Multi-sensor fire monitoring and adaptive control method and system for closed engine room

Abstract

The present invention discloses a method and system for multi-sensor fire monitoring and adaptive control in a closed cabin, specifically relating to the field of fire monitoring and control, including collecting and extracting first disaster characteristics through multi-modal sensors, and then fusing sensor confidence and using a long-short-term memory network combined with a topological structure to predict fire characteristics, and finally generating a fire extinguishing instruction matrix and optimizing task allocation. The method and system for multi-sensor fire monitoring and adaptive control in a closed cabin provides a data basis for dynamically adapting to multi-modal feature changes in different fire source development stages by deploying a multi-modal sensor array and adopting a hardware timestamp protocol; quantifies the fire development stage through a disaster stage identification model, so that fire decisions can be dynamically adjusted according to a real-time stage index; and realizes adaptive prediction of the fire spread path through a long-short-term memory network algorithm combined with cabin topology data, breaking through the limitations of traditional predefined mapping tables.

Description

Technical Field

[0001] The present invention relates to the technical field of fire monitoring and control, and more particularly to a method and system for multi-sensor fire monitoring and adaptive control in a closed cabin. Background Art

[0002] With the improvement of safety standards for enclosed cabins, it is particularly important to achieve accurate monitoring of fires in the early stages. Traditional technology uses independently deployed smoke or heat detectors and a central controller to form a star architecture, and realizes fire judgment through the threshold comparison principle. However, actual measurements show that the false alarm rate of a single sensor is extremely high in the complex electromagnetic and temperature-changing environment of the cabin, and the differences in characteristic parameters of different fire source types lead to a surge in the missed alarm rate.

[0003] To reduce the false alarm rate, dual-modal sensors for temperature and smoke are deployed in key areas such as the cabin top, equipment compartment, and ventilation ducts, and edge computing nodes are added to perform sliding window filtering on the original signals to eliminate instantaneous interference noise.

[0004] However, it still has some shortcomings in actual use. For example, fire situation decisions cannot dynamically adapt to the changes in multimodal characteristics of different fire source development stages; the correspondence between fire extinguishing instructions and fire levels is still a predefined mapping table, and there is a lack of adaptive prediction capabilities for the fire spread path, resulting in insufficient accuracy in the release of fire extinguishing agents and causing secondary damage. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, the present invention provides a closed cabin multi-sensor fire monitoring and adaptive control method and system, which solves the problems raised in the above-mentioned background technology through the following scheme.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A multi-sensor fire monitoring and adaptive control method for a closed engine room includes:

[0008] S1: Collect the first disaster information by deploying a multimodal sensor array in the target cabin area;

[0009] S2: performing noise reduction processing on the first disaster information using an adaptive wavelet packet decomposition algorithm, and simultaneously extracting the first disaster characteristics of each mode;

[0010] S3: Obtain a disaster stage identification model, input the first disaster feature into the disaster stage identification model, and obtain a second disaster feature;

[0011] S4: Based on the second disaster characteristics, dynamically adjust the weight of each sensor in the multimodal sensor array, and fuse the confidence of the multimodal sensors through evidence theory;

[0012] S5: applying the fusion result of the multimodal sensor to a long short-term memory network algorithm and combining it with the topological structure of the target cabin to obtain a third disaster feature;

[0013] S6: Dynamically generate a fire extinguishing instruction matrix based on the second disaster characteristic and the third disaster characteristic;

[0014] S7: Based on the fire extinguishing instruction matrix, optimizing the allocation of fire extinguishing tasks by using the Hungarian algorithm;

[0015] S8: Collecting the data change rate of the multimodal sensor after the fire is extinguished in real time, and triggering the parameter update of the disaster stage recognition model and the retraining of the long short-term memory network.

[0016] Preferably, the first disaster information in S1 includes temperature per unit time, smoke concentration, CO concentration, light radiation pulse sequence, and auxiliary data;

[0017] The auxiliary data includes a spatiotemporal tag consisting of a timestamp and three-dimensional coordinates, a self-test tag of each sensor, and environmental parameters in the target cabin;

[0018] The environmental parameters in the target cabin include the risk factor of the deployment area, the volume of the target enclosed cabin, the effective coverage area of ​​a single sensor, and the mass and calorific value of combustibles.

[0019] Preferably, the step S1, calculating the deployment density of the multimodal sensor array based on the environmental parameters in the first disaster information, specifically includes:

[0020] ,

[0021] in, Expressed as the deployment density of the multimodal sensor array, Expressed as the risk factor of the deployment area, Expressed as the equipment fire load in the target engine room, Expressed as the target enclosed cabin volume, Expressed as the effective coverage area of ​​a single sensor; the equipment fire load The calculation is specifically expressed as:

[0022] ,

[0023] in, Indicates the equipment in the target cabin The mass of the combustible material, Expressed as Calorific value of combustible materials, Expressed as the floor area of ​​the area occupied by the equipment, Expressed as the total number of combustible types in the equipment in the target engine room, The index of the combustible type in the equipment of the target engine room.

