Artificial intelligence backdraft prediction method and system based on multi-source heterogeneous data
Through the artificial intelligence reignitment prediction method of multi-source heterogeneous data, combined with high-definition cameras, sensors and Internet of Things technology, the deep learning model is used to perform real-time data analysis at the fire scene, solving the problems of low prediction accuracy and missing data in traditional methods, realizing the problem of More accurate and efficient return fire prediction.
Patent Information
- Application Number
- CN202411112329.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-14
- Publication Date
- 2024-11-26
AI Technical Summary
Traditional single-modal artificial intelligence fire recognition methods have low prediction accuracy in the fire field, and cannot effectively predict the return-ignition phenomenon when data is missing.
Using artificial intelligence reignitment prediction method based on multi-source heterogeneous data, multi-dimensional information is collected in real time through high-definition cameras, sensors and Internet of Things technology, data preprocessing is carried out, and multi-source heterogeneous databases and artificial intelligence reignitment prediction models are constructed, and a multi-source heterogeneous database and artificial intelligence reignitment prediction model is used. Deep learning models such as CNN, GANs and Transformer make predictions.
The accuracy and timeliness of the return fire prediction are improved, and the model has the ability to optimize itself and can continuously improve the prediction performance as data accumulates.
Smart Images

Figure CN119027781A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of intelligent environmental prediction, and in particular to an artificial intelligence flashback prediction method and system based on multi-source heterogeneous data. Background Art
[0002] With the development of industrialization and urbanization, fire and explosion accidents occur frequently. In a fire scene, backdraft is a common extreme fire phenomenon, which is highly sudden and destructive, and seriously threatens the safety of life and property of people in the fire scene. The methods of backdraft prediction include statistical prediction, numerical simulation prediction and single-modal artificial intelligence fire prediction. Statistical prediction estimates the time and location where backdraft may occur by analyzing and processing fire data. Numerical simulation prediction simulates and predicts the evolution process and state of fire by computer, and combines the characteristics and laws of backdraft for prediction. Single-modal artificial intelligence fire prediction identifies fire conditions through image data or sensor data.
[0003] Statistical predictions are often based on historical data and certain assumptions, which simplify the complex physical process of fire and may ignore some key details of fire dynamics. They are subject to the representativeness of the data set. If the historical data is not sufficient to cover all possible fire scenarios, the model may fail when faced with unseen situations. Numerical simulations require solving complex sets of physical equations and have high requirements for computing resources, especially when dealing with large-scale, high-resolution fire scenes. Its accuracy and reliability are highly dependent on the accuracy of initial and boundary conditions, which are often uncertain in practical applications. Traditional single-modal artificial intelligence fire recognition methods have low prediction accuracy. When data is missing due to the influence of the fire scene, its fire recognition ability is hindered and the occurrence of backdraft cannot be predicted. Therefore, in actual fire scenes, it is often necessary to predict the backdraft phenomenon based on the multi-faceted fire development of the specific scene.
[0004] In order to solve the above problems, the present invention proposes an artificial intelligence flashback prediction method and system based on multi-source heterogeneous data. Summary of the invention
[0005] The purpose of the present invention is to propose an artificial intelligence flashback prediction method and system based on multi-source heterogeneous data to solve the problems raised in the background technology:
[0006] The traditional single-modal artificial intelligence fire recognition method has low prediction accuracy. When the fire scene causes data loss, its fire recognition ability is hindered and it is impossible to predict the occurrence of backdraft.
[0007] In order to achieve the above object, the present invention adopts the following technical solutions:
[0008] The artificial intelligence flashback prediction method based on multi-source heterogeneous data includes the following steps:
[0009] S1: Real-time capture of multi-dimensional information of the fire scene through high-definition cameras, sensors and IoT technology;
[0010] S2: Use data processing technology to pre-process the acquired data stream;
[0011] S3: Build a multi-source heterogeneous database and integrate the pre-processed multi-source heterogeneous data;
[0012] S4: Build an artificial intelligence flashback prediction model based on multi-source heterogeneous data and conduct training;
[0013] S5: Use the pre-trained flashback prediction model to perform real-time analysis and calculation of the monitoring information of the fire scene to predict the probability and intensity of flashback;
[0014] S6: Update the data of multi-source heterogeneous databases based on actual fire scene data, and train and update the artificial intelligence flashback prediction model.
