Chemical storage fire early warning method based on data fusion

Abstract

The invention discloses a chemical storage fire early warning method based on data fusion. A data acquisition module, a data fusion module and a data transmission module are included. The data acquisition module is composed of a plurality of sensor terminals, data acquired by sensors are uploaded to an OneNET platform through a core chip and a WiFi module, and when it is detected that the data are abnormal or do not meet a preset value, an alarm gives an alarm; the data fusion module is composed of a sensor terminal needing to be collected and a data fusion algorithm LHEA-CNN-DELM-SEAttention, and the data fusion algorithm is used for carrying out fire prediction on the collected chemical storage real-time environment data; and training the established LHEA-CNN-DELM-SEAttention model, and verifying the effectiveness of the model in the prediction of the chemical storage fire hazard. Compared with the prior art, through cooperative work of multiple modules, the accuracy and reliability are improved, the requirement for manual monitoring is reduced, and the risk of human errors is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of chemical storage fire early warning, and in particular to a chemical storage fire early warning method based on data fusion. Background Art

[0002] The technical principle of chemical storage fire warning mainly includes multi-sensor data collection, data transmission, data fusion and other steps. First, by deploying multiple types of sensor terminals in the chemical storage environment, real-time environmental data related to fire is collected, and then the collected sensor data is transmitted from the site to the cloud platform. Then the data fusion algorithm comprehensively analyzes the data from different sensors and predicts and warns the fire. Finally, the effectiveness of the algorithm model in chemical storage fire prediction is verified.

[0003] The chemical industry is extremely dangerous and risky. During chemical storage and production, the storage, handling and transportation of chemicals may cause fire or explosion accidents. Fire may not only cause huge economic losses, but also pose a serious threat to the safety of life and the environment. Therefore, chemical companies should improve their fire safety protection capabilities and reduce fire risks by installing advanced fire warning systems, strengthening employee training, implementing strict safety management, regular equipment maintenance, optimizing plant design and establishing emergency response mechanisms.

[0004] Traditional fire warning methods usually rely on a single type of sensor, such as smoke detectors or temperature sensors, which have limitations, such as the inability to integrate multiple factors for early warning, delayed response, high false alarm rate, lack of intelligent data analysis capabilities, difficult maintenance, and slow response. These deficiencies may lead to inaccurate and in-time identification of fire risks, affecting the overall effectiveness and reliability of the system. Summary of the invention

[0005] Purpose of the invention: In view of the shortcomings of the prior art, the present invention provides a chemical storage fire warning method based on data fusion. By integrating multiple sensors (such as temperature, humidity, smoke and gas sensors) and data fusion algorithms, it can comprehensively analyze different types of data, monitor the fire risks in the storage environment in real time, and reduce false alarms and missed alarms.

[0006] Technical solution: The present invention provides a chemical storage fire early warning method based on data fusion, comprising the following steps:

[0007] Step 1: Conduct real-time monitoring of environmental temperature and humidity, smoke and CO concentration, and flames in the chemical storage area, and perform personnel detection at the chemical storage access control;

[0008] Step 2: Display the collected real-time data on the display screen and upload it to the OneNET platform; when the data is detected to be abnormal or does not meet the preset value, the alarm will sound an alarm;

[0009] Step 3: Perform fire prediction on the collected real-time environmental data of chemical storage, normalize the real-time environmental data, use the processed data as the data set for fire prediction, and divide it into a training set and a test set; the real-time environmental data includes temperature, humidity, CO and smoke concentration, and open fire probability;

[0010] Step 4: Construct a data fusion algorithm model LHEA-CNN-DELM-SE_Attention, with temperature, humidity, CO and smoke concentration as input data and open fire probability as output data; the data fusion algorithm model LHEA-CNN-DELM-SE_Attention optimizes the hyperparameters of the CNN-DELM-SE_Attention model based on the dragonfly optimization algorithm LHEA to obtain an optimized CNN-DELM-SE_Attention model; the CNN-DELM-SE_Attention model combines the convolutional neural network CNN and the deep extreme learning machine DELM, and introduces the SE attention mechanism;

[0011] Step 5: Train the data fusion algorithm model LHEA-CNN-DELM-SE_Attention to verify the effectiveness of the model in chemical storage fire prediction, and use the trained data fusion algorithm model LHEA-CNN-DELM-SE_Attention for fire prediction.

[0012] Furthermore, in step 1, data collection uses temperature and humidity sensors, smoke and gas sensors, flame sensors, and human infrared sensors.

[0013] Furthermore, in the step 2, the OneNET platform interface can view the fire hazard level of the storage area. The fire hazard level is divided into three levels. When only smoke or CO is detected, it is level one; when both smoke and CO are detected, it is level two; when only flame is detected, it is the highest hazard level three; the OneNET platform changes the monitored temperature and humidity thresholds according to the storage requirements of the chemical storage; when an abnormality is detected, the OneNET platform interface will pop up a warning and an alarm will sound.

[0014] Furthermore, in step 3, the real-time environmental data is normalized. During normalization, the characteristic parameter values ​​are clipped. If the values ​​exceed the specified clipping value, they are all 1 after normalization. The range of the data values ​​is normalized to [0, 1]. The calculation formula is:

[0015]

[0016] Among them, A j Represents the normalized value of parameter j input; LI j Indicates the limit value of parameter j determined by real-time data; FM j represents the reference threshold of parameter j; x j represents the input value of parameter j; k represents the environmental compensation correction value.

