Intelligent fire detection method and fire-fighting internet-of-things device thereof

Abstract

The invention relates to the technical field of intelligent fire detection, and discloses an intelligent fire detection method and a fire-fighting internet-of-things device thereof, and intelligent fire detection is carried out by constructing a parameter detection module and a fire detection subsystem; the system is composed of a sensor data acquisition node, a fire-fighting equipment control node, a gateway node, a remote monitoring terminal, a cloud platform and a mobile phone APP, the sensor data acquisition node, the fire-fighting equipment control node and the gateway node communicate in a Lora communication module ad hoc network mode, and communication is realized through the gateway node. The remote monitoring end and the cloud platform carry out environment big data storage and processing, and the mobile phone APP carries out two-way communication with the gateway node through the cloud platform. Compared with the prior art, the system utilizes the sensor, the Internet of Things and the intelligent control technology to monitor environmental parameters in real time and perform data analysis, so that the accuracy of environmental fire detection and the timeliness of fire extinguishing are effectively improved, and the system has good practical application value.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent fire detection and automation equipment of fire-fighting device Internet of Things devices, and in particular to an intelligent fire detection method and a fire-fighting Internet of Things device. Background Art

[0002] With the rapid development of my country's industry, various fire hazards have also increased, and the probability and degree of harm of fire have also increased accordingly. The industrial production process involves a large number of electrical equipment, flammable items, chemicals, etc., which are the source of fire. The industrial production process often requires long-term, high-intensity, and high-frequency use of electrical equipment. If the equipment is not properly maintained and the operation is not standardized, it is easy to cause fire accidents. The industrial production process also produces a large amount of waste gas, waste liquid, waste residue and other pollutants. If not handled properly, it is also easy to cause fire accidents.

[0003] Traditional fire detection includes manual inspection and the placement of fire alarm sensors at fixed points. Manual inspection may result in the inspectors being unable to detect fires in time, and the inspection quality may be unstable. In addition, the cost of arranging fire alarm sensors is high, and regular maintenance and inspection are required. Moreover, they cannot cover the entire building or site, resulting in blind spots and failure to detect fires in time. Therefore, the research and development and application of fire detection technology and IoT firefighting devices are of great significance to ensuring industrial production safety. Through the combination of IoT technology and intelligent control technology, real-time monitoring and data collection of industrial production sites can be achieved, fire hazards can be discovered in time, and corresponding preventive measures can be taken to reduce the probability and degree of fire. By processing and analyzing the collected data through artificial intelligence algorithms, it is possible to accurately determine whether there is a fire risk, improve the real-time and accuracy of fire detection, and timely extinguishing of fires by IoT firefighting devices can better ensure industrial production safety and promote economic development.

[0004] With the continuous emergence of new fire detection technologies and the characteristics of intelligent fire detection and extinguishing devices, intelligent fire detection methods and fire-fighting Internet of Things devices deploy various sensors at the production site to monitor key indicators such as combustible gas concentration, smoke concentration and temperature in real time, and combine artificial intelligence algorithms to further analyze the data collected on site to reduce the false alarm rate of fire and improve detection accuracy. Fire-fighting devices based on Internet of Things technology are one of the important means to ensure industrial production safety, ensuring the real-time nature of fire detection and extinguishing, better ensuring industrial production safety and promoting economic development. Summary of the invention

[0005] Purpose of the invention: In response to the problems pointed out in the background technology, the present invention provides an intelligent fire detection method and a fire-fighting Internet of Things device. By utilizing sensor technology, Internet of Things technology and intelligent control technology, an intelligent fire detection method and a fire-fighting Internet of Things device can monitor environmental parameters in real time and perform data analysis, effectively improving the accuracy of environmental fire detection and the timeliness of fire extinguishing, and having good practical application value.

[0006] Technical solution: The present invention discloses an intelligent fire detection method, comprising the following steps:

[0007] Step 1: Construct a parameter detection module, which is composed of a CNN-BiGRU neural network model, a NARX neural network model, a PI regulator, an Elman neural network model of GA, and multiple BiGRU neural network models; the CNN-BiGRU neural network model, the NARX neural network model, the PI regulator, and the Elman neural network model of GA are connected in sequence, and the output of the Elman neural network model of GA is used as the output of the parameter detection module; the parameter detection module is shown in Figure 2 shown.

[0008] Step 2: Construct a fire detection subsystem, which consists of multiple sensor groups, multiple parameter detection modules, multiple NARX-BiGRU neural network models, multiple GCN-BiGRU neural network models, TS fuzzy neural network models and TDL beat delay lines; see the fire detection subsystem Figure 3 shown.

