Fire prediction and evaluation method, system and equipment assisted by interpretable machine learning based on multiple types of sensors, and medium

Abstract

A fire prediction and evaluation method, system, device and medium based on explainable machine learning assistance of multiple types of sensors, uses multiple types of sensors for different scenes, collects and integrates a large amount of fire data sets, covers multiple different fire scenes, namely different time, places and the like, and divides the data sets into a training set and a test set; constructing a convolutional neural network sub-model, and training the model by using a data set to realize universal modeling of fire prediction; a principal component analysis method is used and an entropy weight method is combined to carry out disaster situation evaluation on a place where a fire is predicted to occur, so that accurate universality evaluation on the disaster degree is realized; according to the method, disaster assessment is carried out by combining the convolutional neural network sub-model with the principal component analysis (PCA) and the entropy weight method, fire situations of different types and scales can be accurately analyzed and predicted, the accuracy and operability of disaster assessment are improved, and meanwhile the data processing capacity and the output interpretability of the model are enhanced.

Description

Technical Field

[0001] The present invention belongs to the technical field of fire prediction and assessment, and in particular relates to a fire prediction and assessment method, system, device and medium assisted by interpretable machine learning based on multiple types of sensors. Background Art

[0002] Effective fire prediction and assessment technology is particularly important. With the development of science and technology, a series of advanced technologies have been applied to the early prediction and real-time assessment of fires, showing great potential.

[0003] Some existing fire detection systems use multi-sensor fusion technology to expand the spatiotemporal range of fire monitoring by monitoring parameters such as light, smoke, temperature, gas and humidity. This technology combines the signals of multiple sensors to improve the accuracy and timeliness of fire alarms. For example, a multi-sensor fire alarm system based on the Dempster-Shafer evidence theory can effectively handle conflicting evidence between sensors and improve the reliability of fire warnings by processing and fusing data from multiple sensors.

[0004] In traditional methods, sensors are often used to monitor environmental variables such as smoke, temperature, and gas, which are important indicators for early detection of fires. However, these methods often rely on simple threshold judgments and are difficult to adapt to the increasing complexity of the environment. In order to improve the accuracy of predictions and the meticulousness of evaluations, researchers have begun to explore more complex data analysis models. Specifically, machine learning technology has been introduced into the field of fire monitoring due to its powerful data processing capabilities. Mert Nakip et al. studied the application of recursive trend prediction neural network (rTPNN) in multi-sensor fire detection in the article "Recurrent Trend Predictive Neural Network for Multi-Sensor Fire Detection", and verified the effectiveness of ML technology in fire prediction and evaluation modeling.

[0005] Existing traditional fire prediction technology will be interfered by various factors in complex environments, making it difficult to give full play to its advantages.

[0006] Current research has developed fire prediction models in a variety of environments and has made significant progress with the support of machine learning techniques, which has played a positive role in the development of fire prediction. However, due to the limitations of experimental data, these models are only applicable to specific proposed scenarios, but may face some challenges when trying to extend them to other environments.

[0007] In complex environments, fire prediction has significant randomness due to the combined influence of multiple environmental factors. For example, meteorological conditions (such as temperature, humidity, wind speed, etc.), terrain characteristics (such as slope, vegetation density, etc.), and human activities (such as agricultural burning, camping fires, etc.) may all have an impact on the occurrence and spread of fires. In addition, the rapid spread and uncertainty of fires further amplify the difficulty and randomness of fire prediction. It is worth noting that compared with traditional fixed monitoring systems, the occurrence and spread of fires usually have strong temporal and spatial heterogeneity. In this case, the monitoring equipment may not be able to capture the key data of the fire in real time, resulting in insufficient accuracy and timeliness of the prediction.

[0008] At the same time, in a changing environment, computer vision has strong environmental dependence in fire prediction and dynamic assessment, is highly dependent on ambient lighting conditions, perspective and occlusion problems, has high computing resource requirements, and has high false alarm rates. In order to analyze and improve the reliability of fire prediction and assessment, it is imperative to build an environmentally universal fire prediction and assessment model.

[0009] The patent application document with publication number CN202410324749 discloses a fire prediction system and method based on multi-source information fusion, which includes collecting fire data through a sensor system inside a building, using LSTM and ELM neural networks for data prediction and feature processing, and using MATLAB's fuzzy control toolbox to establish fuzzy rules for final fire prediction. However, this method lacks the ability to capture spatial features; feature selection is black-boxed, lacks transparency, and has poor interpretability; fuzzy rule setting is highly subjective; high-dimensional data processing is difficult; and fuzzy rules have limitations in weighting multiple indicators. Summary of the invention

[0010] In order to overcome the deficiencies of the above-mentioned prior art, the purpose of the present invention is to provide a fire prediction and assessment method, system, device and medium assisted by interpretable machine learning based on multiple types of sensors, which can accurately analyze and predict fire situations of different types and scales by combining convolutional neural network sub-models with principal component analysis (PCA) and entropy weight method for disaster assessment, thereby improving the accuracy and operability of disaster assessment, and enhancing the data processing capability of the model and the interpretability of the output. This method effectively improves the response speed and processing accuracy to complex fire situations by optimizing data processing and evaluation algorithms, thereby providing more effective technical support for fire prevention and emergency response.

