A kitchen fire intelligent early warning and disposal method and system

Abstract

The present invention discloses a method and system for intelligent early warning and disposal of kitchen fires, and relates to the technical field of intelligent early warning and disposal of kitchen fires. The kitchen data acquisition terminal is connected to an edge intelligent computing box network; the edge intelligent computing box is electrically connected to an electronically controlled gas valve, and is connected to a cloud server network; the cloud server is connected to an interactive terminal network; the edge intelligent computing box can perform federated learning training with the cloud server and use a graph neural network model; the kitchen data acquisition terminal collects data and transmits it to the edge intelligent computing box, which can warn and dispose of the kitchen fire status through the graph neural network model, and can push kitchen abnormal status information to the cloud server; the present invention improves the accuracy and efficiency of fire warning and disposal through the corresponding artificial intelligence early warning and disposal settings.

Description

Technical Field

[0001] The present invention relates to the technical field of kitchen fire intelligent early warning and disposal, and in particular to a kitchen fire intelligent early warning and disposal method and system. Background Art

[0002] Household fires have become one of the most serious threats to life and property safety. Especially with the continuous increase in the height of urban buildings and the growth of population density, the difficulty of fire fighting and disaster relief is also increasing, and the economic losses caused by fires are increasing sharply. Among them, kitchen fires are the most common type of household fires and the most serious threat to life and property safety.

[0003] Therefore, it is very important to raise people's awareness of preventing home and kitchen fires. Various types of fire detection and alarm devices for families are gradually becoming popular, such as temperature sensors, humidity sensors, catalytic combustion gas sensors and dual-band smoke detectors. In addition, with the development of Internet technology, a series of fire warning systems based on the Internet of Things have emerged, greatly improving the real-time monitoring, warning and handling capabilities of fires. However, most of the current fire detection and alarm devices are passive warnings, which cannot really prevent fires before they happen, and the fire handling capabilities are weak. There are also the following problems:

[0004] (1) Timeliness of fire warning: Existing fire warning devices usually use smoke detectors or video surveillance equipment for monitoring, and usually respond through sound and light alarms and other equipment after the fire occurs, which has a certain lag. From the perspective of firefighting, the golden time for fire fighting is about 3 minutes, so research on more real-time or even early warning methods is of great significance to reducing the difficulty and loss of fire fighting;

[0005] (2) The accuracy of fire warning: Existing fire detection equipment usually independently measures various environmental parameters such as temperature, humidity and gas indicators, and issues alarms based on threshold values. However, the accuracy of fire warning is poor if it relies solely on a single parameter. If the threshold is set too high, the false alarm rate will be high, which may easily lead to serious fire consequences; on the contrary, lowering the threshold may lead to a high false alarm rate, increasing unnecessary disposal costs;

[0006] (3) Issues with the effectiveness of fire handling: During the fire handling process, the effectiveness of fire handling is greatly reduced due to factors such as delayed detection leading to the spread of the disaster and weak equipment linkage processing capabilities.

[0007] Therefore, it is very necessary to comprehensively utilize digital technologies such as the Internet of Things and artificial intelligence, study new active fire prevention and alarm methods, detect and accurately perceive disasters as early as possible, and deal with fires automatically and effectively. Summary of the invention

[0008] The purpose of the present invention is to provide a method and system for intelligent early warning and handling of kitchen fires, which realizes feature extraction and alignment through multi-sensor information fusion, builds a task-driven graph neural network model in combination with fire scenes, and then solves the current technical problems of low kitchen fire early warning accuracy, poor real-time performance and poor handling through cloud-edge collaborative learning and reasoning operations.

[0009] The present invention relates to an intelligent kitchen fire early warning and disposal system, which is implemented by the following technical solutions:

[0010] The kitchen data collection end is connected to the edge intelligent computing box network; the edge intelligent computing box is electronically connected to the electronically controlled gas valve and is connected to the cloud server network; the cloud server is connected to the interactive terminal network; the edge intelligent computing box can perform federated learning training with the cloud server and use the graph neural network model; the kitchen data collection end collects data and transmits it to the edge intelligent computing box, which performs fire warning and disposal through the graph neural network model and pushes kitchen abnormal status information to the cloud server, and the cloud server interacts with the interactive terminal.