[0009] The outputs of sensor group 1 are respectively used as the corresponding inputs of parameter detection module 1, NARX-BiGRU neural network model 1 and GCN-BiGRU neural network model a; the outputs of parameter detection module 1 are respectively used as the corresponding inputs of NARX-BiGRU neural network model 1 and GCN-BiGRU neural network model b; the outputs of NARX-BiGRU neural network model 1 are used as the corresponding inputs of GCN-BiGRU neural network model c;

[0010] The outputs of sensor group 2 are respectively used as the corresponding inputs of parameter detection module 2, NARX-BiGRU neural network model 2 and GCN-BiGRU neural network model a; the outputs of parameter detection module 2 are respectively used as the corresponding inputs of NARX-BiGRU neural network model 2 and GCN-BiGRU neural network model b; the output of NARX-BiGRU neural network model 2 is used as the corresponding input of GCN-BiGRU neural network model c;

[0011] The outputs of sensor group 3 are respectively used as the corresponding inputs of parameter detection module 3, NARX-BiGRU neural network model 3 and GCN-BiGRU neural network model a; the outputs of parameter detection module 3 are respectively used as the corresponding inputs of NARX-BiGRU neural network model 3 and GCN-BiGRU neural network model b; the output of NARX-BiGRU neural network model 3 is used as the corresponding input of GCN-BiGRU neural network model c;

[0012] The outputs of GCN-BiGRU neural network model a, GCN-BiGRU neural network model b and GCN-BiGRU neural network model c are respectively used as the corresponding inputs of the TS fuzzy neural network model, the output of the TS fuzzy neural network model is used as the input of the TDL beat-by-beat delay line, the output of the TDL beat-by-beat delay line is respectively used as the corresponding inputs of the GCN-BiGRU neural network model a, GCN-BiGRU neural network model b and GCN-BiGRU neural network model c, and the output value of the TS fuzzy neural network model represents the probability of fire occurrence.

[0013] Furthermore, three BiGRU neural network models are arranged in the parameter detection module, namely BiGRU neural network model a, BiGRU neural network model b and BiGRU neural network model c; the output of the sensor, the output of the BiGRU neural network model a, the output of the BiGRU neural network model b and the BiGRU neural network model c are respectively used as the corresponding inputs of the CNN-BiGRU neural network model, the output of the CNN-BiGRU neural network model is respectively used as the input of the NARX neural network model and the corresponding input of the Elman neural network model of the GA, the output of the NARX neural network model is respectively used as the input of the BiGRU neural network model a and the corresponding input of the Elman neural network model of the GA, the difference and the rate of change of the difference between the output of the NARX neural network model and the output of the Elman neural network model of the GA are respectively used as the input of the PI regulator, the input of the BiGRU neural network model b and the corresponding input of the Elman neural network model of the GA, the output of the PI regulator is respectively used as the input of the BiGRU neural network model c and the corresponding input of the Elman neural network model of the GA, and the output of the Elman neural network model of the GA is used as the output of the parameter detection module.

[0014] The CNN-BiGRU neural network model is a series connection of a CNN convolutional neural network model and a BiGRU neural network model, and the output of the CNN convolutional neural network model serves as the input of the BiGRU neural network model.

[0015] Furthermore, the Elman neural network model of the GA, based on the Elman neural network, utilizes the global search capability and wide adaptability of the genetic algorithm GA to globally optimize the initial weights and thresholds of the Elman neural network, and then feeds the optimized initial weights and thresholds back to the Elman neural network, and then uses the Elman neural network to perform local precise optimization to optimize the network model.

[0016] Furthermore, the steps of building the model of the Elman neural network of GA are as follows:

[0017] ① Determine the number of hidden layer neurons and other network-related parameters of the Elman neural network;

[0018] ② Initialize the population, encode the weights and thresholds of the Elman neural network, determine the initial population size, and generate the initial population;

[0019] ③Construct the fitness function of individuals and sort them. Select network individuals, fi is the fitness value of the i-th individual;

[0020] ④ Perform crossover operation with crossover probability Pc and mutation operation with mutation probability Pm, insert the newly generated individuals into the original population to generate a new population, and calculate the fitness value of the new individuals;

[0021] ⑤ Determine whether an individual that meets the requirements is found. If found, the loop ends; otherwise, repeat steps ② to ④ until the network structure corresponding to an individual in the current population meets the requirements;

[0022] ⑥Assign the optimized network weights and thresholds to the Elman neural network for training.

[0023] Furthermore, the multiple sensor groups are respectively a temperature sensor group, a CO sensor group, and a smoke sensor group.

[0024] Furthermore, the NARX-BiGRU neural network model is a series connection of a NARX neural network model and a BiGRU neural network model, and the GCN-BiGRU neural network model is a series connection of a GCN graph convolutional neural network model and a BiGRU neural network model.

[0025] Furthermore, the multiple sensor groups are respectively a temperature sensor group, a CO sensor group, and a smoke sensor group.