[0011] In order to achieve the above object, the technical solution adopted by the present invention is:

[0012] A fire prediction and assessment method based on interpretable machine learning assisted by multiple types of sensors includes the following steps:

[0013] Step 1: Collect a fire dataset T. The fire dataset T consists of time series data collected by multiple types of sensors and real labels of whether a fire occurs at the corresponding moment. The fire dataset T is preprocessed and divided into a training set and a test set.

[0014] Step 2, construct a convolutional neural network sub-model, use the training set obtained in step 1 to train the convolutional neural network sub-model, input the test set obtained in step 1 into the trained convolutional neural network sub-model, and determine whether the output result matches the corresponding true label by calculating the accuracy and loss value. After multiple rounds of training and testing, complete the model training and obtain the optimal convolutional neural network sub-model after training; obtain the sensor data set A' currently deployed at multiple locations, input the current sensor data set A' into the optimal convolutional neural network sub-model after training, and obtain the output result. Use this as a basis to filter the current sensor data set A', delete the data corresponding to the location where fire will not occur, and obtain the sensor data set A corresponding to the location where fire will occur;

[0015] Step 3: For the sensor data set A corresponding to the location where the fire will occur in step 2, use the principal component analysis (PCA) method to determine the main factors that affect the occurrence of the fire at this location. Based on this, data screening is performed on the basis of the sensor data set A corresponding to the location where the fire will occur obtained in step 2 to obtain a data set M containing only the main factors that affect the occurrence of the fire at this location.

[0016] Step 4: Use the entropy weight method to evaluate the disaster situation after the fire occurs at the location for the data set M in step 3 that only contains the main factors affecting the occurrence of fire at the location.

[0017] In step 1, the collected fire data set T is preprocessed by the following method:

[0018]

[0019] Among them, μ and σ represent the mean and standard deviation of each sensor data, respectively. The fire dataset T is then divided into a training set and a test set.

[0020] The specific method of step 2 is:

[0021] In step 2.1, the convolutional neural network sub-model is used to detect potential signals of fire from environmental monitoring data; the convolutional neural network sub-model architecture is as follows:

[0022] f(x)=σ(w2ReLU(w1*x+b1)+b2)

[0023] Where * represents the convolution operation, σ is the Sigmoid activation function, which is used to output the probability of predicting the positive class (i.e., detecting a fire); w1 and w2 are the weights of the model, and b1 and b2 are the bias terms;

[0024] Step 2.2: Input the training set obtained in step 1 into the convolutional neural network sub-model constructed in step 2.1 to obtain the final prediction probability, i.e., p. If p ≥ 0.5, then the prediction label at the corresponding moment is on the contrary,

[0025] Step 2.3, input the final predicted probability p obtained in step 2.2 and the true label of the corresponding moment in the fire data set T in step 1 into the minimization binary cross entropy loss function for training. During the training process, each batch of data undergoes forward propagation, loss calculation, and back propagation;

[0026] The loss calculation formula is as follows:

[0027]

[0028] Among them, N train is the number of samples in the training set, y i is the true label of the i-th sample, whose value is 0 or 1, p i is the probability that the model predicts the i-th sample as a positive class;

[0029] Step 2.4: Perform performance evaluation on the convolutional neural network sub-model trained in step 2.3 on the test set obtained in step 1. The evaluation indicators are accuracy and loss value. The calculation formula is as follows:

[0030]

[0031] Among them, N test is the number of samples in the test set, y i is the true label of the i-th sample;

[0032] The loss value is calculated as follows:

[0033]

[0034] Among them, N test is the number of samples in the test set, y i is the true label of the i-th sample, whose value is 0 or 1, p i is the probability that the model predicts the i-th sample as a positive class;

[0035] Step 2.5, repeat steps 2.3 and 2.4 to obtain multiple rounds of test data, draw the loss curve and accuracy curve during training and testing, verify the performance and stability of the convolutional neural network sub-model, and obtain the optimal convolutional neural network sub-model;

[0036] Step 2.6, obtain the sensor data set A' currently deployed at multiple locations. The current sensor data set A' consists of multiple sets of multi-type sensor data at different locations at the same time. Input the data in the current sensor data set A' into the optimal convolutional neural network sub-model of step 2.5 one by one to obtain the output value p, that is, the probability of fire. If p ≥ 0.5, the prediction label at the corresponding time Do not delete, otherwise, Delete the corresponding data group in the current sensor data set A', repeat the above operation until all groups of data in the current sensor data set A' have completed the input and deletion operations, and the final current sensor data set A' is the sensor data set A corresponding to the location where the fire will occur.

[0037] The specific method of step 3 is:

[0038] Step 3.1, standardize the sensor data set A corresponding to the location where the fire will occur in step 2 to obtain a standardized data matrix Z. The standardization formula is:

[0039]

[0040] Among them, X ij is the value of the i-th sample on the j-th variable, is the mean of the jth variable, s j is the standard deviation of the jth variable;

[0041] Step 3.2, calculate the covariance matrix: Based on the data matrix Z standardized in step 3.1, calculate the covariance matrix C, whose elements C jk Represents the covariance between variables j and k:

[0042]

[0043] Step 3.3, eigenvalue decomposition: Perform eigenvalue decomposition on the covariance matrix C obtained in step 3.2, specifically through the following formula:

[0044] CV=VD

[0045] Get the eigenvalue D and eigenvector V, where V is a matrix containing all eigenvectors of C; D is a diagonal matrix whose diagonal elements are the corresponding eigenvalues, representing the variance contribution of each feature of C;

[0046] Step 3.4, according to the size of the eigenvalue D obtained in step 3.3, based on the Kaiser criterion, select the principal component matrix P composed of the eigenvectors V corresponding to the first m eigenvalues ​​whose eigenvalues ​​are greater than 1. According to the principal component matrix P, obtain the data set M containing only the main factors affecting the occurrence of fire at this location.