[0011] In general, the kitchen fire intelligent early warning and disposal system obtains data through the kitchen data collection terminal and transmits it to the edge intelligent computing box. The edge intelligent computing box uses the artificial intelligence model to perform fire early warning and disposal (or cooperates with the cloud server for collaborative disposal), and controls the electronically controlled gas valve at the same time. When necessary, it pushes abnormal status information to the cloud server, and the cloud server further notifies the interactive terminal to achieve real-time early warning and intelligent disposal. The artificial intelligence model is obtained by the edge intelligent computing box and the cloud server through federated learning.

[0012] The present invention also relates to a kitchen fire intelligent early warning and disposal method, which is implemented by using the kitchen fire intelligent early warning and disposal system, and comprises the following steps:

[0013] S1, the kitchen data collection terminal collects various kitchen environment parameters in real time and reports them to the edge intelligent computing box;

[0014] S2, the edge intelligent computing box determines the corresponding finite state machine mode according to the kitchen environment parameters, completes the graph structure construction, and loads the corresponding graph neural network model, where the graph nodes represent different kitchen environment states and the edges represent the switching between states;

[0015] S3, the edge intelligent computing box preprocesses the input kitchen environment parameters, generates corresponding feature parameter vectors as the input of each graph node, constructs a training data set using a weak supervision method, and provides two data channels for model training and prediction; the weak supervision method is: the edge intelligent computing box periodically scans and determines the historical attributes (states) of the graph nodes (i.e., the finite state machine states) to obtain the corresponding kitchen environment states (labels), thereby solving the problem of small sample data of the kitchen early warning model;

[0016] S4, through the model training channel, the edge intelligent computing box performs model training in a horizontal federated learning manner;

[0017] S5, through the model prediction channel, the edge intelligent computing box performs early warning and disposal based on the trained graph neural network model.

[0018] Preferably, each state of the finite state machine is mapped to a graph node, and the state relationship of the finite state machine is mapped to the edge of the graph. A corresponding relationship is established between the finite state machine and the graph neural network model to mine the spatial correlation relationship; and a time series neural network is used to mine the temporal correlation relationship of the changes in the kitchen environment state at different time points.

[0019] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0020] 1. There are five major shortcomings in the fire warning systems in the prior art: First, they rely on a single or simple combination of sensors, resulting in untimely warnings. Secondly, due to the lack of in-depth mining of complex environmental features, the warning is not accurate and prone to false alarms or missed reports. Third, the data transmission and processing efficiency is low, and a large amount of raw data needs to be transmitted to the cloud, which increases communication overhead and data privacy risks. Fourth, there is a lack of automated disposal capabilities, mainly focusing on monitoring and alarming. Finally, it is difficult to achieve local optimization, model training and optimization rely on the cloud, and cannot respond quickly to local environmental changes. Correspondingly, the present invention designs a kitchen fire intelligent early warning and disposal system in a cloud-edge-end collaborative mode, which is mainly composed of a front-end device responsible for data collection, an edge intelligent computing box responsible for data analysis and processing, and a cloud server responsible for business applications, realizing the full process control of fire monitoring, early warning and disposal. The system uses multi-sensor information fusion technology to extract and align features using data from different sensors to obtain rich feature expressions, thereby improving the accuracy of state transition judgment. By building a time-series neural network, deeply mining and utilizing the implicit knowledge of the state machine, and combining the federated learning mechanism to achieve local optimization, the communication overhead and privacy risks of transmitting raw data to the cloud for training are avoided. The system can effectively capture dynamic changes in the kitchen environment and achieve accurate early warning, significantly improving the timeliness and accuracy of fire warnings;

[0021] 2. The present invention maps typical kitchen scenes into finite state machines, uses collaborative sensing nodes for real-time signal collection and analysis, and uses finite state machines to determine state transitions, thereby overcoming the shortcomings of the prior art, such as untimely and inaccurate early warning, low data processing efficiency, lack of automated handling capabilities, and difficulty in achieving localized optimization. Through the fusion of multi-dimensional features such as multi-sensor nodes, multi-states, and time domain sequence inputs, the system can accurately identify fire risks and automatically take measures when risks are identified, such as closing gas valves, pushing alarm information, and notifying relevant personnel, thereby significantly improving the accuracy and efficiency of fire warning and handling, and reducing potential losses. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, constitute a part of this application, and do not constitute a limitation of the embodiments of the present invention. In the drawings:

[0023] Figure 1 is a system architecture diagram of this embodiment;

