Fire early warning system for firefighting facilities based on multi-channel gas analysis

Abstract

The present invention discloses a fire early warning system for fire-fighting facilities with multi-channel gas analysis, which belongs to the field of artificial intelligence technology. The present invention locates the corresponding floors and corresponding equipment inside the building, obtains fire-fighting equipment information, hospital fire records and equipment alarm records, and analyzes the equipment status; monitors the data inside the building in real time, unifies the time axis, forms fire warning data, trains long-term and short-term memory network models, analyzes fire warning data, generates fire warning levels, and locates in the three-dimensional model of the building; collects real-time gas composition data in the fire warning location area, trains a fire degree assessment model, predicts the newly collected data, determines whether there is a fire in the current environment, assesses the severity of the fire, identifies false alarms, and adjusts the warning level; according to the confirmed warning level, triggers the corresponding alarm mechanism, automatically links the fire-fighting equipment, generates evacuation instructions in the case of advanced warnings, and guides the evacuation of personnel.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a fire early warning system for fire fighting facilities using multi-channel gas analysis. Background Art

[0002] With the acceleration of urbanization, the size and complexity of buildings are increasing, especially in crowded places such as hospitals, shopping malls, and office buildings, where the risk of fire has increased significantly. In order to reduce the loss of life and property caused by fire, modern buildings are usually equipped with a variety of fire-fighting equipment and monitoring systems, such as smoke alarms, temperature alarms, manual alarm buttons, and automatic fire-extinguishing systems. Although these devices can provide fire warning and fire-fighting support to a certain extent, there are still many deficiencies in practical applications.

[0003] Existing fire warning systems usually rely on a single or a few data sources, making it difficult to effectively integrate multi-source data, resulting in insufficient accuracy and reliability of fire detection. Due to the lack of multi-channel data analysis and intelligent fire warning algorithms, existing systems often have problems with delayed warnings or frequent false alarms. Delayed warnings can cause fires to spread and be difficult to control in a timely manner, while false alarms can cause unnecessary panic and waste of resources. Current automatic fire extinguishing and smoke exhaust systems often rely on pre-set fixed parameters. During the development of a fire, it is impossible to dynamically adjust the fire extinguishing and smoke exhaust strategies based on real-time data, resulting in unsatisfactory fire extinguishing effects and even exacerbating fire hazards. Existing evacuation guidance systems are mostly fixed routes, and it is impossible to generate the best evacuation route in real time based on the actual situation of the fire. For complex building structures or scenes where fire spreads rapidly, fixed evacuation routes may cause secondary injuries. Summary of the invention

[0004] The purpose of the present invention is to provide a fire early warning system for fire fighting facilities with multi-channel gas analysis to solve the problems raised in the prior art.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A fire early warning system for fire-fighting facilities with multi-channel gas analysis, the system comprising:

[0007] Fire equipment management module: including: positioning unit, data acquisition unit and equipment status analysis unit; the positioning unit uses the floor plan and three-dimensional model of the building to locate the corresponding floors and corresponding equipment inside the building; the data acquisition unit is responsible for obtaining fire equipment information, including the total number of equipment, equipment operation rate, offline rate and failure rate, and obtaining hospital fire records and equipment alarm records; the equipment status analysis unit is responsible for analyzing the equipment status using a decision tree algorithm in combination with the fire equipment information;

[0008] Fire warning module: including: data monitoring unit, data integration unit and warning level generation unit; the data monitoring unit is responsible for real-time monitoring of equipment status data, smoke alarm data, temperature alarm data and manual alarm data inside the building; the data integration unit is responsible for unifying the timeline, synchronizing data from different sources to the same timeline, and integrating data through weighted fusion to form fire warning data; the warning level generation unit is responsible for training the long short-term memory network model, analyzing fire warning data, generating fire warning levels, and positioning in the three-dimensional model of the building;

[0009] Multi-channel gas analysis module: including: data preprocessing unit, fire severity assessment unit, fire spread speed calculation unit and false alarm identification unit; the data preprocessing unit is responsible for obtaining the fire warning level and the fire warning location data in the three-dimensional model, collecting the real-time gas composition data of the fire warning location area through the multi-channel gas sensors deployed at the warning location, and preprocessing the collected gas data; the fire severity assessment unit uses historical fire data and false alarm data to train the fire severity assessment model of the convolutional neural network combined with the long short-term memory network, predicts the newly collected data, determines whether there is a fire in the current environment, and assesses the severity of the fire; the fire spread speed calculation unit uses the multivariate regression algorithm to construct a fire spread model based on the fire warning data, the severity of the fire, the collected gas data and the three-dimensional model of the building, and generates the fire spread speed; the false alarm identification unit is responsible for identifying false alarms based on the output of the fire severity assessment model and adjusting the warning level;

[0010] Early warning processing module: including: alarm trigger unit, fire equipment linkage unit, evacuation guidance unit and evacuation monitoring unit; the alarm trigger unit triggers the corresponding alarm mechanism according to the confirmed early warning level. If the early warning level reaches the intermediate or advanced level, the fire equipment linkage unit will automatically link the fire equipment, trigger the automatic fire extinguishing system in the corresponding area according to the fire warning location and the severity of the fire, and adjust the ventilation and smoke exhaust system in the building. In the case of advanced early warning, the evacuation guidance unit generates evacuation instructions and guides personnel evacuation through the public broadcasting system in the building. The evacuation monitoring unit tracks the evacuation progress in real time through monitoring cameras and sensors.

[0011] In the fire equipment management module, the initial position of each fire equipment is marked on the floor plan, and a unique equipment ID is used for marking. The equipment position is determined by on-site measurement to ensure that the marking of each equipment is consistent with the corresponding actual physical location; according to the layout in the floor plan, the interior of the building is divided into areas, each area corresponds to a certain number of fire equipment, and the fire equipment is associated with the area information;

[0012] Map the equipment annotations in the plan view to the three-dimensional model, and determine the spatial position of the equipment by combining the two-dimensional coordinates of the plan view with the Z-axis coordinates of the three-dimensional model. For equipment that requires a special installation location, indicate the specific installation height and angle in the three-dimensional model. Ensure that the equipment position data in the plan view is consistent with that in the three-dimensional model. If the equipment position is updated in the plan view, it will be updated synchronously in the three-dimensional model. If the equipment position is updated in the three-dimensional model, it will be updated synchronously in the plan view. Regularly calibrate the position of equipment inside the building, use the positioning error correction algorithm Kalman filtering to optimize the accuracy of the collected data, and reduce the deviation between the actual position of the equipment and the position in the plan view and three-dimensional model.