[0047] The specific method of step 4 is:

[0048] Step 4.1: For each type of data collected by multiple types of sensors in the data set M that only contains the main factors affecting the occurrence of fire at this location in step 3, calculate its information entropy to evaluate the dispersion of the data. The higher the information entropy, the more irregular the data change of this type of sensor, and vice versa, the more regular the change. The information entropy E j Calculated by the following formula:

[0049]

[0050] Among them, p ij is the ratio of the value of the i-th sample on the j-th index to the total value of all samples of the index; k is a constant, n is the number of samples;

[0051] Step 4.2: Use the information entropy E obtained in step 4.1 j , calculate the weight of each data collected by multiple types of sensors in the data set M of the main factors affecting the occurrence of fire at this location, weight w j By calculating the complement of information entropy, the formula is as follows:

[0052]

[0053] Where m is the total number of indicators;

[0054] Step 4.3, use the weight w obtained in step 4.2 j , weighted sum of each type of data collected by multiple types of sensors in the data set M that only contains the main factors affecting the occurrence of fire at this location is performed to obtain a comprehensive score for disaster assessment:

[0055]

[0056] Among them, x ij is the value of the i-th sample on the j-th index;

[0057] Step 4.4, determine the severity of the fire based on the comprehensive score of the disaster assessment in step 4.3. The higher the score, the greater the potential risk and severity of the fire.

[0058] The present invention also provides a fire prediction and assessment system assisted by interpretable machine learning based on multiple types of sensors, comprising:

[0059] The fire data set T collection and processing module is used to collect the fire data set T. The fire data set T consists of time series data collected by multiple types of sensors and the real label of whether a fire occurs at the corresponding moment. The fire data set T is preprocessed and divided into a training set and a test set.

[0060] A sensor data set A acquisition module corresponding to the location where a fire will occur is used to construct a convolutional neural network sub-model, use a training set to train the convolutional neural network sub-model, input the test set into the trained convolutional neural network sub-model, and determine whether the output result matches the corresponding true label by calculating the accuracy and loss value. After multiple rounds of training and testing, the model training is completed to obtain the optimal convolutional neural network sub-model after training; obtain the sensor data set A' currently deployed at multiple locations, input the current sensor data set A' into the optimal convolutional neural network sub-model after training, and obtain the output result. Based on this, the current sensor data set A' is screened, and the data corresponding to the location where a fire will not occur is deleted to obtain the sensor data set A corresponding to the location where a fire will occur;

[0061] A module for acquiring a data set M containing only the main factors affecting the occurrence of a fire at the location, for determining the main factors affecting the occurrence of a fire at the location using a principal component analysis (PCA) method for a sensor data set A corresponding to the location where a fire will occur, and based on this, performing data screening on the basis of the sensor data set A corresponding to the location where a fire will occur, to obtain a data set M containing only the main factors affecting the occurrence of a fire at the location;

[0062] The disaster assessment module is used to use the entropy weight method to conduct a disaster assessment after a fire occurs at a location on a data set M that only contains the main factors that affect the occurrence of a fire at that location.

[0063] The present invention also provides a fire prediction and assessment device assisted by interpretable machine learning based on multiple types of sensors, comprising:

[0064] Memory: a computer program storing the above-mentioned fire prediction and assessment method assisted by interpretable machine learning based on multiple types of sensors, which is a computer-readable device;

[0065] Processor: used to implement the fire prediction and assessment method assisted by interpretable machine learning based on multiple types of sensors when executing the computer program.

[0066] The present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it can implement the fire prediction and assessment method assisted by interpretable machine learning based on multiple types of sensors.

[0067] Compared with the prior art, the present invention has the following beneficial effects:

[0068] First, the present invention uses a convolutional neural network sub-model, which has a strong ability to extract spatial features and is particularly suitable for processing disaster data with spatial distribution characteristics (such as fire images or geographic information). In contrast, LSTM and ELM are mainly good at processing time series data, and their ability to process spatial features is relatively weak.

[0069] Second, the present invention uses principal component analysis (PCA) for dimensionality reduction, which can reduce data redundancy and improve the computational efficiency of the model, while the entropy weight method can reasonably distribute the importance of each feature, thereby avoiding human bias and further improving the accuracy and scientificity of the model. In contrast, LSTM and ELM usually perform feature learning adaptively through network training, but overfitting or underfitting may occur, and the importance of features is not transparent enough.

[0070] Third, the present invention adopts the entropy weight method to combine the data discreteness of each feature, which can assign more reasonable weights to different factors in disaster assessment, thereby improving the rationality of model decision-making. Although the fuzzy control method in the prior art can set fuzzy rules through expert experience, the rules are highly subjective and it is difficult to handle the relative importance of features when weighting multiple indicators.