[0024] Figure 2 This is a schematic diagram of the kitchen fire intelligent early warning and handling process of this embodiment;

[0025] Figure 3 This is a schematic diagram of the state switching of a typical kitchen scene in this embodiment;

[0026] Figure 4 This is a schematic diagram of the kitchen multi-sensor information fusion process of this embodiment;

[0027] Figure 5 Schematic diagram of the conversion from a finite state machine to a graph in this embodiment;

[0028] Figure 6 This is a schematic diagram of the timing neural network workflow of this embodiment;

[0029] Figure 7 Schematic diagram of the overall architecture of the kitchen fire warning model based on the timing graph neural network model in this embodiment. DETAILED DESCRIPTION

[0030] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0031] like Figures 1 to 7As shown, this embodiment relates to a kitchen fire intelligent early warning and disposal method and system, wherein the kitchen fire intelligent early warning and disposal system involved in this embodiment adopts a cloud-edge-end collaborative mode design, and is mainly composed of a front-end device responsible for data collection (various data collection sensors), a front-end device network connected to an edge intelligent computing box responsible for data analysis and processing, and an edge intelligent computing box network connected to a cloud server responsible for business applications (see Figure 1 ), the cloud server network is connected to interactive terminals (mobile interactive terminals such as mobile application Apps, etc.), and the edge intelligent computing box is also electrically connected to the electronically controlled gas valve (the edge intelligent computing box controls the opening and closing of the electronically controlled gas valve).

[0032] Furthermore, the front-end equipment is mainly composed of various sensors, including temperature sensors, humidity sensors, catalytic combustion gas sensors, dual-band smoke detectors, manual triggers and video monitoring equipment, etc., which are responsible for the real-time collection of kitchen environment parameters. Specifically, the sensors can be deployed in a corresponding manner according to actual needs. They can be integrated into kitchen hoods and other equipment, or designed as a separate fusion data collection equipment, or each sensor can even exist independently for distributed data collection.

[0033] Furthermore, the collected videos, images and structured data are uniformly imported into the edge intelligent computing box for storage, and graph neural network model training and reasoning operations are performed online. Once the model finds an abnormal situation, it is promptly pushed to the distant cloud server and interactive terminal for corresponding danger notification and remote fire extinguishing operations. Through the cloud-edge-end collaborative model, an efficient and intelligent kitchen fire warning and disposal system is built to comprehensively improve the accuracy and rapid response capabilities of fire prevention and control. The front-end equipment is equipped with a variety of sensors, including temperature sensors, humidity sensors, catalytic combustion gas sensors, dual-band smoke detectors, manual triggers and video monitoring equipment, etc., which collect various parameters of the kitchen environment in real time to ensure the comprehensiveness and flexibility of data collection.

[0034] The collected data is fed into the edge intelligent computing box for storage and processing, and the graph neural network model is used for real-time training and reasoning. This edge computing method (it connects to the network and communicates with the remote cloud server to receive and execute various tasks. After the device is connected to the network, the edge computing box first collects and processes the data. These data may include images, sounds, text, etc. Through the AI ​​chip embedded in the device, the data will be preliminarily processed for identification, analysis, etc. The edge computing box processes and analyzes the data on the device side, which is edge computing, and finally uploads it) significantly reduces the delay in data transmission, ensuring that fire hazards can be detected quickly and responded to in a timely manner. Edge computing also reduces the load on the cloud server, allowing the system to operate efficiently when facing large amounts of data. Once the edge intelligent computing box detects an abnormal situation, it will immediately push an alarm to the cloud server and interactive terminal (such as APP) in the cloud to achieve remote monitoring and control, including remote fire extinguishing operations.

[0035] Through this design, the entire system not only improves the accuracy and timeliness of fire warnings, but also significantly improves the safety of the kitchen through intelligent and automated means. In addition, users can monitor the kitchen situation anytime and anywhere through mobile devices, which greatly improves the convenience and peace of mind of use. In an emergency, the system can also automatically control the electronically controlled gas valve to further reduce the risk of fire and ensure the safety of the family and property. In general, the system integrates advanced sensing, computing and communication technologies to provide a comprehensive, intelligent and reliable kitchen fire warning and disposal solution.