[0015] Preferably, the data preprocessing in S2 includes data cleaning and dimensionality unification processing, wherein the data cleaning includes missing value processing, outlier detection and processing operations; the dimensionality unification includes unit conversion, data standardization or normalization, and data encoding processing.
[0016] Preferably, the artificial intelligence flashback prediction model in S4 analyzes the feature information of the fire image data obtained by the high-definition camera by the CNN network, and enhances the image data based on the GANs model, and then uses the encoder and fusion self-attention mechanism in the Transformer to extract the sensor data features, and then inputs the acquired multimodal data features into the improved multi-layer Bi-LSTM network to capture the relationship between the multimodal data streams, and finally completes the flashback output prediction result through the fully connected layer.
[0017] Preferably, the CNN network extracts image features from temporal neighborhoods of different spans based on a number of convolution kernels of different sizes.
[0018] Preferably, the GANs model consists of 4 generator subnets with multiple resolutions, and each generator subnet is added with a spatial attention module, as follows:
[0019] Based on the self-attention operation, we obtain the query vector matrix Q1, key vector matrix K1 and value vector matrix V1 of the image features.
[0020]
[0021] The output of the adjacent space self-attention mechanism Atten(Q1,K1,V1) is as follows:
[0022]
[0023] in, and are all learning parameters; d1 is the dimension of the query vector matrix Q1, key vector matrix K1 and value vector matrix V1 in the adjacency space self-attention mechanism; T is the transpose of the vector;
[0024] The four generator subnetworks enhance the input image features in turn, and the learning parameters of the attention module in each generator subnetwork are obtained based on the improved particle swarm algorithm.
[0025] The improved particle swarm algorithm is specifically as follows:
[0026] v a+1,b =ω t v a,b +c1r1(p b -u a,b )+c2r2(g b -u a,b )
[0027] u a+1,b =u a,b +v a+1,b
[0028]
[0029] Among them, v a,b 、v a+1,b are the speeds of the ath and a+1th iterations of the bth parameter respectively; u a,b 、u a+1,b is the value of the ath and a+1th iterations of the bth parameter; ω t is the inertia weight; c1 and c2 are acceleration factors; r1 and r2 are random variables uniformly distributed between [0,1]; p b is the individual optimal value; g b is the global optimal value; τ is the maximum number of iterations; ω max ,ω min are the maximum and minimum values of the inertia weight respectively.
[0030] Preferably, the encoder is composed of Transformer layers, each of which includes a BN layer and a ReLU activation function, as well as a feedforward neural network part and a residual connection. The encoder also adds a multi-head fusion attention mechanism, on this basis, introduces a dynamic gating mechanism, and the dynamic gating weights are obtained based on the dynamic optimization of the improved particle swarm algorithm according to the characteristics and context of the input data.
[0031] Preferably, the fusion attention mechanism fuses three attention mechanisms and obtains the final attention output based on a weighted manner, as follows:
[0032] The first attention mechanism is to integrate the temporal self-attention mechanism, as follows:
[0033] The input feature matrix is formed based on the time step. For each time step, the query vector matrix Q2, key vector matrix K2 and value vector matrix V2 of the corresponding features are formed.