[0026] The present invention also discloses a fire fighting Internet of Things device, including a sensor data acquisition node, a fire fighting equipment control node, a gateway node, a remote monitoring terminal, a cloud platform and a mobile phone APP. The sensor data acquisition node, the fire fighting equipment control node and the gateway node communicate in a self-organizing network manner through a Lora communication module. The communication between the sensor data acquisition node, the fire fighting equipment control node, the remote monitoring terminal, the cloud platform and the mobile phone APP is realized through the gateway node. The remote monitoring terminal and the cloud platform store and process environmental big data. The mobile phone APP performs two-way communication with the gateway node through the cloud platform, so that the management personnel can remotely monitor the operating status of the sensor data acquisition node and the fire fighting equipment control node for monitoring; the system is loaded with computer program steps for realizing the above-mentioned intelligent fire detection method.

[0027] Compared with the prior art, the present invention has the following advantages:

[0028] 1. The CNN-BiGRU neural network model used in the present invention is a series connection of a CNN convolutional neural network and a BiGRU neural network model. CNN can effectively extract local features from input data, and the BiGRU model can consider historical information when processing sequence data, and use a bidirectional RNN structure, so that the model can make fuller use of historical information and reduce information loss. By fusing the feedback outputs of multiple BiGRU models, the detection accuracy and robustness of the parameter detection module are improved.

[0029] 2. The GCN neural network model in the GCN-BiGRU neural network model of the present invention is a powerful graph neural network model that can automatically learn the features and structural information of graph data, and can automatically learn the features of each node without manual definition. This enables the model to better capture the differences between nodes. At the same time, during the training process, only a small number of parameters need to be saved and updated, which makes the model more computationally efficient and reduces the risk of overfitting. Using the GCN neural network model to fuse sensor group data from multiple spatial nodes can better utilize structural information and improve the predictive performance of the model.

[0030] 3. The NARX neural network used in the present invention is a nonlinear autoregressive neural network with external input. This network can be obtained by using input and feedback connections under the condition of time delay, and can use the output signal at the input end in the next iterative training. At the same time, due to its own strong memory system, it has a stronger ability to describe time-varying systems with complex mapping relationships. Compared with traditional recurrent neural networks (RNNs), this network has the characteristics of strong learning ability, fast convergence speed, high generalization performance and high prediction accuracy.

[0031] Fourth, the GA-Elman neural network model used in the present invention is a typical dynamic recursive neural network. By adding a successor layer in the hidden layer as a one-step delay operator, the purpose of memory is achieved, the global stability of the network is enhanced, and it has stronger computing power. The Elman neural network model optimized by the GA genetic algorithm has the characteristics of a dynamic recursive neural network and the advantages of genetic algorithm optimization, and has the advantages of fast optimization capability, strong robustness, adaptability, and strong generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 The fire fighting Internet of Things device of the present invention;

[0033] Figure 2 A parameter detection module of the present invention;

[0034] Figure 3 The fire detection subsystem of the present invention;

[0035] Figure 4 It is the sensor data collection node of the present invention;

[0036] Figure 5 It is a fire fighting equipment control node of the present invention;

[0037] Figure 6 A gateway node of the present invention;

[0038] Figure 7 It is the remote monitoring terminal platform software of the present invention. DETAILED DESCRIPTION

[0039] In order to better explain the present invention, the following Figure 1-7 The technical solution of the present invention is described in detail. The following embodiments are explanations of the invention, and the present invention is not limited to the following embodiments.

[0040] The present invention discloses an intelligent fire detection method and a fire protection Internet of Things device thereof. The specific steps of the intelligent fire detection method are as follows:

[0041] 1. Parameter detection module design

[0042] The parameter detection module is composed of a CNN-BiGRU neural network model, a NARX neural network model, a PI regulator, a GA Elman neural network model, and multiple BiGRU neural network models. Figure 2 shown.

[0043] The CNN-BiGRU neural network model, the NARX neural network model, the PI regulator, and the Elman neural network model of the GA are connected in sequence, and the output of the Elman neural network model of the GA is used as the output of the parameter detection module.

[0044] In this embodiment, three BiGRU neural network models are provided in the parameter detection module, namely BiGRU neural network model a, BiGRU neural network model b and BiGRU neural network model c. The output of the sensor, the output of BiGRU neural network model a, BiGRU neural network model b and BiGRU neural network model c are respectively used as the corresponding inputs of the CNN-BiGRU neural network model, the output of the CNN-BiGRU neural network model is respectively used as the input of the NARX neural network model and the corresponding input of the Elman neural network model of the GA, the output of the NARX neural network model is respectively used as the input of the BiGRU neural network model a and the corresponding input of the Elman neural network model of the GA, the difference and the rate of change of the difference between the output of the NARX neural network model and the output of the Elman neural network model of the GA are respectively used as the input of the PI regulator, the input of the BiGRU neural network model b and the corresponding input of the Elman neural network model of the GA, the output of the PI regulator is respectively used as the input of the BiGRU neural network model c and the corresponding input of the Elman neural network model of the GA, and the output of the Elman neural network model of the GA is used as the output of the parameter detection module.