[0071] Fourth, the present invention adopts principal component analysis (PCA) and entropy weight method to make the feature selection process of the evaluation model more interpretable, which can clarify the contribution of each factor to the disaster assessment results. However, the black box characteristics of LSTM and ELM are strong, making it difficult to explain the intermediate process of the model and the reasons for the formation of the prediction results.

[0072] Fifth, principal component analysis (PCA) is used for dimensionality reduction, which can effectively deal with multi-dimensional data problems that may be involved in disaster assessment, reduce data dimensions and model complexity, and reduce computing costs. However, when LSTM and ELM process high-dimensional data, the model complexity increases, the training time will be greatly extended, and performance may be reduced.

[0073] In summary, the present invention provides a comprehensive and accurate solution by combining the convolutional neural network sub-model with principal component analysis (PCA) and entropy weight method, integrating advanced data fusion technology and innovative scoring mechanism, and significantly improving the efficiency and accuracy of fire prediction and assessment. The application of these technologies not only enhances the performance of the prediction model, but also provides a powerful tool for disaster management and emergency response. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1It is a flow chart for implementing the present invention.

[0075] Figure 2 This is a graph showing the application results of the prediction model of the present invention in a real scenario.

[0076] Figure 3 Visualize the training results of the model of the present invention (loss value is less than 0.1).

[0077] Figure 4 Visualization results for the model training of the present invention (accuracy up to 96.1%).

[0078] Figure 5 PCA is used for the present invention to obtain the cumulative explained variance. DETAILED DESCRIPTION

[0079] The technical solution adopted by the present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0080] The present invention aims to solve the defects of the prior art and proposes a fire prediction and assessment method assisted by interpretable machine learning based on multiple types of sensors. The model takes into account the complexity and variability of fire scenes, the correlation between sensor data and other influencing factors. Its implementation scheme is: using multiple types of sensors for different scenes, collecting and integrating a large number of fire data sets, covering a variety of different fire scenes, that is, different times, locations, etc., and dividing the data sets into training sets and test sets; constructing a convolutional neural network sub-model, using the data set to train the model, and realizing universal modeling for fire prediction; using principal component analysis combined with entropy weight method to conduct disaster assessment on the places where fire is predicted to occur, and realizing accurate universal assessment of the degree of disaster.

[0081] The universal fire prediction and dynamic assessment model based on convolutional neural network, principal component analysis (PCA) and entropy weight method proposed in the present invention can accurately and efficiently predict and assess fire conditions in various fire scenarios, providing strong support for fire protection and timely rescue.

[0082] The present invention aims to meet the challenges of fire risk monitoring and assessment in complex and changing environments. The device integrates various types of sensor data (such as temperature, humidity, smoke concentration, carbon dioxide concentration, etc.) and uses advanced machine learning algorithms to achieve deep fusion and feature extraction of multi-dimensional data. Compared with traditional single sensor systems, multi-sensor fusion can capture more comprehensive environmental information, effectively reduce false alarms and missed alarms, and improve the accuracy and response speed of fire detection.

[0083] The device can not only accurately monitor the possibility of fire in different scenarios (such as indoors, outdoors, and industrial environments), but also dynamically evaluate the expansion trend and severity of the fire. Through real-time analysis and fusion of multiple environmental variables, the device can provide more reliable early warning information under complex conditions, providing a scientific basis for emergency decision-making and resource allocation. The application of multi-sensor fusion technology ensures the high adaptability and stability of the device in various environments, effectively improving the overall performance of fire warning and the efficiency of emergency response.

[0084] Reference Figure 1 In this example, a fire prediction and assessment method based on interpretable machine learning assisted by multiple types of sensors is implemented as follows:

[0085] Step 1: Collect the fire dataset T. The fire dataset T consists of time series data collected by multiple types of sensors and the real labels of whether a fire occurs at the corresponding moment. This ensures that the dataset covers the comprehensiveness of fire scenes. This process must fully consider various key factors, such as the diversity of fire sources, site environment conditions, etc., and divide the fire dataset T into a training set and a test set.

[0086] Step 1.1, collect a fire data set. This embodiment uses the fire sensor data set obtained by Stefan Blattmann (Yu R, Kim KA Study of Novel Initial Fire Detection Algorithm Based on Deep Learning Method [J]. Journal of Electrical Engineering & Technology, 2024, (prepublish): 1-12.), marked as T.

[0087]

[0088] in, represents the input sensor data, Indicates the output value corresponding to the input sensor data, that is, whether a fire occurs;

[0089] The types of sensors used are shown in Table 1:

[0090] Table 1 Sensor types

[0091]

[0092]

[0093] Step 1.2: Preprocess the fire dataset T in step 1.1 by the following method:

[0094]

[0095] Among them, μ and σ represent the mean and standard deviation of each sensor data, respectively. The fire dataset T is then divided into training set and test set with the ratio of 80% and 20%.

[0096] Step 2: Construct a convolutional neural network sub-model, use the training set obtained in step 1 to train the convolutional neural network sub-model, input the test set obtained in step 1 into the trained convolutional neural network sub-model, and determine whether the output result matches the corresponding true label by calculating the accuracy and loss value. Figure 2 As shown; after multiple rounds of training and testing, the model training is completed, and the optimal convolutional neural network sub-model after training is obtained; the sensor data set A' currently deployed in multiple locations is obtained, and the current sensor data set A' is input into the optimal convolutional neural network sub-model after training to obtain the output result, which is used as a basis for screening the current sensor data set A', deleting the data corresponding to the location where the fire will not occur, and obtaining the sensor data set A corresponding to the location where the fire will occur; and realizing the universal modeling of fire prediction.