[0034]
[0035] The output of the temporal self-attention mechanism Atten(Q2,K2,V2) is as follows:
[0036]
[0037] in, and are all learning parameters; d2 is the dimension of the query vector matrix Q2, key vector matrix K2, and value vector matrix V2 in the temporal self-attention mechanism;
[0038] The second attention mechanism is to integrate the global self-attention mechanism, as follows:
[0039] Based on two MLPs, the input features are mapped to obtain the query vector matrix Q3, key vector matrix K3 and value vector matrix V3:
[0040]
[0041] The output of the global self-attention mechanism Atten(Q3,K3,V3) is as follows:
[0042] Atten(Q3,K3,V3)=Drop[(K3-Q3)·MLP(K3)]·V3
[0043] in, and are all learning parameters; Drop(·) is a random dropout operation;
[0044] The third attention mechanism is the local self-attention mechanism, which is as follows:
[0045] Based on two MLPs, the input features are mapped to obtain the query vector matrix Q4, key vector matrix K4 and value vector matrix V4:
[0046]
[0047]
[0048] For each position i, define the window range And extract the sub-vector matrix k corresponding to the key vector matrix K4 j , calculate the attention weight within the window range:
[0049]
[0050] The output of the local self-attention mechanism Atten(Q4,K4,V4) is as follows:
[0051] Atten(Q4,K4,V4)=w ij V4
[0052] Among them, w ij represents the local attention weight; and All are learning parameters;
[0053] The learning parameters and weighting parameters in the three attention mechanisms are obtained based on the improved particle swarm algorithm.
[0054] Preferably, the improved multi-layer Bi-LSTM network adds a Dropout layer after each hidden layer, and simultaneously obtains various network parameters based on an improved particle swarm algorithm.
[0055] The artificial intelligence flashback prediction system based on multi-source heterogeneous data includes:
[0056] Data acquisition module: used to capture multi-dimensional information of the fire scene in real time through high-definition cameras, sensors and Internet of Things technology;
[0057] Data preprocessing module: used to preprocess the acquired data stream using data processing technology;
[0058] Multi-source heterogeneous database module: used to build a multi-source heterogeneous database and integrate pre-processed multi-source heterogeneous data;
[0059] Artificial intelligence flashback prediction module: used to build and train an artificial intelligence flashback prediction model based on multi-source heterogeneous data;
[0060] Result analysis and visualization module: used to convert the prediction results into charts and reports, and to make corresponding risk visualization warnings based on the prediction results;
[0061] Self-update module: used to update the data of multi-source heterogeneous databases based on actual fire scene data, and to train and update the artificial intelligence flashback prediction model;
[0062] The data preprocessing module includes a data cleaning unit and a data dimension unification unit, which are used for data cleaning and data dimension unification respectively;
[0063] The artificial intelligence flashback prediction module includes:
[0064] Feature extraction unit: used to extract features from images and data information collected by high-definition cameras, sensors, and Internet of Things technologies;
[0065] Image enhancement unit: used to enhance the extracted image features based on the GANs model;
[0066] Multimodal data fusion prediction unit: used to input image features and data features into the improved multi-layer Bi-LSTM network to capture the relationship between multimodal data streams, and then complete the prediction results of the back-burning output through the fully connected layer.
[0067] Compared with the prior art, the present invention provides an artificial intelligence flashback prediction method and system based on multi-source heterogeneous data, which has the following beneficial effects:
[0068] The present invention realizes real-time data collection through cameras and sensor Internet of Things technologies, proposes an artificial intelligence flashback prediction model based on multi-source heterogeneous data, adopts a new network model architecture, and performs real-time calculation and analysis on the multi-source heterogeneous data collected in the fire scene. Through advanced data processing technology and deep learning models, the probability and intensity of flashback in actual fire scenes are predicted, thereby improving the accuracy and timeliness of flashback prediction. The model has self-optimization capabilities and can continuously improve the prediction performance as data accumulates. BRIEF DESCRIPTION OF THE DRAWINGS
[0069] Figure 1 This is a flow chart of the method mentioned in Example 1 of the present invention;
[0070] Figure 2 This is a schematic diagram of the enhancement result of GANs mentioned in Example 1 of the present invention;
[0071] Figure 3 It is a schematic diagram of the on-site temperature prediction result mentioned in Example 1 of the present invention;
[0072] Figure 4 This is a system block diagram mentioned in Embodiment 2 of the present invention;
[0073] Figure 5 This is the artificial intelligence flashback prediction flow chart based on multi-source heterogeneous data mentioned in Example 2 of the present invention.
[0074] Meaning of the symbols in the figure:
[0075] 100. Data acquisition module; 200. Data preprocessing module; 210. Data cleaning unit; 220. Data dimension unification unit; 300. Multi-source heterogeneous database module; 400. Artificial intelligence flashback prediction module; 410. Feature extraction unit; 420. Image enhancement unit; 430. Multimodal data fusion prediction unit; 500. Result analysis and visualization module; 600. Self-update module. DETAILED DESCRIPTION
[0076] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.