[0017] Furthermore, the CNN-DELM-SE_Attention model in step 4 includes a sequence folding layer and an anti-folding layer, a convolution layer Conv_1, a convolution layer Conv_2, a ReLU_1 activation layer, a ReLU_2 activation layer, a ReLU_3 activation layer, a global average pooling layer GAP, a fully connected layer 1, a fully connected layer 2, and a Sigmoid activation function, and also includes a deep extreme learning machine DELM, a fully connected layer 3, and a regression layer;

[0018] The output of the folding layer is connected to the input of the convolution layer Conv_1. The output of the convolution layer Conv_1 is connected to the ReLU_1 activation layer and the global average pooling layer GAP. After the ReLU_1 activation layer, the convolution layer Conv_2 and the ReLU_2 activation layer are connected in series.

[0019] The output of the global average pooling layer GAP is then connected to the fully connected layer 1 in series with the ReLU_3 activation layer and the fully connected layer 2, and finally processed by the Sigmoid activation function;

[0020] After Sigmoid activation function processing, the output of the ReLU_2 activation layer is connected to the dot product layer, and the output of the dot product is connected to the input of the anti-folding layer;

[0021] The output of the folding layer is connected to the deep extreme learning machine DELM, and the output of the deep extreme learning machine DELM is predicted through the fully connected layer 3 and the regression layer.

[0022] Furthermore, the dragonfly optimization algorithm LHEA in step 4 optimizes the hyperparameters of the CNN-DELM-SE_Attention model, including the number of convolutional layers, the range of the number of convolutional kernels and the size of the convolutional kernels, the number of deep extreme learning machine DELM units, and the number of neurons in the fully connected layer.

[0023] 7. A chemical storage fire early warning method based on data fusion according to claim 6, characterized in that the dragonfly optimization algorithm LHEA is specifically as follows:

[0024] 1) Description of the separation process in the LHEA exploration phase:

[0025]

[0026] in, It is represented as the separation process of individual i at position t in the evolutionary iteration; represents the position of the individual with current index i in evolutionary iteration t; represents the position of the individual j in the neighborhood in evolutionary iteration t; N represents the number of individuals in the neighborhood;

[0027] 2) Alignment equations for the LHEA exploration phase:

[0028]

[0029] in, It is represented as the alignment process of individual i at position t in the evolutionary iteration; represents the position of the individual j in the neighborhood in evolutionary iteration t; N represents the number of individuals in the neighborhood;

[0030] 3) Cohesion calculation equation in the LHEA exploration phase:

[0031]

[0032] in, The process of calculating the cohesion of individual i at position t in the evolutionary iteration is represented as; represents the position of the individual with current index i in evolutionary iteration t; represents the position of the individual j in the neighborhood in evolutionary iteration t; N represents the number of individuals in the population; N represents the number of individuals in the neighborhood;

[0033] 4) Description of the food attraction calculation equation in the exploration phase of LHEA:

[0034]

[0035] Among them, F i t The food attraction process of individual i at position t in the evolutionary iteration is represented as; represents the position of the individual with current index i in evolutionary iteration t; represents the target source position generated by the current evolutionary iteration t;

[0036] 5) The calculation equation for disrupting enemy targets during the LHEA exploration phase is described as follows:

[0037]

[0038] in, It is represented as the process of disrupting the enemy target by the individual i at position t in the evolutionary iteration; Indicates the enemy position generated by the current evolution iteration t; represents the target source position generated by the current evolutionary iteration t;

[0039] 6) Step vector calculation equation for LHEA exploration phase:

[0040]

[0041] in, The step vector update position of individual i in the evolutionary iteration; α represents the calibration coefficient; It is the alignment of individuals i; c represents the cohesion coefficient; It is represented by the cohesion of the individual i; e represents the enemy factor; i represents the enemy individual; f represents the food factor; F i t It is represented by the food source of individual i; s represents the separation coefficient; It represents the iterative separation of individual i during the evolution process; w represents the inertia weight; The position of the step vector of individual i in the evolutionary iteration;

[0042] 7) Position vector calculation equation in the LHEA development phase:

[0043]

[0044] in, The position of the step vector of individual i in the evolutionary iteration; represents the position of the individual with the current index i in the evolution iteration t; R represents the growth region; g* represents the best position in the evolution iteration; t represents the evolution iteration of the current algorithm; L represents the final number of iterations of the algorithm;

[0045] 8) Calculation equation of deep pit capacity during the LHEA intensification phase:

[0046]

[0047] in, It is represented by the pit capacity of individual i at position t in the evolutionary iteration; f i t It is represented by the pit size of individual i at position t in the evolutionary iteration; f max Indicates the maximum fitness in the pit; f min Indicates the minimum fitness in the deep pit; Const indicates the maximum capacity constant in the deep pit;

[0048] 9) Deep pit selection probability during the LHEA strengthening phase:

[0049]

[0050] in, It is represented as the probability of pit selection of individual i at position t in the evolutionary iteration; It represents the capacity of the deep pit of individual i at position t in the evolutionary iteration; k represents the number of pits with a capacity higher than the overflow pit;

[0051] 10) The velocity and position equations of water droplets during the local search in the LHEA enhancement phase are described as follows:

[0052] V i t+1 =q×V i t

[0053]

[0054] Among them, V i t+1 V is the speed of the water droplet i during local search in the evolutionary iteration; i t denoted as the current velocity of the water droplet i at position t in the evolutionary iteration; q represents the velocity increment coefficient; The position of the water droplet i during the local search in the evolutionary iteration; The current position of the water droplet denoted as i at position t in the evolutionary iteration; Represented as the current position of the deepest pit in evolutionary iteration t.