[0024] Preferably, the S2 performs noise reduction processing by an adaptive wavelet packet decomposition algorithm, specifically including:

[0025] Based on the kurtosis coefficient of the fire source development stage corresponding to the first disaster information, the adaptive threshold method is used to dynamically adjust the adaptive threshold Denoising the pre-processed first disaster information; The calculation is specifically expressed as:

[0026] ,

[0027] in, Expressed as The standard deviation of the layer wavelet coefficients, It represents the signal length corresponding to the first disaster information, Expressed as the critical transition value of the fire source development stage, Expressed as the kurtosis coefficient, Expressed as the normal distribution value corresponding to the kurtosis coefficient.

[0028] Preferably, the disaster stage identification model in S3 is a temporal convolutional network, which outputs the second disaster feature as a quantitative index of the fire development stage in the interval [0, 1], and defines:

[0029] The second disaster characteristic ≤0.4 corresponds to the smoldering period;

[0030] 0.4<Second disaster characteristic≤0.7 corresponds to the open fire period;

[0031] The second disaster characteristic >0.7 corresponds to the explosion period.

[0032] Preferably, the rule for dynamically adjusting the weight of each sensor in the multimodal sensor array in S4 is:

[0033] When the second disaster characteristic is ≤0.4, corresponding to the smoldering period, the CO sensor is activated to dominate;

[0034] When 0.4<the second disaster characteristic≤0.7 corresponds to the open fire period, the activation temperature sensor dominates;

[0035] When the second disaster characteristic is greater than 0.7, which corresponds to the deflagration period, the optical radiation sensor is activated as an auxiliary. Except for the optical radiation sensor, the weights of the other modal sensors are reduced in equal proportion.

[0036] Preferably, the S5, injects physical constraints , calculate the probability of fire spread, which can be expressed as:

[0037] ,

[0038] in, Expressed as sigmoid activation function, Represented as an obstacle matrix, It is represented as the weight matrix composed of the proportion of each sensor data after S4 adjustment. Represented as the hidden state of LSTM at time t;

[0039] The obstacle matrix sets the obstacle position to 1 and the non-obstacle position to 0 based on the current prediction area.

[0040] Preferably, the S5 maps the predicted fire intensity coordinate points within the preset prediction time to the voxel space and calculates the risk value of each voxel. , specifically expressed as:

[0041] ,

[0042] in, Represents the preset forecast period, Represents the index of the moment in the preset prediction time period, Expressed as the first Whether the voxel is covered by fire, if covered, it is 1, otherwise it is 0, It is the predicted temperature at time t.

[0043] To achieve the above objectives, the present invention provides the following technical solution: a closed cabin multi-sensor fire monitoring and adaptive control system, which implements the above closed cabin multi-sensor fire monitoring and adaptive control method, comprising:

[0044] Multimodal sensor acquisition module: used to collect the first disaster information by deploying a multimodal sensor array in the target cabin area;

[0045] Dynamic feature extraction module: used to perform noise reduction processing on the first disaster information through an adaptive wavelet packet decomposition algorithm, and simultaneously extract the first disaster features of each mode;

[0046] Fire situation quantitative analysis module: used to obtain a disaster situation stage identification model, input the first disaster situation feature into the disaster situation stage identification model, and obtain a second disaster situation feature;

[0047] A dynamic decision module is configured to dynamically adjust the weight of each sensor in the multimodal sensor array based on the second disaster characteristics, and fuse the confidence of the multimodal sensors through evidence theory;

[0048] Fire spread prediction module: used to obtain the third disaster feature by applying the fusion result of the multimodal sensor to the long short-term memory network algorithm and combining it with the topological structure of the target cabin;

[0049] Hierarchical fire extinguishing strategy generation module: used to dynamically generate a fire extinguishing instruction matrix based on the second disaster characteristics and the third disaster characteristics;

[0050] Dynamic scheduling module: used to optimize the allocation of fire extinguishing tasks through the Hungarian algorithm based on the fire extinguishing instruction matrix;

[0051] Feedback learning module: used to collect the data change rate of the multimodal sensor after the fire is extinguished in real time, and trigger the parameter update of the disaster stage identification model and the retraining of the long short-term memory network.

[0052] Preferably, the enclosed cabin multi-sensor fire monitoring and adaptive control system further includes:

[0053] At least one central processing unit, communicatively connected to the multimodal sensor acquisition module and the dynamic scheduling module, for coordinating the execution of information text instructions output by each module and dynamically allocating computing power resources of edge computing nodes;

[0054] At least one historical database stores the three-dimensional point cloud topology data of the target cabin, fire spread trajectory samples and model training parameters, and interacts with the fire situation quantitative analysis module and the fire spread prediction module in a two-way manner;

[0055] The visualization interface renders the spatial heat map of the third disaster characteristic in real time and overlays the fire extinguishing device dispatch path.

[0056] Technical effects and advantages of the present invention:

[0057] 1. By deploying a multimodal sensor array and adopting a hardware timestamp protocol, this invention achieves millisecond-level spatiotemporal synchronous data acquisition. This provides a high-precision spatiotemporal alignment data foundation for dynamically adapting to multimodal feature changes at different fire development stages, and resolves the feature analysis bias problem caused by asynchrony of multimodal data.

[0058] 2. This invention builds a disaster stage recognition model based on a temporal convolutional network to quantify the fire development stage, enabling fire situation decisions to be dynamically adjusted based on the real-time stage index, thus solving the problem that fire situation decisions cannot dynamically adapt to changes in characteristics at different stages.