[0036] This embodiment involves a kitchen fire intelligent early warning and handling method, the following steps include (see Figure 2 ):

[0037] S1, the kitchen data collection terminal collects various kitchen environment parameters in real time and reports them to the edge intelligent computing box;

[0038] S2, the edge intelligent computing box determines the corresponding finite state machine mode according to the kitchen environment parameters, completes the graph structure construction, and loads the corresponding graph neural network model (heterogeneous graph neural network model). Figure 5 , where graph nodes correspond to and represent different states, and edges represent the switching between states);

[0039] S3, the edge intelligent computing box preprocesses the input kitchen environment parameters, generates the corresponding feature parameter vector as the input of each graph node, constructs the training data set using a weak supervision method, and provides two data channels for model training and prediction. The weak supervision method is that the edge intelligent computing box periodically scans and determines the historical attributes (states) of the graph nodes (corresponding to the different states of the finite state machine) to obtain the corresponding kitchen environment state (label), thereby solving the problem of small sample data of the kitchen early warning model;

[0040] S4, through the model training channel, the edge intelligent computing box performs model training through horizontal federated learning. On the one hand, the cloud server uses the gradient data reported by each edge intelligent computing box to aggregate and obtain the global basic model to improve the generalization ability; on the other hand, the edge intelligent computing box performs local optimization based on the global basic model to improve the pertinence to the local scene;

[0041] S5, through the model prediction channel, the edge intelligent computing box performs early warning and disposal based on the trained graph neural network model;

[0042] Furthermore, the specific steps of S4 are:

[0043] S411, the edge intelligent computing box uses the training data set channel to retrain the graph neural network model to obtain a global optimized neural network model, and stores the training results in the local model pool;

[0044] S412, the edge intelligent computing box compresses and encrypts the trained parameters and reports them to the cloud server. The parameters include the first-order derivative of the loss function of the graph neural network model to the model parameters, that is, the gradient parameter;

[0045] S413, after decrypting and decompressing the data reported by each edge intelligent computing box, the cloud server aggregates it based on the global optimum to obtain an updated graph neural network model, which is sent to each edge intelligent computing box at a regular interval.

[0046] Furthermore, the specific steps of S5 are:

[0047] S511, a graph neural network model selected and loaded from a local model pool by an edge intelligent computing box;

[0048] S512, the edge intelligent computing box uses the kitchen environment data collected in real time to perform reasoning, and maps and determines the type of abnormal kitchen environment state through the switching relationship between the graph nodes;

[0049] S513, the edge intelligent computing box pushes the abnormal kitchen environment status type to the cloud server for remote processing. At the same time, based on the abnormal kitchen environment status type, it is linked with the electric switch and air lock equipment to take emergency measures for the abnormal kitchen environment status.

[0050] The kitchen fire intelligent early warning and disposal method involved in this embodiment realizes an efficient and accurate kitchen fire early warning and disposal mechanism by building a cloud-edge-end collaborative intelligent system. It first uses the kitchen data acquisition terminal to collect a variety of environmental parameters in real time and report them to the edge intelligent computing box to ensure the comprehensiveness and real-time nature of the data. The edge intelligent computing box then uses these parameters to determine the finite state machine mode, build a graph structure, and load the graph neural network model, thereby achieving accurate representation and analysis of complex states.

[0051] The edge intelligent computing box preprocesses the input data, generates feature parameter vectors, and uses weak supervision to build a training data set, solving the problem of small sample data. By periodically scanning and determining historical attributes, state category labels are generated, which improves the effectiveness of model training. In the model training channel, the edge intelligent computing box performs model training through horizontal federated learning. The cloud server aggregates the gradient data reported by each edge device to obtain the global basic model, and sends it to the edge device for local optimization, achieving a balance between global generalization capabilities and local scenario targeting.

[0052] In the model prediction channel, the edge intelligent computing box uses the trained graph neural network model to make early warning predictions, and maps different warning types by inferring node relationships through real-time environmental data. The abnormal kitchen environment status type is not only pushed to the cloud server in the cloud for remote processing, but also linked with the electric switch and air lock equipment for emergency disposal, ensuring that a quick response can be made in the early stage of the fire to prevent the fire from spreading.