[0055] Furthermore, in step 5, the training parameters of the established LHEA-CNN-DELM-SE_Attention model are set, the training cycle is set to 80 rounds in total, 10 iterations in each round, 800 iterations in total, the initial learning rate is 0.01, the learning rate reduction factor is 0.1, and the learning rate reduction cycle is 400. The data is shuffled at the beginning of each training cycle. The data set used for model training consists of normalized data on temperature, humidity, CO concentration, smoke concentration, and open fire probability of chemical storage. The mean absolute error MAE, mean square error MSE, root mean square error RMSE, and determination coefficient R are selected. 2 As the basis for evaluating the prediction performance of the model, the effectiveness of the model in predicting chemical storage fires was verified.

[0056] Beneficial effects:

[0057] The present invention adopts a chemical storage fire early warning method based on data fusion. The data fusion module fuses multiple data from different sensors through an algorithm, reducing the false alarm or missed alarm phenomenon that may exist in a single sensor and improving the detection accuracy of the overall system. This fusion strategy can make more accurate fire predictions based on multi-source data. Through the collaborative work of multiple modules, the present invention improves accuracy and reliability, reduces the need for manual monitoring, and reduces the risk of human error.

[0058] The OneNET platform enables real-time data transmission and sharing, while the host computer can send instructions to interact with the platform. This real-time communication capability enhances the system's response speed, ensuring that warnings can be quickly issued and corresponding measures can be taken when fire risks are detected.

[0059] The present invention analyzes time series data and uses the principle of extreme learning machine to construct a deep network. CNN focuses on processing and learning features in data through convolution operations. The architecture of DELM usually does not include convolution layers, and the core feature of CNN is convolution layers. After combining DELM and CNN, the powerful feature extraction ability of CNN and the advantages of easy implementation and efficient classification of DELM can be fully utilized. By introducing the SE attention mechanism, SE_Attention performs weighted synthesis of different feature channels to highlight important features.

[0060] The present invention further optimizes the model based on the dragonfly optimization algorithm. LHEA is used as an optimization algorithm to optimize the CNN-DELM-SE_Attention model, thereby improving the accuracy and stability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 This is a structural diagram of a chemical storage fire warning method based on data fusion;

[0062] Figure 2 A physical diagram of a multi-sensor data transmission module in an embodiment of the present invention;

[0063] Figure 3 The OneNET platform in the embodiment of the present invention monitors and obtains chemical storage environment information in real time;

[0064] Figure 4 This is a humidity (Humi) data graph obtained in real time by the OneNET platform in an embodiment of the present invention;

[0065] Figure 5 It is a structural diagram of the CNN-DELM-SE_Attention model in an embodiment of the present invention;

[0066] Figure 6This is a model effect diagram of LHEA in an embodiment of the present invention;

[0067] Figure 7 This is a comparison chart of the prediction results of the training set and the test set of the LHEA-CNN-DELM-SE_Attention model algorithm in an embodiment of the present invention;

[0068] Figure 8 This is a prediction error graph of the LHEA-CNN-DELM-SE_Attention model algorithm test set in an embodiment of the present invention. DETAILED DESCRIPTION

[0069] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0070] The present invention discloses a chemical storage fire early warning method based on data fusion, which specifically comprises the following steps:

[0071] Step (1), such as Figure 2 As shown, the data acquisition module uses temperature and humidity sensors, smoke and gas sensors, and flame sensors to monitor the temperature and humidity, smoke and CO concentration, and flames of the chemical storage area in real time; and uses human infrared sensors to detect personnel at the chemical storage access control.

[0072] The temperature and humidity sensors, smoke and gas sensors, flame sensors, and human infrared sensors used for data collection are: SHTC3 digital sensor, MQ-2 smoke sensor, MQ-7 carbon monoxide sensor, Risym infrared flame sensor, and HC-SR505 human infrared sensing module. The temperature measurement range is -40 to 125°C, the humidity measurement range is 0 to 100% RH, the smoke concentration measurement range is 10-1000ppm, and the CO concentration measurement range is 20-2000ppm. The SHTC3 temperature and humidity sensor converts the original value into a physical quantity.

[0073] The relative temperature conversion formula is (the result is ℃):

[0074]

[0075] The relative humidity conversion formula is (the result is %RH):

[0076]

[0077] Where T is relative temperature; RH is relative humidity; S T and S RH Indicates the temperature and humidity of the original sensor.

[0078] Step (2), such as Figure 3 As shown, the data transmission module displays the collected real-time data on the display screen and uploads it to the OneNET platform at the same time, wherein the uploaded data values ​​include: temperature alarm value (Temp_THR), humidity alarm value (Humi_THR), administrator setting (Admi), real-time personnel detection in chemical warehouse (HCSR), real-time flame monitoring in chemical warehouse (FIRE), real-time CO concentration in chemical warehouse (CO), real-time temperature in chemical warehouse (Temp), real-time humidity in chemical warehouse (Humi), real-time smoke concentration in chemical warehouse (SMOG), real-time danger level in chemical warehouse (LEVEL); when abnormal data is detected or the preset value is not met, the alarm sounds an alarm.

[0079] The OneNET platform interface can view the fire hazard level of the storage area. The fire hazard level is divided into three levels. When only smoke or CO is detected, it is level one; when both smoke and CO are detected, it is level two; as long as flame is detected, it is the highest hazard level three; the OneNET platform can change the monitored temperature and humidity thresholds according to the storage needs of chemical storage; when an abnormality is detected, there will be a corresponding pop-up warning on the platform interface and an alarm will sound, which can be cleared through the administrator mode; when the sensor detects someone approaching, the alarm will sound and transmit real-time data to the OneNET platform, showing 1 if someone is approaching and 0 if no one is approaching, and the alarm can be cleared through the administrator mode; ESP8266 is selected as the WiFi communication module. The module has two working modes: STA mode (Station) and AP mode (Access Point), and can also switch between the two modes; the parameter display uses an IPS-LCD display, which has a more uniform brightness distribution, higher contrast, faster response speed, longer service life and no risk of screen burn-in compared to an OLED display.

