A machine large model-based fire prediction system and method

Abstract

The application discloses a fire prediction system and method based on a machine large model, relates to the technical field of intelligent fire fighting, and solves the technical problem that the prediction result is compared with a fixed threshold in the prior art, resulting in low accuracy of the fire prediction method; scene data is input into a scene recognition model to obtain a scene type label; a prediction label and a prediction confidence are generated according to the abnormal parameters and the scene type label, and a fire confidence corresponding to a fire label is obtained by weighted fusion of a plurality of prediction confidences; a confidence threshold corresponding to the fire label is generated based on fire data; when the fire confidence is greater than the confidence threshold, a fire occurrence early warning signal is generated, and the confidence threshold is dynamically adjusted under different scene labels and corresponding fire data, so that the prediction result can be compared with the confidence threshold to fit the actual scene and improve the accuracy of fire prediction.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent fire fighting, and particularly relates to a fire prediction system and method based on a machine large model. BACKGROUND

[0002] As one of the most common disasters, fire has the characteristics of strong burst, fast spreading speed, wide harm range, etc., and brings great threat to people's life and property safety and social stability. In recent years, with the continuous expansion of city size, the increase of population density and the emergence of various complex building structures, fire hazards are increasing. In order to effectively prevent the occurrence of fire accidents and improve the disaster prevention and control capability, the research on fire prediction method has gradually become an important topic in the field of public safety.

[0003] When obtaining the prediction result of fire prediction, the prior art often compares it with a fixed threshold, lacks consideration of the great difference in fire characteristics of different scenes, causes false alarm or missed alarm to occur due to the fixed threshold, and further causes the accuracy of the fire prediction method to be low; therefore, the fire prediction method still needs to be further improved. SUMMARY

[0004] The application aims to at least solve one of the technical problems existing in the prior art; for this purpose, the application provides a fire prediction system and method based on a machine large model, which is used to solve the technical problem that the accuracy of the fire prediction method is low due to the fact that the prior art often compares the prediction result with a fixed threshold.

[0005] To achieve the above-mentioned purpose, a first aspect of the application provides a fire prediction method based on a machine large model, comprising:

[0006] Obtaining scene data, environment data, equipment operation data and fire fighting data; the scene data refers to scene data that needs to be monitored for fire warning;

[0007] Inputting the scene data into a scene recognition model to obtain a scene type label; the scene recognition model is constructed by an artificial intelligence model, and is used to identify the scene type label corresponding to the scene data;

[0008] Extracting a plurality of abnormal parameters in the environment data and the equipment operation data; the abnormal parameters refer to parameters in the environment data and the equipment operation data that deviate from the normal range;

[0009] Generating a prediction label and a prediction confidence according to the abnormal parameters and the scene type label; the prediction label includes a fire label and a no-fire label;

[0010] Fusing the prediction confidence corresponding to a plurality of fire labels through a confidence weight coefficient to obtain a fire confidence corresponding to the fire label;

[0011] generate a confidence threshold corresponding to the fire label based on the fire data;

[0012] generate a fire occurrence early warning signal when the fire confidence is greater than the confidence threshold.

[0013] The present application dynamically adjusts the confidence threshold under different scene labels and corresponding fire data through the above steps, so that when the prediction result is compared with the confidence threshold, it can fit the actual scene and improve the accuracy of fire prediction.

[0014] Further, the scene recognition model is constructed by an artificial intelligence model, comprising:

[0015] Obtain a plurality of historical scene data corresponding to a plurality of scene type labels;

[0016] Obtain a plurality of enhanced scene data through data enhancement technology for a plurality of historical scene data;

[0017] Integrate a plurality of historical scene data and a plurality of enhanced scene data under the same scene type label into historical scene data corresponding to the scene type label;

[0018] Divide the historical scene data into training data, validation data and test data according to the scene type label; and perform data preprocessing on the training data, validation data and test data to obtain a training set, a validation set and a test set;

[0019] Select an artificial intelligence model as a base model;

[0020] Train the base model through the training set, and adjust the learning rate and other hyperparameters on the validation set to obtain a pre-trained model;

[0021] Verify the pre-trained model on the test set to finally obtain a scene recognition model with scene data as input and scene type label as output.

[0022] Further, the generation of the prediction label and the corresponding prediction confidence according to the abnormal parameter and the scene type label comprises:

[0023] Extract a plurality of historical abnormal time points corresponding to a plurality of abnormal parameters; the historical abnormal time point refers to a historical time point that does not exist between the abnormal parameters of the historical time point and the abnormal parameters of the current time point. Time point disconnection;

[0024] Obtain a parameter prediction sequence corresponding to the abnormal parameter based on the historical abnormal time point;

[0025] Generate a sequence weight coefficient based on the scene type label;

[0026] Several parameter prediction sequences are fused using sequence weight coefficients to obtain a fused prediction sequence.

[0027] Obtain analysis tags; the analysis tags are tags corresponding to the anomaly prediction sequences, and the anomaly prediction sequences include parameter prediction sequences and fusion prediction sequences;

[0028] Several analysis labels and their corresponding anomaly prediction sequences are input into a multimodal fire prediction model to obtain the prediction labels and prediction confidence levels corresponding to the analysis labels; the multimodal fire prediction model is constructed using a large machine model.

