Fire identification method based on LSTM model and multi-sensor data fusion

Abstract

The application provides a fire identification method based on an LSTM model and multi-sensor data fusion, which comprises the following steps: collecting smoke concentration data, temperature data and combustible gas concentration data in a scene to form multivariate time series data; sequentially performing data cleaning, data normalization and data correlation enhancement operations to obtain preprocessed data; calculating characteristic values and characteristic vectors based on a Gaussian radial basis kernel function, screening characteristic vectors with variance contribution rates satisfying a preset condition to obtain fusion features; performing MK test on the time series in the preprocessed data to calculate a standard normal statistical variable as a trend factor; inputting the fusion features and the trend factor into an LSTM model combined with an attention mechanism, processing long sequence information through a gate structure, giving high weights to key information combined with the attention mechanism, and outputting a fire identification result. In the application, the sensitivity of fire monitoring can be effectively improved and the false alarm rate can be reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to a fire identification method based on an LSTM model and multi-sensor data fusion. BACKGROUND

[0002] The loss caused by fire is often huge. Therefore, improving the early warning, detection and extinguishing ability of fire has important significance for reducing the loss of life and property caused by fire. The occurrence of fire usually has the characteristics of suddenness and rapid spread. Although the traditional fire detector can monitor the fire, it has certain limitations in the accuracy and rapidity of flame identification. These detectors often need the fire source to reach a certain area or intensity to trigger the alarm, thereby limiting their identification ability of the initial fire. In the initial stage of fire occurrence, the detector may not detect the presence of flame in time, especially in the case where the fire occurrence site is hidden or the environment is complex, it is more difficult for the traditional technology to accurately determine the source and location of the fire, and it is difficult to start the fire control equipment in the shortest time, which delays the rescue opportunity and leads to the spread of fire and the expansion of loss. Therefore, an efficient and accurate fire warning and identification system becomes the key to solving this problem.

[0003] The existing achievements still have many limitations. Although the traditional physical sensor is widely used, it is easy to be disturbed by the environment, has a high false alarm rate, and has a lagging response, which is particularly prominent in complex scenes. The motion trajectory of smoke changes quickly, and the shape and color are diverse, which leads to the fact that the accuracy of smoke detection does not meet the requirements of real-time detection, and false positives and false negatives are prone to occur. Using deep learning algorithm for fire identification can obtain good identification effect, but the algorithm complexity is high, the network parameter quantity is large, and it is difficult to deploy on embedded devices, which also limits further development. These problems affect the application of fire warning and identification technology in actual scenes. SUMMARY

[0004] The main purpose of the present application is to provide a fire identification method based on an LSTM model and multi-sensor data fusion, which aims to overcome the defects of false positives and false negatives in the current fire identification method.

[0005] To achieve the above purpose, the present application provides a fire identification method based on an LSTM model and multi-sensor data fusion, comprising the following steps:

[0006] Collecting smoke concentration data, temperature data and flammable gas concentration data in the scene to form multi-variable time series data;

[0007] Performing data cleaning, data normalization and data correlation enhancement operations on the multi-variable time series data in sequence to obtain preprocessed data;

[0008] Dimensionality reduction fusion is performed on the preprocessed data, eigenvalues and eigenvectors are calculated based on a Gaussian radial basis kernel function, and an eigenvector meeting a preset condition in terms of variance contribution rate is screened to obtain fused features;

[0009] Mann-Kendall test is performed on the time series in the preprocessed data, and a standard normal statistical variable is calculated as a trend factor;

[0010] The fused features and the trend factor are input into an LSTM model combined with an attention mechanism, long sequence information is processed through the gate structure of the LSTM model, key information is given a high weight through the attention mechanism, and a fire identification result is output.

[0011] Further, a smoke sensor, a temperature sensor and a flammable gas sensor are used to collect smoke concentration data, temperature data and flammable gas concentration data in the scene at a preset sampling frequency.

[0012] Further, the data cleaning includes using forward filling or mean replacement to process missing values, removing outliers through Z-score method or interquartile range method, and applying moving average filtering to smooth data;

[0013] Min-Max normalization or standardization method is used for data normalization;

[0014] The data correlation enhancement includes constructing a sliding time series window, calculating combined features and temperature change rate, concatenating multi-sensor data into a vector, calculating a Pearson correlation coefficient and adding data enhancement disturbance.

[0015] Further, the calculation formula of the Gaussian radial basis kernel function is:

[0016]

[0017] wherein, are feature parameter sequences, denote the Euclidean distance between them: is a parameter controlling the width of the Gaussian distribution;

[0018] Eigenvalues and eigenvectors are calculated, and the most important eigenvectors are selected according to the principal component variance contribution rate; the preset condition is that the sum of the variance contribution rates is greater than 95%, and The value is 15.

[0019] Further, the Mann-Kendall test on the time series in the preprocessed data includes:

[0020] Based on the time series in the preprocessed data , calculate the statistic S,

[0021]

[0022]

[0023] Calculate the variance Var(s) = n(n-1)(2n+5) / 18, when n≥10, calculate the standard normal statistic variable Z;

[0024]

[0025] By , determine whether there is a significant trend in the time series, for the confidence level;

[0026] If Z is greater than 0, the time series is an upward trend; if Z is less than 0, the time series is a downward trend.

[0027] Further, the calculation formula of the LSTM model comprises:

[0028] forget gate ;

[0029] input gate ;

[0030] alternative cell state ;

[0031] output gate ;

[0032] cell state update ;

[0033] output ;

[0034] wherein, is the weight matrix of the forget gate to leave information, is the weight matrix of the input gate to measure the importance of the input, is the weight matrix of the input gate to adjust the comprehensive input, is the weight matrix of the output gate to measure how much comprehensive input is transmitted to the hidden layer;

[0035] , , , respectively represent the bias of the forget gate, the input gate, the comprehensive input and the output gate;

[0036] represents the result of the forget gate at time t; represents the input of the cell at time t; For the softmax activation function:

[0037] Indicates an alternative cell state: Indicates the cell state at the previous time; Indicates the updated state of the cell; Indicates the hidden layer output at the previous time; Indicates the output at this time t.