[0097] In step 2.1, a convolutional neural network sub-model is used to process the time series data collected by multiple sensors. This convolutional neural network sub-model is designed to detect potential signals of fire from environmental monitoring data. The architecture of this convolutional neural network sub-model is as follows:

[0098] f(x)=σ(w2ReLU(w1*x+b1)+b2)

[0099] Where * represents the convolution operation, σ is the Sigmoid activation function, which is used to output the probability of predicting the positive class (i.e., detecting a fire); w1 and w2 are the weights of the model, and b1 and b2 are the bias terms;

[0100] Step 2.2: Input the training set obtained in step 1 into the convolutional neural network sub-model constructed in step 2.1 to obtain the final prediction probability, i.e., p. If p ≥ 0.5, then the prediction label at the corresponding moment is on the contrary,

[0101] Step 2.3, input the final predicted probability p obtained in step 2.2 and the true label at the corresponding time in the fire data set T in step 1 into the minimization of the binary cross entropy loss function for training to optimize the accuracy of fire detection. The Adam optimizer is used for parameter optimization, and the initial learning rate is set to 0.00001. During the training process, each batch of data undergoes forward propagation, loss calculation, and back propagation;

[0102] The loss calculation formula is as follows:

[0103]

[0104] Among them, N train is the number of samples in the training set, y i is the true label of the i-th sample, whose value is 0 or 1, p i is the probability that the model predicts the i-th sample as a positive class;

[0105] Step 2.4: Perform performance evaluation on the convolutional neural network sub-model trained in step 2.3 on the test set obtained in step 1. The main evaluation indicators are accuracy and loss value. The calculation formula is as follows:

[0106]

[0107] Among them, N test is the number of samples in the test set, y i is the true label of the i-th sample;

[0108] The loss value is calculated as follows:

[0109]

[0110] Among them, N test is the number of samples in the test set, y i is the true label of the i-th sample, whose value is 0 or 1, p i is the probability that the model predicts the i-th sample as a positive class;

[0111] Step 2.5, repeat steps 2.3 and 2.4 to obtain multiple rounds of test data, draw the loss curve and accuracy curve during training and testing, verify the performance and stability of the convolutional neural network sub-model, and obtain the optimal convolutional neural network sub-model;

[0112] Figure 3 The figure shows the change of model loss during training. The blue curve represents the training loss and the orange curve represents the test loss. As the number of iterations increases, both curves show a downward trend, which means that the model is gradually learning, the loss is gradually decreasing, and the model performance is improving. It can be observed that the training loss is always slightly higher than the test loss, which may indicate that the model has good generalization ability and no obvious overfitting phenomenon occurs.

[0113] Figure 4 The accuracy of the model during training is shown. As the number of iterations increases, the accuracy gradually increases from 87.5% to around 96.1%, showing a steady growth trend. This shows that the model is constantly optimizing and is gradually able to correctly classify data, and the training effect is good.

[0114] In general, from Figure 3 and Figure 4 It can be seen that the convolutional neural network performs stably during the training process, the loss is reduced, the accuracy is improved, and the model has good learning ability and generalization performance.

[0115] Step 2.6, obtain the sensor dataset A' currently deployed in multiple locations. This dataset consists of multiple sets of multi-type sensor data at different locations at the same time. Input the data in the current sensor dataset A' into the optimal convolutional neural network sub-model in step 2.5 one by one to obtain the output value p, that is, the probability of fire. If p ≥ 0.5, the prediction label at the corresponding moment is Do not delete, otherwise, Delete the corresponding data group in the current sensor data set A', repeat the above operation until all groups of data in the current sensor data set A' have completed the input and deletion operations, and the final current sensor data set A' is the sensor data set A corresponding to the location where the fire will occur.

[0116] Step 3: For the sensor data set A corresponding to the location where the fire will occur in step 2, use the principal component analysis (PCA) method to determine the main factors that affect the occurrence of the fire at this location. Based on this, data screening is performed on the basis of the sensor data set A corresponding to the location where the fire will occur obtained in step 2 to obtain a data set M containing only the main factors that affect the occurrence of the fire at this location.

[0117] The principal component analysis method is used to screen the main influencing factors to reduce the impact of the deviation and distortion of the weight results caused by the fact that the main factors affecting fire in different environments are different and there may be a high degree of multicollinearity between many factors (such as the concentration of various chemical gases).

[0118] The specific method of step 3 is:

[0119] Step 3.1, in order to eliminate the influence of dimension, firstly, the sensor data set A corresponding to the location where the fire will occur in step 2 is standardized to obtain the standardized data matrix Z. The standardized formula is:

[0120]

[0121] Among them, X ij is the value of the i-th sample on the j-th variable, is the mean of the jth variable, s j is the standard deviation of the jth variable;

[0122] Step 3.2, calculate the covariance matrix: Based on the data matrix Z standardized in step 3.1, calculate the covariance matrix C, whose elements Cjk Represents the covariance between variable j and variable k:

[0123]

[0124] Step 3.3, eigenvalue decomposition: Perform eigenvalue decomposition on the covariance matrix C obtained in step 3.2, specifically through the following formula:

[0125] CV=VD

[0126] Get the eigenvalue D and eigenvector V, where V is a matrix containing all eigenvectors of C; D is a diagonal matrix whose diagonal elements are the corresponding eigenvalues, representing the variance contribution of each feature of C;

[0127] Step 3.4, according to the size of the eigenvalue D obtained in step 3.3, based on the Kaiser criterion, select the principal component matrix P composed of the eigenvectors V corresponding to the first 7 eigenvalues ​​with eigenvalues ​​greater than 1, and obtain the data set M of the main factors affecting the occurrence of fire at this location according to the principal component matrix P. These principal components, that is, the matrix composed of the eigenvectors corresponding to the first m eigenvalues, explain most of the variance in the original data; specifically, these eigenvectors represent the weight (i.e., load) of the original variables on each principal component, and each principal component represents a set of linear combinations that aim to explain the most part of the data variance.

[0128] Step 4: Use the entropy weight method to evaluate the disaster situation after the fire occurs at the location for the data set M in step 3 that only contains the main factors affecting the occurrence of fire at the location.

[0129] The specific method of step 4 is:

[0130] Step 4.1: For each type of data collected by multiple types of sensors in the data set M that only contains the main factors affecting the occurrence of fire at this location in step 3, calculate its information entropy to evaluate the dispersion of the data. The higher the information entropy, the more irregular the data change of this type of sensor, and vice versa, the more regular the change. The information entropy E j Calculated by the following formula:

[0131]

[0132] Among them, p ij is the ratio of the value of the i-th sample on the j-th index to the total value of all samples of the index; k is a constant, n is the number of samples;

[0133] Step 4.2: Use the information entropy E obtained in step 4.1 j, calculate the weight of each data collected by multiple types of sensors in the data set M of the main factors affecting the occurrence of fire at this location, weight w j By calculating the complement of information entropy, the formula is as follows:

[0134]

[0135] Where m is the total number of indicators;

[0136] like Figure 5 As shown in the figure, it is the result of weight calculation. It can be seen that Raw Ethanol has the largest weight, reaching 42.92, followed by Raw H2, which is 25.57. The weights of the remaining indicators are relatively small. Among them, the weight of Pressure is 4.41, the weight of Temperature is 11.85, the weight of Humidity is 8.87, the weight of TVOC is 2.94, and the weight of eCO2 is 3.44.

[0137] Step 4.3, use the weight w obtained in step 4.2 j , weighted sum of each type of data collected by multiple types of sensors in the data set M that only contains the main factors affecting the occurrence of fire at this location is performed to obtain a comprehensive score for disaster assessment:

[0138]

[0139] Among them, x ij is the value of the i-th sample on the j-th index;

[0140] Step 4.4, determine the severity of the fire based on the comprehensive score of the disaster assessment in step 4.3. The higher the score, the greater the potential risk and severity of the fire.

[0141] The key points and protection points of the present invention are:

[0142] (1) Application of multi-sensor fusion technology: The present invention uses multiple types of sensors (such as temperature sensors, humidity sensors, smoke sensors, air pressure sensors, etc.) to monitor environmental data in real time. By fusing the data of these different sensors, the present invention can perform fire detection in a variety of complex environments, thereby improving the accuracy and reliability of fire warning.

[0143] (2) Multi-environment adaptability: The present invention designs a fire detection model with strong universality, which can adapt to a variety of different fire scenarios (such as indoor, outdoor, industrial environments, etc.). Through a large amount of data training and testing, the model performs well in different environments and has high adaptability and generalization ability.

[0144] (3) Innovative data fusion and processing methods: In order to improve the accuracy of fire detection, the present invention combines principal component analysis (PCA) and entropy weight method to process and analyze sensor data. These methods can reduce data redundancy, solve multicollinearity problems, and improve the interpretability and accuracy of the model.

[0145] (4) Efficient emergency management support: Through multi-sensor data fusion and intelligent analysis, the present invention can provide more accurate fire information for emergency management, optimize resource allocation, and help formulate more effective emergency response plans.

[0146] (5) Feature extraction and prediction model based on deep learning: This paper introduces a convolutional neural network sub-model deep learning algorithm to automatically extract effective features from multi-dimensional sensor data. Compared with traditional rules and threshold judgment methods, these algorithms can more accurately predict the occurrence and development trend of fires and reduce false alarms and missed alarms.

[0147] The present invention also provides a fire prediction and assessment system assisted by interpretable machine learning based on multiple types of sensors, comprising:

[0148] A fire data set T collection and processing module is used to collect the fire data set T in step 1. The fire data set T consists of time series data collected by multiple types of sensors and a real label of whether a fire occurs at a corresponding moment. The fire data set T is preprocessed and divided into a training set and a test set.