[0077] The present invention realizes real-time data collection through cameras and sensor Internet of Things technologies, proposes an artificial intelligence flashback prediction model based on multi-source heterogeneous data, adopts a new network model architecture, performs real-time calculation and analysis on multi-source heterogeneous data collected in the fire scene, and realizes the prediction of the probability and intensity of flashback in actual fire scenes through advanced data processing technology and deep learning models, thereby improving the accuracy and timeliness of flashback prediction; and the model has self-optimization capabilities, and can continuously improve the prediction performance as data accumulates. Specifically including the following contents.
[0078] Embodiment 1:
[0079] See also Figure 1-3 The artificial intelligence flashback prediction method based on multi-source heterogeneous data of the present invention comprises the following steps:
[0080] S1: Real-time capture of multi-dimensional information of the fire scene through high-definition cameras, sensors and IoT technology; details are as follows:
[0081] By integrating high-definition cameras, various sensors and Internet of Things technology, real-time capture of multi-dimensional information at the fire scene is achieved, including fire images, temperature, gas concentration, pressure, meteorological conditions, etc.; high-definition cameras are used to capture real-time image information at the fire scene. These images contain key information such as fire size, spread direction, smoke distribution, etc., providing raw data for subsequent image feature analysis. Various sensors are used to collect multi-dimensional environmental data at the fire scene, providing the location of the fire source and the characteristics of the fire development.
[0082] S2: Use data processing technology to pre-process the acquired data stream; the details are as follows:
[0083] Use data processing techniques to preprocess the acquired data stream, including data cleaning: 1) Remove duplicate data: Ensure that there are no duplicate records in the data set. 2) Missing value processing: Fill or delete data points with missing values. 3) Outlier detection and processing: Identify data points that deviate significantly from the average value of the data set. Outliers may be caused by measurement errors or extreme events and can be corrected or deleted. Dimension unification: 1) Unit conversion: Convert all data to a unified unit, such as converting temperature from Celsius to Fahrenheit. 2) Data standardization or normalization: Scale numerical features to the same range or standard deviation to eliminate dimensional effects and make model training more stable. 3) Data encoding: Convert non-numerical data (such as image data) into numerical vector form for easy algorithm processing.
[0084] S3: Build a multi-source heterogeneous database and integrate the pre-processed multi-source heterogeneous data; the details are as follows:
[0085] Record the source, type, format, update time, version and other information of the data, provide training data for the model, and update the collected data in real time.
[0086] S4: Build an artificial intelligence flashback prediction model based on multi-source heterogeneous data and train it; the details are as follows:
[0087] Construct an artificial intelligence flashback prediction model based on multi-source heterogeneous data and train it. The prediction model structure uses CNN (convolutional neural network) to analyze the feature information of fire image data obtained by high-definition cameras. The CNN network extracts image features from time series neighborhoods of different spans based on several convolution kernels of different sizes. Specifically, time series feature extraction units with convolution kernel sizes of 1, 3, 5, 7, and 9 can be used to extract image features. Among them, data dimension is increased based on the time series feature extraction unit with a convolution kernel size of 1, and the image features extracted by the convolution kernels of other sizes are spliced to obtain the feature information of the fire image data. The image data is then enhanced based on the GANs model, which consists of four generation subnetworks with multiple resolutions. A spatial attention module is added to each generator subnetwork, as follows:
[0088] Based on the self-attention operation, we obtain the query vector matrix Q1, key vector matrix K1 and value vector matrix V1 of the image features.
[0089]
[0090] The output of the adjacent space self-attention mechanism Atten(Q1,K1,V1) is as follows:
[0091]
[0092] in, and are all learning parameters; d1 is the dimension of the query vector matrix Q1, key vector matrix K1 and value vector matrix V1 in the adjacency space self-attention mechanism; T is the transpose of the vector;
[0093] The four generator subnetworks enhance the input image features in turn, and the learning parameters of the attention module in each generator subnetwork are obtained based on the improved particle swarm algorithm.