[0038] Further, the attention mechanism converts the input sequence element encoding into a query, key and value vector through a weight matrix , , , calculates the similarity and normalizes, and obtains the attention output according to the formula

[0039]

[0040] Indicates the i-th element of the input sequence, is obtained by encoding , indicating the value obtained after the original i-th element is encoded, and then multiplied by the weight matrix , , , , matrix, is used to measure the association degree between the current and other , represents the association degree between and the query, indicates the key information of ;

[0041] is obtained by calculating and and then normalizing, indicating the similarity between and , multiplied by to obtain the output .

[0042] Further, the LSTM model adopts a multi-layer stacking structure to extract input data features through different layers.

[0043] The application also provides a fire identification device based on an LSTM model and multi-sensor data fusion, comprising:

[0044] ​​The collection unit is used for collecting smoke concentration data, temperature data and combustible gas concentration data in a scene to form multivariate time series data;

[0045] The processing unit is used for sequentially performing data cleaning, data normalization and data correlation enhancement operations on the multivariate time series data to obtain preprocessed data;

[0046] The fusion unit is used for dimension reduction fusion on the preprocessed data, calculating characteristic values and characteristic vectors based on a Gaussian radial basis kernel function, screening characteristic vectors with variance contribution rates satisfying preset conditions to obtain fusion features;

[0047] The inspection unit is used for performing Mann-Kendall inspection on time series in the preprocessed data to calculate a standard normal statistical variable as a trend factor;

[0048] The output unit is used for inputting the fusion features and the trend factor into an LSTM model combined with an attention mechanism, processing long sequence information through a gate structure of the LSTM model, giving high weights to key information combined with the attention mechanism, and outputting a fire identification result.

[0049] The fire identification method based on the LSTM model and the multi-sensor data fusion provided by the application comprises the following steps: collecting smoke concentration data, temperature data and combustible gas concentration data in a scene to form multivariate time series data; sequentially performing data cleaning, data normalization and data correlation enhancement operations on the multivariate time series data to obtain preprocessed data; performing dimension reduction fusion on the preprocessed data, calculating characteristic values and characteristic vectors based on a Gaussian radial basis kernel function, screening characteristic vectors with variance contribution rates satisfying preset conditions to obtain fusion features; performing Mann-Kendall inspection on time series in the preprocessed data to calculate a standard normal statistical variable as a trend factor; inputting the fusion features and the trend factor into an LSTM model combined with an attention mechanism, processing long sequence information through a gate structure of the LSTM model, giving high weights to key information combined with the attention mechanism, and outputting a fire identification result. In the application, multiple sensors are used to collect information including gas, temperature and smoke, and an MK inspection and an LSTM algorithm combined with an attention mechanism are used to process the multiple types of information provided by the sensors, which can effectively improve the sensitivity of fire monitoring and reduce the false alarm rate. Dimension reduction fusion is performed on multi-source heterogeneous sensor information, which can reduce information redundancy and reduce the amount of calculation. The Mann-Kendall inspection method is added before the prediction of the LSTM model, which is beneficial to the judgment of fire conditions in the initial stage. The attention mechanism is added to the LSTM model to better retain long-time information and ignore irrelevant information, thereby forcing the network to automatically extract important information, and further improving the fire identification and prediction accuracy of the model. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 is a fire identification method based on an LSTM model and multi-sensor data fusion in an embodiment of the present application, and a step schematic diagram thereof is shown in FIG. 1;

[0051] Figure 2 is a fire identification device structure block diagram based on an LSTM model and multi-sensor data fusion in an embodiment of the present application;

[0052] Figure 3 is a structural schematic block diagram of a computer device in an embodiment of the present application.

[0053] The implementation, functional features and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0054] Referring to Figure 1 In an embodiment of the present application, a fire identification method based on an LSTM model and multi-sensor data fusion is provided, which includes the following steps:

[0055] Step S1, collecting smoke concentration data, temperature data, and combustible gas concentration data in a scene to form multi-variable time series data;

[0056] Step S2, sequentially performing data cleaning, data normalization and data correlation enhancement operations on the multi-variable time series data to obtain preprocessed data;

[0057] Step S3, performing dimensionality reduction fusion on the preprocessed data, calculating eigenvalues and eigenvectors based on a Gaussian radial basis kernel function, and screening eigenvectors with variance contribution rates satisfying a preset condition to obtain fusion features;

[0058] Step S4, performing Mann-Kendall test on the time series in the preprocessed data, and calculating a standard normal statistical variable as a trend factor;

[0059] Step S5, inputting the fusion features and the trend factor into an LSTM model combined with an attention mechanism, processing long sequence information through a gate structure of the LSTM model, giving high weights to key information combined with the attention mechanism, and outputting a fire identification result.

[0060] In the present embodiment, as described above in step S1, first, key physical feature data at the time of fire occurrence is collected through multiple types of sensors to provide comprehensive and continuous raw data support for subsequent analysis, and the specific implementation details are as follows:

[0061] Three types of core sensors are selected: smoke sensor, temperature sensor, and flammable gas sensor. These three types of sensors correspond to the characteristic features of fire. Smoke generation, temperature rise, and flammable gas leakage or release can avoid misjudgment caused by the one-sidedness of single sensor data and capture fire signals from multiple dimensions.

[0062] The output data of each sensor is adapted to the actual application scenario. The smoke sensor outputs an analog voltage of 0-5V (the voltage size is positively correlated with the smoke concentration) or a concentration value of 0-1000ppm (the value is close to 0 under normal air, and will rise sharply in a fire); the temperature sensor outputs a real-time temperature reading of -20-100°C (the value is stable under normal room temperature, and will quickly rise in a fire); the flammable gas sensor outputs an analog voltage of 0-5V or a concentration value of 0-5000ppm (mainly for CO, CH4, and other gases related to fire, the value is low under normal conditions, and will rise in a fire or gas leakage).

[0063] Considering the real-time demand and noise control demand of fire, the sampling frequency is set to collect data every 1-5 seconds. Low frequency may miss the subtle changes of the initial characteristics of the fire, and high frequency may introduce a large amount of redundant noise data. This frequency range can balance the timeliness and reliability of the data.