[0080] In step (3), the data fusion module is composed of the sensor terminals that need to be collected and the data fusion algorithm. The data fusion algorithm performs fire prediction on the collected real-time environmental data of chemical storage: the real-time data collected by the required multiple sensors (temperature, humidity, CO and smoke concentration) are exported from the OneNET platform, and the real-time data are normalized; the processed data is used as the data set of the data fusion algorithm model LHEA-CNN-DELM-SE_Attention, and the data set is divided into a training set and a prediction set to test the model's fire prediction accuracy and other indicators.

[0081] Take obtaining the real-time humidity value in a chemical storage as an example. Figure 4 As shown, the OneNET platform obtains the humidity (Humi) data graph from multiple sensors in real time. Through the above method, other data sets for predicting fires are obtained.

[0082] First, the collected data on the OneNET platform is exported and read, and the data is analyzed: the first four columns of data (temperature, humidity, CO and smoke concentration) are used as input judgment basis, and the last column of data (open fire probability) is used as output result; the data and order are shuffled to ensure the randomness and robustness of the data.

[0083] Secondly, all input data are normalized. Data normalization is to prevent large values ​​from drowning small values. During normalization, the characteristic parameter values ​​are limited. If they exceed the specified limit value, they are all 1 after normalization. The range of data values ​​is normalized to [0,1]. The calculation formula is:

[0084]

[0085] Among them, A j Represents the normalized value of parameter j input; LI j Indicates the limit value of parameter j determined by real-time data; FM j represents the reference threshold of parameter j; x j represents the input value of parameter j; k represents the environmental compensation correction value.

[0086] Then, the processed 336 sets of data are divided into training sets and test sets in a ratio of 7:3, and the divided data are flattened into specific dimensions to facilitate input into the network for training.

[0087] Then establish the LHEA-CNN-DELM-SE_Attention model, such as Figure 5 As shown, the convolutional neural network CNN and the deep extreme learning machine DELM are combined, and the SE attention mechanism is introduced, including sequence folding layer and anti-folding layer, convolution layer Conv_1, convolution layer Conv_2, ReLU_1 activation layer, ReLU_2 activation layer, ReLU_3 activation layer, global average pooling layer GAP, fully connected layer 1, fully connected layer 2 and Sigmoid activation function, as well as deep extreme learning machine DELM, fully connected layer 3 and regression layer.

[0088] The output of the folding layer is connected to the input of the convolution layer Conv_1, and the output of the convolution layer Conv_1 is connected to the ReLU_1 activation layer and the global average pooling layer GAP. After the ReLU_1 activation layer, the convolution layer Conv_2 and the ReLU_2 activation layer are connected in series.

[0089] The global average pooling layer GAP outputs to the fully connected layer 1, which is then connected in series with the ReLU_3 activation layer and the fully connected layer 2, and finally processed by the Sigmoid activation function.

[0090] After being processed by the Sigmoid activation function, the output of the ReLU_2 activation layer is connected to the dot product layer, and the output of the dot product is connected to the input of the anti-folding layer.

[0091] The output of the folding layer is connected to the deep extreme learning machine DELM, and the output of the deep extreme learning machine DELM is predicted through the fully connected layer 3 and the regression layer.

[0092] 1) Use the Sequence Folding Layer and Sequence Unfolding Layer to process time series data: This model mainly analyzes time series data. The sequence folding layer folds the input sequence data and merges each element in the sequence with its previous element to reduce the length of the sequence; the sequence unfolding layer restores the folded sequence data and restores each element in the sequence to its original position. This operation can be used to restore the length of the sequence and retain detailed information in the sequence.

[0093] 2) Add convolutional layer and ReLU activation layer to the network: Create a convolutional layer with a convolution kernel size of [1,1] and a channel number of 32. Extract data features through the convolutional layer. The convolutional layer is the core part of the convolutional neural network CNN. Its calculation formula is:

[0094]

[0095] Among them, X represents input data; W represents weight matrix; Represents the convolution operation; s(i,j) represents the output data, (i,j) represents an element in the input matrix X and the weight matrix W, x(i,j) represents the i-th row and j-th column element of the input data x; w(m,n) represents the m-th row and n-th column element of the weight matrix W.

[0096] The activation function ReLU(x)=max(0,x) is then used to enhance the nonlinearity of the network. Furthermore, another convolutional layer (with 64 channels) and a ReLU activation layer are added to increase the complexity of feature extraction.

[0097] 3) Add SE attention mechanism (SE_Attention) to the network: SE_Attention weights different feature channels to highlight important features. The overall process of SE attention mechanism includes:

[0098] Global average pooling layer (GAP): Perform global average pooling on the feature map output by the convolution layer to obtain the average value of each channel; Fully connected layer: Input the result of global average pooling to two fully connected layers, first a fully connected layer that reduces the number of channels (reducing 64 channels to 16), and then restores it to the original number of channels (64); Activation function (ReLU and Sigmoid): Add nonlinearity through the ReLU activation function, and then use the Sigmoid function to compress the result to between 0 and 1 to form the attention weight of each channel. Point product: Multiply the obtained attention weight with the feature map output by the convolution layer channel by channel to adjust the importance of each channel. The specific steps are as follows:

[0099] First, perform global average pooling (GAP) on the output of the convolutional layer. The global average pooling operation averages all spatial positions of each feature map. The calculation formula is:

[0100]

[0101] Where GAP(x) represents the value of global average pooling of input parameter x; H and W represent the height and width of the feature map respectively, x ij Represents the value of position (i, j) in the feature map.

