:

[0099]

[0100] in is the input vector, Indicates the normalized value of the parameter input, y min Represents the minimum value of the input vector, y max Represents the maximum value of the input vector;

[0101] The model will The normalized data is used as the input of the neural network, and the fire probability y is used as the output of the neural network;

[0102] The model considers the input of the neural network as:

[0103] They represent the i-th group of temperature, CO concentration, smoke concentration, temperature benchmark error, CO concentration benchmark error, smoke concentration benchmark error, temperature prediction benchmark error, CO concentration prediction benchmark error, smoke concentration prediction benchmark error, temperature cycle trend index, CO concentration cycle trend index, and smoke concentration cycle trend index, respectively, which is conducive to the formation of the model;

[0104] The algorithm flow of the robust fire warning model based on multi-source time series data fusion includes initializing the network, taking the input data as the input vector of the BP neural network linear model, each group of input data, i.e. each row, corresponds to a fire probability label, and determining the network structure as a three-layer structure, namely, input layer, hidden layer, and output layer, to facilitate subsequent operations;

[0105] The following is the information of each layer:

[0106] Expected output Y i =[y i ] represents the fire probability of the i-th group of expected output;

[0107] The input to the hidden layer is:

[0108] The output of the hidden layer is:

[0109] The input to the output layer is:

[0110] The output of the output layer is:

[0111] The weights of the input layer and the hidden layer are w ij , the threshold is θ j ;

[0112] The weights v of the hidden layer and the output layer jt , the threshold is γ t ;

[0113] Randomly select a group of samples as input and target samples: the kth group of input And the corresponding expected output Y i = [y], select the error function e, set the calculation accuracy and the number of learning times;

[0114] Forward propagation: pass the input data through the network, calculate the output of each neuron, and calculate the weighted input and the output of the activation function in each layer;

[0115] Calculate the input of each unit in the hidden layer:

[0116]

[0117] Output of the hidden layer:

[0118]

[0119] Input of each neuron in the output layer:

[0120]

[0121] Output of each neuron in the output layer:

[0122]

[0123] Calculate the error: Compare the output of the neural network with the actual target and use the error function to calculate the error;

[0124] Calculate the unit error between the actual output layer and the target output layer:

[0125] d k =(y k -c)c(1-c) (13)

[0126] Calculate the error of each unit in the middle layer:

[0127]

[0128] Backward propagation: Starting from the output layer, calculate the error gradient of each layer:

[0129] Use the chain rule to calculate the contribution of each parameter to the error, propagate the error gradient back to each layer, and update each parameter to minimize the error:

[0130] Correct the weights and thresholds from the output layer to the hidden layer:

[0131] v jt (N+1)=v jt (N)+αd t b j Y (15)

[0132] γ j (N+1)=γ j (N)+αd t (16)

[0133] Correct the weights and thresholds from the hidden layer to the input layer:

[0134]

[0135] θ j (N+1)=θ j (N)+α[(1-η)e j +ηe j ] (18)

[0136] Among them, α is the learning rate, 0<α<1;

[0137] Update iteration: Determine the number of training iterations based on the loss of training data, and repeat forward propagation and back propagation until the number of learning times reaches the preset value;

[0138] The model uses Python to train and simulate neural networks. The public data set of the National Institute of Standards and Technology (NIST) is selected as the training sample. The learning rate is set to 0.001, the training batch is set to 32, and the cycle is repeated 1000 times to obtain the final model.

[0139] In view of the problems of low accuracy, high false alarm rate and poor robustness faced by traditional fire warning methods, the present invention proposes a robust fire warning model based on multi-source time series data fusion. The model uses the time series characteristics of sensor data to train a model with high accuracy, low false alarm rate and low missed alarm rate through a BP neural network. At the same time, it estimates and replaces the abnormal data collected by the sensor, effectively improving the robustness of the fire warning.