[0064] The collected data form a multivariate time series data in the form of "timestamp + multi-sensor readings", which can reflect the dynamic change process of smoke, temperature, and flammable gas concentration over time (such as the trend of slow rise of each parameter in the initial stage of fire), and meets the input requirements of the subsequent LSTM model for time series data.

[0065] As described in step S2 above, for the problems such as missing, noise, and scale difference in the multi-sensor raw data, a series of operations are used to optimize the data quality and enhance the correlation between the data of each sensor, laying a foundation for subsequent fusion and modeling. The technical logic of each sub-operation is as follows:

[0066] Data cleaning: mainly solves the completeness and accuracy of the original data. For missing values, forward filling (using the valid data of the adjacent previous time) or mean replacement (filling with the average value of the historical data of the sensor) is used, and if the missing values exceed the set threshold, a sensor failure alarm is triggered; for abnormal values (such as temperature suddenly exceeding 200°C, which is likely to be a sensor failure or interference), Z-score method (based on the overall distribution of data to judge abnormalities) or interquartile range method (based on the four quartile range to filter extreme values) is used to remove; at the same time, moving average filtering technology is applied to smooth the data sequence and eliminate the interference of transient fluctuations (such as temporary jump of smoke concentration caused by air flow) on subsequent analysis.

[0067] Data normalization: solve the problem of inconsistent scales of multi-sensor data. Two common methods are used: one is Min-Max normalization, which scales all data to the interval of 0-1, ensuring that the numerical range of different sensor data is unified; the other is standardization, which processes data into a distribution with mean 0 and variance 1, enhancing the statistical correlation between different sensor data. And perform normalization operation independently for each sensor data, to avoid that the extreme value of a certain sensor (such as the large fluctuation of combustible gas concentration) covers the effective change of other sensors, and ensure that each feature has balanced weight in model training.

[0068] Data correlation enhancement: solve the problem of isolation of multi-sensor data, and mine the implicit correlation between features. Specifically, it is achieved in three ways: first, build a sliding time series window to convert discrete single-time data into continuous sequence data, so that smoke, temperature, and combustible gas data at the same time node are associated; second, carry out feature engineering to calculate combined features such as smoke value x combustible gas value (reflecting the synergistic change effect of the two types of parameters), add temperature change rate (capture the dynamic rising trend of temperature over time), and concatenate multi-sensor data into a vector containing smoke, temperature, and combustible gas values; third, calculate the Pearson correlation coefficient of any two sensor data (quantify the correlation strength between features), and add a small amount of noise or slight time shift to achieve data enhancement, improving the model's ability to adapt to data variations in real scenarios.

[0069] As in step S3 above, to address the information redundancy and large computational load of the pre-processed multi-source heterogeneous data (smoke, temperature, and combustible gas data), kernel principal component analysis (KPCA) is used to achieve dimensionality reduction and fusion, which not only preserves key feature information but also reduces the computational burden of subsequent models. The technical details are as follows:

[0070] When a fire occurs, features such as smoke concentration and combustible gas concentration often exhibit nonlinear changes (e.g., slow initial rise and rapid growth in the middle period). Traditional linear dimensionality reduction methods (such as ordinary PCA) cannot effectively extract such nonlinear correlation features. KPCA, however, maps data to a high-dimensional space through a kernel function and then performs linear dimensionality reduction in the high-dimensional space, which better adapts to the nonlinear characteristics of fire features and extracts more effective fusion features.

[0071] The Gaussian radial basis function (GRBF) is chosen as the kernel function of KPCA. This kernel function can effectively map nonlinearly separable fire features to a high-dimensional space, achieving effective separation of features. During calculation, the similarity between two feature parameter sequences is measured (represented by the Euclidean distance), and a hyperparameter that controls the width of the Gaussian distribution is used (experimentally verified to be optimal at a value of 15, balancing feature discrimination and generalization).

[0072] The characteristic values and characteristic vectors obtained after KPCA processing are screened, and the characteristic vectors are selected according to the preset condition that the sum of the variance contribution rates is greater than 95%. This standard can ensure that more than 95% of the original data information is retained, while redundant features are removed, greatly reducing the input dimension and computational complexity of the subsequent LSTM model, and improving the efficiency of the model.

[0073] As described in step S4 above, to address the difficulty of identifying early-stage fires (features do not reach the traditional alarm threshold but show an upward trend), the Mann-Kendall (MK) test is used to extract the trend features of the time series, generate trend factors, and assist the LSTM model in early fire detection. The technical principles and operational details are as follows:

[0074] MK test is a non-parametric statistical method that does not require data to follow a normal distribution or have a linear trend. It can adapt to weak trends and non-linear changes in temperature, smoke concentration, and other features at the early stage of a fire. At the same time, it is not sensitive to periodic disturbances in time series data (such as daily temperature changes), and can stably capture the core upward trend, avoiding false positives caused by periodic fluctuations in the environment.

[0075] First, for the three types of time series after preprocessing: temperature, smoke concentration, and flammable gas concentration, compare each data point with all previous data points, count the number of subsequent data points greater than, equal to, and less than the previous data points, and obtain the statistical quantity S reflecting the upward / downward trend of the sequence. Then, based on the total amount of data (which must satisfy data amount ≥ 10 to ensure statistical reliability), calculate the variance of statistical quantity S. Finally, based on statistical quantity S and variance, calculate the standard normal statistical variable Z, which is used as the trend factor. Z greater than 0 represents an upward trend in the sequence (consistent with the characteristics of the early stage of a fire), and Z less than 0 represents a downward trend. If the absolute value of Z is greater than the critical value corresponding to the confidence level, it indicates a significant trend. Even if the temperature, smoke concentration, and other features do not reach the alarm threshold of traditional detectors, if the MK test shows a significant upward trend in Z, it may indicate the presence of early-stage fire, overcoming the lagging defect of traditional detectors that require reaching the threshold to trigger an alarm, and achieving earlier fire detection.

[0076] As described in step S5 above, the fusion features (core static features) and trend factors (dynamic trend features) are combined and input into the model. Through the long sequence processing capability of LSTM and the key information focusing capability of attention mechanism, accurate fire detection is achieved. The model design and working principle are as follows:

[0077] The fire feature data is a sequence data that changes over time (such as the continuous process of temperature rising from normal to high), and the traditional neural network is difficult to handle the long-term dependence problem in long sequence (such as the correlation between the slow rise of temperature in the early stage of fire for 10 minutes and the subsequent concentration explosion). While LSTM can selectively retain or discard time series information through the gating structure of forget gate, input gate and output gate, and can better adapt to the learning needs of fire time series features.