[0102] Next, the fully connected layer maps the pooled features to 16 neurons, and after a ReLU activation layer, another fully connected layer maps them to 64 neurons, applying the Sigmoid activation function. Calculate the feature weights. Finally, multiply the obtained attention weights by the feature map output by the convolutional layer channel by channel through dot multiplication to adjust the importance of each channel.

[0103] 4) Adding Deep Extreme Learning Machine (DELM): DELM mainly uses the principle of extreme learning machine to build deep networks, while CNN focuses on processing and learning features in data through convolution operations. DELM architecture usually does not include convolution layers, while the core feature of CNN is convolution layers. After combining DELM and CNN, the advantages of CNN's powerful feature extraction ability and DELM's easy implementation and efficient classification can be fully utilized.

[0104] DELM builds a deep network by stacking multiple ELM units. Each layer of ELM units learns the high-order feature representation of the input data and uses these features as the input of the next layer. ELM is a feedforward neural network training algorithm that avoids the gradient descent process in traditional neural network training by randomly generating input weights and biases and then calculating the analytical solution of the output weights. Each layer of ELM units in DELM independently learns the representation of the input data until the last layer, and then uses the standard ELM algorithm to learn the mapping from the feature representation of the last layer to the target output. By selecting the depth of the network (i.e., the number of ELM layers) and the number of hidden nodes in each layer, the input weights W are randomly initialized for each layer. i and bias B i For each layer of ELM, the output of the previous layer (the input data for the first layer) is used as the hidden layer output of the current layer. The specific formula is:

[0105] H i =σ(W i X+B i )

[0106] Among them, H i represents the output of the current hidden layer; σ represents the Sigmoid activation function; W i represents the weight matrix; B i represents the bias vector; X represents the output or input data.

[0107]

[0108] Among them, β represents the output weight; H represents the hidden layer output matrix of the last layer; represents the Moore-Penrose generalized inverse of H; T represents the target output.

[0109]

[0110] in, Represents the predicted output value of the model; W i represents weight; X represents output or input data; B i represents the bias vector; σ represents the Sigmoid activation function.

[0111] 5) Output regression: Add a fully connected layer and a regression layer for prediction. The regression layer is used for regression tasks and calculates the error between the predicted value and the actual value.

[0112] 6) Add the dragonfly optimization algorithm LHEA: This algorithm further optimizes the model based on the dragonfly optimization algorithm. In the LHEA optimization algorithm, the performance and feature extraction ability of the convolution layer are optimized by adjusting the step vector during the iteration process; the learning and memory ability of the model on sequence data is optimized by adjusting the number of DELM units; the number of neurons in the fully connected layer is adjusted to optimize the feature integration and propagation ability of the model. The main improvements include dynamically adjusting the search position and other parameters to balance the local position optimization. The position update formula adjusts the position of the search agent based on the neighborhood information and the local optimal position. LHEA is added as an optimization algorithm to optimize the CNN-DELM-SE_Attention model, thereby improving the accuracy and stability of the model. The optimized hyperparameters are specifically set as follows: the number of convolution layers ranges from 1 to 3; the number of convolution kernels ranges from 4 to 16; the size of the convolution kernel ranges from 4 to 5, the number of DELM units, the number of neurons in the fully connected layer, etc. These parameters are continuously adjusted during the optimization process to find the combination that performs best on the validation set. In this way, the hyperparameters of the model are optimized and the accuracy and stability of the prediction model are improved. Figure 6 This is the improved model effect diagram of LHEA. The specific steps are as follows:

[0113] ① LHEA introduces the separation process of individual positions in the exploration phase, taking into account the position of individuals in the neighborhood and optimizing according to the number of individuals in the neighborhood. Individuals are adjusted according to the positions of other individuals in the neighborhood during the iteration process. In the model, due to the differences between different parameters, the optimization of the separation process helps to explore a wider solution space and find a better parameter configuration. Description of the separation process in the exploration phase of LHEA:

[0114]

[0115] in, It is represented as the separation process of individual i at position t in the evolutionary iteration; represents the position of the individual with current index i in evolutionary iteration t; represents the position of individual j in the neighborhood in evolutionary iteration t; N represents the number of individuals in the neighborhood.

[0116] ② In the exploration phase, LHEA uses the alignment equation to optimize the position of individuals in the evolutionary iteration, considering the influence of the position of individuals in the neighborhood on the current individual position. In the evolutionary iteration, individuals adjust their positions according to the alignment of other individuals in the neighborhood; in the model optimization, they adjust their positions according to the performance or characteristics of other individuals in the neighborhood. The alignment process helps individuals learn and coordinate with each other during the optimization process to improve the overall performance. Alignment equation in the exploration phase of LHEA:

[0117]

[0118] in, It is represented as the alignment process of individual i at position t in the evolutionary iteration; represents the position of individual j in the neighborhood in evolutionary iteration t; N represents the number of individuals in the neighborhood.

[0119] ③ In the exploration phase, LHEA optimizes the cohesion of individuals through the cohesion calculation equation, taking into account the number of individuals in the population and the neighborhood to improve the efficiency and stability of the algorithm; in model optimization, individuals form a stable and efficient configuration in the parameter space, and this cohesion helps maintain the stability and consistency of the optimization process. The cohesion calculation equation of the LHEA exploration phase is:

[0120]

[0121] in, The process of calculating the cohesion of individual i at position t in the evolutionary iteration is represented as; represents the position of the individual with current index i in evolutionary iteration t; represents the position of the individual j in the neighborhood in evolutionary iteration t; N represents the number of individuals in the population; N represents the number of individuals in the neighborhood.