[0140] First, anomaly detection is performed based on the time series data collected by the temperature, CO concentration, and smoke concentration sensors. If the data value of a certain time period is not within the confidence interval we set, it is necessary to use the interpolation algorithm to estimate the value of the time period to replace the abnormal value. Then, the measurement data, benchmark error, predicted benchmark error, and cyclic trend index of the three sensors are used as neural network inputs, and the fire probability is used as a label for training. The trained model can eventually achieve the purpose of predicting whether a fire will occur with high accuracy. According to the temperature, CO concentration, smoke concentration and other data collected by multiple sensors, a highly robust and high-precision prediction algorithm based on multi-source time series data fusion is proposed by making full use of the time series evolution law of sensor data and the preprocessing of collection anomalies. The key to the fire warning model with high accuracy and low false alarm rate is as follows: for the temperature, CO concentration, smoke concentration and other data collected by the sensor, a baseline value is set for each type of data, and then the difference between each time point and the baseline value (benchmark error) is calculated based on the collected data, the difference between the data at the next time point and the baseline value (prediction baseline error) and their cyclic trend indicators are predicted to reflect the relevant characteristics of this type of data at the current moment, and use it as the input data of the BP neural network for training; for the situation where abnormal data may appear in the temperature, CO concentration and smoke concentration collected by the sensor, the autoregressive function is used to predict the abnormal data and replace the original abnormal data, so as to reduce the false alarm probability of the fire warning.

[0141] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A robust fire warning model based on multi-source time series data fusion, characterized in that: include Data exception processing module: The fire warning system contains three types of sensors in hardware: temperature sensor, CO concentration sensor and smoke concentration sensor. The time series data collected at a certain time point is used Indicates, where l represents the sensor type and k represents the time series; First, the collected time series data is used for anomaly detection and judgment. Is it within the confidence interval? k1 and k2 are interval values ​​set according to actual conditions. Beyond this range, it means The data is abnormal and needs to be processed. According to the data missing experiment, the autoregressive model has a good effect on the prediction of unknown data. So the autoregressive formula is used to predict the abnormal value, that is: In the autoregressive formula, ε(t) is the random error value. The autocorrelation coefficient represents the linear relationship between the observed value at the current moment and the observed value at the past p moments. After predicting the data at the abnormal moment, the predicted value Replace the original abnormal data. The above is the abnormal data processing module of the entire model; Neural Network Module: Obtain time series data of factors with obvious changes in the fire process from known data sets, including temperature, carbon monoxide concentration, and smoke concentration, and extract characteristic values ​​that can accurately reflect the data change process; The input data is represents the value of sensor data l at time k, L represents the number of sensors of different types, and K is the time label sequence of the data set; Information extraction is performed on it. The entire information extraction content includes: moving average, benchmark error, predicted benchmark error, and cyclic trend index. Moving average is one of the well-known technical indicators, which is used to predict future data in time series analysis. Moving average is used to enhance the judgment of data changes. The formula of moving average is: u l (k) represents the moving average of sensor data l at time k; Moving average is a key element in extracting time series changes. On this basis, the benchmark error and the benchmark error in the prediction stage can be calculated. The benchmark error is calculated by the difference between the current value of the sensor and the benchmark value of the environment. This can directly reflect the degree of deviation between the current data and the environmental benchmark. The benchmark value of the environment is replaced by the moving average. The benchmark error formula is expressed as: Indicates the l sensor data at time k with u l The difference between (k) and (k) is used as an input to the neural network; In order to better predict the occurrence of fire in advance, the model judges by predicting the difference between the data at the next moment and the set benchmark. The predicted benchmark error uses the exponential smoothing formula, namely: in and The two constants have no direct relationship and are used to weight the historical value and the current value. It represents the benchmark error of sensor data at the prediction stage at time k. Then the model extracts the cyclic trend of the sensor's previous benchmark error, current benchmark error, and next predicted benchmark error. By using the feature that sgn(x) only returns -1, 0, and 1, it returns the cyclic trend index of the same sensor: The cyclic trend index is the sum of the previous moment's benchmark error, the benchmark error, and the benchmark error in the forecast stage under the sgn(x) function. The cyclic trend index is expressed as: It represents the cyclic trend index of sensor data l at time k, which is a comprehensive reflection of the trend change degree of the same sensor at the previous moment, the current moment and the next moment; Since the method adopted by the model includes BP neural network, in the BP neural network, the model uses the S-type function as the activation function. In order to prevent the smaller values ​​in the output data from being swallowed by the larger data, the data after abnormal processing needs to be And the extracted information and Normalization is performed, and the specific normalization formula is as follows: in is the input vector, Indicates the normalized value of the parameter input, y min Represents the minimum value of the input vector, y max Represents the maximum value of the input vector; The model will The normalized data is used as the input of the neural network, and the fire probability y is used as the output of the neural network.