[0078] The forget gate is responsible for deciding how much of the core information at the previous time (such as the previous temperature trend) is retained to the current time, and the output value is controlled between 0-1 through a specific activation function. The closer the value is to 1, the more historical information is retained; the input gate cooperates with the candidate cell state to determine how much new information (such as the fusion features and trend factors at the current time) is included in the current core information; the cell state update link integrates the historical information retained by the forget gate and the new information included by the input gate to form new core information; the output gate determines how much of the current core information is output to the hidden layer to provide feature support for subsequent classification. At the same time, a multi-layer LSTM stacking structure is adopted to extract single sensor change features in shallow layers and multi-feature collaborative change features in deep layers, improving the depth of feature learning.

[0079] There is a lot of irrelevant information in fire time series data (such as normal fluctuations in the environment), and attention mechanism can improve model recognition accuracy by giving high weight to key information and low weight to irrelevant information. The implementation is as follows: after encoding each element in the LSTM input sequence, it is converted into a query vector, a key vector, and a value vector. The query vector is used to measure the association between the current element and other elements, the key vector reflects the matching degree of the element and the query, and the value vector stores the key information of the element; by calculating the similarity between the query vector and the key vector, the weight of each element is obtained and normalized; finally, the attention output is obtained by combining the weight and the value vector. For example, give high weight to the collaborative sequence of "temperature rise + combustible gas concentration rise" in the early stage of fire, and give low weight to the accidental fluctuations of single sensor, to strengthen the influence of core features on the recognition result.

[0080] The model finally outputs the recognition result of "fire" or "non-fire", combining the static information of the fusion features (the current state of each parameter) and the dynamic information of the trend factor (the change trend of each parameter), to realize accurate judgment of the early stage and development stage of fire. At the same time, through multi-sensor fusion and attention mechanism, the false positive rate is greatly reduced, and the recognition reliability in complex scenarios is improved.

[0081] In the embodiment, a plurality of sensors are used to collect information including gas, temperature, smoke, etc., and the MK test and the LSTM algorithm with attention mechanism are used to process the multi-class information provided by the sensors, so as to effectively improve the sensitivity of fire monitoring and reduce the false alarm rate; the dimensionality reduction and fusion of the multi-source heterogeneous sensor information can reduce information redundancy and reduce the amount of calculation; the Mann-Kendall test method is added before the prediction of the LSTM model, which is beneficial to the fire condition judgment in the initial stage; the attention mechanism is added to the LSTM model, so that the model can better retain long-time information and ignore irrelevant information, thereby forcing the network to automatically extract important information, and further improving the fire identification and prediction accuracy of the model.

[0082] In an embodiment, a smoke sensor, a temperature sensor and a flammable gas sensor are used to collect smoke concentration data, temperature data and flammable gas concentration data in the scene at a preset sampling frequency.

[0083] In an embodiment, the data cleaning includes using forward filling or mean replacement to process missing values, removing outliers by Z-score method or interquartile range method, and applying moving average filtering to smooth the data.

[0084] The data normalization is performed by using Min-Max normalization or standardization method.

[0085] The data correlation enhancement includes constructing a sliding time series window, calculating the combined features and temperature change rate, concatenating the multi-sensor data into a vector, calculating the Pearson correlation coefficient, and adding data enhancement disturbance.

[0086] In the embodiment, the missing values in the fire scene may be caused by sensor failure, and the data continuity and failure warning need to be considered. The forward filling is to fill the missing value with the reading of the previous effective time (for example, the temperature value of the previous 1 second is used to replace the missing value), which is suitable for short-term sensor failure and smooth data fluctuation, and can maximize the retention of time series continuity. The mean replacement is to fill the missing value with the average value of a period of historical effective data (for example, the missing value is filled with the average temperature of the past 1 minute), which is suitable for the case that the missing time is slightly longer but the data is stable overall. At the same time, a threshold alarm needs to be set, if the missing value exceeds the set number of times (for example, 5 consecutive sampling missing), the sensor failure is determined and the alarm is triggered, so as to avoid invalid data entering the subsequent process.

[0087] Readings of temperature greater than 200°C are likely to be noise (normal fire scenarios typically have temperatures in the range of -20-100°C) and need to be removed through outlier detection. The Z-score method calculates the deviation of each data point from the overall data mean (in units of standard deviation) and determines an outlier if the deviation is more than ±3 standard deviations. This method is suitable for data that is approximately normally distributed (e.g. temperature data in a normal environment). The interquartile range method sorts the data and removes extreme values that are more than 1.5 interquartile ranges below the lower quartile or more than 1.5 interquartile ranges above the upper quartile. This method is more suitable for non-normally distributed data (e.g. occasional sudden jumps in smoke concentration). Both methods can be selected based on the characteristics of the sensor data to ensure that invalid noise is removed.

[0088] Sensor data can have transient fluctuations (e.g. a sudden rise in smoke sensor readings caused by air flow). These fluctuations are not fire signals and need to be removed through filtering. Moving average filtering selects a fixed length sliding window (e.g. 3 consecutive sampling data) and calculates the average of the data in the window as the smoothed value at the current time. This method processes the entire sequence by sliding the window, preserving the overall trend of temperature and smoke concentration (e.g. slow rise in the early stages of a fire), while filtering out transient disturbances and avoiding misleading the model's judgment with such noise.

[0089] Different sensor data have large scale differences (e.g. smoke concentration is in ppm, temperature is in °C, and voltage is in V). If not handled, one indicator may have too much influence on the model, while another indicator may be masked, so normalization is needed to achieve scale unification:

[0090] Min-Max normalization scales all data from each sensor to the [0, 1] interval. For example, temperature data is converted from "-20-100°C" to a value between 0 and 1. The core formula is to calculate the proportion by the maximum and minimum values of the data, preserving the relative size relationship between the data. This method is suitable for all sensors, especially for smoke concentration and flammable gas concentration, which have a clear numerical range. It can intuitively represent the relative height of data at different times (e.g. a normalized value of 0.8 represents a value close to the upper limit of the sensor range).