[0122] ④ LHEA takes into account the calculation process of food attraction in the exploration phase and adjusts the individual position according to the target source position generated by the current evolutionary iteration; in model optimization, the individual is adjusted towards the target optimization direction. The calculation of food attraction helps individuals move towards a better parameter configuration during the optimization process. The calculation equation of food attraction in the exploration phase of LHEA is described as:

[0123]

[0124] Among them, F i t The food attraction process of individual i at position t in the evolutionary iteration is represented as; represents the position of the individual with current index i in evolutionary iteration t; Represents the target source position produced by the current evolutionary iteration t.

[0125] ⑤LHEA introduces the calculation process of disrupting enemy targets in the exploration phase, considering the disturbance adjustment of the enemy position in the current evolutionary iteration to optimize the algorithm's search ability. In model optimization, it is used to learn how individuals avoid falling into local optimal solutions and explore parameter space more extensively. The calculation equation for disrupting enemy targets in the exploration phase of LHEA is described as follows:

[0126]

[0127] in, It is represented as the process of disrupting the enemy target by the individual i at position t in the evolutionary iteration; Indicates the enemy position generated by the current evolution iteration t; Represents the target source position produced by the current evolutionary iteration t.

[0128] ⑥ In the exploration phase, LHEA updates the step position of individuals through the step vector calculation equation, taking into account factors such as calibration coefficient, alignment, cohesion, enemy factor and food factor, and optimizes the algorithm's iteration process and search effect; in model optimization, individuals adjust their parameter configuration according to a variety of influencing factors, and the calculation of the step vector helps to balance the relationship between exploration and utilization during the optimization process. The step vector calculation equation of the LHEA exploration phase is:

[0129]

[0130] in, The step vector update position of individual i in the evolutionary iteration; α represents the calibration coefficient; It is the alignment of individuals i; c represents the cohesion coefficient; It is represented by the cohesion of the individual i; e represents the enemy factor; i represents the enemy individual; f represents the food factor; F i t It is represented by the food source of individual i; s represents the separation coefficient; It represents the iterative separation of individual i during the evolution process; w represents the inertia weight; Represented as the position of the step vector of individual i in the evolutionary iteration.

[0131] ⑦LHEA introduced the position vector calculation equation in the development stage, and optimized the position of individuals in the evolutionary iteration by considering factors such as the growth area, the optimal position and the number of iterations; in the model optimization, individuals adjust their parameter configuration according to the current optimization state and iteration progress to converge to the optimal solution faster. The position vector calculation equation in the development stage of LHEA is:

[0132]

[0133] in, The position of the step vector of individual i in the evolutionary iteration; represents the position of the individual with the current index i in evolutionary iteration t; R represents the growth region; g* represents the best position in the evolutionary iteration; t represents the evolutionary iteration of the current algorithm; L represents the final number of iterations of the algorithm.

[0134] ⑧LHEA introduces the deep pit capacity calculation equation in the strengthening stage, taking into account the size of the fitness and capacity constant in the deep pit, and optimizing the efficiency of the algorithm in the deep pit selection process; in the later stage of the optimization process, the individual selects its next moving direction and strategy according to the characteristics of the deep pit (the high fitness area in the solution space); in the model optimization, the deep pit capacity helps the individual to focus on the exploration and utilization of high potential areas in the later stage of optimization. The deep pit capacity calculation equation of LHEA strengthening stage is:

[0135]

[0136] in, It is represented by the pit capacity of individual i at position t in the evolutionary iteration; f i t It is represented by the pit size of individual i at position t in the evolutionary iteration; f max Indicates the maximum fitness in the pit; f min It represents the minimum fitness in the deep pit; Const represents the maximum capacity constant in the deep pit.

[0137] ⑨LHEA considers the calculation of deep pit selection probability in the strengthening stage, adjusts the individual selection strategy according to the deep pit capacity and the number of overflow pits, and improves the search efficiency of the algorithm; in model optimization, the selection probability is conducive to individuals to concentrate on exploring and utilizing areas with high potential in the later stages of optimization. Deep pit selection probability in the strengthening stage of LHEA:

[0138]

[0139] in, It is represented as the probability of pit selection of individual i at position t in the evolutionary iteration; It represents the capacity of the deep pit at position t of individual i in the evolutionary iteration; k represents the number of pits with a capacity higher than the overflow pit.

[0140] ⑩LHEA introduces the description of the velocity and position equations of water droplets during local search in the enhanced stage. Through the velocity increment coefficient and the current position of the deepest pit, the local area in the solution space is explored more deeply, and the performance of the algorithm in the local search process is optimized; in model optimization, local search helps individuals discover and utilize more accurate parameter configurations. The velocity and position equations of water droplets during local search in the enhanced stage of LHEA are described as follows:

[0141] V i t+1 =q×V i t

[0142]

[0143] Among them, Vi t+1 V is the speed of the water droplet i during local search in the evolutionary iteration; i t denoted as the current velocity of the water droplet i at position t in the evolutionary iteration; q represents the velocity increment coefficient; The position of the water droplet i during the local search in the evolutionary iteration; The current position of the water droplet denoted as i at position t in the evolutionary iteration; Represented as the current position of the deepest pit in evolutionary iteration t.