[0029] Furthermore, obtaining the parameter prediction sequence corresponding to the abnormal parameters based on historical abnormal time points includes the following steps:

[0030] Step 1: Obtain several historical abnormal time points corresponding to several abnormal parameters and calculate the total number of abnormal time points corresponding to several abnormal parameters.

[0031] Step 2: Extract the maximum value among the total number of abnormal time points corresponding to several abnormal parameters, and take the maximum value among the total number of abnormal time points as the maximum total number of abnormal time points;

[0032] Determine if the total number of the largest abnormal time points is within the time window size;

[0033] Yes, obtain historical time point data corresponding to several abnormal parameters based on the total number of maximum abnormal time points, and proceed to step four; the historical time point data includes abnormal parameters and normal parameters corresponding to the abnormal parameters. This step is to standardize the length of the abnormal prediction sequence.

[0034] No, when the total number of maximum abnormal time points is greater than the maximum value in the time window size range, extract the historical time point data corresponding to several abnormal parameters based on the maximum value in the time window size range, and proceed to step four; when the total number of maximum abnormal time points is less than the minimum value in the time window size range, extract the historical time point data corresponding to several abnormal parameters based on the minimum value in the time window size range, and proceed to step four.

[0035] Step 4: Integrate the historical time point data into a parameter prediction sequence corresponding to the abnormal parameters according to the chronological order.

[0036] This application obtains historical abnormal time points for each abnormal parameter, and uses the maximum total number of time points as a benchmark, combined with the time window size range, to unify the parameter prediction sequence of each abnormal parameter in terms of time length, providing data support for the subsequent fusion of parameter prediction sequences.

[0037] Furthermore, the generation of sequence weight coefficients based on scene type labels includes:

[0038] The initial weights of several abnormal parameters are obtained from the initial weight table based on the scene type label; the initial weight table is set by experts based on experience, including the initial weight table for office building scene, the initial weight table for chemical storage scene, and the initial weight table for residential community scene;

[0039] Extract the historical abnormal time point data and abnormal duration corresponding to the abnormal parameters;

[0040] An anomaly score is generated based on several historical anomaly time points and anomaly durations.

[0041] The normalized normalized parameter anomaly score GCYP is obtained by performing max-min normalization on the anomaly scores of several parameters.

[0042] Based on the nonlinear relationship between the initial weight CQ, the normalized parameter anomaly score GCYP and the sequence weight coefficient, a sequence weight calculation function XQF(CQ,GCYP) is constructed.

[0043] The sequence weight calculation function satisfies the following formula:

[0044] Where i and n represent the numbers corresponding to the abnormal parameters, N represents the total number of abnormal parameters, and k represents the abnormal gain coefficient, k∈(0,1).

[0045] Substituting the initial weights and normalized parameter anomaly scores into the sequence weight calculation function yields the sequence weight coefficients XQX.

[0046] This application analyzes the initial weights of each abnormal parameter under different scenarios and quantifies the degree of abnormality of each parameter to obtain a parameter abnormality score. Based on this, it further studies the impact of the parameter abnormality score on the initial weights, dynamically adjusts the weight coefficients of each abnormal parameter, and obtains the corresponding sequence weight coefficients. This enables the weights of the parameter prediction sequence to be adaptively determined under different scenarios and different abnormal parameter conditions, thereby improving the accuracy and applicability of multi-parameter fusion prediction.

[0047] Furthermore, the step of generating parameter anomaly scores based on several historical anomaly time points and anomaly durations includes:

[0048] Extract several historical anomaly point data and anomaly duration YS corresponding to the anomaly parameters;

[0049] Several abnormal values ​​are obtained by calculating the difference between several historical outlier data points and the normal data range corresponding to the outlier parameters.

[0050] The average value of several abnormal values ​​is calculated to obtain the abnormal mean value YJ corresponding to the abnormal parameter;

[0051] The abnormal growth rate YZL is calculated based on several abnormal values ​​in chronological order.

[0052] Calculate the standard deviation of the abnormal growth rate YZBC based on several abnormal growth rates;

[0053] Through formula Calculate the parameter anomaly score corresponding to the abnormal parameters. Where i represents the number corresponding to the abnormal parameter; and These are represented as the data anomaly sensitivity coefficient and the duration response coefficient, respectively. and All are greater than 0.

[0054] This application analyzes the correlation between abnormal parameters and fire occurrence, selects multi-source data of abnormal parameters to quantify the degree of abnormality of abnormal parameters, makes the quantification of parameter abnormality scores more accurate, and provides accurate data support for calculating sequence weight coefficients.

[0055] Furthermore, the multimodal fire prediction model is constructed using a large machine model, including:

[0056] Obtain several analysis labels and their corresponding historical anomaly prediction sequences, historical prediction labels, and prediction confidence levels corresponding to the historical prediction labels;

[0057] Several analysis labels and their corresponding historical anomaly prediction sequences, historical prediction labels, and prediction confidence scores corresponding to historical prediction labels are divided into training data, validation data, and test data corresponding to the analysis labels; and data preprocessing is performed on the training data, validation data, and test data corresponding to the analysis labels to obtain the training set, validation set, and test set corresponding to the analysis labels.

[0058] Choose the large machine model as the base model for analyzing the corresponding labels;

[0059] The basic model corresponding to the analysis label is trained by analyzing the training set corresponding to the analysis label, and the learning rate and other hyperparameters are adjusted on the validation set corresponding to the analysis label to obtain the pre-trained model corresponding to the analysis label.