[0091] Standardization is a process of processing data to have a mean of 0 and a variance of 1. By eliminating the dimensionality of the data, different sensor data are statistically equivalent (e.g. 1 standard deviation of temperature and 1 standard deviation of smoke concentration represent the same degree of fluctuation). The role of standardization is to enhance the correlation between different sensors. For example, after standardization, the synchronous rising trend of temperature and smoke concentration is more easily captured by the model. At the same time, each sensor needs to be processed independently to avoid the influence of extreme values of one sensor (e.g. sudden rise in flammable gas concentration) on the normalization results of other sensors, ensuring that each feature has an equal weight in model training.

[0092] Data correlation is enhanced to solve the problem of multi-sensor data isolation. This is to break the limitations of single sensor data, mine the implicit correlation between smoke, temperature, and flammable gas, and provide a feature set rich in information for subsequent fusion analysis. This is achieved through five types of operations:

[0093] A sliding time series window is constructed, which requires converting discrete single-time data into continuous sequences. For example, set the window length to 5 (corresponding to 5 samples, i.e., 5-25 seconds of data), combine the smoke, temperature, and flammable gas data of every 5 consecutive time points into a sequence, so that the model can learn the time series correlation between the previous few seconds of data and the current data (e.g., the linkage between a slow increase in temperature over the previous 3 seconds and an increase in smoke concentration at the current time).

[0094] Calculate the combined feature and temperature change rate. The combined feature (e.g., "smoke value x flammable gas value") is used to reflect the synergistic effect of the two types of parameters. Smoke and flammable gas often rise simultaneously at the initial stage of a fire. This combined feature can amplify this synergistic signal. The temperature change rate is calculated by dividing the temperature difference between adjacent time points by the time interval (e.g., if the temperature rises from 25°C to 27°C in 1 second, the change rate is 2°C / s). This captures the temperature rise speed and avoids missing the initial fire situation due to relying solely on static temperature values.

[0095] Serial multi-sensor data as a vector. The smoke concentration, temperature, and flammable gas concentration data at the same time are directly concatenated into a multi-dimensional vector (e.g., smoke value, temperature value, flammable gas value). This directly integrates multi-source information and allows the model to learn the combined state of multiple parameters at the same time point (e.g., "low smoke + normal temperature + low flammable gas" corresponds to a normal state, and "medium smoke + high temperature + medium flammable gas" corresponds to the initial stage of a fire).

[0096] Calculate the Pearson correlation coefficient. The Pearson correlation coefficient between any two sensor data is calculated in advance (e.g., the correlation coefficient between temperature and smoke concentration). This coefficient ranges from -1 to 1. A positive number indicates a positive correlation (e.g., when the temperature rises, the smoke concentration also rises, which is consistent with the characteristics of a fire). The absolute value is larger, the correlation is stronger. The correlation coefficient is used as a feature input to the LSTM, providing statistical evidence of the strength of the correlation between parameters to assist the model in determining whether the data changes are consistent with the typical correlation patterns of a fire.

[0097] To consider the variability of actual fire scene data (e.g., different smoke concentration rising speeds in different environments), data augmentation perturbations are added, and an enhancement strategy of adding noise or time offset is proposed. For example, a small amount of random noise (e.g., temperature value ± 0.5°C) is added to the original data, or the time series is slightly offset (e.g., a certain segment of data is delayed by 1 second), to generate more diverse training data, avoid the model from overfitting to fixed data, and ensure stable fire identification in different actual scenarios.

[0098] In an embodiment, the calculation formula of the Gaussian radial basis kernel function is:

[0099]

[0100] wherein, are feature parameter sequences, denote the Euclidean distance between them: is a parameter that controls the width of the Gaussian distribution;

[0101] The characteristic values and eigenvectors are calculated, and the most important eigenvectors are selected according to the principal component variance contribution rate; the preset condition is that the sum of the variance contribution rates is greater than 95%, and The value is 15.

[0102] In this embodiment, the Gaussian radial basis kernel function (GRBF) is the core tool for realizing multi-sensor data dimensionality reduction fusion, mainly used to solve the nonlinear correlation problem of fire characteristics. When a fire occurs, the changes of smoke concentration, temperature, and flammable gas concentration often show a complex nonlinear relationship (e.g., slow rise at the beginning and rapid growth in the middle), and traditional linear methods are difficult to capture this correlation, while GRBF can effectively handle it.

[0103] The Gaussian radial basis kernel function is a tool for measuring the similarity of two feature parameter sequences (x, y). When the two sequences are similar (e.g., both show a trend of temperature rising + smoke increasing in two time periods), the function output value is close to 1; when the features are very different, the output value is close to 0.

[0104] represent two groups of feature parameter sequences to be compared (e.g., may be the temperature sequence from 1st to 10th second, may be the temperature sequence from 11th to 20th second)

[0105] : represents the Euclidean distance between the two sequences, which is used to quantify the degree of feature difference, and the smaller the distance, the more similar the features.

[0106] δ (value 15): control the hyperparameter of Gaussian distribution width, which is verified by experiments that this value can best balance the discrimination and generalization of fire features. If the value of δ is too large, different features will be blurred, and if the value of δ is too small, it will be too sensitive to subtle differences.

[0107] Map the feature difference to 0-1 through the exponential function, so that the originally nonlinearly separable fire features become separable in high-dimensional space, laying the foundation for subsequent dimensionality reduction. After GRBF processing, the core features need to be further screened to reduce redundant information. The specific process is as follows:

[0108] Eigenvalue and eigenvector calculation: After KPCA processing, a series of eigenvalues and corresponding eigenvectors will be obtained. The size of the eigenvalue represents the amount of information contained in the original data by the eigenvector. The larger the eigenvalue, the more important the information contained in the corresponding eigenvector.

[0109] Variance contribution rate screening: select the eigenvector according to the preset condition that the sum of variance contribution rates is greater than 95%. The variance contribution rate is the proportion of a certain eigenvalue to the sum of all eigenvalues, which reflects the retention degree of the original data information by the eigenvector.