[0013] In the fire equipment management module, relevant features are extracted from the collected fire equipment information and used as input to the decision tree algorithm to classify the equipment status. Each equipment status corresponds to a specific feature combination, which will be used as the output of the decision tree;

[0014] The collected equipment status and historical data are used as training sets and input into the decision tree algorithm. The decision tree recursively selects the features that best distinguish the equipment status and gradually builds a tree structure. Each branch node represents the selection of a feature, and the leaf node represents the final prediction result of the equipment status. The decision tree classifies the equipment according to its offline rate. If the offline rate exceeds a certain threshold, the equipment will be marked as "offline". If it does not exceed the threshold, it will be classified according to the operating rate to determine whether the equipment is in a "normal" state or a "fault warning" state. The system continuously monitors and updates the status information of each fire-fighting equipment. New data will be input into the trained decision tree model in real time to analyze the current status of the equipment in real time.

[0015] The decision tree makes decisions along the tree structure from the root node to the leaf node based on the input feature value, and finally outputs the device status prediction result.

[0016] In the fire warning module, the fire warning data is obtained through the data integration unit to form a fire warning vector ,in, arrive It represents the value of the first category fire warning data to the nth category fire warning data at time t, where n is the number of data categories in the fire warning data;

[0017] The long short-term memory network model controls the flow of fire warning data through input gate, forget gate and output gate. The formula is as follows:

[0018]

[0019] in, is the weight matrix of the forget gate, is the bias term of the forget gate, is the activation function, which is the sigmoid function. is the hidden state at the previous moment;

[0020] Input Gate and candidate memory cell states in the long short-term memory network model The formula is as follows:

[0021]

[0022]

[0023] in, and are the weight matrices of the input gate and candidate memory cell states, and are the bias terms of the input gate and candidate memory cell states respectively;

[0024] Update the memory cell state, the formula is as follows:

[0025]

[0026] in, is the current state of the memory cell, is the state of the memory unit at the previous moment;

[0027] Output Gate and hidden state The formula is as follows:

[0028]

[0029]

[0030] in, is the weight matrix of the output gate, is the bias of the output gate;

[0031] The long short-term memory network model is trained using historical fire data and false alarm data. The goal is to input the fire warning vector Predict future fire risk levels and define the loss function L:

[0032]

[0033] in, is the fire warning level predicted by the model, is the actual warning level, T is the total number of time steps; the loss function is minimized using back propagation and gradient descent algorithms to adjust the parameters of the long short-term memory network model;

[0034] The trained LSTM network model can predict the fire warning level based on the real-time input data stream in practical applications. The output of the model It is a real value representing the risk score of fire. The threshold is set according to the score to divide the warning levels into three levels: low, medium and high.

[0035] In the fire warning module, after generating the fire warning level, the warning signal is located at a specific location in combination with the three-dimensional model of the building; the building is divided into several monitoring areas, each of which has corresponding sensor data. , where j represents the region, and the long short-term memory network model will generate an early warning level for each region;

[0036] In the three-dimensional model of the building, the generated warning level is mapped to the specific location in the three-dimensional space based on the physical location and corresponding area of ​​the sensor; the location and risk level of the fire warning signal can be viewed through the visualization system of the three-dimensional model.

[0037] In the multi-channel gas analysis module, the data preprocessing unit obtains real-time gas composition data from the multi-channel gas sensors deployed in the fire warning location area. Each sensor detects different types of gases in real time and inputs them in the form of time series, which is expressed as ,in, arrive Indicates the concentration value of the first gas to the mth gas at time t, where m is the number of gas types; Kalman filtering is used to Smoothing is performed to reduce the impact of noise; a low-pass filter is used to eliminate high-frequency noise and retain the effective low-frequency signal part; Z-score standardization is used to normalize the data of different gases to the same scale; and the data of all sensors are synchronized to a unified time axis through interpolation methods.

[0038] In the multi-channel gas analysis module, for the convolutional neural network part, the input layer receives real-time gas composition data collected from the multi-channel gas sensor, and the input data is a multi-dimensional matrix, each dimension represents a gas sensor, and each element in the matrix represents the concentration value of the corresponding gas at different time points; the convolution layer operates on the input data through a number of filters, each filter is a matrix with a smaller dimension than the input data; during the sliding process of the filter, the convolution operation is performed on the local area of ​​the input matrix, the weighted sum of the local data is calculated, and a new feature map is generated, and the convolution operation will extract the relationship between the gas concentrations;

[0039] The output of the convolution layer is transformed nonlinearly through the activation function ReLU. The pooling layer is used to reduce the dimension of the convolution layer output and aggregate local features, thereby reducing the amount of data and enhancing the robustness of the model. The convolutional neural network contains a combination of several convolutional layers and pooling layers, which extract high-level features of the data layer by layer. The first few layers focus on capturing local changes in gas concentration, and the latter layers are responsible for identifying how local changes combine to form more complex fire signals. It is connected to several fully connected layers to aggregate the extracted high-level features to generate a high-dimensional feature vector to reflect the overall trend and characteristics of the fire. The final output layer maps the high-dimensional feature vector into a fire risk score.