[0060] By validating the pre-trained model corresponding to the analysis label on the test set corresponding to the analysis label, a multimodal fire prediction model is finally obtained with the analysis label and its corresponding anomaly prediction sequence as input and the prediction label and prediction confidence of the analysis label as output.

[0061] Furthermore, the confidence weight coefficients are generated in the following ways:

[0062] Extract the sequence weight coefficients and their corresponding fire label prediction confidence, and fuse the prediction confidence of the fire label corresponding to the prediction sequence.

[0063] The predicted confidence score is calculated as the difference between the predicted confidence score of the fire tag with outlier parameters and the predicted confidence score of the fire tag corresponding to the fused predicted sequence; the outlier parameters correspond one-to-one with the sequence weight coefficients.

[0064] Obtain confidence level labels; each confidence level label corresponds one-to-one with an analysis label and is used to locate the number corresponding to several predicted confidence levels;

[0065] The confidence labels and their corresponding sequence weight coefficients, the predicted confidence of the fire labels, and the predicted confidence of the difference are integrated into confidence analysis data.

[0066] The confidence analysis data is input into the confidence weight generation model to obtain several confidence weight coefficients; the confidence weight generation model is constructed through a machine learning model and is used to evaluate the confidence weight coefficients corresponding to the confidence labels.

[0067] Furthermore, the confidence weight generation model is constructed using a machine learning model, including:

[0068] Obtain several historical confidence analysis data and their corresponding historical confidence weight coefficients;

[0069] Several historical confidence analysis data and their corresponding historical confidence weight coefficients are divided into training data, validation data, and test data; and the training data, validation data, and test data are preprocessed to obtain the training set, validation set, and test set.

[0070] Choose a machine learning model as the base model;

[0071] The base model is trained on the training set, and the learning rate and other hyperparameters are adjusted on the validation set to obtain the pre-trained model.

[0072] By validating the pre-trained model on the test set, a confidence weight generation model is finally obtained, which takes confidence analysis data as input and outputs several confidence weight coefficients.

[0073] This application uses a pre-trained confidence weight generation model to dynamically generate the confidence weight coefficients of the predicted confidence, which enables a more accurate fire confidence of the fire label when performing weighted fusion of several predicted confidences of the fire label, thereby improving the accuracy and computational efficiency of the fire confidence.

[0074] Furthermore, the confidence threshold for generating fire labels based on fire data includes:

[0075] The initial confidence threshold CZY is extracted from the initial threshold table based on the scene type label;

[0076] Extract the historical fire count (HC) and the corresponding fire severity level from the fire data;

[0077] The average severity level (PYD) is obtained by averaging the fire severity levels corresponding to the number of fires.

[0078] A confidence function ZDF(CZY, HC, PYD) is constructed based on the nonlinear relationship between the initial confidence threshold CZY, the number of historical fires HC, the average severity level PYD and the confidence threshold.

[0079] The confidence function satisfies the following calculation formula:

[0080] ;in, and These are respectively represented as the risk sensitivity coefficient and the frequency saturation coefficient; and ∈(0,1);

[0081] The initial confidence threshold, the number of historical fires, and the average severity level are substituted into the confidence function to calculate the confidence threshold corresponding to the fire label.

[0082] Another aspect of the present invention provides a fire prediction system based on a large machine model, comprising: a data acquisition module, a data analysis module, and an early warning module; the data acquisition module and the data analysis module are connected; the data analysis module and the early warning module are connected.

[0083] The data acquisition module acquires scene data, environmental data, equipment operation data, and fire protection data through data acquisition devices; the scene data refers to scene data that requires fire early warning monitoring.

[0084] The data analysis module: inputs scene data into the scene recognition model to obtain scene type labels; extracts several abnormal parameters from environmental data and equipment operation data; generates predicted labels and predicted confidence scores based on the abnormal parameters and scene type labels; obtains the fire confidence score corresponding to the fire label by fusing the predicted confidence scores corresponding to several fire labels through confidence score weighting coefficients; generates a confidence score threshold corresponding to the fire label based on fire protection data; and generates a fire warning signal when the fire confidence score is greater than the confidence score threshold.

[0085] The early warning module issues corresponding signal prompts based on the early warning signal.

[0086] Compared with the prior art, the beneficial effects of this application are:

[0087] 1. This application obtains scene type labels by inputting scene data into a scene recognition model; extracts several abnormal parameters from environmental data and equipment operation data; generates prediction labels and prediction confidence scores based on the abnormal parameters and scene type labels; obtains the fire confidence score corresponding to the fire label by fusing the prediction confidence scores corresponding to several fire labels through confidence score weighting coefficients; generates a confidence score threshold corresponding to the fire label based on fire protection data; when the fire confidence score is greater than the confidence score threshold, a fire warning signal is generated; and the confidence score threshold is dynamically adjusted under different scene labels and their corresponding fire protection data so that when the prediction results are compared with the confidence score threshold, they can fit the actual scene and improve the accuracy of fire prediction.

[0088] 2. This application analyzes the initial weights of each abnormal parameter under different scenarios and quantifies the degree of abnormality of each abnormal parameter to obtain a parameter abnormality score. By analyzing the impact of the parameter abnormality score on the initial weight, the weight coefficients of each abnormal parameter are dynamically adjusted to obtain the sequence weight coefficients. This allows the sequence weight coefficients of the parameter prediction sequence to be obtained adaptively under different scenarios and different abnormal parameters, thereby improving the accuracy of the fused prediction sequence.