[0149] The module for acquiring the sensor data set A corresponding to the location where the fire will occur is used to implement the construction of the convolutional neural network sub-model in step 2, use the training set obtained in step 1 to train the convolutional neural network sub-model, input the test set obtained in step 1 into the trained convolutional neural network sub-model, and determine whether the obtained output result matches the corresponding true label by calculating the accuracy and loss value. After multiple rounds of training and testing, the model training is completed to obtain the optimal convolutional neural network sub-model after training; obtain the sensor data set A' currently deployed at multiple locations, input the current sensor data set A' into the optimal convolutional neural network sub-model after training, and obtain the output result, which is used as a basis for screening the current sensor data set A', deleting the data corresponding to the location where the fire will not occur, and obtaining the sensor data set A corresponding to the location where the fire will occur;

[0150] The module for acquiring the data set M of only the main factors affecting the occurrence of fire at the location is used to implement the sensor data set A corresponding to the location where the fire will occur in step 2 in step 3, use the principal component analysis (PCA) method to determine the main factors affecting the occurrence of fire at the location, and based on this, perform data screening on the basis of the sensor data set A corresponding to the location where the fire will occur obtained in step 2, and obtain the data set M of only the main factors affecting the occurrence of fire at the location;

[0151] The disaster assessment module is used to implement the entropy weight method in step 4 to conduct disaster assessment after a fire occurs at the location on the data set M of step 3 that only contains the main factors affecting the occurrence of fire at the location.

[0152] The present invention also provides a fire prediction and assessment device assisted by interpretable machine learning based on multiple types of sensors, comprising:

[0153] Memory: a computer program storing the above-mentioned fire prediction and assessment method assisted by interpretable machine learning based on multiple types of sensors, which is a computer-readable device;

[0154] Processor: used to implement the fire prediction and assessment method assisted by interpretable machine learning based on multiple types of sensors when executing the computer program.

[0155] The present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it can implement the fire prediction and assessment method assisted by interpretable machine learning based on multiple types of sensors.

Claims

1. A fire prediction and assessment method based on explainable machine learning assisted by multiple types of sensors, characterized in that: The steps include: Step 1: Collect a fire dataset T. The fire dataset T consists of time series data collected by multiple types of sensors and real labels of whether a fire occurs at the corresponding moment. The fire dataset T is preprocessed and divided into a training set and a test set. Step 2, construct a convolutional neural network sub-model, use the training set obtained in step 1 to train the convolutional neural network sub-model, input the test set obtained in step 1 into the trained convolutional neural network sub-model, and determine whether the output result matches the corresponding true label by calculating the accuracy and loss value. After multiple rounds of training and testing, complete the model training and obtain the optimal convolutional neural network sub-model after training; obtain the sensor data set A' currently deployed at multiple locations, input the current sensor data set A' into the optimal convolutional neural network sub-model after training, and obtain the output result. Use this as a basis to filter the current sensor data set A', delete the data corresponding to the location where fire will not occur, and obtain the sensor data set A corresponding to the location where fire will occur; Step 3: For the sensor data set A corresponding to the location where the fire will occur in step 2, use the principal component analysis (PCA) method to determine the main factors that affect the occurrence of the fire at this location. Based on this, data screening is performed on the basis of the sensor data set A corresponding to the location where the fire will occur obtained in step 2 to obtain a data set M containing only the main factors that affect the occurrence of the fire at this location. Step 4: Use the entropy weight method to evaluate the disaster situation after the fire occurs at the location for the data set M in step 3 that only contains the main factors affecting the occurrence of fire at the location.

2. The fire prediction and assessment method based on interpretable machine learning assisted by multi-type sensors according to claim 1 is characterized in that: In step 1, the collected fire data set T is preprocessed by the following method: Among them, μ and σ represent the mean and standard deviation of each sensor data, respectively. The fire dataset T is then divided into a training set and a test set.

3. The fire prediction and assessment method based on interpretable machine learning assisted by multi-type sensors according to claim 1 is characterized in that: The specific method of step 2 is: In step 2.1, the convolutional neural network sub-model is used to detect potential signals of fire from environmental monitoring data; the convolutional neural network sub-model architecture is as follows: f(x)=σ(w2ReLU(w1*x+b1)+b2) Where * represents the convolution operation, σ is the Sigmoid activation function, which is used to output the probability of predicting the positive class (i.e., detecting a fire); w1 and w2 are the weights of the model, and b1 and b2 are the bias terms; Step 2.2: Input the training set obtained in step 1 into the convolutional neural network sub-model constructed in step 2.1 to obtain the final prediction probability, i.e., p. If p ≥ 0.5, then the prediction label at the corresponding moment is on the contrary, Step 2.3, input the final predicted probability p obtained in step 2.2 and the true label of the corresponding moment in the fire data set T in step 1 into the minimization binary cross entropy loss function for training. During the training process, each batch of data undergoes forward propagation, loss calculation, and back propagation; The loss calculation formula is as follows: Among them, N train is the number of samples in the training set, y i is the true label of the i-th sample, whose value is 0 or 1, p i is the probability that the model predicts the i-th sample as a positive class; Step 2.4: Perform performance evaluation on the convolutional neural network sub-model trained in step 2.3 on the test set obtained in step 1. The evaluation indicators are accuracy and loss value. The calculation formula is as follows: Among them, N test is the number of samples in the test set, y i is the true label of the i-th sample; The loss value is calculated as follows: Among them, N test is the number of samples in the test set, y i is the true label of the i-th sample, whose value is 0 or 1, p i is the probability that the model predicts the i-th sample as a positive class; Step 2.5, repeat steps 2.3 and 2.4 to obtain multiple rounds of test data, draw the loss curve and accuracy curve during training and testing, verify the performance and stability of the convolutional neural network sub-model, and obtain the optimal convolutional neural network sub-model; Step 2.6, obtain the sensor data set A' currently deployed at multiple locations. The current sensor data set A' consists of multiple sets of multi-type sensor data at different locations at the same time. Input the data in the current sensor data set A' into the optimal convolutional neural network sub-model of step 2.5 one by one to obtain the output value p, that is, the probability of fire. If p ≥ 0.5, the prediction label at the corresponding time Do not delete, otherwise, Delete the corresponding data group in the current sensor data set A', repeat the above operation until all groups of data in the current sensor data set A' have completed the input and deletion operations, and the final current sensor data set A' is the sensor data set A corresponding to the location where the fire will occur.