[0094] The improved particle swarm algorithm is as follows:
[0095] v a+1,b =ω t v a,b +c1r1(p b -u a,b )+c2r2(g b -u a,b )
[0096] u a+1,b =u a,b +v a+1,b
[0097]
[0098] Among them, v a,b 、v a+1,b are the speeds of the ath and a+1th iterations of the bth parameter respectively; u a,b 、u a+1,b is the value of the ath and a+1th iterations of the bth parameter; ω t is the inertia weight; c1 and c2 are acceleration factors; r1 and r2 are random variables uniformly distributed between [0,1]; p b is the individual optimal value; g b is the global optimal value; τ is the maximum number of iterations; ω max ,ω min are the maximum and minimum values of the inertia weight respectively;
[0099] Based on the fusion output of the four generative subnetworks, each feature on the feature map can be enhanced or suppressed accordingly. The images passing through multiple subnetworks will be weightedly modified to ensure that the effect is significantly improved.
[0100] Reference Figure 2 Specifically, six blurred flame images at different moments in the development of a fire can be selected. The model receives the blurred flame images and transmits them to the first generation subnet for detail reconstruction and resolution restoration. Each generation subnet can pass the enhanced performance to the next subnet, thereby performing deeper enhancement and finally obtaining a clear flame image.
[0101] Use the encoder in Transformer and the multi-head fusion attention mechanism to extract features and dependencies of sensor data such as temperature, gas concentration, pressure, and meteorological conditions.
[0102] The fusion attention mechanism combines the three attention mechanisms and obtains the final attention output based on a weighted approach, as follows:
[0103] The first attention mechanism is to integrate the temporal self-attention mechanism, as follows:
[0104] The input feature matrix is formed based on the time step. For each time step, the query vector matrix Q2, key vector matrix K2 and value vector matrix V2 of the corresponding features are formed.
[0105]
[0106] The output of the temporal self-attention mechanism Atten(Q2,K2,V2) is as follows:
[0107]
[0108] in, and are all learning parameters; d2 is the dimension of the query vector matrix Q2, key vector matrix K2, and value vector matrix V2 in the temporal self-attention mechanism;
[0109] The second attention mechanism is to integrate the global self-attention mechanism, as follows:
[0110] First, two MLPs are used to map the input features respectively to obtain the query matrix, key matrix and value matrix. Then, group aggregation is applied to divide the key matrix into different regions, and the aggregated key matrix is subtracted from the query matrix to extract the spatial relationship between the points in the query matrix and the points in the key matrix. Then, the global attention weight map is obtained by element-by-element multiplication with the key matrix after an MLP layer, and then the attention weight map is randomly deactivated to alleviate overfitting. Then, the value matrix is matrix-multiplied with the global attention weight map after random deactivation to finally obtain the global feature matrix.
[0111] Based on two MLPs, the input features are mapped to obtain the query vector matrix Q3, key vector matrix K3 and value vector matrix V3:
[0112]
[0113] The output of the global self-attention mechanism Atten(Q3,K3,V3) is as follows:
[0114] Atten(Q3,K3,V3)=Drop[(K3-Q3)·MLP(K3)]·V3
[0115] in, and are all learning parameters; Drop(·) is a random dropout operation;
[0116] The third attention mechanism is to integrate the local self-attention mechanism, as follows:
[0117] Based on two MLPs, the input features are mapped to obtain the query vector matrix Q4, key vector matrix K4 and value vector matrix V4:
[0118]
[0119] For each position i, define the window range And extract the sub-vector matrix k corresponding to the key vector matrix K4 j , calculate the attention weight within the window range:
[0120]
[0121] The output of the local self-attention mechanism Atten(Q4,K4,V4) is as follows:
[0122] Atten(Q4,K4,V4)=w ij V4
[0123] Among them, w ij represents the local attention weight; and All are learning parameters;
[0124] The learning parameters and weighting parameters in the three attention mechanisms are obtained based on the improved particle swarm algorithm.
[0125] The fusion encoding information of multimodal data captures the relationship between multimodal data streams through BI-LSTM (bidirectional long short-term memory network), and finally completes the prediction result of the backfire output through the fully connected layer.