[0089] 3. This application dynamically adjusts the confidence threshold under different scenario labels and their corresponding fire protection data, so that the confidence threshold can fit the actual situation of the current scenario. For the same scenario where fires frequently occur, it is necessary to improve the sensitivity of fire early warning, reduce the probability of fire occurrence, and improve the accuracy of fire prediction methods. Attached Figure Description

[0090] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0091] Figure 1 This is a flowchart of a fire prediction method based on a large machine model according to this application;

[0092] Figure 2 This is a schematic diagram illustrating the principle of a fire prediction system based on a large machine model, as described in this application. Detailed Implementation

[0093] The technical solutions of this application will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0094] Please see Figure 1 The first aspect of this application provides a fire prediction method based on a large machine model, comprising:

[0095] Acquire scene data, environmental data, equipment operation data, and fire protection data; scene data refers to scene data that requires fire early warning monitoring.

[0096] Scene data is input into the scene recognition model to obtain scene type labels; the scene recognition model is constructed using an artificial intelligence model to identify the scene type labels corresponding to the scene data;

[0097] Extract several abnormal parameters from environmental data and equipment operation data; abnormal parameters refer to parameters in environmental data and equipment operation data that deviate from the normal range.

[0098] Predictive labels and prediction confidence scores are generated based on abnormal parameters and scene type labels; the predictive labels include fire labels and no-fire labels.

[0099] The fire confidence level corresponding to a fire label is obtained by fusing the predicted confidence levels of several fire labels using confidence weight coefficients.

[0100] Confidence thresholds for fire labels are generated based on fire data.

[0101] When the fire confidence level is greater than the confidence threshold, a fire warning signal is generated.

[0102] The scene recognition model in this embodiment is constructed using an artificial intelligence model, including:

[0103] Retrieve historical scene data corresponding to several scene type tags;

[0104] Several enhanced scenario data were obtained by using data augmentation techniques on several historical scenario data.

[0105] Integrate several historical scene data and several enhanced scene data under the same scene type label into historical scene data corresponding to the scene type label;

[0106] Historical scene data is divided into training data, validation data, and test data according to scene type labels; and the training data, validation data, and test data are preprocessed to obtain training set, validation set, and test set; the ratio between training set, test set, and validation set is 7:2:1;

[0107] Choose an artificial intelligence model as the base model; artificial intelligence models include convolutional neural network models, etc.

[0108] The base model is trained on the training set, and the learning rate and other hyperparameters are adjusted on the validation set to obtain the pre-trained model.

[0109] By validating the pre-trained model on the test set, a scene recognition model is finally obtained, which takes scene data as input and outputs scene type labels.

[0110] This embodiment uses a pre-trained scene recognition model to accurately identify scene data during fire prediction, providing accurate data support for subsequent fire prediction.

[0111] In this embodiment, generating prediction labels and their corresponding prediction confidence scores based on abnormal parameters and scene type labels includes:

[0112] Extract several historical abnormal time points corresponding to several abnormal parameters; historical abnormal time points refer to historical time points where there is no break between the abnormal parameters at historical time points and the abnormal parameters at the current time. In this embodiment, it can be understood that the time point corresponding to the abnormal parameters at the current time is t, and the time points of the abnormal parameters at historical time points are t-1, t-2, and t-4. Then, in this embodiment, the several historical abnormal time points are t, t-1, and t-2. Since there is a time point t-3 between t-4 and t-2, the time points t-4 and t-2 are broken. Therefore, the time point t-4 is not used as one of the several historical abnormal time points of the abnormal parameters.

[0113] The parameter prediction sequence is obtained based on historical anomaly time points; the parameter prediction sequence refers to the sequence for predictive analysis of anomaly parameters.

[0114] Sequence weight coefficients are generated based on scene type labels; sequence weight coefficients are used to integrate the parameter prediction sequences corresponding to several abnormal parameters into a fusion prediction sequence.

[0115] Several parameter prediction sequences are fused using sequence weight coefficients to obtain a fused prediction sequence.

[0116] Obtain analysis tags; analysis tags are the tags corresponding to the anomaly prediction sequences, which include parameter prediction sequences and fusion prediction sequences.

[0117] Several analytical labels and their corresponding anomaly prediction sequences are input into a multimodal fire prediction model to obtain the prediction labels and prediction confidence levels corresponding to the analytical labels; the multimodal fire prediction model is constructed using a large machine model.

[0118] In this embodiment, obtaining the parameter prediction sequence corresponding to the abnormal parameters based on historical abnormal time points includes the following steps:

[0119] Step 1: Obtain several historical abnormal time points corresponding to several abnormal parameters and calculate the total number of abnormal time points corresponding to several abnormal parameters.

[0120] Step 2: Extract the maximum value among the total number of abnormal time points corresponding to several abnormal parameters, and take the maximum value among the total number of abnormal time points as the maximum total number of abnormal time points;

[0121] Determine whether the total number of the largest abnormal time points is within the time window size range; the time window size range is set based on experience, and in this embodiment, the time window size range is set to 10 to 60; that is, the length of the abnormal prediction sequence should be within the range of 10 to 60 time points.

[0122] Yes, obtain historical time point data corresponding to several abnormal parameters based on the total number of maximum abnormal time points, and proceed to step four; the historical time point data includes abnormal parameters and normal parameters corresponding to the abnormal parameters. This step is to standardize the length of the abnormal prediction sequence.