[0053] In this embodiment, the finite state machine , consisting of a finite set of states , Input conditions , state transition function (refers to a mapping relationship that maps input conditions to specific state transitions. This mapping relationship determines the state transition of the node in the graph based on a finite set of states and input conditions), the initial state , accepting state (or terminal state) There are eight typical scenes for daily kitchen scenes, including There are eight scenarios, and the corresponding status of each scenario is: : Idle state; : Normal state with no fire and people present; : Abnormal state of no fire; : Normal state when someone is in the fire; : The fire starts and someone is in an abnormal state; : The fire moves away from the normal state; : The abnormal state of the fire-moving person; : Termination state;

[0054] Each state has different input conditions , mainly including {people appear, people leave, fire on, fire off, abnormality appears, abnormality disappears, abnormality handling}, etc. (see Figure 3 ). In particular, abnormal conditions include {timeout, dry burning, gas monitoring abnormality, temperature abnormality, humidity abnormality, video monitoring abnormality} and many other conditions. The state transition process is as follows Figure 3 As shown, the conversion function The correctness of the decision depends on the state and the accuracy of the input conditions.

[0055] In view of this, the fire intelligent early warning method proposed in this embodiment is committed to improving the accuracy of state transfer from two dimensions (that is, the edge computing method of the edge computing box), and specifically designs the following two methods:

[0056] Method 1 (see Figure 4 ), for each state , whose state feature expression is composed of multiple heterogeneous feature vectors Composition (Y t is the output result, X t is a heterogeneous feature vector. The edge intelligent computing box combines heterogeneous data such as temperature detector data, gas detector data, and video surveillance images through the corresponding graph neural network model to extract values, with multi-dimensional heterogeneous features of the kitchen environment state). The feature vector here does not only come from a specific sensor data collection for independent results, but also achieves feature extraction and alignment through multi-sensor information fusion to obtain rich feature expressions. The basic process is as follows Figure 4 For example, open fire anomaly detection no longer relies solely on the "presence / absence" judgment of video surveillance equipment, but integrates heterogeneous data such as temperature detector data, gas detector data, and video surveillance images through multi-dimensional feature expression extracted through deep learning of the graph neural network model.

[0057] Method 2 (see Figure 5 ) For abnormal judgment (white circles represent graph nodes of normal kitchen environment status, which do not require corresponding fire disposal operations, and black circles represent graph nodes of abnormal kitchen environment status, which require fire disposal operations), further integrate the multi-dimensional features of nodes (kitchen environment status), the time sequence of state transitions, etc., and construct a time series neural network to Figure 3 The implicit knowledge described is represented.

[0058] on the one hand, Figure 3 Each state of the finite state machine is mapped to a graph node, and the state relationship of the finite state machine is mapped to the edge of the graph, so as to establish a correspondence between the finite state machine and the graph neural network model, thereby mining spatial association relationships.

[0059] The two methods of improving accuracy involved in the embodiment improve the state transition accuracy and abnormality detection capability of the kitchen fire intelligent early warning system. First, multi-sensor information fusion ( Figure 4 ) ensures the diversity and comprehensiveness of state feature expression. The feature vector of each state no longer relies on the data of a single sensor, but combines the data of multiple sensors such as temperature detectors, gas detectors, and video surveillance. Through the deep learning network, these heterogeneous data are feature extracted and aligned to generate rich multi-dimensional feature expressions. In this way, even if a sensor fails or misreports, the system can still correct and supplement it through the data of other sensors, greatly improving the reliability and accuracy of state judgment. For example, open fire anomaly detection not only relies on the "presence / absence" judgment of video surveillance, but also integrates the data of temperature and gas detectors to form a more accurate anomaly detection mechanism.

[0060] Secondly, temporal neural network modeling ( Figure 5 ) further improves the accuracy of anomaly detection and the rationality of state transfer. By mapping the states of the finite state machine to graph nodes, and mapping the relationships between states to graph edges, a graph structure containing timing information is constructed, and the system can capture the time dependency and spatial correlation in the state transfer process. In this way, not only can we better understand the conversion logic between states, but also we can mine potential abnormal patterns and correlations through timing neural networks. For example, when the temperature and humidity suddenly change abnormally when the fire is in normal state and someone is in charge, the system can quickly determine possible abnormal situations through timing analysis and issue early warnings.