[0110] Through screening, more than 95% of the original data information can be retained (to ensure that key fire features are not lost), and the feature dimension can be greatly reduced (usually from dozens of dimensions to a few dimensions), reducing the computational complexity of the subsequent LSTM model, while avoiding the decline of recognition accuracy caused by information redundancy.

[0111] The above dimensionality reduction fusion method based on Gaussian radial basis kernel function is particularly suitable for processing the nonlinear correlation characteristics of multi-sensor data in fire scenes, and provides high-quality core features for subsequent model recognition.

[0112] In an embodiment, the Mann-Kendall test is performed on the time series in the preprocessed data, including:

[0113] Based on the time series in the preprocessed data , calculate the statistic S,

[0114]

[0115]

[0116] Calculate the variance Var(s) = n(n-1)(2n+5) / 18, and when n≥10, calculate the standard normal statistic variable Z.

[0117]

[0118] By determining whether a time series has a significant trend, for a confidence level;

[0119] If Z is greater than 0, the time series is an upward trend; if Z is less than 0, the time series is a downward trend.

[0120] In this embodiment, the Mann-Kendall (MK) test is a non-parametric statistical method for detecting trend changes in time series data. It can determine whether the data has a significant upward trend, downward trend or no obvious trend. Due to its non-parametric nature, the MK test does not require the data to be normally distributed, nor does it assume that the trend is linear, so it is very widely applicable. In the initial stage of a fire, the temperature, smoke concentration, and flammable gas volume fraction often change little, but there is an upward trend. If there is an upward trend, the data measured by the detector does not reach the alarm level, and it can also be inferred that there is a fire, so the introduction of the trend factor is beneficial to the judgment of the initial stage of the fire. Since the MK test is a non-parametric test and is not affected by sample values, distribution types, etc., it is widely used in the fields of hydrology and meteorology.

[0121] The principle of the MK test is to compare each data point with its previous data points to detect monotonic trends (upward, downward, or no trend) in time series data. Specifically, the MK test compares each data point in the time series with all previous data points, counts the number of data points that are smaller than it and the number of data points that are larger than it, and then compares the size relationship of the two quantities to determine whether there is a monotonic trend. The null hypothesis of the MK test is that the sequence is independent and identically distributed, and the alternative hypothesis is that there is a monotonic trend. The advantage of the MK test is that it does not require any assumptions about the data, can be used for various types of time series data, including non-normal data, and is not sensitive to periodic changes in time series data.

[0122] In an embodiment, the calculation formula of the LSTM model comprises:

[0123] forget gate ;

[0124] input gate ;

[0125] alternative cell state ;

[0126] output gate ;

[0127] cell state update ;

[0128] output ;

[0129] wherein, is the weight matrix of the forget gate that leaves information, is the weight matrix of the input gate that measures the importance of the input, is the weight matrix of the input gate that adjusts the integrated input, is the weight matrix of the output gate that measures how much integrated input is passed to the hidden layer;

[0130] , , , respectively represent the bias of the forget gate, the input gate, the integrated input and the output gate corresponding to;

[0131] represents the result of the forget gate at time t; represents the input of the unit at time t; is a softmax activation function:

[0132] represents the alternative unit state: represents the unit state at the last time; represents the updated state of the unit; represents the hidden layer output at the last time; represents the output at this time t.

[0133] In this embodiment, as a variant of recurrent neural network (RNN), LSTM can effectively process long sequence input and solve the problem of long-term dependence of recurrent neural network. LSTM selectively allows information to pass through the gate structure to realize the function of long-term memory, which are the forget gate, the input gate and the output gate. The forget gate decides how much of the unit state at the last time is retained in the unit at the current time; the input gate decides how much of the current input is retained in the unit at the current time; and the output gate decides how much of the unit state at the current time will be output.

[0134] In an embodiment, the attention mechanism is a means to enhance the adaptive learning ability of the model, which essentially gives high weight to important information and ignores irrelevant information, so that the model can better mine the key features of the data, thereby improving the accuracy of prediction or classification.

[0135] The attention mechanism converts the input sequence element encoding into query, key and value vectors through the weight matrix , , , calculates the similarity and normalizes it according to the formula

[0136] to obtain the attention output.

[0137] represents the i-th element of the input sequence, is obtained by encoding, representing the value obtained after the original i-th element is encoded, and then multiplied by the weight matrix , , to obtain , , matrix, is used to measure the degree of association between the current and other , represents the degree of association between and the query, represents the key information of ;

[0138] is obtained by normalizing the calculation of and , representing the similarity between and , and multiplying to obtain the output .

[0139] In this embodiment, wherein is obtained by considering all position information of the entire sequence, so that the attention mechanism can adaptively capture long-range dependencies between elements. The attention mechanism is used to enhance the ability of the LSTM model to capture key information, and to solve the problem that important features in fire time series data are covered by redundant information. The core idea is to let the model automatically focus on the most critical features for fire identification (such as the stage where temperature and smoke concentration rise synchronously), while weakening irrelevant information (such as normal fluctuations in the environment).

[0140] In an embodiment, the LSTM model adopts a multi-layer stacking structure to extract input data features through different layers. The multi-layer stacking structure refers to connecting multiple LSTM layers in sequence, with the output of the previous layer as the input of the next layer, forming a layer-by-layer progressive feature extraction link. For example, 3-layer LSTM stacking can be used in the present application:

[0141] The first layer (bottom layer) directly receives the preprocessed multi-sensor data (such as time series sequences of smoke concentration, temperature, and flammable gas concentration); the second layer (middle layer) receives the output features of the first layer; the third layer (top layer) receives the output features of the second layer, and finally outputs high-level features for fire identification. Each layer of LSTM has independent gating parameters (weight matrix and bias), which can learn different levels of features.

[0142] In the fire identification scenario, the feature extraction of each layer of LSTM has significant differences:

[0143] The bottom-layer LSTM extracts basic physical features; it directly processes raw sensor data and focuses on the short-term change characteristics of a single parameter: such as capturing the short-term upward trend of "temperature rising from 25°C to 28°C in 10 seconds"; identifying subtle changes such as the instantaneous fluctuation of smoke concentration appearing 0.5ppm; extracting the stable state of flammable gas concentration maintaining around 10ppm; the above features are atomic-level information for fire identification, reflecting the direct observation results of the sensor.