[0123] No, when the total number of maximum abnormal time points is greater than the maximum value in the time window size range, extract the historical time point data corresponding to several abnormal parameters based on the maximum value in the time window size range, and proceed to step four; when the total number of maximum abnormal time points is less than the minimum value in the time window size range, extract the historical time point data corresponding to several abnormal parameters based on the minimum value in the time window size range, and proceed to step four.

[0124] Step 4: Integrate the historical time point data into a parameter prediction sequence corresponding to the abnormal parameters according to the chronological order.

[0125] This embodiment collects historical abnormal time points for each abnormal parameter and uses the maximum number of time points as a benchmark, combined with a preset time window size range, to uniformly process the predicted sequences of each abnormal parameter in terms of time length, thereby providing a consistent and comparable data foundation for subsequent fusion analysis of parameter predicted sequences.

[0126] In this embodiment, the generation of sequence weight coefficients based on scene type labels includes:

[0127] The initial weights of several abnormal parameters are obtained from the initial weight table based on the scene type label. The initial weight table is set by experts based on experience and includes initial weight tables for office building scenes, chemical storage scenes, and residential community scenes. The initial weight table for the office building scene in this embodiment is shown in the table below:

[0128]

[0129] The initial weighting table for the chemical storage scenario in this embodiment is shown in the table below:

[0130]

[0131] The initial weight table for the residential community scenario in this embodiment is shown in the table below:

[0132]

[0133] Extract the historical abnormal time point data and abnormal duration corresponding to the abnormal parameters;

[0134] A parameter anomaly score is generated based on several historical abnormal time points and abnormal durations; the parameter anomaly score quantifies the degree of abnormality of the abnormal parameters.

[0135] The normalized normalized parameter anomaly score GCYP is obtained by performing max-min normalization on the anomaly scores of several parameters.

[0136] Based on the nonlinear relationship between the initial weight CQ, the normalized parameter anomaly score GCYP and the sequence weight coefficient, a sequence weight calculation function XQF(CQ,GCYP) is constructed.

[0137] The sequence weight calculation function satisfies the following formula:

[0138] Where i and n represent the numbers corresponding to the abnormal parameters, N represents the total number of abnormal parameters; k represents the abnormal gain coefficient, k∈(0,1); the setting of k is to control the increase of the abnormal score on the weight. The larger k is, the more significant the abnormal influence. The value of k is set according to experience. In this embodiment, k is set to 0.3.

[0139] Substituting the initial weights and normalized parameter anomaly scores into the sequence weight calculation function yields the sequence weight coefficients XQX.

[0140] In this embodiment, the parameter anomaly score is generated based on several historical abnormal time points and abnormal durations, including:

[0141] Extract several historical anomaly point data and anomaly duration YS corresponding to the anomaly parameters;

[0142] Several abnormal values ​​are obtained by calculating the difference between several historical outlier data points and the normal data range corresponding to the outlier parameters.

[0143] The average value of several abnormal values ​​is calculated to obtain the abnormal mean value YJ corresponding to the abnormal parameter;

[0144] The abnormal growth rate YZL is calculated based on several abnormal values ​​in chronological order.

[0145] Calculate the standard deviation of the abnormal growth rate YZBC based on several abnormal growth rates;

[0146] Through formula Calculate the parameter anomaly score corresponding to the abnormal parameters. Where i represents the number corresponding to the abnormal parameter; and These are represented as the data anomaly sensitivity coefficient and the duration response coefficient, respectively. and All values ​​are greater than 0. The specific values ​​are set based on experience. In this embodiment, they will be... Set to 1.2; The value of is related to the duration of the anomaly. In this embodiment, Represented as: Where YSY represents the abnormal duration threshold, the specific value of which is set based on experience. In this embodiment, YSY is set to 5 minutes.

[0147] This embodiment analyzes the correlation between abnormal parameters and fire occurrence in depth, and comprehensively selects multi-source data to quantify the degree of abnormality of abnormal parameters, thereby improving the accuracy of parameter anomaly scoring and providing reliable data support for the subsequent calculation of sequence weight coefficients.

[0148] The multimodal fire prediction model in this embodiment is constructed using a large machine model, including:

[0149] Obtain several analysis labels and their corresponding historical anomaly prediction sequences, historical prediction labels, and prediction confidence levels corresponding to the historical prediction labels;

[0150] Several analysis labels and their corresponding historical anomaly prediction sequences, historical prediction labels, and prediction confidence scores corresponding to historical prediction labels are divided into training data, validation data, and test data corresponding to the analysis labels. Data preprocessing is performed on the training data, validation data, and test data corresponding to the analysis labels to obtain the training set, validation set, and test set corresponding to the analysis labels. The ratio between the training set, test set, and validation set is 7:2:1.

[0151] Select a large machine learning model as the base model for analyzing the corresponding labels; large machine learning models include LSTM models, etc.

[0152] The basic model corresponding to the analysis label is trained by analyzing the training set corresponding to the analysis label, and the learning rate and other hyperparameters are adjusted on the validation set corresponding to the analysis label to obtain the pre-trained model corresponding to the analysis label.

[0153] By validating the pre-trained model corresponding to the analysis label on the test set corresponding to the analysis label, a multimodal fire prediction model is finally obtained with the analysis label and its corresponding anomaly prediction sequence as input and the prediction label and prediction confidence of the analysis label as output.