[0061] In addition, this design method is highly scalable and adaptable. Through the application of the graph neural network model, the system can continuously learn and optimize the state transition model to adapt to changes in different kitchen environments and usage habits. Moreover, new sensor data or new types of anomalies can be easily integrated into the existing system without the need to fully reconstruct the system architecture. This not only reduces the complexity of system maintenance, but also improves the flexibility and continuous improvement capabilities of the system. Through multi-sensor information fusion and time series neural network modeling, the finite state machine is combined with the graph neural network model, which not only improves the state transition accuracy and anomaly detection capability of the kitchen fire intelligent warning system, but also enhances the reliability, scalability and adaptability of the system, providing a more intelligent and comprehensive solution for kitchen fire warning.

[0062] On the other hand, considering the changes in state characteristics at different time points, we further use temporal neural networks (such as RNN) to mine temporal correlations (see Figure 6 ).

[0063] Taking into account the different degrees of mutual dependence between graph nodes (finite state machine states), attention is used again in the message passing method of GNN, and the weighted summation of neighboring node features is performed according to the neighboring node weights obtained by attention, that is,

[0064]

[0065] in, To obtain the attention weight coefficient of node i’s neighbor node j; is the vector representation of each node, which is mapped to dimensional feature vector, is a transformable parameter shared among all nodes; then the transformed node feature vector and Concatenate and add a vector Do the inner product and pass the activation function , get an attention coefficient, and use the softmax transformation to get the standardized attention coefficient ; Finally, the attention coefficient As the weight, the vector representation of the neighboring nodes is weighted and summed, and then the activation function is used. Get the characteristic expression of node i . The changeable parameters can be obtained through learning (the kitchen fire intelligent warning model based on the time series graph neural network is obtained as follows Figure 7 As shown in the figure, this function uses the attention mechanism to handle the interdependence between nodes. Specifically, the role of this function is to perform weighted summation of the features of neighboring nodes according to the weights of the neighboring nodes in the message passing method of GNN. The weights in this function are obtained through learning and can be obtained based on the time-series graph neural network to further explore the time correlation relationship). Finally, the feature expression extracted from multiple dimensions such as nodes (multi-sensor nodes), airspace (multi-state), and time domain (time-series input) is input into the classifier, and the abnormal kitchen environment state of the fire is obtained by determining the category of the edge for subsequent disposal.

[0066] The embodiment combines a finite state machine (FSM) with a graph neural network (GNN) and a time series neural network (such as LSTM or GRU), which can significantly improve the performance and intelligence level of the kitchen fire intelligent early warning and disposal system. First, by mapping each state of the FSM to a graph node and mapping the relationship between the states to the edges of the graph, we can use GNN to mine the complex spatial correlation between each state and sensor data in the system. GNN has a strong information integration capability and can capture global information by iteratively updating node features, so that each node (state) not only relies on local information, but also perceives the information of other nodes. This global perspective is crucial for understanding and predicting the mutual influence between various states in the kitchen environment.

[0067] In the time dimension, using time series neural networks to process changes in state features at different time points can effectively capture dynamic changes in the kitchen environment. RNN and its variants (such as LSTM and GRU) are good at processing time series data and can capture long-term and short-term dependencies, thereby helping the system to more accurately predict future states. For example, the changing trends of temperature and smoke concentration in the kitchen are key factors in fire warning. Time series neural networks can identify gradual fire risks and sudden abnormal situations by analyzing these time series data.

[0068] This combination not only enhances the robustness of the system, but also improves the accuracy of early warning. By considering the spatial and temporal correlations at the same time, the system can conduct a more comprehensive analysis and monitoring of the kitchen environment. Even when there is noise or data loss in the sensor data, the system can still make reasonable early warnings through the data of other sensors. In addition, combined with federated learning, the system can continuously optimize itself through distributed learning without centralizing data, adapt to new situations and environmental changes, and improve its intelligence level.

[0069] Ultimately, this combined system can achieve accurate early warning and automated disposal. When a fire risk is identified, the system can close the gas valve in time, push alarm information to the cloud server, and notify relevant personnel. This not only reduces the potential losses caused by fire, but also improves response speed and processing efficiency. Through intelligent and automated early warning and disposal, the system provides a strong guarantee for kitchen safety.