2. A robust fire warning model based on multi-source time series data fusion according to claim 1, characterized in that: The model considers the input of the neural network as: They respectively represent the i-th group of temperature, CO concentration, smoke concentration, temperature benchmark error, CO concentration benchmark error, smoke concentration benchmark error, temperature prediction benchmark error, CO concentration prediction benchmark error, smoke concentration prediction benchmark error, temperature cycle trend index, CO concentration cycle trend index, and smoke concentration cycle trend index.

3. The robust fire warning model based on multi-source time series data fusion according to claim 1 is characterized in that: The algorithm flow of the robust fire warning model based on multi-source time series data fusion includes initializing the network, taking the input data as the input vector of the BP neural network linear model, each group of input data, i.e. each row, corresponds to a fire probability label, and determining the network structure as a three-layer structure, namely, an input layer, a hidden layer, and an output layer.

4. The robust fire warning model based on multi-source time series data fusion according to claim 3 is characterized in that: The following is the information of each layer: Expected output Y i =[y i ] represents the fire probability of the i-th group of expected output; The input to the hidden layer is: The output of the hidden layer is: The input to the output layer is: The output of the output layer is: The weights of the input layer and the hidden layer are w ij , the threshold is θ j ; The weights v of the hidden layer and the output layer jt , the threshold is γ t ; Randomly select a group of samples as input and target samples: the kth group of input And the corresponding expected output Y i =[y], select the error function e, set the calculation accuracy and number of learning times.

5. A robust fire warning model based on multi-source time series data fusion according to claim 4, characterized in that: Forward propagation: pass the input data through the network, calculate the output of each neuron, and calculate the weighted input and the output of the activation function in each layer; Calculate the input of each unit in the hidden layer: Output of the hidden layer: Input of each neuron in the output layer: Output of each neuron in the output layer:

6. A robust fire warning model based on multi-source time series data fusion according to claim 5, characterized in that: Calculate the error: Compare the output of the neural network with the actual target and use the error function to calculate the error; Calculate the unit error between the actual output layer and the target output layer: d k =(y k -c)c(1-c) (13) Calculate the error of each unit in the middle layer:

7. The robust fire warning model based on multi-source time series data fusion according to claim 6 is characterized in that: Backward propagation: Starting from the output layer, calculate the error gradient of each layer: Use the chain rule to calculate the contribution of each parameter to the error, propagate the error gradient back to each layer, and update each parameter to minimize the error: Correct the weights and thresholds from the output layer to the hidden layer: v jt (N+1)=v jt (N)+αd t b j γ (15)γ j (N+1)=γ j (N)+αd t (16) Correct the weights and thresholds from the hidden layer to the input layer: i j (N+1)=θ j (N)+α[(1-η)e j +ηe j ] (18) Among them, α is the learning rate, 0<α<1.

8. The robust fire warning model based on multi-source time series data fusion according to claim 7 is characterized in that: Update iteration: Determine the number of training iterations based on the loss of the training data, and repeat forward propagation and back propagation until the number of learning times reaches the preset value.

9. The robust fire warning model based on multi-source time series data fusion according to claim 8 is characterized in that: The model uses Python to train and simulate neural networks. The public data set of the National Institute of Standards and Technology (NIST) is selected as the training sample. The learning rate is set to 0.001, the training batch is set to 32, and the cycle is repeated 1000 times to obtain the final model.

Images