[0154] The confidence weight coefficients in this embodiment are generated in the following ways:

[0155] Extract the sequence weight coefficients and their corresponding fire label prediction confidence, and fuse the prediction confidence of the fire label corresponding to the prediction sequence.

[0156] The predicted confidence score is calculated by comparing the predicted confidence score of the fire tag with the outlier parameter with the predicted confidence score of the fire tag corresponding to the fused predicted sequence. There is a one-to-one correspondence between the outlier parameter and the sequence weight coefficient; that is, there is a unique sequence weight coefficient corresponding to each outlier parameter.

[0157] Obtain confidence level labels; each confidence level label corresponds one-to-one with an analysis label, used to locate the number corresponding to several prediction confidence levels; the number of analysis labels is one more than the number of outlier parameters, the extra one being the fused prediction sequence; while the number of confidence level labels is consistent with the number of analysis labels.

[0158] The confidence level labels and their corresponding sequence weight coefficients, the predicted confidence level corresponding to the fire label, and the difference predicted confidence level are integrated into confidence level analysis data; the confidence level analysis data is used to analyze the analysis data required for analyzing the confidence level weight coefficients corresponding to several predicted confidence levels;

[0159] The confidence analysis data is input into the confidence weight generation model to obtain several confidence weight coefficients; the confidence weight generation model is constructed through a machine learning model and is used to evaluate the confidence weight coefficients corresponding to the confidence labels.

[0160] The confidence weight generation model in this embodiment is constructed using a machine learning model, including:

[0161] Obtain several historical confidence analysis data and their corresponding historical confidence weight coefficients;

[0162] Several historical confidence analysis data and their corresponding historical confidence weight coefficients are divided into training data, validation data, and test data; and the training data, validation data, and test data are preprocessed to obtain training set, validation set, and test set; the ratio between training set, test set, and validation set is 7:2:1;

[0163] Choose a machine learning model as the base model; machine learning models include the BP model, etc.

[0164] The base model is trained on the training set, and the learning rate and other hyperparameters are adjusted on the validation set to obtain the pre-trained model.

[0165] By validating the pre-trained model on the test set, a confidence weight generation model is finally obtained, which takes confidence analysis data as input and outputs several confidence weight coefficients.

[0166] This embodiment uses a pre-trained confidence weight generation model to dynamically generate the weight coefficients corresponding to the predicted confidence level. This enables a more accurate calculation of the final fire confidence level when weighting and fusing the predicted confidence levels of multiple fire labels. This not only improves the accuracy of fire confidence assessment but also increases computational efficiency, enhancing the feasibility and response speed of the model in practical applications.

[0167] In this embodiment, the confidence threshold for generating fire tags based on fire data includes:

[0168] The initial confidence threshold CZY is extracted from the initial threshold table based on the scene type label; the initial threshold table is set by experts based on experience, as shown in the table below:

[0169]

[0170] Extract the historical fire count (HC) and the corresponding fire severity level from the fire data;

[0171] The average severity level (PYD) is obtained by averaging the fire severity levels corresponding to the number of fires.

[0172] A confidence function ZDF(CZY, HC, PYD) is constructed based on the nonlinear relationship between the initial confidence threshold CZY, the number of historical fires HC, the average severity level PYD and the confidence threshold.

[0173] The confidence function satisfies the following calculation formula:

[0174] ;in, and These are respectively represented as the risk sensitivity coefficient and the frequency saturation coefficient; and ∈(0, 1), the specific value is set according to experience, and in this embodiment it will be and Set them to 0.3 and 0.6 respectively; set This is to control the rate at which the confidence threshold decreases, by setting... This is to avoid the unlimited amplification of the confidence threshold by HC;

[0175] The initial confidence threshold, the number of historical fires, and the average severity level are substituted into the confidence function to calculate the confidence threshold corresponding to the fire label.

[0176] This embodiment dynamically adjusts the confidence threshold based on different scenario labels and their corresponding fire data, making the threshold more closely match the actual operating conditions of the current scenario. For similar scenarios with frequent fires, the early warning sensitivity is appropriately increased, thereby effectively reducing the probability of fires and further improving the accuracy and practicality of the fire prediction method.

[0177] Please see Figure 2 Another embodiment of this application provides a fire prediction system based on a large machine model, including: a data acquisition module, a data analysis module, and an early warning module; the data acquisition module and the data analysis module are electrically and / or communicatively connected; the data analysis module and the early warning module are electrically and / or communicatively connected.

[0178] Data acquisition module: Acquires scene data, environmental data, equipment operation data, and fire protection data through data acquisition equipment; scene data refers to scene data that requires fire early warning monitoring; data acquisition equipment includes several sensors, etc.

[0179] Data Analysis Module: Inputs scene data into the scene recognition model to obtain scene type labels; extracts several abnormal parameters from environmental data and equipment operation data; generates predicted labels and predicted confidence scores based on abnormal parameters and scene type labels; obtains the fire confidence score corresponding to the fire label by fusing the predicted confidence scores corresponding to several fire labels through confidence score weighting coefficients; generates the confidence score threshold corresponding to the fire label based on fire protection data; when the fire confidence score is greater than the confidence score threshold, generates a fire warning signal.

[0180] Early warning module: Issues corresponding signal prompts based on early warning signals, including early warning signals indicating the occurrence of a fire.