[0126] S5: Use the pre-trained flashback prediction model to perform real-time analysis and calculation on the monitoring information of the fire scene to predict the probability and intensity of flashback; the details are as follows:
[0127] The temperature of the fire scene is predicted based on the model built by S4 and compared with the actual measured value. Figure 3 As can be seen from the figure, the predicted temperature value pred is close to the actual measured value real, indicating that the flashback prediction model of this embodiment has a good prediction effect.
[0128] S6: Update the data of multi-source heterogeneous database based on actual fire scene data, and update the training of artificial intelligence flashback prediction model. The details are as follows:
[0129] Whenever new data from an actual fire scene is collected and uploaded to a multi-source heterogeneous database, the model immediately starts a new learning cycle, continuously improving prediction accuracy and robustness through incremental training.
[0130] Embodiment 2:
[0131] Reference Figure 4-5 The present invention provides an artificial intelligence flashback prediction system based on multi-source heterogeneous data, comprising:
[0132] The data acquisition module 100 is composed of a high-definition camera, a temperature sensor, a gas concentration sensor, a pressure sensor, etc. The Internet of Things technology connects the above devices through a wireless or wired network to form an intelligent monitoring network. The data collected by each device is summarized and preliminarily processed through the Internet of Things platform.
[0133] The data preprocessing module 200 receives the raw data from the data acquisition module 100, cleans, formats, standardizes and processes outliers on the raw data to ensure the quality and consistency of the data and prepares a data set for model training or real-time analysis. The data preprocessing module 200 includes a data cleaning unit 210 and a data dimension unification unit 220, which are used for data cleaning and data dimension unification respectively;
[0134] The multi-source heterogeneous database module 300 receives the preprocessed data output by the data preprocessing module 200 and stores various types of preprocessed data, provides data support for model training and real-time analysis, and provides interfaces for data query, update, and management.
[0135] The artificial intelligence flashback prediction module 400 obtains training data from multi-source heterogeneous databases and trains the model through historical data. After receiving real-time data from the fire scene, the model can perform real-time analysis and prediction of the fire development trend and the time and intensity of flashback. The artificial intelligence flashback prediction module 400 includes:
[0136] The feature extraction unit 410 is used to extract features from images and data information collected by high-definition cameras, sensors, and Internet of Things technologies;
[0137] The image enhancement unit 420 is used to enhance the extracted image features based on the GANs model;
[0138] The multimodal data fusion prediction unit 430 is used to input image features and data features into the improved multi-layer Bi-LSTM network to capture the relationship between multimodal data streams, and then complete the prediction results of the backfire output through the fully connected layer.
[0139] The result analysis and visualization module 500 analyzes the fire development in real time, locates the scale and location of the fire source, and converts the prediction results into intuitive charts and reports to provide decision support.
[0140] The self-update module 600 analyzes the prediction results generated by the calculation module in real time, as well as the latest data continuously collected by the data acquisition module 100. The new data is fed back to the multi-source heterogeneous database, the training data set is updated, and the incremental training or retraining of the model is triggered to continuously optimize the prediction accuracy and robustness of the model.
[0141] Reference Figure 5 ,The data acquisition module 100 is responsible for collecting fire scene information from ,multiple source heterogeneous data such as various sensors (such as temperature, smoke, and gas concentration ,sensors), surveillance camera image data, meteorological data, and ,other.
[0142] Specifically, when a certain data collection fails, the artificial intelligence flashback prediction model can also complete the prediction of the flashback phenomenon based on other data.
[0143] The data preprocessing module 200 cleans, fuses and standardizes the collected fire scene data to ensure the consistency of data dimensions for the training and prediction of the artificial intelligence flashback prediction model.
[0144] The multi-source heterogeneous database module 300 stores pre-processed images and fire history data, and continuously receives data from the fire scene and the prediction results of the model, providing reliable data for model training and self-updating.
[0145] The artificial intelligence flashback prediction module 400 uses deep learning technologies, such as convolutional neural networks (CNN) and bidirectional long short-term memory networks (BI-LSTM), combined with the Transformer algorithm, to train the preprocessed data to achieve the prediction of the flashback phenomenon.