[0040] In the multi-channel gas analysis module, for the long short-term memory network part, the data received by the input layer is the fire warning data processed by the data integration unit of the fire warning module; when the equipment state changes abnormally, the input gate writes the abnormal information into the memory unit, and if the smoke alarm at a certain moment returns to normal, the forget gate will reduce the memory strength of the previous alarm; if several sensors send abnormal signals at the same time, the output gate will integrate the abnormal signals and output a comprehensive fire warning signal;

[0041] At each time step, the information in the memory unit will be updated according to the current input and the previous state, and the current output will be generated; the long short-term memory network will stack several layers, and the output of each layer will be used as the input of the next layer to extract higher-level time series features layer by layer; it will be aggregated through the fully connected layer to generate a fire warning indicator; the output layer will map the fire warning indicator to a fire warning level to indicate the severity of the fire;

[0042] Get fire warning vector , Fire severity , Gas data and 3D models of buildings ,in, is the spatial coordinate; from The structural characteristics of the building, the layout of items inside the building, the heat conduction path inside the building and the spatial connectivity of the building are extracted and quantified. In the quantification of structural characteristics, the room area is scored according to the area, and the larger the area, the higher the score. The floor height is scored according to the height, and the higher the floor height, the higher the score. The wall thickness is scored according to the wall thickness and the fire resistance of the material. The number of doors and windows is recorded to affect the speed of fire spread. In the quantification of item layout, the score is set according to the room use category, and the storage of flammable items affects the score. In the quantification of heat conduction paths, the score is set according to the thermal conductivity of the material. The thermal conductive material affects the fire resistance. The higher the ventilation system score, the lower the smoke diffusion in the building. In the quantification of spatial connectivity, the more open areas, the lower the score, and effective partitions can improve the score. A weight is set for each feature to reflect its impact on the spread of fire. The score of each feature is multiplied by the corresponding weight, and the sum is taken to obtain the building's fire spread resistance index u.

[0043] Using the regression model, assuming the fire spread rate is , the mathematical expression is as follows:

[0044]

[0045] in, is the intercept term of the model, They are the building's fire spread resistance index, fire warning vector, fire severity, and regression coefficient of gas data, indicating the impact of each feature on the fire spread speed. is the error term;

[0046] The regression model is trained using historical data, with the actual fire spread rate as the target variable. The training data set contains and The value of is determined using the least squares method, and the performance of the model is verified using cross-validation. In actual applications, the corresponding parameters are input to generate the fire spread rate.

[0047] In the multi-channel gas analysis module, the fire severity assessment model outputs a fire severity score and the probability of fire , will be used as the input of the false positive identification unit; among them, , where 1 means the fire is extremely serious and 0 means there is no fire. , where 1 indicates a fire has occurred and 0 indicates no fire; set the fire severity threshold and fire probability threshold ,if and , it is considered that there may be a false alarm; the false alarm identification unit uses other sensor data for weighted fusion and calculates a false alarm identification comprehensive index K to assist false alarm identification. If K is less than the false alarm identification comprehensive index threshold ,and and , it is considered a false alarm and the warning level is adjusted.

[0048] In the warning processing module, the system monitors the warning level. When the level reaches the preset intermediate or high level, the alarm is triggered. According to the warning level and the location of the fire, the personnel in the key areas are notified first. The fire equipment linkage unit will start the automatic fire extinguishing system and dynamically adjust the ventilation and smoke exhaust system in the building according to the severity and location of the fire.

[0049] Generate the best evacuation routes and instructions based on the location, severity and building structure of the fire, communicate evacuation instructions through the public broadcasting system in the building, and activate emergency lights and escape signs in the building in coordination with evacuation instructions;

[0050] When generating the best evacuation route and instruction process, mark the fire area and impact range, identify and avoid obstacles in the building, apply the Dijkstra algorithm to calculate the shortest path from each room to the nearest safe exit, and design several backup evacuation routes to deal with the situation where the main route is blocked; generate an evacuation route map, including path guidance from each area to the safe exit, adjust instructions in real time and update the evacuation route according to the progress of the fire and the progress of evacuation.

[0051] The interior of a building is modeled as a graph structure, where vertices represent important locations such as rooms, corridors, stairwells, and emergency exits, and edges represent pathways between vertices, such as corridors between rooms and connections between stairwells and floors; the weight of the edge represents the time or distance required to move from one vertex to another, and these weights can be set according to the actual layout and specific circumstances of the building.

[0052] For each room as the starting point, the Dijkstra algorithm is applied to calculate the shortest path to each safety exit in the building. The main steps of the algorithm are as follows: Set the distance from the starting room to itself to 0, and set the initial distance of all other vertices to infinity. Put the starting room into a priority queue, and the priority is determined by the currently known shortest path length. Each time the nearest vertex is taken out of the priority queue, check the neighbors of the vertex, that is, the adjacent rooms or passages that can be reached directly. Through the path of the current vertex, try to update the shortest path length of the neighbors. If the path length from the current vertex to a neighbor is less than the known shortest path length, update the shortest path length of the neighbor and reinsert the neighbor into the priority queue to continue exploring other extension possibilities of the path. The algorithm ends when all vertices have been processed or the shortest path from the starting point to all safety exits has been found.

[0053] After calculating the shortest path, the system generates backup paths based on the results of multiple iterations of the Dijkstra algorithm. These backup paths usually do not overlap with the main path, or only overlap in some sections to ensure that there are other available paths when the main path is blocked. When designing backup paths, priorities can be given to different paths. For example, the shortest path has the highest priority, and the priority of the backup paths decreases successively. In this way, when a fire occurs, the system can give priority to guiding people to take the main path and switch to the backup path when necessary. During the evacuation process, if it is detected that the main path is blocked, for example due to fire spread or structural damage, the system can adjust the strategy in real time, select the backup path for evacuation, and notify the people in the building to change the evacuation direction.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] 1. The present invention collects real-time gas composition data of the fire warning area through multi-channel gas sensors, and uses multi-source data fusion technology to integrate fire equipment data, gas analysis data, smoke and temperature sensing data to form high-precision fire warning data.

[0056] 2. The present invention introduces a fire severity assessment model that combines a convolutional neural network with a long short-term memory network. It uses historical fire data and false alarm data for training, and can accurately determine whether there is a fire in the current environment and assess the severity of the fire, effectively reducing the false alarm rate and providing more timely warnings.

[0057] 3. The present invention can generate the best evacuation route using an optimization algorithm according to the location, severity and building structure of the fire, and provide dynamic evacuation guidance in real time through a public broadcasting system, emergency indicator lights and guides. It can also adjust the evacuation path in real time according to the development of the fire, avoiding the risks that may be caused by fixed paths and ensuring the safe evacuation of personnel. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a schematic diagram of the steps of the fire early warning system for fire fighting facilities of the present invention using multi-channel gas analysis;

[0059] Figure 2 It is a system structure diagram of the fire early warning system for fire fighting facilities of the present invention with multi-channel gas analysis. DETAILED DESCRIPTION

[0060] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] Example: Figure 1-Figure 2 As shown, the present invention provides a technical solution.