[0181] Some of the data in the above formula are calculated by removing dimensions and taking their numerical values. The formula is the closest to the real situation obtained by software simulation of a large amount of collected data. The preset parameters and preset thresholds in the formula are set by those skilled in the art according to the actual situation or obtained through simulation of a large amount of data.

[0182] The working principle of this application is as follows: Scene data, environmental data, equipment operation data, and fire protection data are acquired; the scene data is input into a scene recognition model to obtain scene type labels; several abnormal parameters are extracted from the environmental data and equipment operation data; prediction labels and prediction confidence scores are generated based on the abnormal parameters and scene type labels; the fire confidence scores corresponding to several fire labels are fused using confidence weight coefficients to obtain the fire confidence scores corresponding to the fire labels; a confidence threshold for the fire labels is generated based on the fire protection data; when the fire confidence score is greater than the confidence threshold, a fire warning signal is generated. The confidence threshold is dynamically adjusted under different scene labels and their corresponding fire protection data, so that when the prediction results are compared with the confidence threshold, they can better fit the actual scene, improving the accuracy of fire prediction and avoiding the problem that existing technologies often compare prediction results with fixed thresholds, resulting in low accuracy of fire prediction methods.

[0183] The above embodiments are only used to illustrate the technical methods of this application and are not intended to limit it. Although this application has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of this application without departing from the spirit and scope of the technical methods of this application.

Claims

1. A fire prediction method based on a large machine model, characterized in that, include: Acquire scene data, environmental data, equipment operation data, and fire protection data; the scene data refers to the scene data that needs to be monitored for fire early warning. The fire protection data shall include at least the number of historical fires and the severity level of the fires; The scene data is input into the scene recognition model to obtain the scene type label; The scene recognition model is constructed using an artificial intelligence model and is used to identify the scene type label corresponding to the scene data. Extract several abnormal parameters from environmental data and equipment operation data; the abnormal parameters refer to parameters in the environmental data and equipment operation data that deviate from the normal range. Predictive labels and prediction confidence scores are generated based on abnormal parameters and scene type labels; the predictive labels include fire labels and no-fire labels. The fire confidence level corresponding to a fire label is obtained by fusing the predicted confidence levels of several fire labels using confidence weight coefficients. Confidence thresholds for fire labels are generated based on fire data. When the fire confidence level is greater than the confidence threshold, a fire warning signal is generated. The confidence weight coefficients are generated in the following ways: Extract the sequence weight coefficients and their corresponding fire label prediction confidence, and fuse the prediction confidence of the fire label corresponding to the prediction sequence. The predicted confidence score is calculated as the difference between the predicted confidence score of the fire tag with outlier parameters and the predicted confidence score of the fire tag corresponding to the fused predicted sequence; the outlier parameters correspond one-to-one with the sequence weight coefficients. Obtain confidence level labels; each confidence level label corresponds one-to-one with an analysis label and is used to locate the number corresponding to several predicted confidence levels; The confidence labels and their corresponding sequence weight coefficients, the predicted confidence of the fire labels, and the predicted confidence of the difference are integrated into confidence analysis data. The confidence analysis data is input into the confidence weight generation model to obtain several confidence weight coefficients; the confidence weight generation model is constructed through a machine learning model and is used to evaluate the confidence weight coefficients corresponding to the confidence labels.

2. The fire prediction method based on a large machine model according to claim 1, characterized in that, The scene recognition model is constructed using an artificial intelligence model, including: Retrieve historical scene data corresponding to several scene type tags; Several enhanced scenario data were obtained by using data augmentation techniques on several historical scenario data. Integrate several historical scene data and several enhanced scene data under the same scene type label into historical scene data corresponding to the scene type label; Historical scene data is divided into training data, validation data, and test data according to scene type labels; and data preprocessing is performed on the training data, validation data, and test data to obtain training set, validation set, and test set; Choose an artificial intelligence model as the base model; The base model is trained on the training set, and the learning rate and other hyperparameters are adjusted on the validation set to obtain the pre-trained model. By validating the pre-trained model on the test set, a scene recognition model is finally obtained, which takes scene data as input and outputs scene type labels.

3. The fire prediction method based on a large machine model according to claim 1, characterized in that, The step of generating prediction labels and their corresponding prediction confidence scores based on abnormal parameters and scene type labels includes: Extract several historical abnormal time points corresponding to several abnormal parameters; Based on historical anomaly time points, the parameter prediction sequence corresponding to the anomaly parameters is obtained; Generate sequence weight coefficients based on scene type labels; Several parameter prediction sequences are fused using sequence weight coefficients to obtain a fused prediction sequence. Obtain analysis tags; the analysis tags are tags corresponding to the anomaly prediction sequences, and the anomaly prediction sequences include parameter prediction sequences and fusion prediction sequences; Several analysis labels and their corresponding anomaly prediction sequences are input into a multimodal fire prediction model to obtain the prediction labels and prediction confidence levels corresponding to the analysis labels; the multimodal fire prediction model is constructed using a large machine model.