[0059] 3. The present invention uses a long short-term memory network algorithm combined with cabin topology data to achieve adaptive prediction of fire spread paths, breaking through the limitations of traditional predefined mapping tables, providing a spatial dimension prediction basis for the accurate generation of fire extinguishing instructions, and solving the problem of lack of fire spread prediction capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 This is a flowchart of the steps for implementing the method for multi-sensor fire monitoring and adaptive control in a closed cabin according to an embodiment of the present application.

[0061] Figure 2 This is a module block diagram of a multi-sensor fire monitoring and adaptive control system for a closed cabin provided according to an embodiment of the present application.

[0062] Figure 3 This is a collaborative workflow diagram of a multi-sensor fire monitoring and adaptive control system for a closed cabin provided according to an embodiment of the present application. DETAILED DESCRIPTION

[0063] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.

[0064] The terms used in the following examples of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in the specification of the present application, the singular expressions "a", "an", "said", "above", "the", and "this" are intended to include plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used in the present application refers to and includes any or all possible combinations of one or more of the listed items.

[0065] In the following, the terms "first," "second," and "third" are used for descriptive purposes only and should not be understood to imply or suggest relative importance or implicitly indicate the number of the technical features indicated. Therefore, a feature defined as "first," "second," and "third" may explicitly or implicitly include one or more of the features. In the description of the embodiments of this application, unless otherwise specified, "plurality" means two or more.

[0066] As attached Figure 1The multi-sensor fire monitoring and adaptive control method for a closed cabin shown in the figure uses multimodal sensors to collect and extract the first disaster characteristics, then fuses the sensor confidence and uses a long short-term memory network combined with a topological structure to predict the fire characteristics, ultimately generating a fire extinguishing command matrix and optimizing task allocation. The method specifically includes the following steps:

[0067] S1: Collect the first disaster information by deploying a multimodal sensor array in the target cabin area;

[0068] S2: performing noise reduction processing on the first disaster information using an adaptive wavelet packet decomposition algorithm, and simultaneously extracting the first disaster characteristics of each mode;

[0069] S3: Obtain a disaster stage identification model, input the first disaster feature into the disaster stage identification model, and obtain a second disaster feature;

[0070] S4: Based on the second disaster characteristics, dynamically adjust the weight of each sensor in the multimodal sensor array, and fuse the confidence of the multimodal sensors through evidence theory;

[0071] S5: applying the fusion result of the multimodal sensor to a long short-term memory network algorithm and combining it with the topological structure of the target cabin to obtain a third disaster feature;

[0072] S6: Dynamically generate a fire extinguishing instruction matrix based on the second disaster characteristic and the third disaster characteristic;

[0073] S7: Based on the fire extinguishing instruction matrix, optimizing the allocation of fire extinguishing tasks by using the Hungarian algorithm;

[0074] S8: Collecting the data change rate of the multimodal sensor after the fire is extinguished in real time, and triggering the parameter update of the disaster stage recognition model and the retraining of the long short-term memory network.

[0075] Specifically, in S1, after analyzing the internal structure of the cabin, potential fire source types, airflow characteristics and fire development laws in the target enclosed cabin area, a multimodal sensor array is deployed to collect information and parameters of different fire stages to represent the first disaster information. The first disaster information includes temperature, smoke concentration, CO concentration and light radiation pulse sequence per unit time, and auxiliary data, wherein the auxiliary metadata includes a spatiotemporal label composed of a timestamp and three-dimensional coordinates, a self-test label of each sensor, and environmental parameters in the target cabin; in this embodiment, the self-test label includes but is not limited to a normal label, a drift label, a fault label, etc.; the environmental parameters in the target cabin include but are not limited to the risk factor of the deployment area, the volume of the target enclosed cabin, the effective coverage area of ​​a single sensor, the mass and calorific value of combustible materials, etc.

[0076] Furthermore, the multimodal sensor array includes but is not limited to temperature sensors, smoke sensors, CO sensors, optical radiation sensors, etc.; the temperature sensor is configured with a wide dynamic range of 0-800°C and an accuracy of ±0.5°C; the smoke sensor monitors PM2.5 concentration and is highly sensitive to tiny smoke particles produced during the smoldering period; the CO sensor monitors CO concentration in the range of 0-2000ppm; and the optical radiation sensor monitors the light radiation intensity in the 400-700nm band.

[0077] In this embodiment, the deployment area of ​​the multimodal sensor array is selected based on the cabin top area, the equipment compartment area, and the ventilation duct area.

[0078] It should be noted that the top area of ​​the cabin is the core channel for rising hot air and smoke, and the multimodal sensor array is staggeredly deployed in the top area of ​​the cabin at a height of 0.3 to 0.5 meters below the ceiling; the equipment compartment area is a high-incidence area for fire sources such as oil leakage and electrical short circuit, and the multimodal sensor array is arranged in a ring with a radius of 1 meter around the potential fire source in the equipment compartment area; the ventilation duct area is the core influencing factor of the fire spread path, and the multimodal sensor array is deployed near the air inlet and outlet at intervals of 0.5 meters, and special attention is paid to monitoring abnormal disturbances with air flow speeds greater than 2 meters / second.