[0144] The middle-layer LSTM extracts parameter coordination features, based on the single-parameter features output by the bottom layer, further learning the correlation patterns between multiple parameters. Such as identifying the coordinated changes of temperature rising while smoke concentration also increasing synchronously; capturing the time sequence correlation that flammable gas concentration suddenly increases and temperature starts to rise within 30 seconds; distinguishing the difference between "only temperature rising (possibly equipment heating)" and "temperature + smoke rising simultaneously (possibly fire)" ; the above features are correlation information for fire identification, reflecting the linkage relationship between different physical parameters.

[0145] The top-layer LSTM extracts fire pattern features, based on the coordination features of the middle layer, extracting high-level patterns consistent with the development law of fire. Such as identifying the typical fire development sequence of "temperature slowly rising → smoke appearing → flammable gas concentration increasing"; distinguishing the stage difference between initial small fire (slow feature change) and fire outbreak (rapid feature change); capturing the essential difference between non-fire disturbance (such as cooking smoke, no temperature continuously rising) and real fire; the above features are decision-level information for fire identification, directly serving the final fire judgment.

[0146] In this embodiment, the advantages of multiple LSTM layers compared to a single LSTM layer are:

[0147] Hierarchical expression of features, from raw sensor data to fire patterns, achieving gradual abstraction of features, consistent with the cognitive logic of fire from physical phenomena to essential judgment.

[0148] Complex pattern capturing ability, fire features often have complexity of multiple scales and multiple correlations (such as coexistence of short-term fluctuations and long-term trends), and the multi-layer structure can process features of different scales at different levels.

[0149] Enhanced anti-interference ability, the bottom layer may be affected by environmental noise, but the high layer can filter isolated noise (such as false alarms of a single sensor) through learning of coordination features, improving the robustness of the model.

[0150] Adapt to the characteristics of fire development: fire development from initial to outbreak is a gradual process, and the multi-layer structure can better learn this stage change (such as capturing instantaneous changes at the bottom layer and capturing trend evolution at the high layer).

[0151] Through the multi-layer stacking design, the LSTM model can more accurately mine deep features in the fire time series data, provide more effective input for subsequent recognition and judgment combined with the attention mechanism, and finally improve the accuracy and early warning ability of fire identification.

[0152] With reference Figure 2 In another embodiment of the present application, a fire identification device based on an LSTM model and multi-sensor data fusion is also provided, comprising:

[0153] The acquisition unit is configured to acquire smoke concentration data, temperature data, and combustible gas concentration data in a scene to form multivariate time series data.

[0154] The processing unit is configured to sequentially perform data cleaning, data normalization, and data correlation enhancement operations on the multivariate time series data to obtain preprocessed data.

[0155] The fusion unit is configured to perform dimensionality reduction fusion on the preprocessed data, calculate eigenvalues and eigenvectors based on a Gaussian radial basis kernel function, and select eigenvectors whose variance contribution rates satisfy a preset condition to obtain fusion features.

[0156] The test unit is configured to perform Mann-Kendall testing on time series in the preprocessed data and calculate a standard normal statistical variable as a trend factor.

[0157] The output unit is configured to input the fusion features and the trend factor into an LSTM model combined with an attention mechanism, process long sequence information through a gate structure of the LSTM model, assign a high weight to key information combined with the attention mechanism, and output a fire identification result.

[0158] In the present embodiment, the specific implementation of each unit in the above device embodiment is described with reference to the above method embodiment, and will not be described here.

[0159] With reference Figure 3 In the present embodiment, a computer device is also provided, which can be a server, and the internal structure thereof can be as shown in Figure 3The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected by a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium. The database of the computer device is configured to store corresponding data in the embodiment. The network interface of the computer device is configured to communicate with an external terminal through a network connection. The computer program is executed by the processor to implement the above method.

[0160] Those skilled in the art can understand that, Figure 3 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied.

[0161] The embodiment of the present application also provides a computer readable storage medium, which stores a computer program. The computer program is executed by the processor to implement the above method. It can be understood that the computer readable storage medium in the embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0162] In summary, the fire identification method based on the LSTM model and multi-sensor data fusion provided in the embodiments of the present application comprises: collecting smoke concentration data, temperature data, and combustible gas concentration data in a scene to form multivariate time series data; sequentially performing data cleaning, data normalization, and data correlation enhancement operations on the multivariate time series data to obtain preprocessed data; performing dimensionality reduction fusion on the preprocessed data, calculating eigenvalues and eigenvectors based on a Gaussian radial basis kernel function, screening eigenvectors with variance contribution rates satisfying a preset condition to obtain fusion features; performing Mann-Kendall test on the time series in the preprocessed data, calculating a standard normal statistical variable as a trend factor; inputting the fusion features and the trend factor into an LSTM model combined with an attention mechanism, processing long sequence information through the gate structure of the LSTM model, giving high weights to key information combined with the attention mechanism, and outputting a fire identification result. In the present application, multiple sensors are used to collect information including gas, temperature, smoke, etc., and the MK test and the LSTM algorithm combined with the attention mechanism are used to process the multiple types of information provided by the sensors, which can effectively improve the sensitivity of fire monitoring and reduce the false alarm rate; dimensionality reduction fusion is performed on the multi-source heterogeneous sensor information, which can reduce information redundancy and reduce the amount of calculation; the Mann-Kendall test method is added before the prediction of the LSTM model, which is beneficial to the fire condition judgment in the initial stage; the attention mechanism is added to the LSTM model, which makes the model better retain long-time information and ignore irrelevant information, thereby forcing the network to autonomously capture important information, thereby further improving the fire identification and prediction accuracy of the model.

[0163] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when executed, can include the processes of the above-mentioned embodiment methods. Any reference to memory, storage, database or other medium provided by the present application and used in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM, etc.

[0164] It should be noted that in this document, the terms "comprising", "including", or any other variant thereof are intended to cover a non-exclusive inclusion, such that a process, device, article or method that comprises a list of elements does not only include those elements, but can also include other elements not expressly listed or inherent to such process, device, article or method. Without more limitations, the element defined by the statement "comprising a" does not exclude the presence of additional identical elements in the process, device, article or method that includes the element.