[0146] When receiving real-time multi-source data of the fire scene, the model can warn the location and size of the fire source, predict the temperature development changes in real time, and determine the time of backdraft. If backdraft will not occur, the fire development information is sent to the result analysis and visualization module 500.
[0147] The result analysis and visualization module 500 converts the prediction results into intuitive charts and reports, which facilitate the fire command center to quickly understand the situation. The interface displays the prediction results, risk level, and emergency response measures based on the prediction results. In this intuitive way, the command personnel can quickly understand the on-site situation and take corresponding actions.
[0148] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.
Claims
1. An artificial intelligence flashback prediction method based on multi-source heterogeneous data, characterized in that: The steps include: S1: Real-time capture of multi-dimensional information of the fire scene through high-definition cameras, sensors and IoT technology; S2: Use data processing technology to pre-process the acquired data stream; S3: Build a multi-source heterogeneous database and integrate the pre-processed multi-source heterogeneous data; S4: Build an artificial intelligence flashback prediction model based on multi-source heterogeneous data and conduct training; S5: Use the pre-trained flashback prediction model to perform real-time analysis and calculation of the monitoring information of the fire scene to predict the probability and intensity of flashback; S6: Update the data of multi-source heterogeneous databases based on actual fire scene data, and train and update the artificial intelligence flashback prediction model.
2. The artificial intelligence flashback prediction method based on multi-source heterogeneous data according to claim 1 is characterized in that: The data preprocessing in S2 includes data cleaning and dimensionality unification processing. The data cleaning includes missing value processing, outlier detection and processing operations; the dimensionality unification includes unit conversion, data standardization or normalization, and data encoding processing.
3. The artificial intelligence flashback prediction method based on multi-source heterogeneous data according to claim 1 is characterized in that: The artificial intelligence flashback prediction model in S4 uses a CNN network to analyze the feature information of the fire image data obtained by the high-definition camera, and enhances the image data based on the GANs model, and then uses the encoder and fusion self-attention mechanism in the Transformer to extract the sensor data features, and then inputs the acquired multimodal data features into the improved multi-layer Bi-LSTM network to capture the relationship between the multimodal data streams, and finally completes the flashback output prediction results through the fully connected layer.
4. The artificial intelligence flashback prediction method based on multi-source heterogeneous data according to claim 3 is characterized in that: The CNN network extracts image features from temporal neighborhoods of different spans based on a number of convolution kernels of different sizes.
5. The artificial intelligence flashback prediction method based on multi-source heterogeneous data according to claim 3 is characterized in that: The GANs model consists of four generator subnetworks with multiple resolutions, and each generator subnetwork has a spatial attention module added, as follows: Based on the self-attention operation, we obtain the query vector matrix Q1, key vector matrix K1 and value vector matrix V1 of the image features. The output of the adjacent space self-attention mechanism Atten(Q1,K1,V1) is as follows: in, and are all learning parameters; d1 is the dimension of the query vector matrix Q1, key vector matrix K1 and value vector matrix V1 in the adjacency space self-attention mechanism; T is the transpose of the vector; The four generator subnetworks enhance the input image features in turn, and the learning parameters of the attention module in each generator subnetwork are obtained based on the improved particle swarm algorithm. The improved particle swarm algorithm is specifically as follows: v a+1,b =ω t v a,b +c1r1(p b -u a,b )+c2r2(g b -u a,b ) u a+1,b =u a,b +v a+1,b Among them, v a,b 、v a+1,b are the speeds of the ath and a+1th iterations of the bth parameter respectively; u a,b 、u a+1,b is the value of the ath and a+1th iterations of the bth parameter; ω t is the inertia weight; c1 and c2 are acceleration factors; r1 and r2 are random variables uniformly distributed between [0,1]; p b is the individual optimal value; g b is the global optimal value; τ is the maximum number of iterations; ω max ,ω min are the maximum and minimum values of the inertia weight respectively.