[0045] 1. CNN-BiGRU neural network model design

[0046] The CNN-BiGRU neural network model is a series connection of the CNN convolutional neural network model and the BiGRU neural network model. The output of the CNN convolutional neural network model is used as the input of the BiGRU neural network model. The CNN convolutional neural network model (Convolutional Neural Network, CNN) is a type of feedforward neural network with a deep structure that includes convolution calculations. It performs convolution and pooling operations on the input data through local connections and shared weights to extract relevant features. The basic framework of the CNN convolutional neural network includes: ① Input layer: When processing time series data, the input layer is a one-dimensional feature matrix, and each element represents the feature value at a certain time point; ② Convolution layer: The convolution layer is the core part of CNN and is responsible for convolution operations. Each neuron in the convolution layer performs convolution operations with a part of the input feature matrix and passes the result to the next layer; ③ Pooling layer: The pooling layer is used to reduce the dimension of the feature matrix to reduce the amount of calculation and avoid overfitting. The pooling layer performs statistical operations on the local area of ​​the input feature matrix, such as maximum pooling, average pooling, etc. The purpose of the pooling layer is to extract the key information of the input feature matrix to reduce noise and redundancy; ④ Fully connected layer: The fully connected layer connects all neurons in the previous layer to all neurons in the current layer to form a fully connected network. The output of the fully connected layer is a vector that represents the important features of the input data. In classification tasks, the fully connected layer is used to map feature vectors to label vectors. One-dimensional convolutional networks are often used to process time series signals and calculate the delayed accumulation of input signals. Let the input data sequence for a period of time be (X1, X2,…, X n ) T ∈R a×b , where a represents the length of the input time series, and b represents the number of features of the input data. The time series data can be obtained through convolution operation:

[0047]

[0048] Among them, X represents the input data, o c Indicates the output, represents the convolution operation, W c is the weight matrix of the one-dimensional convolution kernel, b c represents the bias matrix of this layer, f c () represents the activation function of the convolutional layer. The ReLU function is usually used as the activation function, which takes advantage of its easy derivative during gradient back propagation to speed up convergence. The ReLU function only needs a threshold to get the activation value and has good sparsity. Its calculation formula is:

[0049]

[0050] BiGRU (Bidirectional Gated Recurrent Unit) is an improved recurrent neural network (RNN) structure that consists of two independent GRU layers, one processing the sequence in the forward direction and the other in the reverse direction. This bidirectional structure enables BiGRU to capture long-range dependencies in the sequence, thereby improving the performance of the model. In BiGRU, the input sequence is first processed by two networks, the forward GRU and the backward GRU. The forward GRU processes the input sequence in forward order, and the backward GRU processes the input sequence in reverse order. In this way, the hidden state at each moment contains both past and future information. Then, the forward and backward hidden states are concatenated as the final output. The GRU network contains two gates inside, using the update gate and the reset gate. The reset gate determines how the new input information is combined with the previous memory, and the update gate defines the amount of the previous memory saved to the current time step. If we set the reset gate to 1 and the update gate to 0, we will get the standard RNN model again. These two gating vectors determine which information can eventually be used as the output of the gated recurrent unit. The special thing about these two gating mechanisms is that they can preserve information in long-term sequences and will not be cleared over time or removed because it is not relevant to the prediction. The mathematical expression of the GRU neural network is:

[0051]

[0052] In the formula, h t-1 is the output of the previous neuron, x t is the input of this neuron, h t is the output of this neuron; Wz is the weight of the update gate; σ is the sigmoid function; the update gate Zt is composed of h t-1 +x t Multiply by the update gate weight Wz, and then use the sigmoid function to calculate. For the update gate Zt, when the value is larger, it means that the current neuron has to retain more information, and the previous neuron has to retain less information; r t The reset gate can control the current input x t With the previous state h t-1 The combination of r t The closer it is to 0, the smaller the proportion of the output state at the previous moment; Wr is the weight of the reset gate; W is the candidate output value The weight of is the output candidate value after being processed by the reset gate.