4. The fire prediction and assessment method based on explainable machine learning assisted by multiple types of sensors according to claim 1 is characterized in that: The specific method of step 3 is: Step 3.1, standardize the sensor data set A corresponding to the location where the fire will occur in step 2 to obtain a standardized data matrix Z. The standardization formula is: Among them, X ij is the value of the i-th sample on the j-th variable, is the mean of the jth variable, s j is the standard deviation of the jth variable; Step 3.2, calculate the covariance matrix: Based on the data matrix Z standardized in step 3.1, calculate the covariance matrix C, whose elements C jk Represents the covariance between variables j and k: Step 3.3, eigenvalue decomposition: Perform eigenvalue decomposition on the covariance matrix C obtained in step 3.2, specifically through the following formula: CV=VD Get the eigenvalue D and eigenvector V, where V is a matrix containing all eigenvectors of C; D is a diagonal matrix whose diagonal elements are the corresponding eigenvalues, representing the variance contribution of each feature of C; Step 3.4, according to the size of the eigenvalue D obtained in step 3.3, based on the Kaiser criterion, select the principal component matrix P composed of the eigenvectors V corresponding to the first m eigenvalues ​​whose eigenvalues ​​are greater than 1. According to the principal component matrix P, obtain the data set M containing only the main factors affecting the occurrence of fire at this location.

5. The fire prediction and assessment method based on explainable machine learning assisted by multiple types of sensors according to claim 1 is characterized in that: The specific method of step 4 is: Step 4.1: For each type of data collected by multiple sensors in the data set M that only contains the main factors affecting the occurrence of fire at this location in step 3, calculate its information entropy to evaluate the dispersion of the data. The higher the information entropy, the more irregular the data change of this type of sensor, and vice versa, the more regular the change. The information entropy E j Calculated by the following formula: Among them, p ij is the ratio of the value of the i-th sample on the j-th index to the total value of all samples of the index; k is a constant, n is the number of samples; Step 4.2: Use the information entropy E obtained in step 4.1 j , calculate the weight of each data collected by multiple types of sensors in the data set M of the main factors affecting the occurrence of fire at this location, weight w j By calculating the complement of information entropy, the formula is as follows: Where m is the total number of indicators; Step 4.3, use the weight w obtained in step 4.2 j , weighted sum of each type of data collected by multiple types of sensors in the data set M that only contains the main factors affecting the occurrence of fire at this location is performed to obtain a comprehensive score for disaster assessment: Among them, x ij is the value of the i-th sample on the j-th index; Step 4.4, determine the severity of the fire based on the comprehensive score of the disaster assessment in step 4.

3. The higher the score, the greater the potential risk and severity of the fire.

6. A fire prediction and assessment system based on interpretable machine learning assisted by multiple types of sensors based on the method according to any one of claims 1 to 5, characterized in that: include: The fire data set T collection and processing module is used to collect the fire data set T. The fire data set T consists of time series data collected by multiple types of sensors and the real label of whether a fire occurs at the corresponding moment. The fire data set T is preprocessed and divided into a training set and a test set. A sensor data set A acquisition module corresponding to the location where a fire will occur is used to construct a convolutional neural network sub-model, use a training set to train the convolutional neural network sub-model, input the test set into the trained convolutional neural network sub-model, and determine whether the output result matches the corresponding true label by calculating the accuracy and loss value. After multiple rounds of training and testing, the model training is completed to obtain the optimal convolutional neural network sub-model after training; obtain the sensor data set A' currently deployed at multiple locations, input the current sensor data set A' into the optimal convolutional neural network sub-model after training, and obtain the output result. Based on this, the current sensor data set A' is screened, and the data corresponding to the location where a fire will not occur is deleted to obtain the sensor data set A corresponding to the location where a fire will occur; A module for acquiring a data set M containing only the main factors affecting the occurrence of a fire at the location, for determining the main factors affecting the occurrence of a fire at the location using a principal component analysis (PCA) method for a sensor data set A corresponding to the location where a fire will occur, and based on this, performing data screening on the basis of the sensor data set A corresponding to the location where a fire will occur, to obtain a data set M containing only the main factors affecting the occurrence of a fire at the location; The disaster assessment module is used to use the entropy weight method to conduct a disaster assessment after a fire occurs at a location on a data set M that only contains the main factors that affect the occurrence of a fire at that location.

7. A fire prediction and assessment device based on interpretable machine learning assisted by multiple types of sensors, characterized in that: include: Memory: a computer program storing a fire prediction and assessment method based on explainable machine learning assisted by multiple types of sensors as described in any one of claims 1 to 5, which is a computer-readable device; Processor: used to implement the fire prediction and assessment method based on explainable machine learning assisted by multiple types of sensors as described in any one of claims 1-5 when executing the computer program.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which, when executed by a processor, can implement a fire prediction and assessment method based on explainable machine learning assisted by multiple types of sensors as described in any one of claims 1 to 5.

Images