[0070] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A kitchen fire intelligent early warning and disposal method, characterized in that: The following steps are involved: S1, the kitchen data collection terminal collects various kitchen environment parameters in real time and reports them to the edge intelligent computing box; S2, the edge intelligent computing box determines the corresponding finite state machine mode according to the kitchen environment parameters, completes the graph structure construction, and loads the corresponding graph neural network model, where the graph nodes represent different kitchen environment states and the edges represent the switching between states; S3, the edge intelligent computing box preprocesses the input kitchen environment parameters, generates corresponding feature parameter vectors as the input of each graph node, constructs a training data set using a weak supervision method, and provides two data channels for model training and prediction; The weak supervision method is as follows: the edge intelligent computing box periodically scans and determines the historical attributes of the graph nodes to obtain the corresponding kitchen environment status, wherein the kitchen environment status includes idle state, normal state with no fire, abnormal state with no fire, normal state with fire, abnormal state with fire, normal state with fire and no one, abnormal state with fire and no one, and terminated state; S4, through the model training channel, the edge intelligent computing box performs model training in a horizontal federated learning manner; S5, through the model prediction channel, the edge intelligent computing box performs early warning and disposal based on the trained graph neural network model.

2. The kitchen fire intelligent early warning and handling method according to claim 1, characterized in that: The specific steps of S4 include: S411, the edge intelligent computing box uses the training data set channel to retrain the graph neural network model to obtain a global optimized neural network model, and stores the training results in the local model pool; S412, the edge intelligent computing box compresses and encrypts the trained parameters and reports them to the cloud server. The parameters include the first-order derivative of the loss function of the graph neural network model to the model parameters, that is, the gradient parameter; S413, after decrypting and decompressing the data reported by each edge intelligent computing box, the cloud server aggregates it based on the global optimum to obtain an updated graph neural network model, which is sent to each edge intelligent computing box at a regular interval.

3. The kitchen fire intelligent early warning and handling method according to claim 1, characterized in that: The specific steps of S5 include: S511, a graph neural network model selected and loaded from a local model pool by an edge intelligent computing box; S512, the edge intelligent computing box uses the kitchen environment data collected in real time to perform reasoning, and maps and determines the abnormal kitchen environment state type through the switching relationship between the graph nodes. The abnormal kitchen environment state types include abnormal state of no fire, abnormal state of someone in the fire, and abnormal state of no one in the fire. S513, the edge intelligent computing box pushes the abnormal kitchen environment status type to the cloud server for remote processing, and based on the abnormal kitchen environment status type, it links with the electronically controlled gas valve to take emergency measures for the abnormal kitchen environment status.

4. The kitchen fire intelligent early warning and handling method according to claim 1, characterized in that: Each state of the finite state machine is mapped to a graph node, and the state relationship of the finite state machine is mapped to the edge of the graph. A corresponding relationship is established between the finite state machine and the graph neural network model, and the spatial correlation between the two is mined; the temporal correlation between different time points and the changes in the kitchen environment state is mined using the time series neural network.

5. The kitchen fire intelligent early warning and handling method according to claim 4, characterized in that: The graph neural network corresponding to the graph node performs weighted summation of the neighboring node features according to the neighboring node weights obtained by attention, that is, in, To obtain the attention weight coefficient of node i’s neighbor node j; is each node Vector representation, mapped to dimensional feature vector, is a transformable parameter shared among all nodes; then the transformed node feature vector and Concatenate and add a vector Do the inner product and pass the activation function , get an attention coefficient, and use the softmax transformation to get the standardized attention coefficient ; Finally, the attention coefficient As the weight, the vector representation of the neighboring nodes is weighted and summed, and then the activation function is used. Get the characteristic expression of node i .

6. A kitchen fire intelligent early warning and disposal system, used to implement a kitchen fire intelligent early warning and disposal method according to any one of claims 1 to 5, comprising a kitchen data acquisition terminal, an edge intelligent computing box, a cloud server, an interactive terminal and an electronically controlled gas valve, characterized in that: The kitchen data collection terminal is connected to the edge intelligent computing box network; The edge intelligent computing box is electronically connected to the electronically controlled gas valve and is network-connected to the cloud server; The cloud server is connected to the interactive terminal network; The edge intelligent computing box can conduct federated learning training with the cloud server and use the graph neural network model; the kitchen data collection end collects data and transmits it to the edge intelligent computing box, which performs fire warning and disposal through the graph neural network model, and pushes kitchen abnormal status information to the cloud server, which interacts with the interactive terminal.

7. The kitchen fire intelligent early warning and handling system according to claim 6, characterized in that: The kitchen data collection terminal includes temperature sensors, humidity sensors, catalytic combustion gas sensors, dual-band smoke detectors, manual triggers and video monitoring equipment.

Images