[0144] Step (4), set the training parameters of the established LHEA-CNN-DELM-SE_Attention model. The training cycle is set to 80 rounds, 10 iterations per round, 800 iterations in total, the initial learning rate is 0.01, the learning rate reduction factor is 0.1, the learning rate reduction cycle is 400, and the data is shuffled at the beginning of each training cycle. The data set used for model training consists of 336 sets of normalized data on temperature, humidity, CO concentration, smoke concentration, and open fire probability of chemical storage.

[0145] The present invention uses the mean absolute error MAE, mean square error MSE, root mean square error RMSE and determination coefficient R 2 As the basis for evaluating the prediction performance of the model, the relevant calculation formula is:

[0146]

[0147] Among them, y i represents the actual value of the i-th observation; represents the predicted value of the i-th observation; represents the actual value mean of the ith observation; n represents the total number of observations.

[0148] In order to verify the effectiveness of the LHEA-CNN-DELM-SE_Attention chemical storage fire prediction model proposed in this invention to improve the original algorithm, the CNN and CNN-BiLSTM-SE_Attention models were selected for comparative verification. The results are shown in Table 1.

[0149] Table 1 Comparative analysis of chemical storage fire prediction models

[0150]

[0151] like Figure 7As shown in the figure, the prediction results of the training set and test set of the LHEA-CNN-DELM-SE_Attention model algorithm. Among them, the black line represents the actual true value, the red line represents the predicted value of the model training set, and the blue line represents the predicted value of the test set of the model algorithm. The predicted values ​​of the model training set and the predicted values ​​of the test set are closely following the true values, which shows that the model has a good fit on both training and test data. The closeness between the predicted value of the test set and the true value can be used as an indicator of the generalization ability of the model. It can be seen from the figure that the predicted value of the test set is very close to the true value, indicating that the model has good generalization ability and meets the expectations of chemical storage fire prediction.

[0152] like Figure 8 As shown in the figure, the prediction error results of the LHEA-CNN-DELM-SE_Attention model algorithm test set. The error can be positive or negative, indicating the deviation between the predicted value and the actual value. From the figure, it can be seen that the error range is between -0.05 and 0.06, which shows that the prediction error of the model is relatively small, indicating that the prediction accuracy of the model is high; the error distribution does not have obvious large fluctuations, which means that the model is relatively stable in terms of prediction performance.

[0153] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with the technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent transformation or modification made according to the spirit of the invention should be included in the protection scope of the present invention.

Claims

1. A chemical storage fire early warning method based on data fusion, characterized by: The following steps are involved: Step 1: Conduct real-time monitoring of environmental temperature and humidity, smoke and CO concentration, and flames in the chemical storage area, and perform personnel detection at the chemical storage access control; Step 2: Display the collected real-time data on the display screen and upload it to the OneNET platform; when the data is detected to be abnormal or does not meet the preset value, the alarm will sound an alarm; Step 3: Perform fire prediction on the collected real-time environmental data of chemical storage, normalize the real-time environmental data, use the processed data as the data set for fire prediction, and divide it into a training set and a test set; the real-time environmental data includes temperature, humidity, CO and smoke concentration, and open fire probability; Step 4: Construct a data fusion algorithm model LHEA-CNN-DELM-SE_Attention, with temperature, humidity, CO and smoke concentration as input data and open fire probability as output data; the data fusion algorithm model LHEA-CNN-DELM-SE_Attention optimizes the hyperparameters of the CNN-DELM-SE_Attention model based on the dragonfly optimization algorithm LHEA to obtain an optimized CNN-DELM-SE_Attention model; the CNN-DELM-SE_Attention model combines the convolutional neural network CNN and the deep extreme learning machine DELM, and introduces the SE attention mechanism; Step 5: Train the data fusion algorithm model LHEA-CNN-DELM-SE_Attention to verify the effectiveness of the model in chemical storage fire prediction, and use the trained data fusion algorithm model LHEA-CNN-DELM-SE_Attention for fire prediction.

2. A chemical storage fire early warning method based on data fusion according to claim 1, characterized in that: In step 1, data collection uses temperature and humidity sensors, smoke and gas sensors, flame sensors, and human infrared sensors.

3. A chemical storage fire early warning method based on data fusion according to claim 1, characterized in that: In step 2, the OneNET platform interface can view the fire hazard level of the storage area. The fire hazard level is divided into three levels. When only smoke or CO is detected, it is level one; when both smoke and CO are detected, it is level two; as long as flame is detected, it is the highest hazard level three; the OneNET platform changes the monitored temperature and humidity thresholds according to the storage requirements of the chemical warehouse; when an abnormality is detected, the OneNET platform interface will pop up a warning window and the alarm will sound.

4. A chemical storage fire early warning method based on data fusion according to claim 1, characterized in that: In step 3, the real-time environmental data is normalized. During normalization, the characteristic parameter values ​​are limited. If the values ​​exceed the specified limit value, they are all 1 after normalization. The range of data values ​​is normalized to [0, 1]. The calculation formula is: Among them, A j Represents the normalized value of parameter j input; LI j Indicates the limit value of parameter j determined by real-time data; FM j represents the reference threshold of parameter j; x j represents the input value of parameter j; k represents the environmental compensation correction value.