[0079] In a possible implementation, the deployment density of the multimodal sensor array is specifically expressed as:

[0080] ,

[0081] in, Expressed as the deployment density of the multimodal sensor array, Expressed as the risk factor of the deployment area, Expressed as the equipment fire load in the target engine room, Expressed as the target enclosed cabin volume, Expressed as the effective coverage area of ​​a single sensor; in this embodiment, the risk factor of the cabin top area is set to 1.2, the risk factor of the equipment compartment area is set to 1.5, and the risk factor of the ventilation duct area is set to 1.8; the equipment fire load The calculation is specifically expressed as:

[0082] ,

[0083] in, Indicates the equipment in the target cabin The mass of the combustible material, Expressed as Calorific value of combustible materials, Expressed as the floor area of ​​the area occupied by the equipment, Expressed as the total number of combustible types in the equipment in the target engine room, The index of the combustible type in the equipment of the target engine room.

[0084] Furthermore, to ensure that the first disaster information acquired by the multimodal sensor array has high-precision spatiotemporal synchronization, a hardware timestamp protocol is adopted to achieve millisecond-level spatiotemporal synchronization data acquisition; this embodiment selects IEEE 1588 PTPv2 as the hardware timestamp protocol.

[0085] It should be noted that the implementation steps of the spatiotemporal synchronization are as follows: when the data packet passes through the MAC layer, the PHY chip built into each sensor adds a 64-bit hardware timestamp to each frame of data, in a format that includes seconds and nanosecond offsets; following the PTPv2 protocol standard, the offset between the data obtained by each sensor and the master clock is calculated; and the precise hardware timestamp and the spatial coordinate information of the sensor are attached to the header of each frame of sensor data.

[0086] Specifically, in S2, the first disaster information collected by the multimodal sensor array is subjected to refined preprocessing, noise reduction and adaptive feature extraction to obtain a first disaster feature that can accurately characterize the fire status. The first disaster feature includes time domain features and frequency domain features, wherein the time domain features include the signal slope mean and the peak mutation rate, and the frequency domain features include the energy spectrum entropy and the main frequency offset after adaptive wavelet packet decomposition.

[0087] It should be noted that the preprocessing includes but is not limited to cleaning and normalization processing. The cleaning rules include but are not limited to detecting physical range violations and marking them as invalid data, detecting drastic mutations in the signal, and applying a median filter for smoothing. The normalization method uses Z-score standardization based on historical data.

[0088] It should be noted that the noise reduction process uses an adaptive wavelet packet decomposition algorithm to match the characteristics of different fire source development stages; based on the kurtosis coefficient of the fire source development stage corresponding to the first disaster information, an adaptive threshold method is used to dynamically adjust the adaptive threshold. Denoising the pre-processed first disaster information; The calculation is specifically expressed as:

[0089] ,

[0090] in, Expressed as The standard deviation of the layer wavelet coefficients, It represents the signal length corresponding to the first disaster information, Expressed as the critical transition value of the fire source development stage, Expressed as the kurtosis coefficient, It is expressed as the normal distribution value corresponding to the kurtosis coefficient. .

[0091] In this embodiment, the adaptive wavelet packet decomposition algorithm is based on the kurtosis coefficient of the input signal; in this embodiment, when the kurtosis coefficient is greater than 3, it indicates that there are spike pulses in the input signal, and the db4 wavelet basis function is selected and decomposed into 4 layers; when 1≤kurtosis coefficient≤3, it indicates that the input signal has a gradual trend, and the sym5 wavelet basis function is selected and decomposed into 6 layers; when the kurtosis coefficient is less than 1, it indicates that the input signal is relatively flat, and the coif3 wavelet basis function is selected and decomposed into 3 layers.

[0092] It should be noted that the adaptive feature extraction synchronously extracts time domain features and frequency domain features from each modal signal after noise reduction to form the first disaster feature.

[0093] Furthermore, the extraction of the time domain features includes: using a dynamic window mechanism to calculate the signal slope within the window , specifically expressed as:

[0094] ,

[0095] in, It is expressed as the number of data points contained in the time period selected by the dynamic window mechanism. Represents the time index of the sampling point, Represents the value collected by the sensor at time index t, Expressed as the average level over the time period, Expressed as the average value of the time index in the window, and the slope mean after the window slides M times is calculated , specifically expressed as:

[0096] ,

[0097] in, Expressed as the number of times the window slides, Represents the starting position of the i+1th window, Represents the starting position index when the window slides. Represented as the starting position of the i-th window; the frequency change rate of the peak occurrence per unit time is calculated through the peak value with a prominence greater than 3 times the standard deviation in each signal in the first disaster information to represent the peak mutation rate.

[0098] Furthermore, the extraction of frequency domain features includes: performing FFT transformation on the noise reduction signal to obtain the spectrum, calculating the energy of different frequency bands , construct the energy probability distribution and calculate the energy spectrum entropy, which is specifically expressed as:

[0099] ,

[0100] in, Expressed as the energy of the bi-th frequency band, is represented as the index of the frequency band, Expressed as The end frequency of the band, Expressed as a frequency variable, Expressed as The starting frequency of the frequency band, Expressed as a signal at frequency The complex amplitude value at the location; locate the main frequency of each signal in the first disaster information, and calculate the average main frequency offset per unit time.