[0053] 2. NARX neural network model design

[0054] NARX neural network is a nonlinear autoregressive neural network with external input. This network can be obtained by using time delays in both input and feedback connections, and can use the output signal at the input end in the next iterative training. At the same time, due to its strong memory system, it has a stronger ability to describe time-varying systems with complex mapping relationships. Compared with traditional recurrent neural networks (RNNs), NARX has shown better results in learning ability, convergence speed, generalization performance, and prediction accuracy. A typical NARX neural network mainly includes input layer, hidden layer, output layer, and input and output delay layers. The mathematical formula of the NARX neural network model is as follows:

[0055] y(t)=f(x(t-1),x(t-2),…,x(td),y(t-1),y(t-2),…,y(td)) (4)

[0056] Where: f is a nonlinear function; x(t) represents the influencing factor of the sensor at the current moment, which is the input variable; y(t) represents the sensor value at the current moment, which is the output variable; d represents the time delay. y(t) can be obtained by nonlinear mapping from the previous d values ​​of x(t) and the previous d values ​​of y(t). The input layer has n neurons, and the network input is x1, x2, ..., x n ,The calculation formula of each layer output is:

[0057]

[0058] Where: x i represents the input of the neuron; w ij Represents the weight between layers; b j represents the threshold of this layer; f(f1,f2) represents the activation function.

[0059] 3. GA-Elman neural network model design

[0060] The Elman neural network is a feedback neural network with local regression characteristics, and is similar to the BP neural network in structure. It can adapt to time-varying characteristics, has short-term memory capabilities, and uses information output at past moments to feedback and store within the model. Unlike the BP network, in addition to the basic input layer, hidden layer, and output layer, the Elman neural network also has an additional receiving layer in the hidden layer to form local feedback, thereby enhancing the accuracy of the model. The feedback connection consists of a group of "structural" units that remember the previous output value, and the connection weights are fixed. All nodes in the hidden layer have corresponding receiving layer nodes connected to them. The function of the receiving layer is to use the state of the hidden layer at the previous moment and the input at the current moment as the input of the hidden layer through the connection memory, which is a state feedback. The mathematical model formula of the Elman neural network is as follows:

[0061]

[0062] In the formula, O is the output vector; h is the hidden layer vector; c is the receiving layer vector; u is the input vector; ω is the weight, where ω j ,q is the weight from the hidden layer to the output layer, W i,j is the weight from the input layer to the hidden layer, and ωj,k is the weight from the receiving layer to the hidden layer.

[0063] The GA-Elman neural network model is based on the Elman neural network. It uses the global search capability and wide adaptability of the genetic algorithm (GA) to globally optimize the initial weights and thresholds of the Elman neural network, and then feeds the optimized initial weights and thresholds back to the Elman neural network. The Elman neural network is then used for local precise optimization to optimize the network model and improve the accuracy of model prediction. The general steps for establishing the GA-Elman neural network model are:

[0064] ① Determine the network-related parameters such as the number of hidden layer neurons of the Elman neural network.

[0065] ② Initialize the population. Encode the weights and thresholds of the Elman neural network, determine the initial population size, and generate the initial population;

[0066] ③Construct the fitness function of individuals and sort them. Select network individuals, fi is the fitness value of the i-th individual;

[0067] ④ Perform crossover operation with crossover probability Pc and mutation operation with mutation probability Pm, insert the newly generated individuals into the original population to generate a new population, and calculate the fitness value of the new individuals;

[0068] ⑤ Determine whether an individual that meets the requirements is found. If found, the loop ends; otherwise, repeat steps ② to ④ until the network structure corresponding to an individual in the current population meets the requirements;

[0069] ⑥Assign the optimized network weights and thresholds to the Elman neural network for training.

[0070] 2. Fire Detection Subsystem Design

[0071] The fire detection subsystem consists of multiple parameter detection modules, multiple NARX-BiGRU neural network models, multiple GCN-BiGRU neural network models, TS fuzzy neural network models, and TDL beat delay lines. The NARX-BiGRU neural network model is a series connection of the NARX neural network model and the BiGRU neural network model, and the GCN-BiGRU neural network model is a series connection of the GCN graph convolutional neural network model and the BiGRU neural network model. Figure 3 This embodiment is described by taking three sensor groups as an example, and the three sensor groups are a temperature sensor group, a CO sensor group, and a smoke sensor group.

[0072] The outputs of the temperature sensor group are used as the corresponding inputs of parameter detection module 1, NARX-BiGRU neural network model 1 and GCN-BiGRU neural network model a, respectively. The outputs of parameter detection module 1 are used as the corresponding inputs of NARX-BiGRU neural network model 1 and GCN-BiGRU neural network model b, respectively. The outputs of NARX-BiGRU neural network model 1 are used as the corresponding inputs of GCN-BiGRU neural network model c.

[0073] The outputs of the CO sensor group are used as the corresponding inputs of parameter detection module 2, NARX-BiGRU neural network model 2 and GCN-BiGRU neural network model a, respectively. The outputs of parameter detection module 2 are used as the corresponding inputs of NARX-BiGRU neural network model 2 and GCN-BiGRU neural network model b, respectively. The outputs of NARX-BiGRU neural network model 2 are used as the corresponding inputs of GCN-BiGRU neural network model c.