6. The artificial intelligence flashback prediction method based on multi-source heterogeneous data according to claim 5 is characterized in that: The encoder is composed of Transformer layers, each of which includes a BN layer and a ReLU activation function, as well as a feedforward neural network part and a residual connection. The encoder also adds a multi-head fusion attention mechanism. On this basis, a dynamic gating mechanism is introduced. The dynamic gating weight is obtained by dynamic optimization based on the improved particle swarm algorithm according to the characteristics and context of the input data.
7. The artificial intelligence flashback prediction method based on multi-source heterogeneous data according to claim 6 is characterized in that: The fusion attention mechanism combines three attention mechanisms and obtains the final attention output based on a weighted approach, as follows: The first attention mechanism is to integrate the temporal self-attention mechanism, as follows: The input feature matrix is formed based on the time step. For each time step, the query vector matrix Q2, key vector matrix K2 and value vector matrix V2 of the corresponding features are formed. The output of the temporal self-attention mechanism Atten(Q2,K2,V2) is as follows: in, and are all learning parameters; d2 is the dimension of the query vector matrix Q2, key vector matrix K2, and value vector matrix V2 in the temporal self-attention mechanism; The second attention mechanism is to integrate the global self-attention mechanism, as follows: Based on two MLPs, the input features are mapped to obtain the query vector matrix Q3, key vector matrix K3 and value vector matrix V3: The output of the global self-attention mechanism Atten(Q3,K3,V3) is as follows: Atten(Q3,K3,V3)=Drop[(K3-Q3)·MLP(K3)]·V3 in, and are all learning parameters; Drop(·) is a random dropout operation; The third attention mechanism is the local self-attention mechanism, which is as follows: Based on two MLPs, the input features are mapped to obtain the query vector matrix Q4, key vector matrix K4 and value vector matrix V4: For each position i, define the window range And extract the sub-vector matrix k corresponding to the key vector matrix K4 j , calculate the attention weight within the window range: The output of the local self-attention mechanism Atten(Q4,K4,V4) is as follows: Atten(Q4,K4,V4)=w ij V4 Among them, w ij represents the local attention weight; and All are learning parameters; The learning parameters and weighting parameters in the three attention mechanisms are obtained based on the improved particle swarm algorithm.
8. The artificial intelligence flashback prediction method based on multi-source heterogeneous data according to claim 7 is characterized in that: The improved multi-layer Bi-LSTM network adds a Dropout layer after each hidden layer, and simultaneously obtains various network parameters based on the improved particle swarm algorithm.
9. The artificial intelligence flashback prediction system based on multi-source heterogeneous data according to any one of claims 1 to 8, characterized in that: include: Data acquisition module (100): used to capture multi-dimensional information of the fire scene in real time through high-definition cameras, sensors and Internet of Things technology; Data preprocessing module (200): used to preprocess the acquired data stream using data processing technology; Multi-source heterogeneous database module (300): used to construct a multi-source heterogeneous database and integrate pre-processed multi-source heterogeneous data; Artificial intelligence flashback prediction module (400): used to construct an artificial intelligence flashback prediction model based on multi-source heterogeneous data and perform training; Result analysis and visualization module (500): used to convert the prediction results into charts and reports, and to perform corresponding risk visualization warning according to the prediction results; A self-update module (600): used to update the data of a multi-source heterogeneous database based on actual fire scene data, and to update the training of an artificial intelligence flashback prediction model; The data preprocessing module (200) comprises a data cleaning unit (210) and a data dimension unification unit (220), which are used for data cleaning and data dimension unification respectively; The artificial intelligence flashback prediction module (400) comprises: Feature extraction unit (410): used to extract features from images and data information collected by high-definition cameras, sensors and Internet of Things technology; An image enhancement unit (420): used for performing enhancement processing on the extracted image features based on a GANs model; Multimodal data fusion prediction unit (430): used to input image features and data features into the improved multi-layer Bi-LSTM network to capture the relationship between multimodal data streams, and then complete the prediction result of the back-burning output through the fully connected layer.
Citation Information
Cited By
Res-Transform urban solid waste incineration flame state identification method based on multi-feature enhancement
CN119693719A
Coal spontaneous combustion temperature prediction method based on numerical value and image fusion
CN120375954A