[0101] Specifically, in S3, a disaster stage identification model is constructed based on a temporal convolutional network. The disaster stage identification model receives the first disaster feature after adaptive wavelet packet decomposition and feature extraction, and outputs a quantitative index of the fire development stage. The quantitative index is the second disaster feature.

[0102] In this embodiment, the disaster stage identification model selects TCN as the core model, and the input channel is defined as 16, corresponding to 4 features of each of the 4 modal sensors output by S2; 3 convolutional layers are used, and the expansion rates are defined as 1, 2, and 4, respectively, which cover a 25-second time window through combination; the convolution kernel size is defined as 3; the convolution output channel is defined as 256; a residual connection is set from the input layer to the last convolution output layer; the 256-dimensional features are mapped to a 1-dimensional output through a fully connected layer, and the output is constrained to the [0,1] interval through the Sigmoid activation function.

[0103] In one possible implementation, the disaster stage identification model outputs a quantitative index of the fire development stage with the second disaster characteristic in the interval [0,1], and defines: the second disaster characteristic ≤0.4 corresponds to the smoldering period; 0.4<second disaster characteristic ≤0.7 corresponds to the open flame period; the second disaster characteristic >0.7 corresponds to the deflagration period.

[0104] Specifically, in S4, based on the quantitative index of the continuous fire development stage in the interval [0,1] of S3, the weight of each sensor in the multimodal sensor array is dynamically adjusted, and the confidence of the multimodal sensor is fused through evidence theory.

[0105] In one possible implementation, dynamically adjusting the weights of the sensors in the multimodal sensor array includes: when the second disaster characteristic is in the range of [0, 0.4], determining it to be in the smoldering period, activating the CO sensor as the primary; when it is in the range of (0.4, 0.7], determining it to be in the open flame period, activating the temperature sensor as the primary; and when it is in the range of (0.7, 1.0], determining it to be in the deflagration period, activating the light radiation sensor as the auxiliary.

[0106] It should be noted that the weight adjustment of each sensor in the multimodal sensor array is specifically expressed as follows:

[0107] ,

[0108] in, It is expressed as the weight of the data proportion corresponding to the ni-th sensor, It is expressed as the initial weight of the data proportion corresponding to the CO sensor, It is expressed as the safety threshold of the adjustable weight of the data proportion corresponding to the CO sensor, It is expressed as the initial weight of the data proportion corresponding to the temperature sensor, It is expressed as the safety threshold of the adjustable weight of the data proportion corresponding to the temperature sensor, It represents the dividing point between the smoldering period and the open flame period. It is expressed as the base number of the time span of the open fire period. It is represented as the second disaster characteristic. Indicates the remaining sensors except CO sensor and temperature sensor, Expressed as the total number of sensor types.

[0109] Furthermore, the confidence of the multimodal sensor is obtained by constructing a BPA rule. For each sensor, based on its monitoring parameters and values ​​and combined with the dynamic weight at the current stage, a basic probability distribution for different fire stages is generated.

[0110] Furthermore, the confidence of the multimodal sensor is fused using the Dempster synthesis rule in evidence theory, combining the BPAs from different sensors, specifically expressed as:

[0111] ,

[0112] Among them, L represents the conflict factor, which is , and It is expressed as the basic probability distribution function from two different sensors, that is, the confidence of the sensor for different fire stages, Expressed as the target set, that is, the fire stage set of joint probability, and It is represented as a set of possible fire stages considered by two different sensors.

[0113] Specifically, in S5, the fusion result of the confidence of the multimodal sensor is input into the pre-trained long short-term memory network, and combined with the three-dimensional point cloud topology data of the target cabin, the fire coverage heat map and the risk level of key equipment within the preset prediction time. The fire coverage heat map and the risk level of key equipment within the preset prediction time are the third disaster feature, and the output dimension of the third disaster feature includes three-dimensional coordinate information and the corresponding risk value.

[0114] In this embodiment, the long short-term memory network includes two LSTM layers and one fully connected layer, the input dimension is defined as 20, and the output dimension is defined as 3; and the input of each LSTM unit includes the fusion features and corresponding spatial coordinates at the current moment.

[0115] Furthermore, the acquisition of the third disaster feature is completed through the spatiotemporal state transfer of the LSTM network. At each time step, the LSTM network receives the fusion features and corresponding spatial coordinates of the current moment, and injects physical constraints into the prediction process. , calculate the probability of fire spread, which can be expressed as:

[0116] ,

[0117] in, Expressed as sigmoid activation function, Represented as an obstacle matrix, It is represented as the weight matrix composed of the proportion of each sensor data after S4 adjustment. It is represented as the hidden state of LSTM at time t; the obstacle matrix sets the obstacle position to 1 and the non-obstacle position to 0 based on the current prediction area.