[0062] According to one embodiment of the present invention, Figure 1 As shown in the schematic diagram of the steps of the fire early warning system for fire fighting facilities using multi-channel gas analysis, the fire early warning system for fire fighting facilities using multi-channel gas analysis comprises:

[0063] Fire equipment management module: including: positioning unit, data acquisition unit and equipment status analysis unit; the positioning unit uses the floor plan and three-dimensional model of the building to locate the corresponding floors and corresponding equipment inside the building; the data acquisition unit is responsible for obtaining fire equipment information, including the total number of equipment, equipment operation rate, offline rate and failure rate, and obtaining hospital fire records and equipment alarm records; the equipment status analysis unit is responsible for analyzing the equipment status using a decision tree algorithm in combination with the fire equipment information;

[0064] Fire warning module: including: data monitoring unit, data integration unit and warning level generation unit; the data monitoring unit is responsible for real-time monitoring of equipment status data, smoke alarm data, temperature alarm data and manual alarm data inside the building; the data integration unit is responsible for unifying the timeline, synchronizing data from different sources to the same timeline, and integrating data through weighted fusion to form fire warning data; the warning level generation unit is responsible for training the long short-term memory network model, analyzing fire warning data, generating fire warning levels, and positioning in the three-dimensional model of the building;

[0065] Multi-channel gas analysis module: including: data preprocessing unit, fire severity assessment unit, fire spread speed calculation unit and false alarm identification unit; the data preprocessing unit is responsible for obtaining the fire warning level and the fire warning location data in the three-dimensional model, collecting the real-time gas composition data of the fire warning location area through the multi-channel gas sensors deployed at the warning location, and preprocessing the collected gas data; the fire severity assessment unit uses historical fire data and false alarm data to train the fire severity assessment model of the convolutional neural network combined with the long short-term memory network, predicts the newly collected data, determines whether there is a fire in the current environment, and assesses the severity of the fire; the fire spread speed calculation unit uses the multivariate regression algorithm to construct a fire spread model based on the fire warning data, the severity of the fire, the collected gas data and the three-dimensional model of the building, and generates the fire spread speed; the false alarm identification unit is responsible for identifying false alarms based on the output of the fire severity assessment model and adjusting the warning level;

[0066] Early warning processing module: including: alarm trigger unit, fire equipment linkage unit, evacuation guidance unit and evacuation monitoring unit; the alarm trigger unit triggers the corresponding alarm mechanism according to the confirmed early warning level. If the early warning level reaches the intermediate or advanced level, the fire equipment linkage unit will automatically link the fire equipment, trigger the automatic fire extinguishing system in the corresponding area according to the fire warning location and the severity of the fire, and adjust the ventilation and smoke exhaust system in the building. In the case of advanced early warning, the evacuation guidance unit generates evacuation instructions and guides personnel evacuation through the public broadcasting system in the building. The evacuation monitoring unit tracks the evacuation progress in real time through monitoring cameras and sensors.

[0067] In the fire equipment management module, the initial position of each fire equipment is marked on the floor plan, and a unique equipment ID is used for marking. The equipment position is determined by on-site measurement to ensure that the marking of each equipment is consistent with the corresponding actual physical location; according to the layout in the floor plan, the interior of the building is divided into areas, each area corresponds to a certain number of fire equipment, and the fire equipment is associated with the area information;

[0068] Map the equipment annotations in the plan view to the three-dimensional model, and determine the spatial position of the equipment by combining the two-dimensional coordinates of the plan view with the Z-axis coordinates of the three-dimensional model. For equipment that requires a special installation location, indicate the specific installation height and angle in the three-dimensional model. Ensure that the equipment position data in the plan view is consistent with that in the three-dimensional model. If the equipment position is updated in the plan view, it will be updated synchronously in the three-dimensional model. If the equipment position is updated in the three-dimensional model, it will be updated synchronously in the plan view. Regularly calibrate the position of equipment inside the building, use the positioning error correction algorithm Kalman filtering to optimize the accuracy of the collected data, and reduce the deviation between the actual position of the equipment and the position in the plan view and three-dimensional model.

[0069] In the fire equipment management module, relevant features are extracted from the collected fire equipment information and used as input to the decision tree algorithm to classify the equipment status. Each equipment status corresponds to a specific feature combination, which will be used as the output of the decision tree;

[0070] The collected equipment status and historical data are used as training sets and input into the decision tree algorithm. The decision tree recursively selects the features that best distinguish the equipment status and gradually builds a tree structure. Each branch node represents the selection of a feature, and the leaf node represents the final prediction result of the equipment status. The decision tree classifies the equipment according to its offline rate. If the offline rate exceeds a certain threshold, the equipment will be marked as "offline". If it does not exceed the threshold, it will be classified according to the operating rate to determine whether the equipment is in a "normal" state or a "fault warning" state. The system continuously monitors and updates the status information of each fire-fighting equipment. New data will be input into the trained decision tree model in real time to analyze the current status of the equipment in real time.

[0071] The decision tree makes decisions along the tree structure from the root node to the leaf node based on the input feature value, and finally outputs the device status prediction result.

[0072] In this embodiment, relevant data of 100 fire-fighting equipment in a hospital are collected. The data of each device contains the following features: Equipment ID: uniquely identifies each device. Operation rate: the proportion of the normal operation time of the equipment to the total time, ranging from 0-100%. Offline rate: the proportion of the offline time of the equipment to the total time, ranging from 0-100%. Failure rate: the ratio of the number of times the equipment fails to the total number of operations. The historical data also includes the status mark of each device, which is used to train the decision tree model: Normal: The equipment is in normal working condition, the operation rate is high, and the offline rate and failure rate are low. Fault warning: The equipment may fail, the operation rate decreases, and the failure rate increases. Offline: The equipment has not been working for a long time and the offline rate is high.