[0074] The outputs of the smoke sensor group are used as the corresponding inputs of the parameter detection module 3, NARX-BiGRU neural network model 3 and GCN-BiGRU neural network model a respectively; the outputs of the parameter detection module 3 are used as the corresponding inputs of the NARX-BiGRU neural network model 3 and GCN-BiGRU neural network model b respectively; the outputs of the NARX-BiGRU neural network model 3 are used as the corresponding inputs of the GCN-BiGRU neural network model c.

[0075] The outputs of GCN-BiGRU neural network model a, GCN-BiGRU neural network model b and GCN-BiGRU neural network model c are respectively used as the corresponding inputs of the TS fuzzy neural network model, the output of the TS fuzzy neural network model is used as the input of the TDL beat-by-beat delay line, the output of the TDL beat-by-beat delay line is respectively used as the corresponding inputs of the GCN-BiGRU neural network model a, GCN-BiGRU neural network model b and GCN-BiGRU neural network model c, and the output value of the TS fuzzy neural network model represents the probability of fire occurrence.

[0076] 1. GCN graph convolutional neural network model design

[0077] The GCN (Graph Convolutional Networks) neural network model is a method for deep learning based on graph structured data. It is based on the spectral convolution theory in graph signal processing and generalizes the traditional convolutional neural network (CNN) to graph structured data. The core of the GCN model is to aggregate the feature vector of a node with the feature vectors of its neighboring nodes to obtain a new node representation. This process is implemented through graph convolution operations. The output of each layer is obtained by multiplying the output of the previous layer by the adjacency matrix, and then processed by a nonlinear activation function. This process can be regarded as an operation on the node feature attributes and the adjacency matrix. Through layer-by-layer aggregation and activation, a new node representation is obtained.

[0078] 2. TS fuzzy neural network model design

[0079] As a nonlinear fuzzy reasoning model, the TS fuzzy neural network model has the characteristics of expressing fuzzy reasoning rules, being easy to calculate, and being helpful for mathematical analysis. The network is divided into two parts: the antecedent network and the consequent network. Its input parameters are the fusion values ​​of the sensor groups at each stage, and the output result is the probability value of fire occurrence.

[0080] The antecedent network consists of 4 layers, namely the input layer, the fuzzification layer, the fuzzy rule calculation layer and the output layer. Its function is to perform fuzzy matching and determine the influence weight of each input parameter. Fuzzification layer: Each node is a semantic variable, which is used to calculate the membership degree u of each input component. j (i=1,2,3…,n,j=1,2,3…,m), n is the number of input nodes, m is the number of fuzzy partitions. Fuzzy rule calculation layer: Each node represents a match of a fuzzy rule, and its function is to calculate the fitness of each rule. Output layer: realize the normalization operation of the output result of the fuzzy layer, and use this value as the weight of each attribute.

[0081] The subsequent network consists of three layers: input layer, fully connected layer and output layer. Its function is to receive input data for learning and output the final result in combination with the weight of the previous network. Fully connected layer: realizes the training of input data. Output layer: The output result is the weighted value of the weight in the previous network and the output result xi of the subsequent network.

[0082]

[0083] Among them, xi is the output of the fully connected layer, and y is the output layer result.

[0084] The fire probability values ​​output by the TS fuzzy neural network model correspond to four fire risk types: no fire risk, low fire risk, medium fire risk and high fire risk. The corresponding relationship between the fire probability values ​​output by the TS fuzzy neural network model and the fire risk level is shown in Table 1.

[0085] Table 1 Correspondence between fire risk level and TS fuzzy neural network model output value

[0086] Serial number Fire risk level TS fuzzy neural network model output 1 No fire risk [0.0，0.2] 2 Low risk of fire (0.2，0.5] 3 Fire risk (0.5，0.8] 4 High fire risk (0.8，1.0]

[0087] The present invention applies the above-mentioned fire monitoring and intelligent fire detection method to a firefighting Internet of Things device, and discloses a firefighting Internet of Things device, which is composed of a sensor data acquisition node, a firefighting equipment control node, a gateway node, a host computer monitoring terminal platform, a cloud platform, and a mobile phone APP. The sensor data acquisition node is responsible for collecting environmental parameters that affect the fire, and the gateway node is responsible for packaging the parameters collected by all terminal data nodes, and sending them to the host computer monitoring terminal and the cloud platform through the field bus and the wireless communication module for data storage and analysis. At the same time, it receives and issues remote commands to control the start and stop of each firefighting equipment in the firefighting equipment control node. The sensor data acquisition nodes, firefighting equipment control nodes, and gateway nodes in the system are connected in a self-organizing network manner through the Lora communication module to complete the functions of online and offline and data transmission of the terminal equipment. The mobile phone APP receives information from the cloud platform and communicates with the gateway node in a two-way manner, so that the management personnel can remotely monitor the system operation status and control the terminal equipment. In a firefighting Internet of Things device such as Figure 1 shown.