[0118] Furthermore, the predicted fire coordinate points within the preset prediction time are mapped to the voxel space, and the risk value of each voxel is calculated. , specifically expressed as:

[0119] ,

[0120] in, Represents the preset forecast period, Represents the index of the moment in the preset prediction time period, Expressed as the first Whether the voxel is covered by fire, if covered, it is 1, otherwise it is 0, It represents the predicted temperature at time t; in this embodiment, the fire coverage heat map adopts visual coding: the risk value in the range of [0, 0.3) is displayed in blue, indicating low risk; the risk value in [0.3, 0.6) is yellow, indicating medium risk; the risk value in [0.6, 0.8) is orange, indicating high risk; the risk value in [0.8, 1.0] is red, indicating urgent risk.

[0121] In this embodiment, the risk assessment indicators include equipment value, fire approach speed, fire resistance level and associated risks, and a risk level list of key equipment is calculated using a weighted formula.

[0122] Specifically, in S6, a fire extinguishing instruction matrix is ​​dynamically generated based on the fire coverage heat map in the third disaster characteristic and the quantitative index of the fire development stage in the second disaster characteristic.

[0123] In this embodiment, the fire extinguishing instruction matrix uses a structured table format to store instruction ID, instruction action object, location information, fire extinguishing execution action, specific conditions for action execution, and priority.

[0124] In this embodiment, if the second disaster characteristic is ≤0.4, it is in the smoldering stage, and the local ventilation and smoke exhaust and ultra-fine dry powder fixed-point spraying device are started. The target coordinates are the risk values ​​in the thermal map. >0.6; if 0.4<the second disaster characteristic≤0.7, it is in the open fire period, triggering water mist curtain isolation and flooding spraying in key areas. The key areas are the risk values ​​in the thermal map. >0.8 and adjacent to the coordinates of key equipment; if the second disaster characteristic is >0.7, it is in the deflagration period, execute full cabin inert gas injection, and cut off the risk value in the thermal map >Power supply of the device at coordinate 0.9.

[0125] Specifically, in S7, based on the fire extinguishing instruction matrix, image factors including spatial distance, priority and waiting time are introduced to optimize the allocation of fire extinguishing tasks through the Hungarian algorithm.

[0126] In this embodiment, an optimization goal is set, and the optimization goal is defined as minimizing the total cost.

[0127] Specifically, in S8, the multimodal sensor array is used to collect fire data of the target cabin area after fire extinguishing in real time. The fire data includes temperature, smoke concentration, CO concentration and light radiation pulse sequence per unit time, and auxiliary data, wherein the auxiliary metadata includes a spatiotemporal tag consisting of a timestamp and three-dimensional coordinates, a self-test tag of each sensor, and environmental parameters in the target cabin. In this embodiment, the self-test tag includes but is not limited to a normal tag, a drift tag, a fault tag, etc.; the environmental parameters in the target cabin include but are not limited to a risk factor of the deployment area, an effective coverage area of ​​a single sensor, the mass and calorific value of the remaining combustible material, etc.

[0128] In this embodiment, the fire control efficiency coefficient is obtained by calculating the weighted ratio of the measured and expected change rates of the three indicators of temperature, smoke, and CO. In this embodiment, when the fire control efficiency coefficient is less than 0.8, the disaster stage identification model parameter update of S3 and the retraining of the LSTM network of S5 are triggered.

[0129] As attached Figure 2 The multi-sensor fire monitoring and adaptive control system for a closed cabin shown in the figure includes a multi-modal sensor acquisition module, a dynamic feature extraction module, a fire situation quantitative analysis module, a dynamic decision module, a fire spread prediction module, a graded fire extinguishing strategy generation module, a dynamic scheduling module, and a feedback learning module.

[0130] Multimodal sensor acquisition module: used to collect the first disaster information by deploying a multimodal sensor array in the target cabin area;

[0131] Dynamic feature extraction module: used to perform noise reduction processing on the first disaster information through an adaptive wavelet packet decomposition algorithm, and simultaneously extract the first disaster features of each mode;

[0132] Fire situation quantitative analysis module: used to obtain a disaster situation stage identification model, input the first disaster situation feature into the disaster situation stage identification model, and obtain a second disaster situation feature;

[0133] A dynamic decision module is configured to dynamically adjust the weight of each sensor in the multimodal sensor array based on the second disaster characteristics, and fuse the confidence of the multimodal sensors through evidence theory;

[0134] Fire spread prediction module: used to obtain the third disaster feature by applying the fusion result of the multimodal sensor to the long short-term memory network algorithm and combining it with the topological structure of the target cabin;

[0135] Hierarchical fire extinguishing strategy generation module: used to dynamically generate a fire extinguishing instruction matrix based on the second disaster characteristics and the third disaster characteristics;

[0136] Dynamic scheduling module: used for optimizing the task allocation of fire extinguishing devices through the Hungarian algorithm based on the fire extinguishing instruction matrix;

[0137] Feedback learning module: used to collect the data change rate of the multimodal sensor after the fire is extinguished in real time, and trigger the parameter update of the disaster stage identification model and the retraining of the long short-term memory network.

[0138] In one possible implementation, as shown in the attached Figure 3 The enclosed cabin multi-sensor fire monitoring and adaptive control system shown also includes at least one central processor, at least one historical database and a visualization interface.