[0073] The data of 80 devices, i.e. 80% of the data, is used as the training set to train the decision tree model. The decision tree algorithm is based on information gain, gradually selecting the features that can best distinguish the device status and building a tree structure. Each node represents a feature selection, and the final leaf node represents the predicted result of the device status. The data of the remaining 20 devices, i.e. 20% of the data, is used as the test set to verify the classification accuracy of the model. By inputting the test set into the trained decision tree model, the status of each device is predicted, and the predicted results are compared with the actual status to evaluate the accuracy of the model.

[0074] The results show that the decision tree model has an accuracy of 95% in classifying the device status on the test set, and most device status can be accurately predicted. Among the 20 test devices, the status of one device was misclassified as "fault warning", while the actual status was "normal". The misclassification was mainly due to the failure rate of the device being slightly higher than the average level, but not meeting the warning standard. As time goes by, the system continues to collect new device data and input it into the trained decision tree model, and the model can be dynamically updated to maintain a high classification accuracy.

[0075] In the fire warning module, the fire warning data is obtained through the data integration unit to form a fire warning vector ,in, arrive It represents the value of the first category fire warning data to the nth category fire warning data at time t, where n is the number of data categories in the fire warning data;

[0076] The long short-term memory network model controls the flow of fire warning data through input gate, forget gate and output gate. The formula is as follows:

[0077]

[0078] in, is the weight matrix of the forget gate, is the bias term of the forget gate, is the activation function, which is the sigmoid function. is the hidden state at the previous moment;

[0079] Input Gate and candidate memory cell states in the long short-term memory network model The formula is as follows:

[0080]

[0081]

[0082] in, and are the weight matrices of the input gate and candidate memory cell states, and are the bias terms of the input gate and candidate memory cell states respectively;

[0083] Update the memory cell state, the formula is as follows:

[0084]

[0085] in, is the current state of the memory cell, is the state of the memory unit at the previous moment;

[0086] Output Gate and hidden state The formula is as follows:

[0087]

[0088]

[0089] in, is the weight matrix of the output gate, is the bias of the output gate;

[0090] The long short-term memory network model is trained using historical fire data and false alarm data. The goal is to input the fire warning vector Predict future fire risk levels and define the loss function L:

[0091]

[0092] in, is the fire warning level predicted by the model, is the actual warning level, T is the total number of time steps; the loss function is minimized using back propagation and gradient descent algorithms to adjust the parameters of the long short-term memory network model;

[0093] The trained LSTM network model can predict the fire warning level based on the real-time input data stream in practical applications. The output of the model It is a real value representing the risk score of fire. The threshold is set according to the score to divide the warning levels into three levels: low, medium and high.

[0094] In the fire warning module, after generating the fire warning level, the warning signal is located at a specific location in combination with the three-dimensional model of the building; the building is divided into several monitoring areas, each of which has corresponding sensor data. , where j represents the region, and the long short-term memory network model will generate an early warning level for each region;

[0095] On the test set, the LSTM model predicted the fire risk score with an average error of 0.15 based on the standardized score range of 0 to 1. The prediction accuracy of low-level warnings was 97%, that of medium-level warnings was 89%, and that of high-level warnings was 92%. Through training with historical false alarm data, the model correctly identified 83% of false alarm events in the test set, significantly reducing the false alarm rate.

[0096] Some data examples are as follows: Area 1 (equipment room): smoke alarm data: 85%, temperature alarm data: 45°C, manual alarm data: 0, equipment status data: 1 (normal), model-predicted fire risk score: 0.85 (high-level warning); Area 2 (corridor): smoke alarm data: 10%; temperature alarm data: 23°C; manual alarm data: 0; equipment status data: 1 (normal); model-predicted fire risk score: 0.2 (low-level warning); Area 3 (storage area): smoke alarm data: 40%; temperature alarm data: 33°C; manual alarm data: 1 (manual alarm); equipment status data: 1 (normal); model-predicted fire risk score: 0.6 (medium-level warning).

[0097] In the three-dimensional model of the building, the generated warning level is mapped to the specific location in the three-dimensional space based on the physical location and corresponding area of ​​the sensor; the location and risk level of the fire warning signal can be viewed through the visualization system of the three-dimensional model.

[0098] According to another embodiment of the present invention, Figure 2 As shown in the system structure diagram of the fire early warning system of fire fighting facilities based on multi-channel gas analysis, in the multi-channel gas analysis module, the data preprocessing unit obtains real-time gas composition data from the multi-channel gas sensors deployed in the fire warning location area. Each sensor detects different types of gases in real time and inputs them in the form of time series, which is expressed as ,in, arrive Indicates the concentration value of the first gas to the mth gas at time t, where m is the number of gas types; Kalman filtering is used to Smoothing is performed to reduce the impact of noise; a low-pass filter is used to eliminate high-frequency noise and retain the effective low-frequency signal part; Z-score standardization is used to normalize the data of different gases to the same scale; interpolation methods are used to synchronize the data of all sensors to a unified time axis. The data contains time series of multiple different types of gas concentrations, such as CO2, CO, CH4, O2, NOx, etc. The sampling frequency is once per second. The data has been continuously monitored for one week, including data of normal operating conditions, fire event simulations, and false alarm scenarios.

[0099] In the multi-channel gas analysis module, for the convolutional neural network part, the input layer receives real-time gas composition data collected from the multi-channel gas sensor, and the input data is a multi-dimensional matrix, each dimension represents a gas sensor, and each element in the matrix represents the concentration value of the corresponding gas at different time points; the convolution layer operates on the input data through a number of filters, each filter is a matrix with a smaller dimension than the input data; in this embodiment, the first convolution layer uses a 3×3 filter to extract the change trend of each gas concentration in a local time window and generate a primary feature map; the second convolution layer uses a larger filter (5×5) to further extract more complex relationships between gas concentrations; during the sliding process of the filter, a convolution operation is performed on a local area of ​​the input matrix, the weighted sum of the local data is calculated, and a new feature map is generated, and the convolution operation will extract the relationship between the gas concentrations;

[0100] The output of the convolution layer is transformed nonlinearly through the activation function ReLU. The pooling layer is used to reduce the dimension of the convolution layer output and aggregate local features, thereby reducing the amount of data and enhancing the robustness of the model. The convolutional neural network contains a combination of several convolutional layers and pooling layers, which extract high-level features of the data layer by layer. The first few layers focus on capturing local changes in gas concentration, and the latter layers are responsible for identifying how local changes combine to form more complex fire signals. It is connected to several fully connected layers to aggregate the extracted high-level features to generate a high-dimensional feature vector to reflect the overall trend and characteristics of the fire. The final output layer maps the high-dimensional feature vector into a fire risk score.