[0088] 1. Sensor data collection node design

[0089] The sensor data acquisition node is composed of a multi-source sensor group that affects environmental safety, such as temperature sensor, flame sensor, CO sensor, smoke sensor, harmful gas sensor, STM32 microprocessor, GPS positioning module, Lora communication module and power module, to realize the collection of environmental data and equipment positioning functions. The power module realizes the power supply of all hardware devices of the environmental parameter detection node. The Lora communication module realizes two-way communication with the gateway node. The structure of the sensor data acquisition node is as follows: Figure 4 shown.

[0090] 2. Fire equipment control node design

[0091] The fire equipment control node is composed of Lora communication module, STM32 microprocessor, water pump controller, smoke exhaust controller, fire extinguishing controller, alarm controller and power module. The STM32 microprocessor receives control instructions from the remote monitoring platform or mobile phone APP through the data transfer of the gateway node, and completes the control relay of the specified fire equipment. The action of closing and releasing the control relay realizes the start and stop control function of various fire equipment in the park, and avoids economic losses through the ability to prevent fire. Two-way communication with the gateway node is realized through the Lora communication module. The structure of the fire equipment control node is as follows: Figure 5 shown.

[0092] 3. Gateway node design

[0093] The gateway node is composed of a Lora communication module, an STM32 microprocessor, an NB-IOT module, and a power supply module. The Lora communication module is used to build an ad hoc network with the sensor data acquisition node and the fire equipment control node. The NB-IOT module is used to realize two-way data interaction with the cloud platform, and the USB interface is used to realize two-way data communication with the remote monitoring terminal. The hardware structure of the gateway node is as follows: Figure 6 shown.

[0094] 4. Cloud platform and mobile APP design

[0095] The cloud platform is developed and designed based on the OneNet IoT cloud platform to achieve data interaction with gateway nodes and mobile APPs, and to realize remote recording and monitoring of terminal node data. The mobile APP is developed on the Qt Creator software platform based on Qt for Android, and subscribes to cloud platform devices through the MQTT protocol to achieve data subscription and publishing, so that managers can monitor the environmental parameters of each terminal node and the results of the park fire risk assessment in real time on the mobile APP.

[0096] 5. Remote monitoring terminal design

[0097] The remote monitoring software is developed in C++ based on the Qt Creator platform and deployed in industrial computers. The remote monitoring software is divided into four functional modules, including communication parameter configuration, data management, data analysis and early warning notification. The communication parameter configuration module includes USB serial port configuration; the data management module includes data storage and data query, which facilitates managers to query historical data and provide raw data for the intelligent fire risk early warning subsystem; the data analysis module includes a parameter detection module and a fire detection subsystem, which realizes the accurate detection and prediction of sensor parameters and the function of environmental fire risk detection; the early warning notification module is divided into local alarms and remote alarms to ensure that managers can receive early warning notifications and deal with risks in a timely manner. The software structure of the on-site monitoring platform is as follows: Figure 7 shown.

[0098] During the operation of the fire-fighting Internet of Things device, the above-mentioned intelligent fire detection method is set in the sensor data collection node and the fire-fighting equipment control node.

[0099] 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 present invention should be included in the protection scope of the present invention.

Claims

1. A fire intelligent detection method, characterized in that: The steps include: Step 1: construct a parameter detection module, which is composed of a CNN-BiGRU neural network model, a NARX neural network model, a PI regulator, an Elman neural network model of GA, and multiple BiGRU neural network models; the CNN-BiGRU neural network model, the NARX neural network model, the PI regulator, and the Elman neural network model of GA are connected in sequence, and the output of the Elman neural network model of GA is used as the output of the parameter detection module; Step 2: construct a fire detection subsystem, which consists of multiple sensor groups, multiple parameter detection modules, multiple NARX-BiGRU neural network models, multiple GCN-BiGRU neural network models, TS fuzzy neural network models and TDL beat delay lines; The outputs of sensor group 1 are respectively used as the corresponding inputs of parameter detection module 1, NARX-BiGRU neural network model 1 and GCN-BiGRU neural network model a; the outputs of parameter detection module 1 are respectively used as the corresponding inputs of NARX-BiGRU neural network model 1 and GCN-BiGRU neural network model b; the outputs of NARX-BiGRU neural network model 1 are used as the corresponding inputs of GCN-BiGRU neural network model c; The outputs of sensor group 2 are respectively used as the corresponding inputs of parameter detection module 2, NARX-BiGRU neural network model 2 and GCN-BiGRU neural network model a; the outputs of parameter detection module 2 are respectively used as the corresponding inputs of NARX-BiGRU neural network model 2 and GCN-BiGRU neural network model b; the output of NARX-BiGRU neural network model 2 is used as the corresponding input of GCN-BiGRU neural network model c; The outputs of sensor group 3 are respectively used as the corresponding inputs of parameter detection module 3, NARX-BiGRU neural network model 3 and GCN-BiGRU neural network model a; the outputs of parameter detection module 3 are respectively used as the corresponding inputs of NARX-BiGRU neural network model 3 and GCN-BiGRU neural network model b; the output of NARX-BiGRU neural network model 3 is used as the corresponding input of GCN-BiGRU neural network model c; The outputs of GCN-BiGRU neural network model a, GCN-BiGRU neural network model b and GCN-BiGRU neural network model c are respectively used as the corresponding inputs of the TS fuzzy neural network model, the output of the TS fuzzy neural network model is used as the input of the TDL beat-by-beat delay line, the output of the TDL beat-by-beat delay line is respectively used as the corresponding inputs of the GCN-BiGRU neural network model a, GCN-BiGRU neural network model b and GCN-BiGRU neural network model c, and the output value of the TS fuzzy neural network model represents the probability of fire occurrence.