[0139] Among them, the central processing unit is the core control unit of the entire enclosed cabin multi-sensor fire monitoring and adaptive control system; it can be communicatively connected with the multimodal sensor acquisition module and the dynamic scheduling module, and is used to coordinate the execution of information text instructions output by each module, receive real-time data from the edge computing node and the temporal convolutional network stage recognition model, and dynamically allocate the computing power resources of the edge computing node; the central processing unit is used to perform collaborative control functions. When it is detected that the second disaster feature has a significant mutation in a short period of time, the central processing unit can interrupt the conventional data fusion process and directly trigger the generation of fire extinguishing strategies to achieve a rapid response to the rapid deterioration of the fire situation; the central processing unit continuously monitors the CPU utilization of the edge computing node and migrates the computationally intensive fire spread prediction task to the backup computing resources when the load is too high, ensuring that the response delay of the key prediction task meets the requirements; the central processing unit is also used to trigger and manage the execution of core algorithm modules such as data fusion, fire prediction, instruction generation and task optimization, and interact with the historical database for data to support model updating and retraining.

[0140] Among them, the historical database is used to store and manage historical data related to fire monitoring and control, including the three-dimensional point cloud topological data of the target cabin, fire spread trajectory samples and model training parameters; the fire spread trajectory samples are the fire development process data occurring under different cabin topological structures, including but not limited to the timestamp of the fire source, the spatial coordinate set, the temperature gradient matrix, the diffusion speed, and the related environmental parameters and sensor data snapshots; the historical database provides the long-term and short-term memory network with model training parameters with physical space labels, and supports the workflow of the model incremental training engine. When the fire extinguishing efficiency is not ideal, the model incremental training engine will retrieve historical fire samples similar to the current cabin topological structure from the historical database, and fuse the actual efficiency data of the fire extinguishing action with the retrieved sample data to generate an enhanced training set; the enhanced training set is used to retrain the disaster stage recognition model and the long-term and short-term memory network, and has two-way data interaction with the fire quantitative analysis module and the fire spread prediction module.

[0141] Among them, the visualization interface is used to render the spatial heat map of the third disaster characteristic in real time, and superimpose the fire-fighting task scheduling path; in this embodiment, different visualization schemes are adopted according to the risk value corresponding to the third disaster characteristic: for high-risk areas, it is displayed with red contour lines superimposed on pulse animation, and supports the operator to click on the coordinates to activate the equipment emergency stop; for medium-risk areas, it is displayed in orange thermal shading, and allows the operator to adjust the coverage of the fire-fighting device by dragging; for low-risk areas, it is displayed with a translucent white outline; in terms of dynamic optimization of scheduling paths, the allocation of fire-fighting tasks is superimposed. When the operator manually adjusts the preset path or target area of ​​the fire-fighting task through the interface, the visualization interface will feed back the manual adjustment information to the central processing unit, and the central processing unit will trigger the Hungarian algorithm to recalculate the cost matrix and perform task optimization.

[0142] Secondly: The drawings of the embodiments disclosed in the present invention only involve structures related to the embodiments disclosed in the present invention. Other structures may refer to conventional designs. The same embodiment and different embodiments of the present invention may be combined with each other without conflict.

[0143] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit 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 multi-sensor fire monitoring and adaptive control method for a closed cabin, characterized by: include: S1: Collect the first disaster information by deploying a multimodal sensor array in the target cabin area; The first disaster information includes temperature per unit time, smoke concentration, CO concentration, and optical radiation pulse sequence, and auxiliary data; the auxiliary data includes a spatiotemporal tag consisting of a timestamp and three-dimensional coordinates, a self-test tag for each sensor, and environmental parameters within the target cabin; the environmental parameters within the target cabin include the risk factor of the deployment area, the volume of the target enclosed cabin, the effective coverage area of ​​a single sensor, and the mass and calorific value of combustible materials; S2: performing noise reduction processing on the first disaster information using an adaptive wavelet packet decomposition algorithm, and simultaneously extracting the first disaster characteristics of each mode; S3: Obtain a disaster stage identification model, input the first disaster feature into the disaster stage identification model, and obtain a second disaster feature; The disaster stage recognition model is a temporal convolutional network, which outputs the second disaster feature as a quantitative index of the fire development stage in the interval [0,1], and defines: the second disaster feature ≤ 0.4 corresponds to the smoldering stage; 0.4 < the second disaster feature ≤ 0.7 corresponds to the open flame stage; the second disaster feature > 0.7 corresponds to the deflagration stage; S4: Based on the second disaster characteristics, dynamically adjust the weight of each sensor in the multimodal sensor array, and fuse the confidence of the multimodal sensors through evidence theory; S5: Applying the fusion result of the multimodal sensor to a long short-term memory network algorithm and combining it with the topological structure of the target cabin to obtain a third disaster feature; wherein the fire coverage heat map and the risk level of key equipment within a preset prediction time are the third disaster feature, and the output dimension of the third disaster feature includes three-dimensional coordinate information and a corresponding risk value; S6: Dynamically generate a fire extinguishing instruction matrix based on the second disaster characteristic and the third disaster characteristic; S7: Based on the fire extinguishing instruction matrix, optimizing the allocation of fire extinguishing tasks by using the Hungarian algorithm; S8: Collecting the data change rate of the multimodal sensor after the fire is extinguished in real time, and triggering the parameter update of the disaster stage recognition model and the retraining of the long short-term memory network.