[0101] In this embodiment, when the CO2 concentration gradually increased from the normal 400ppm to 1000ppm, the fire risk score given by the model increased from 0.2 to 0.8, indicating that the fire risk increased significantly. Finally, the system triggered an intermediate warning. In the simulated gas leak incident, the CH4 concentration quickly increased from 1ppm to 50ppm. The fire risk score of the CNN model jumped from 0.1 to 0.9 in a short period of time. The system triggered an advanced warning and automatically started related equipment. In a set of false alarm data, the NOx concentration fluctuated abnormally, but the concentrations of other gases remained within the normal range. After comprehensively analyzing all gas concentrations, the model only gave a low risk score of 0.3, avoiding false alarm triggers.

[0102] In the multi-channel gas analysis module, for the long short-term memory network part, the data received by the input layer is the fire warning data processed by the data integration unit of the fire warning module; when the equipment state changes abnormally, the input gate writes the abnormal information into the memory unit, and if the smoke alarm at a certain moment returns to normal, the forget gate will reduce the memory strength of the previous alarm; if several sensors send abnormal signals at the same time, the output gate will integrate the abnormal signals and output a comprehensive fire warning signal;

[0103] At each time step, the information in the memory unit will be updated according to the current input and the previous state, and the current output will be generated; the long short-term memory network will stack several layers, and the output of each layer will be used as the input of the next layer to extract higher-level time series features layer by layer; it will be aggregated through the fully connected layer to generate a fire warning indicator; the output layer will map the fire warning indicator to a fire warning level to indicate the severity of the fire;

[0104] Get fire warning vector , Fire severity , Gas data and 3D models of buildings ,in, is the spatial coordinate; from The structural characteristics of the building, the layout of items inside the building, the heat conduction path inside the building and the spatial connectivity of the building are extracted and quantified. In the quantification of structural characteristics, the room area is scored according to the area, and the larger the area, the higher the score. The floor height is scored according to the height, and the higher the floor height, the higher the score. The wall thickness is scored according to the wall thickness and the fire resistance of the material. The number of doors and windows is recorded to affect the speed of fire spread. In the quantification of item layout, the score is set according to the room use category, and the storage of flammable items affects the score. In the quantification of heat conduction paths, the score is set according to the thermal conductivity of the material. The thermal conductive material affects the fire resistance. The higher the ventilation system score, the lower the smoke diffusion in the building. In the quantification of spatial connectivity, the more open areas, the lower the score, and effective partitions can improve the score. A weight is set for each feature to reflect its impact on the spread of fire. The score of each feature is multiplied by the corresponding weight, and the sum is taken to obtain the building's fire spread resistance index u.

[0105] Using the regression model, assuming the fire spread rate is , the mathematical expression is as follows:

[0106]

[0107] in, is the intercept term of the model, They are the building's fire spread resistance index, fire warning vector, fire severity, and regression coefficient of gas data, indicating the impact of each feature on the fire spread speed. is the error term;

[0108] The regression model is trained using historical data, with the actual fire spread rate as the target variable. The training data set contains and The value of is determined using the least squares method, and the performance of the model is verified using cross-validation. In actual applications, the corresponding parameters are input to generate the fire spread rate.

[0109] In the multi-channel gas analysis module, the fire severity assessment model outputs a fire severity score and the probability of fire , will be used as the input of the false positive identification unit; among them, , where 1 means the fire is extremely serious and 0 means there is no fire. , where 1 indicates a fire has occurred and 0 indicates no fire; set the fire severity threshold and fire probability threshold ,if and , it is considered that there may be a false alarm; the false alarm identification unit uses other sensor data for weighted fusion and calculates a false alarm identification comprehensive index K to assist false alarm identification. If K is less than the false alarm identification comprehensive index threshold ,and and , it is considered a false alarm and the warning level is adjusted.

[0110] In the warning processing module, the system monitors the warning level. When the level reaches the preset intermediate or high level, the alarm is triggered. According to the warning level and the location of the fire, the personnel in the key areas are notified first. The fire equipment linkage unit will start the automatic fire extinguishing system and dynamically adjust the ventilation and smoke exhaust system in the building according to the severity and location of the fire.

[0111] Generate the best evacuation routes and instructions based on the location, severity and building structure of the fire, communicate evacuation instructions through the public broadcasting system in the building, and activate emergency lights and escape signs in the building in coordination with evacuation instructions;

[0112] When generating the best evacuation route and instruction process, mark the fire area and impact range, identify and avoid obstacles in the building, apply the Dijkstra algorithm to calculate the shortest path from each room to the nearest safe exit, and design several backup evacuation routes to deal with the situation where the main route is blocked; generate an evacuation route map, including path guidance from each area to the safe exit, adjust instructions in real time and update the evacuation route according to the progress of the fire and the progress of evacuation.

[0113] The interior of a building is modeled as a graph structure, where vertices represent important locations such as rooms, corridors, stairwells, and emergency exits, and edges represent pathways between vertices, such as corridors between rooms and connections between stairwells and floors; the weight of the edge represents the time or distance required to move from one vertex to another, and these weights can be set according to the actual layout and specific circumstances of the building.