4. The fire prediction method based on a large machine model according to claim 3, characterized in that, The method of obtaining the parameter prediction sequence corresponding to the abnormal parameters based on historical abnormal time points includes the following steps: Step 1: Obtain several historical abnormal time points corresponding to several abnormal parameters and calculate the total number of abnormal time points corresponding to several abnormal parameters. Step 2: Extract the maximum value among the total number of abnormal time points corresponding to several abnormal parameters, and take the maximum value among the total number of abnormal time points as the maximum total number of abnormal time points; Determine if the total number of the largest abnormal time points is within the time window size; Yes, obtain historical time point data corresponding to several abnormal parameters based on the total number of maximum abnormal time points, and proceed to step four; the historical time point data includes abnormal parameters and normal parameters corresponding to the abnormal parameters. This step is to standardize the length of the abnormal prediction sequence. No, when the total number of maximum abnormal time points is greater than the maximum value in the time window size range, extract the historical time point data corresponding to several abnormal parameters based on the maximum value in the time window size range, and proceed to step four; when the total number of maximum abnormal time points is less than the minimum value in the time window size range, extract the historical time point data corresponding to several abnormal parameters based on the minimum value in the time window size range, and proceed to step four. Step 4: Integrate the historical time point data into a parameter prediction sequence corresponding to the abnormal parameters according to the chronological order.

5. The fire prediction method based on a large machine model according to claim 3, characterized in that, The generation of sequence weight coefficients based on scene type labels includes: The initial weights of several abnormal parameters are obtained from the initial weight table based on the scene type label; Extract the historical abnormal time point data and abnormal duration corresponding to the abnormal parameters; An anomaly score is generated based on several historical anomaly time points and anomaly durations. The normalized normalized parameter anomaly score GCYP is obtained by performing max-min normalization on the anomaly scores of several parameters. Based on the nonlinear relationship between the initial weight CQ, the normalized parameter anomaly score GCYP and the sequence weight coefficient, a sequence weight calculation function XQF(CQ,GCYP) is constructed. Substitute the initial weights and normalized parameter anomaly scores into the sequence weight calculation function to obtain the sequence weight coefficients.

6. The fire prediction method based on a large machine model according to claim 5, characterized in that, The method of generating parameter anomaly scores based on several historical anomaly time points and anomaly durations includes: Extract several historical anomaly point data and anomaly duration YS corresponding to the anomaly parameters; Several abnormal values ​​are obtained by calculating the difference between several historical outlier data points and the normal data range corresponding to the outlier parameters. The average value of several abnormal values ​​is calculated to obtain the abnormal mean value YJ corresponding to the abnormal parameter; The abnormal growth rate YZL is calculated based on several abnormal values ​​in chronological order. Calculate the standard deviation of the abnormal growth rate YZBC based on several abnormal growth rates; Through formula Calculate the parameter anomaly score corresponding to the abnormal parameters. Where i represents the number corresponding to the abnormal parameter; and These are represented as the data anomaly sensitivity coefficient and the duration response coefficient, respectively. and All are greater than 0.

7. The fire prediction method based on a large machine model according to claim 3, characterized in that, The multimodal fire prediction model is constructed using a large machine learning model, including: Obtain several analysis labels and their corresponding historical anomaly prediction sequences, historical prediction labels, and prediction confidence levels corresponding to the historical prediction labels; Several analysis labels and their corresponding historical anomaly prediction sequences, historical prediction labels, and prediction confidence scores corresponding to historical prediction labels are divided into training data, validation data, and test data corresponding to the analysis labels; and data preprocessing is performed on the training data, validation data, and test data corresponding to the analysis labels to obtain the training set, validation set, and test set corresponding to the analysis labels. Choose the large machine model as the base model for analyzing the corresponding labels; The basic model corresponding to the analysis label is trained by analyzing the training set corresponding to the analysis label, and the learning rate and other hyperparameters are adjusted on the validation set corresponding to the analysis label to obtain the pre-trained model corresponding to the analysis label. By validating the pre-trained model corresponding to the analysis label on the test set corresponding to the analysis label, a multimodal fire prediction model is finally obtained with the analysis label and its corresponding anomaly prediction sequence as input and the prediction label and prediction confidence of the analysis label as output.

8. The fire prediction method based on a large machine model according to claim 1, characterized in that, The confidence threshold for generating fire labels based on fire data includes: The initial confidence threshold CZY is extracted from the initial threshold table based on the scene type label; Extract the historical fire count (HC) and the corresponding fire severity level from the fire protection data; The average severity level (PYD) is obtained by averaging the fire severity levels corresponding to the number of fires. A confidence function ZDF(CZY, HC, PYD) is constructed based on the nonlinear relationship between the initial confidence threshold CZY, the number of historical fires HC, the average severity level PYD and the confidence threshold. The initial confidence threshold, the number of historical fires, and the average severity level are substituted into the confidence function to calculate the confidence threshold corresponding to the fire label.

9. A fire prediction system based on a large machine model, applied to the fire prediction method based on a large machine model as described in any one of claims 1-8, characterized in that, include: Data acquisition module, data analysis module, and early warning module; The data acquisition module is connected to the data analysis module; the data analysis module is connected to the early warning module. The data acquisition module acquires scene data, environmental data, equipment operation data, and fire protection data through data acquisition devices; the scene data refers to scene data that requires fire early warning monitoring. The data analysis module: inputs scene data into the scene recognition model to obtain scene type labels; extracts several abnormal parameters from environmental data and equipment operation data; generates predicted labels and predicted confidence scores based on the abnormal parameters and scene type labels; obtains the fire confidence score corresponding to the fire label by fusing the predicted confidence scores corresponding to several fire labels through confidence score weighting coefficients; and generates the confidence threshold corresponding to the fire label based on fire protection data. When the fire confidence level is greater than the confidence threshold, a fire warning signal is generated. The early warning module issues corresponding signal prompts based on the early warning signal.

Images