2. The closed cabin multi-sensor fire monitoring and adaptive control method according to claim 1, characterized in that: The step S1, calculating the deployment density of the multimodal sensor array based on the environmental parameters in the first disaster information, specifically includes: ， in, Expressed as the deployment density of the multimodal sensor array, Expressed as the risk factor of the deployment area, Expressed as the equipment fire load in the target engine room, Expressed as the target enclosed cabin volume, Expressed as the effective coverage area of ​​a single sensor; the equipment fire load The calculation is specifically expressed as: ， in, Indicates the equipment in the target cabin The mass of the combustible material, Expressed as Calorific value of combustible materials, Expressed as the floor area of ​​the area occupied by the equipment, Expressed as the total number of combustible types in the equipment in the target engine room, The index of the combustible type in the equipment of the target engine room.

3. The closed cabin multi-sensor fire monitoring and adaptive control method according to claim 1, characterized in that: The S2 is to perform noise reduction processing by using an adaptive wavelet packet decomposition algorithm, specifically including: Based on the kurtosis coefficient of the fire source development stage corresponding to the first disaster information, the adaptive threshold method is used to dynamically adjust the adaptive threshold Denoising the pre-processed first disaster information; The calculation is specifically expressed as: ， in, Expressed as The standard deviation of the layer wavelet coefficients, It represents the signal length corresponding to the first disaster information, Expressed as the critical transition value of the fire source development stage, Expressed as the kurtosis coefficient, Expressed as the normal distribution value corresponding to the kurtosis coefficient.

4. The closed cabin multi-sensor fire monitoring and adaptive control method according to claim 1, characterized in that: The rule for dynamically adjusting the weight of each sensor in the multimodal sensor array in S4 is: When the second disaster characteristic is ≤0.4, corresponding to the smoldering period, the CO sensor is activated to dominate; When 0.4<the second disaster characteristic≤0.7 corresponds to the open fire period, the activation temperature sensor dominates; When the second disaster characteristic is greater than 0.7, which corresponds to the deflagration period, the optical radiation sensor is activated as an auxiliary. Except for the optical radiation sensor, the weights of the other modal sensors are reduced in equal proportion.

5. The closed cabin multi-sensor fire monitoring and adaptive control method according to claim 1, characterized in that: The S5, injection of physical constraints , calculate the probability of fire spread, which can be expressed as: ， in, Expressed as sigmoid activation function, Represented as an obstacle matrix, It is represented as the weight matrix composed of the proportion of each sensor data after S4 adjustment. Represented as the hidden state of LSTM at time t; The obstacle matrix sets the obstacle position to 1 and the non-obstacle position to 0 based on the current prediction area.

6. The closed cabin multi-sensor fire monitoring and adaptive control method according to claim 1, characterized in that: The S5 maps the predicted fire intensity coordinate points within the preset prediction time to the voxel space and calculates the risk value of each voxel , specifically expressed as: ， in, Represents the preset forecast period, Represents the index of the moment in the preset prediction time period, Expressed as the first Whether the voxel is covered by fire, if covered, it is 1, otherwise it is 0, It is the predicted temperature at time t.

7. A closed cabin multi-sensor fire monitoring and adaptive control system, according to the closed cabin multi-sensor fire monitoring and adaptive control method according to any one of claims 1 to 6, characterized in that: include: Multimodal sensor acquisition module: used to collect the first disaster information by deploying a multimodal sensor array in the target cabin area; Dynamic feature extraction module: used to perform noise reduction processing on the first disaster information through an adaptive wavelet packet decomposition algorithm, and simultaneously extract the first disaster features of each mode; Fire situation quantitative analysis module: used to obtain a disaster situation stage identification model, input the first disaster situation feature into the disaster situation stage identification model, and obtain a second disaster situation feature; A dynamic decision module is configured to dynamically adjust the weight of each sensor in the multimodal sensor array based on the second disaster characteristics, and fuse the confidence of the multimodal sensors through evidence theory; Fire spread prediction module: used to obtain the third disaster feature by applying the fusion result of the multimodal sensor to the long short-term memory network algorithm and combining it with the topological structure of the target cabin; Hierarchical fire extinguishing strategy generation module: used to dynamically generate a fire extinguishing instruction matrix based on the second disaster characteristics and the third disaster characteristics; Dynamic scheduling module: used to optimize the allocation of fire extinguishing tasks through the Hungarian algorithm based on the fire extinguishing instruction matrix; Feedback learning module: used to collect the data change rate of the multimodal sensor after the fire is extinguished in real time, and trigger the parameter update of the disaster stage identification model and the retraining of the long short-term memory network.

8. The closed cabin multi-sensor fire monitoring and adaptive control system according to claim 7, characterized in that: Also includes: At least one central processing unit, communicatively connected to the multimodal sensor acquisition module and the dynamic scheduling module, for coordinating the execution of information text instructions output by each module and dynamically allocating computing power resources of edge computing nodes; At least one historical database stores the three-dimensional point cloud topology data of the target cabin, fire spread trajectory samples and model training parameters, and interacts with the fire situation quantitative analysis module and the fire spread prediction module in a two-way manner; The visualization interface renders the spatial heat map of the third disaster characteristic in real time and overlays the fire extinguishing device dispatch path.

Images