[0114] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. The fire early warning system for fire fighting facilities with multi-channel gas analysis is characterized by: The system comprises: Fire equipment management module: including: positioning unit, data acquisition unit and equipment status analysis unit; the positioning unit uses the floor plan and three-dimensional model of the building to locate the corresponding floors and corresponding equipment inside the building; the data acquisition unit is responsible for obtaining fire equipment information, including the total number of equipment, equipment operation rate, offline rate and failure rate, and obtaining hospital fire records and equipment alarm records; the equipment status analysis unit is responsible for analyzing the equipment status using a decision tree algorithm in combination with the fire equipment information; Fire warning module: including: data monitoring unit, data integration unit and warning level generation unit; the data monitoring unit is responsible for real-time monitoring of equipment status data, smoke alarm data, temperature alarm data and manual alarm data inside the building; the data integration unit is responsible for unifying the timeline, synchronizing data from different sources to the same timeline, and integrating data through weighted fusion to form fire warning data; the warning level generation unit is responsible for training the long short-term memory network model, analyzing fire warning data, generating fire warning levels, and positioning in the three-dimensional model of the building; Multi-channel gas analysis module: including: data preprocessing unit, fire severity assessment unit, fire spread speed calculation unit and false alarm identification unit; the data preprocessing unit is responsible for obtaining the fire warning level and the fire warning location data in the three-dimensional model, collecting the real-time gas composition data of the fire warning location area through the multi-channel gas sensors deployed at the warning location, and preprocessing the collected gas data; the fire severity assessment unit uses historical fire data and false alarm data to train the fire severity assessment model of the convolutional neural network combined with the long short-term memory network, predicts the newly collected data, determines whether there is a fire in the current environment, and assesses the severity of the fire; the fire spread speed calculation unit uses the multivariate regression algorithm to construct a fire spread model based on the fire warning data, the severity of the fire, the collected gas data and the three-dimensional model of the building, and generates the fire spread speed; the false alarm identification unit is responsible for identifying false alarms based on the output of the fire severity assessment model and adjusting the warning level; In the multi-channel gas analysis module, the data preprocessing unit obtains real-time gas composition data from the multi-channel gas sensors deployed in the fire warning location area. Each sensor detects different types of gases in real time and inputs them in the form of time series, which is expressed as ,in, arrive Indicates the concentration value of the first gas to the mth gas at time t, where m is the number of gas types; Kalman filtering is used to Smoothing is performed to reduce the impact of noise; a low-pass filter is used to eliminate high-frequency noise and retain the effective low-frequency signal part; Z-score standardization is used to normalize the data of different gases to the same scale; interpolation methods are used to synchronize the data of all sensors to a unified time axis; In the multi-channel gas analysis module, for the convolutional neural network part, the input layer receives real-time gas composition data collected from the multi-channel gas sensor, and the input data is a multi-dimensional matrix, each dimension represents a gas sensor, and each element in the matrix represents the concentration value of the corresponding gas at different time points; the convolution layer operates on the input data through a number of filters, each filter is a matrix with a smaller dimension than the input data; during the sliding process of the filter, the convolution operation is performed on the local area of ​​the input matrix, the weighted sum of the local data is calculated, and a new feature map is generated, and the convolution operation will extract the relationship between the gas concentrations; The output of the convolution layer is transformed nonlinearly through the activation function ReLU. The pooling layer is used to reduce the dimension of the convolution layer output and aggregate local features, thereby reducing the amount of data and enhancing the robustness of the model. The convolutional neural network contains a combination of several convolutional layers and pooling layers, extracting high-level features of the data layer by layer. The first several layers focus on capturing local changes in gas concentration, and the latter several layers are responsible for identifying how local changes combine to form more complex fire signals. It is connected to several fully connected layers to aggregate the extracted high-level features to generate a high-dimensional feature vector to reflect the overall trend and characteristics of the fire. The final output layer maps the high-dimensional feature vector into a fire risk score. In the multi-channel gas analysis module, for the long short-term memory network part, the data received by the input layer is the fire warning data processed by the data integration unit of the fire warning module; when the equipment state changes abnormally, the input gate writes the abnormal information into the memory unit, and if the smoke alarm at a certain moment returns to normal, the forget gate will reduce the memory strength of the previous alarm; if several sensors send abnormal signals at the same time, the output gate will integrate the abnormal signals and output a comprehensive fire warning signal; At each time step, the information in the memory unit will be updated according to the current input and the previous state, and the current output will be generated; the long short-term memory network will stack several layers, and the output of each layer will be used as the input of the next layer to extract higher-level time series features layer by layer; it will be aggregated through the fully connected layer to generate a fire warning indicator; the output layer will map the fire warning indicator to a fire warning level to indicate the severity of the fire; Get fire warning vector , Fire severity , Gas data and 3D models of buildings ,in, is the spatial coordinate; from The structural characteristics of the building, the layout of items inside the building, the heat conduction path inside the building and the spatial connectivity of the building are extracted and quantified. In the quantification of structural characteristics, the room area is scored according to the area, and the larger the area, the higher the score. The floor height is scored according to the height, and the higher the floor height, the higher the score. The wall thickness is scored according to the wall thickness and the fire resistance of the material. The number of doors and windows is recorded to affect the speed of fire spread. In the quantification of item layout, the score is set according to the room use category, and the storage of flammable items affects the score. In the quantification of heat conduction paths, the score is set according to the thermal conductivity of the material. The thermal conductive material affects the fire resistance. The higher the ventilation system score, the lower the smoke diffusion in the building. In the quantification of spatial connectivity, the more open areas, the lower the score, and effective partitions can improve the score. A weight is set for each feature to reflect its impact on the spread of fire. The score of each feature is multiplied by the corresponding weight, and the sum is taken to obtain the building's fire spread resistance index u. Using the regression model, assuming the fire spread rate is , the mathematical expression is as follows: ； in, is the intercept term of the model, They are the building's fire spread resistance index, fire warning vector, fire severity, and regression coefficient of gas data, indicating the impact of each feature on the fire spread speed. is the error term; The regression model is trained using historical data, with the actual fire spread rate as the target variable. The training data set contains and The value of is used to determine the corresponding regression coefficient using the least square method, and the performance of the model is verified using cross validation; in actual applications, the corresponding parameters are input to generate the spread rate of the fire; In the multi-channel gas analysis module, the fire severity assessment model outputs a fire severity and the probability of fire , will be used as the input of the false positive identification unit; among them, , where 1 means the fire is extremely serious and 0 means there is no fire. , where 1 indicates a fire has occurred and 0 indicates no fire; set the fire severity threshold and fire probability threshold ,if and , it is considered that there may be a false alarm; the false alarm identification unit uses other sensor data for weighted fusion and calculates a false alarm identification comprehensive index K to assist false alarm identification. If K is less than the false alarm identification comprehensive index threshold ,and and , it is considered that there is a false alarm and the warning level is adjusted; Early warning processing module: including: alarm trigger unit, fire equipment linkage unit, evacuation guidance unit and evacuation monitoring unit; the alarm trigger unit triggers the corresponding alarm mechanism according to the confirmed early warning level. If the early warning level reaches the intermediate or advanced level, the fire equipment linkage unit will automatically link the fire equipment, trigger the automatic fire extinguishing system in the corresponding area according to the fire warning location and the severity of the fire, and adjust the ventilation and smoke exhaust system in the building. In the case of advanced early warning, the evacuation guidance unit generates evacuation instructions and guides personnel evacuation through the public broadcasting system in the building. The evacuation monitoring unit tracks the evacuation progress in real time through monitoring cameras and sensors.