[0165] The above description is only the preferred embodiment of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation, or direct or indirect application in other related technical fields, based on the content of the present application specification and drawings, are also included in the patent protection scope of the present application.

Claims

1. A fire identification method based on an LSTM model and multi-sensor data fusion, characterized in that, The method comprises the following steps: Collecting smoke concentration data, temperature data, and combustible gas concentration data in a scene to form multivariate time series data; Performing data cleaning, data normalization, and data correlation enhancement on the multivariate time series data in sequence to obtain preprocessed data; the data correlation enhancement comprises constructing a sliding time series window, converting discrete single-time data into continuous sequence data, and associating smoke, temperature, and combustible gas data at the same time node; performing feature engineering, calculating a combination feature of smoke value x combustible gas value, adding temperature change rate, and concatenating the multi-sensor data into a vector containing smoke, temperature, and combustible gas values; calculating the Pearson correlation coefficient of any two sensor data to quantify the correlation strength between features, and realizing data enhancement by adding noise or time offset; Performing dimensionality reduction fusion on the preprocessed data, calculating eigenvalues and eigenvectors based on a Gaussian radial basis kernel function, and screening eigenvectors with a variance contribution rate meeting a preset condition to obtain fusion features; Performing Mann-Kendall test on the time series in the preprocessed data to calculate a standard normal statistical variable as a trend factor; Inputting the fusion features and the trend factor into an LSTM model combined with an attention mechanism, processing long sequence information through the gate structure of the LSTM model, giving high weights to key information combined with the attention mechanism, and outputting a fire identification result. 2.The fire identification method based on the LSTM model and the multi-sensor data fusion according to claim 1, wherein, A smoke sensor, a temperature sensor, and a combustible gas sensor are used to collect smoke concentration data, temperature data, and combustible gas concentration data in a scene at a preset sampling frequency. 3.The fire identification method based on LSTM model and multi-sensor data fusion according to claim 1, characterized in that, The data cleaning comprises filling missing values with forward filling or mean replacement, removing outliers through Z-score method or interquartile range method, and smoothing data by applying moving average filtering; Data normalization is performed by using Min-Max normalization or standardization method. 4.The fire identification method based on LSTM model and multi-sensor data fusion according to claim 1, wherein, The calculation formula of the Gaussian radial basis kernel function is: wherein, are sequences of characteristic parameters, respectively, denote the Euclidean distance between them: is a parameter that controls the width of the Gaussian distribution; The characteristic value and the characteristic vector are calculated, and the most important characteristic vector is selected according to the principal component variance contribution rate; the preset condition is that the sum of the variance contribution rates is greater than 95%, and is 15. 5.The fire identification method based on LSTM model and multi-sensor data fusion according to claim 1, characterized in that, The Mann-Kendall test on the time series in the preprocessed data comprises: based on the time series in the pre-processed data a statistical quantity S is calculated, Calculating the variance Var(s) = n(n-1)(2n+5) / 18, and calculating the standard normal statistical variable Z when n≥10; By , determining whether a time series has a significant trend, is a confidence level; If Z is greater than 0, the time series is an upward trend; if Z is less than 0, the time series is a downward trend. 6.The fire identification method based on LSTM model and multi-sensor data fusion according to claim 1, characterized in that, The calculation formula of the LSTM model comprises: forgetting gate ; input gate ; Alternative cell states ; Output gate ; Cell state update ; Output ; wherein, is a weight matrix for the forget gate to leave information, is a weight matrix for the input gate to measure the importance of the input, is a weight matrix for the input gate to adjust the integrated input, is a weight matrix for the output gate to measure how much integrated input is passed to the hidden layer; , , , respectively represent the bias corresponding to the forget gate, the input gate, the integrated input and the output gate. denotes the result of the forget gate at time t; denotes the input to this cell at time t; is a softmax activation function: represents the alternative cell state: represents the cell state at the previous time step; represents the updated state of the cell; represents the hidden layer output at the previous time step; represents the output at this time step t. 7.The fire identification method based on LSTM model and multi-sensor data fusion according to claim 1, characterized in that, The attention mechanism converts the input sequence element encoding into a query, key, value vector through a weight matrix , , , calculates the similarity and normalizes it according to the formula , get the attention output; represents the i-th element of the input sequence, is obtained by encoding, represents the value obtained after encoding the originally i-th element, and then multiplied by the weight matrix , , to obtain , , matrix, is used to measure the degree of association between the current and other , represents the degree of association with the query, represents the key information of ; is calculated by and is normalized after calculation, indicating and the similarity of, multiplied by to get the output . 8.The fire identification method based on LSTM model and multi-sensor data fusion according to claim 1, characterized in that, The LSTM model adopts a multi-layer stacking structure to extract input data features through different layers.

9. A fire identification device based on an LSTM model and multi-sensor data fusion, characterized in that, The method comprises the following steps: A collecting unit is configured to collect smoke concentration data, temperature data, and combustible gas concentration data in a scene to form multivariate time series data; The processing unit is used for sequentially performing data cleaning, data normalization and data correlation enhancement operations on the multivariate time series data to obtain preprocessed data; the data correlation enhancement includes constructing a sliding time series window, converting discrete single-time data into continuous sequence data, and making smoke, temperature and combustible gas data at the same time node form a correlation; performing feature engineering, calculating a combination feature of smoke value x combustible gas value, adding temperature change rate, and concatenating the multi-sensor data into a vector containing smoke, temperature and combustible gas values; calculating the Pearson correlation coefficient of any two sensor data to quantify the correlation strength between features, and simultaneously realizing data enhancement by adding noise or time offset; The fusion unit is used for dimension reduction fusion on the preprocessed data, calculating eigenvalues and eigenvectors based on a Gaussian radial basis kernel function, screening the eigenvectors with variance contribution rates meeting a preset condition to obtain fusion features; The test unit is used for performing Mann-Kendall test on the time series in the preprocessed data to calculate a standard normal statistical variable as a trend factor; The output unit is used for inputting the fusion features and the trend factor into an LSTM model combined with an attention mechanism, processing long sequence information through the gate structure of the LSTM model, giving high weights to key information combined with the attention mechanism, and outputting a fire identification result.

Images