5. A chemical storage fire early warning method based on data fusion according to claim 1, characterized in that: The CNN-DELM-SE_Attention model in step 4 includes a sequence folding layer and an anti-folding layer, a convolution layer Conv_1, a convolution layer Conv_2, a ReLU_1 activation layer, a ReLU_2 activation layer, a ReLU_3 activation layer, a global average pooling layer GAP, a fully connected layer 1, a fully connected layer 2, and a Sigmoid activation function, and also includes a deep extreme learning machine DELM, a fully connected layer 3, and a regression layer; The output of the folding layer is connected to the input of the convolution layer Conv_1. The output of the convolution layer Conv_1 is connected to the ReLU_1 activation layer and the global average pooling layer GAP. After the ReLU_1 activation layer, the convolution layer Conv_2 and the ReLU_2 activation layer are connected in series. The output of the global average pooling layer GAP is then connected to the fully connected layer 1 in series with the ReLU_3 activation layer and the fully connected layer 2, and finally processed by the Sigmoid activation function; After Sigmoid activation function processing, the output of the ReLU_2 activation layer is connected to the dot product layer, and the output of the dot product is connected to the input of the anti-folding layer; The output of the folding layer is connected to the deep extreme learning machine DELM, and the output of the deep extreme learning machine DELM is predicted through the fully connected layer 3 and the regression layer.

6. A chemical storage fire early warning method based on data fusion according to claim 1, characterized in that: In step 4, the dragonfly optimization algorithm LHEA optimizes the hyperparameters of the CNN-DELM-SE_Attention model, including the number of convolution layers, the range of the number of convolution kernels and the size of the convolution kernels, the number of deep extreme learning machine DELM units, and the number of neurons in the fully connected layer.

7. A chemical storage fire early warning method based on data fusion according to claim 6, characterized in that: The dragonfly optimization algorithm LHEA is specifically as follows: 1) Description of the separation process in the LHEA exploration phase: in, It is represented as the separation process of individual i at position t in the evolutionary iteration; represents the position of the individual with current index i in evolutionary iteration t; represents the position of the individual j in the neighborhood in evolutionary iteration t; N represents the number of individuals in the neighborhood; 2) Alignment equations for the LHEA exploration phase: in, It is represented as the alignment process of individual i at position t in the evolutionary iteration; represents the position of the individual j in the neighborhood in evolutionary iteration t; N represents the number of individuals in the neighborhood; 3) Cohesion calculation equation in the LHEA exploration phase: in, The process of calculating the cohesion of individual i at position t in the evolutionary iteration is represented as; represents the position of the individual with current index i in evolutionary iteration t; represents the position of the individual j in the neighborhood in evolutionary iteration t; N represents the number of individuals in the population; N represents the number of individuals in the neighborhood; 4) Description of the food attraction calculation equation in the exploration phase of LHEA: Among them, F i t The food attraction process of individual i at position t in the evolutionary iteration is represented as; represents the position of the individual with current index i in evolutionary iteration t; represents the target source position generated by the current evolutionary iteration t; 5) The calculation equation for disrupting enemy targets during the LHEA exploration phase is described as follows: in, It is represented as the process of disrupting the enemy target by the individual i at position t in the evolutionary iteration; Indicates the enemy position generated by the current evolution iteration t; represents the target source position generated by the current evolutionary iteration t; 6) Step vector calculation equation for LHEA exploration phase: in, The step vector update position of individual i in the evolutionary iteration; α represents the calibration coefficient; It is the alignment of individuals i; c represents the cohesion coefficient; It is represented by the cohesion of the individual i; e represents the enemy factor; i represents the enemy individual; f represents the food factor; F i t It is represented by the food source of individual i; s represents the separation coefficient; It represents the iterative separation of individual i during the evolution process; w represents the inertia weight; The position of the step vector of individual i in the evolutionary iteration; 7) Position vector calculation equation in the LHEA development phase: in, The position of the step vector of individual i in the evolutionary iteration; represents the position of the individual with the current index i in the evolution iteration t; R represents the growth region; g* represents the best position in the evolution iteration; t represents the evolution iteration of the current algorithm; L represents the final number of iterations of the algorithm; 8) Calculation equation of deep pit capacity during the LHEA intensification phase: in, It is represented by the pit capacity of individual i at position t in the evolutionary iteration; f i t It is represented by the pit size of individual i at position t in the evolutionary iteration; f max Indicates the maximum fitness in the pit; f min Indicates the minimum fitness in the deep pit; Const indicates the maximum capacity constant in the deep pit; 9) Deep pit selection probability during the LHEA strengthening phase: in, It is represented as the probability of pit selection of individual i at position t in the evolutionary iteration; It represents the capacity of the deep pit of individual i at position t in the evolutionary iteration; k represents the number of pits with a capacity higher than the overflow pit; 10) The velocity and position equations of water droplets during the local search in the LHEA enhancement phase are described as follows: Among them, V i t+1 V is the speed of the water droplet i during local search in the evolutionary iteration; i t denoted as the current velocity of the water droplet i at position t in the evolutionary iteration; q represents the velocity increment coefficient; The position of the water droplet i during the local search in the evolutionary iteration; The current position of the water droplet denoted as i at position t in the evolutionary iteration; Represented as the current position of the deepest pit in evolutionary iteration t.

8. A chemical storage fire early warning method based on data fusion according to claim 1, characterized in that: In step 5, the training parameters of the established LHEA-CNN-DELM-SE_Attention model are set, the training cycle is set to 80 rounds in total, the number of iterations in each round is 10 times, the number of iterations in total is 800 times, the initial learning rate is 0.01, the learning rate reduction factor is 0.1, and the learning rate reduction cycle is 400. The data is shuffled at the beginning of each training cycle. The data set used for model training consists of the data of temperature, humidity, CO concentration, smoke concentration and open fire probability of chemical storage after normalization. The mean absolute error MAE, mean square error MSE, root mean square error RMSE and determination coefficient R 2 As the basis for evaluating the prediction performance of the model, the effectiveness of the model in predicting chemical storage fires was verified.

Images