2. The fire early warning system for fire fighting facilities based on multi-channel gas analysis according to claim 1 is characterized by: In the fire equipment management module, the initial position of each fire equipment is marked on the floor plan, and a unique equipment ID is used for marking. The equipment position is determined by on-site measurement to ensure that the marking of each equipment is consistent with the corresponding actual physical location; according to the layout in the floor plan, the interior of the building is divided into areas, each area corresponds to a certain number of fire equipment, and the fire equipment is associated with the area information; Map the equipment annotations in the plan view to the three-dimensional model, and determine the spatial position of the equipment by combining the two-dimensional coordinates of the plan view with the Z-axis coordinates of the three-dimensional model. For equipment that requires a special installation location, indicate the specific installation height and angle in the three-dimensional model. Ensure that the equipment position data in the plan view is consistent with that in the three-dimensional model. If the equipment position is updated in the plan view, it will be updated synchronously in the three-dimensional model. If the equipment position is updated in the three-dimensional model, it will be updated synchronously in the plan view. Regularly calibrate the position of equipment inside the building, use the positioning error correction algorithm Kalman filtering to optimize the accuracy of the collected data, and reduce the deviation between the actual position of the equipment and the position in the plan view and three-dimensional model.

3. The fire early warning system for fire fighting facilities based on multi-channel gas analysis according to claim 2 is characterized by: In the fire equipment management module, relevant features are extracted from the collected fire equipment information and used as input of the decision tree algorithm to classify the equipment status. Each equipment status corresponds to a feature combination, which is used as the output of the decision tree. The collected device status and historical data are used as training sets and input into the decision tree algorithm. The decision tree recursively selects the features that best distinguish the device status and gradually builds a tree structure. Each branch node represents the selection of a feature, and the leaf node represents the final prediction result of the device status. The decision tree classifies the device according to its offline rate. If the offline rate exceeds a certain threshold, the device will be marked as "offline". If it does not exceed the limit, the equipment is classified according to the operation rate to determine whether it is in the "normal" state or the "fault warning" state; The system continuously monitors and updates the status information of each fire-fighting device. New data will be input into the trained decision tree model in real time to analyze the current status of the device in real time. The decision tree makes decisions along the tree structure from the root node to the leaf node based on the input feature value, and finally outputs the device status prediction result.

4. The fire early warning system for fire fighting facilities based on multi-channel gas analysis according to claim 1 is characterized by: In the fire warning module, the fire warning data is obtained through the data integration unit to form a fire warning vector ,in, arrive It represents the value of the first category fire warning data to the nth category fire warning data at time t, where n is the number of data categories in the fire warning data; The long short-term memory network model controls the flow of fire warning data through input gate, forget gate and output gate. The formula is as follows: ； in, is the weight matrix of the forget gate, is the bias term of the forget gate, is the activation function, which is the sigmoid function. is the hidden state at the previous moment; Input Gate and candidate memory cell states in the long short-term memory network model The formula is as follows: ； ； in, and are the weight matrices of the input gate and candidate memory cell states, and are the bias terms of the input gate and candidate memory cell states respectively; Update the memory cell state, the formula is as follows: ； in, is the current state of the memory cell, is the state of the memory unit at the previous moment; Output Gate and hidden state The formula is as follows: ； ； in, is the weight matrix of the output gate, is the bias of the output gate; The long short-term memory network model is trained using historical fire data and false alarm data. The goal is to input the fire warning vector Predict future fire risk levels and define the loss function L: ； in, is the fire warning level predicted by the model, is the actual warning level, T is the total number of time steps; the loss function is minimized using back propagation and gradient descent algorithms to adjust the parameters of the long short-term memory network model; The trained LSTM network model can predict the fire warning level based on the real-time input data stream in practical applications. The output of the model It is a real value representing the risk score of fire. The threshold is set according to the score to divide the warning levels into three levels: low, medium and high.

5. The fire early warning system for fire fighting facilities based on multi-channel gas analysis according to claim 4 is characterized by: In the fire warning module, after generating the fire warning level, the warning signal is located at a specific location in combination with the three-dimensional model of the building; the building is divided into several monitoring areas, each of which has corresponding sensor data. , where j represents the region, and the long short-term memory network model will generate an early warning level for each region; In the three-dimensional model of the building, the generated warning level is mapped to the specific location in the three-dimensional space based on the physical location and corresponding area of ​​the sensor; the location and risk level of the fire warning signal can be viewed through the visualization system of the three-dimensional model.

6. The fire early warning system for fire fighting facilities based on multi-channel gas analysis according to claim 1 is characterized by: In the warning processing module, the system monitors the warning level. When the level reaches the preset intermediate or high level, the alarm is triggered. According to the warning level and the location of the fire, the personnel in the key areas are notified first. The fire equipment linkage unit will start the automatic fire extinguishing system and dynamically adjust the ventilation and smoke exhaust system in the building according to the severity and location of the fire. Generate the best evacuation routes and instructions based on the location, severity and building structure of the fire, communicate evacuation instructions through the public broadcasting system in the building, and activate emergency lights and escape signs in the building in coordination with evacuation instructions; When generating the best evacuation route and instruction process, mark the fire area and impact range, identify and avoid obstacles in the building, apply the Dijkstra algorithm to calculate the shortest path from each room to the nearest safe exit, and design several backup evacuation routes to deal with the situation where the main route is blocked; generate an evacuation route map, including path guidance from each area to the safe exit, adjust instructions in real time and update the evacuation route according to the progress of the fire and the progress of evacuation.

Images