2. The intelligent fire detection method according to claim 1, characterized in that: The parameter detection module is provided with three BiGRU neural network models, namely BiGRU neural network model a, BiGRU neural network model b and BiGRU neural network model c; the output of the sensor, the output of the BiGRU neural network model a, the output of the BiGRU neural network model b and the BiGRU neural network model c are respectively used as the corresponding inputs of the CNN-BiGRU neural network model, the output of the CNN-BiGRU neural network model is respectively used as the input of the NARX neural network model and the corresponding input of the Elman neural network model of the GA, the output of the NARX neural network model is respectively used as the input of the BiGRU neural network model a and the corresponding input of the Elman neural network model of the GA, the difference between the output of the NARX neural network model and the output of the Elman neural network model of the GA and the rate of change of the difference are respectively used as the input of the PI regulator, the input of the BiGRU neural network model b and the corresponding input of the Elman neural network model of the GA, the output of the PI regulator is respectively used as the input of the BiGRU neural network model c and the corresponding input of the Elman neural network model of the GA, and the output of the Elman neural network model of the GA is used as the output of the parameter detection module.

3. A fire intelligent detection method according to claim 1 or 2, characterized in that: The CNN-BiGRU neural network model is a series connection of a CNN convolutional neural network model and a BiGRU neural network model, and the output of the CNN convolutional neural network model serves as the input of the BiGRU neural network model.

4. A fire intelligent detection method according to claim 1 or 2, characterized in that: The Elman neural network model of the GA is based on the Elman neural network, and utilizes the global search capability and wide adaptability of the genetic algorithm GA to globally optimize the initial weights and thresholds of the Elman neural network, and then feeds the optimized initial weights and thresholds back to the Elman neural network, and then uses the Elman neural network to perform local precise optimization to optimize the network model.

5. The intelligent fire detection method according to claim 4, characterized in that: The steps to build the GA Elman neural network model are as follows: ① Determine the number of hidden layer neurons and other network-related parameters of the Elman neural network; ② Initialize the population, encode the weights and thresholds of the Elman neural network, determine the initial population size, and generate the initial population; ③Construct the fitness function of individuals and sort them. Select network individuals, fi is the fitness value of the i-th individual; ④ Perform crossover operation with crossover probability Pc and mutation operation with mutation probability Pm, insert the newly generated individuals into the original population to generate a new population, and calculate the fitness value of the new individuals; ⑤ Determine whether an individual that meets the requirements is found, and end the loop if found; Otherwise, repeat steps ② to ④ until the network structure corresponding to an individual in the current population meets the requirements; ⑥Assign the optimized network weights and thresholds to the Elman neural network for training.

6. The intelligent fire detection method according to claim 1, characterized in that: The multiple sensor groups are respectively a temperature sensor group, a CO sensor group, and a smoke sensor group.

7. The intelligent fire detection method according to claim 1, characterized in that: The NARX-BiGRU neural network model is a series connection of the NARX neural network model and the BiGRU neural network model, and the GCN-BiGRU neural network model is a series connection of the GCN graph convolutional neural network model and the BiGRU neural network model.

8. A fire fighting Internet of Things device, characterized in that: It includes sensor data acquisition nodes, fire equipment control nodes, gateway nodes, remote monitoring terminals, cloud platforms and mobile phone APPs. The sensor data acquisition nodes, fire equipment control nodes and gateway nodes communicate in a self-organizing network manner through the Lora communication module. The communication between the sensor data acquisition nodes, fire equipment control nodes, remote monitoring terminals, cloud platforms and mobile phone APPs is realized through the gateway nodes. The remote monitoring terminals and cloud platforms store and process environmental big data. The mobile phone APP performs two-way communication with the gateway nodes through the cloud platform, so that the management personnel can remotely monitor the operating status of the sensor data acquisition nodes and the fire equipment control nodes. The system is loaded with computer program steps for realizing the intelligent fire detection method as described in any one of claims 1 to 7.

Images