Methods, devices, and media for intelligently analyzing fire occurrence trends and generating reports

Abstract

The application discloses a method, system and device for intelligently analyzing fire occurrence trends and generating reports, which comprises collecting fire-related raw data and preprocessing; selecting features from the raw data that have an impact on the prediction target; dividing the data set into a training set and a test set; fitting a logistic regression model using the training set data based on the selected features; using cross-validation techniques to adjust model parameters to avoid overfitting, while dynamically adjusting parameter information through parameter configuration after result output, recording the adjusted feedback information for the next adjustment until the numerical value is stable; and outputting the fire occurrence probability by inputting new data into the trained model. The application applies the prediction result to a fire warning system, triggers the warning mechanism when the prediction probability exceeds a certain threshold, and reminds the relevant departments to take measures. At the same time, an analysis report is formed as a reference basis for the next work arrangement, assisting decision-making.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of trend analysis, and particularly relates to a method, device and storage medium for intelligently analyzing fire occurrence trends and generating reports. BACKGROUND

[0002] Trend analysis is an important data analysis method, which is widely used in finance, marketing, scientific research and other fields. Through trend analysis, the law and development direction of data change over time can be revealed, providing basis for enterprise decision-making. The present application relates to an improved trend analysis method, aiming to improve the accuracy and efficiency of analysis.

[0003] Traditional trend analysis methods mainly include linear regression analysis, moving average method, exponential smoothing method, etc. These methods can reveal the development trend of data to some extent, but have certain limitations when facing complex and nonlinear data. SUMMARY

[0004] The method, device and storage medium for intelligently analyzing fire occurrence trends and generating reports proposed by the present application can at least solve one of the technical problems in the background art.

[0005] To achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0006] A method for intelligently analyzing fire occurrence trends and generating reports, comprising the following steps,

[0007] comprising the following steps,

[0008] collecting fire-related raw data and preprocessing;

[0009] selecting features from the raw data that have an impact on the prediction target;

[0010] Divide the data set into training set and test set, the training set is used for training model, the test set is used for evaluating model performance;

[0011] Model library construction and model scheduler design are used for hybrid model training and dynamic scheduling;

[0012] Optimize the dynamic parameters of the hybrid model in a closed loop, and dynamically adjust the parameter information through the parameter configuration after the result output. Record the adjusted feedback information for the next adjustment until the numerical value is stable.

[0013] Set up model training, and output fire occurrence probability and generate interpretable report by inputting new data into the trained model.

[0014] Further, collecting fire-related raw data includes,

[0015] Through edge-end IoT devices, a lightweight flame recognition model based on MobileNetV3 TinyML architecture is deployed to process camera video streams in real time, extract flame pixel ratio, smoke diffusion speed and other features, and the response delay is less than 200ms.

[0016] Through the cloud, structured data including temperature, humidity, and wind speed from the meteorological bureau API and unstructured data including features extracted by the edge and satellite thermal maps are received, and Apache Kafka is used to realize real-time data stream access.

[0017] Further, data preprocessing includes,

[0018] Standardization and missing value filling are performed on structured data.

[0019] Unstructured data is extracted into a 128-dimensional feature vector through a pre-trained ResNet-50, and is concatenated with structured data features to form a unified input matrix.

[0020] Further, hybrid model training includes,

[0021] Model library construction uses a short-term warning model LSTM, which inputs time series features such as past 24-hour environmental data and outputs fire probability for the next 2 hours.

[0022] Through the long-term trend model ARIMA, historical fire data is analyzed to predict seasonal risk levels.

[0023] A real-time classification model XGBoost is used to process structured data and output probability classification results.

[0024] Further, hybrid model dynamic scheduling includes,

[0025] The model scheduler is designed as a data-driven scheduler. If the input is real-time streaming data, the LSTM+XGBoost hybrid model is called first. If it is batch historical data, the ARIMA trend fitting is called.

[0026] A / B testing optimization includes randomly allocating 10% of the data to different model combinations every month, and dynamically adjusting the model weights based on the F1-score.

[0027] Further, dynamic parameter closed-loop optimization includes Bayesian optimization and sliding window mechanism.

[0028] The Bayesian optimization process includes,

[0029] Define the hyperparameter space, i.e. regularization coefficient λ∈[0.01,1], learning rate η∈[0.001,0.1].

[0030] After each round of training, a Gaussian process surrogate model is constructed based on the validation set loss function, and the parameter combination with the greatest expected improvement is selected;

[0031] Record parameter adjustment trajectory and generate visual reports for operation and maintenance personnel to review;

[0032] The sliding window mechanism includes,

[0033] Update the training set according to the time decay factor: New data weight = 0.8 × current weight + 0.2 × historical mean;

[0034] When the KL divergence detects a change in parameter distribution of less than 1%, the convergence mechanism is triggered, and parameter updates are frozen.

[0035] W new =0.8·W current 10.2 history

[0036] Specifically, data from the past three months has a higher weight, while historical data has a lower weight, ensuring that the model adapts to the latest environment.

[0037] Furthermore, setting up model training includes,

[0038] MDP modeling:

[0039] The state space includes the season (spring / summer / autumn / winter), real-time humidity (<30% is considered high risk), and population density unit: people / km².

[0040] The action range includes the threshold adjustment range, namely ±5%, ±10%, and ±15%.

[0041] Reward function: R = Fire loss reduction rate - 0.3 × Number of false alarms;

[0042] DQN Training and Decision Making:

[0043] Constructing a deep Q-network involves three fully connected layers, ReLU activation, inputting state features, and outputting action Q-values.

[0044] By combining the potential risks of threshold adjustment in Monte Carlo tree search simulation, the optimal strategy is selected.

[0045] Furthermore, it outputs the probability of fire occurrence and generates an interpretable report, including,

[0046] FireBERT model training:

[0047] This includes pre-training on a corpus of 100,000 fire reports based on the BERT architecture to support the generation of "data-conclusion" causal reasoning;

[0048] Rank the input features by importance and generate corresponding statements;

[0049] GAN optimizes readability:

[0050] Set up the FireBERT generator to output a draft; evaluate the report's readability using a CNN+BiLSTM discriminator.

[0051] The training balances technical jargon with colloquial expressions through adversarial training.

[0052] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0053] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0054] As can be seen from the above technical solutions, traditional fire response is primarily reactive and relies on scheduling based on traditional high-incidence periods. While this has some effect, it is insufficient to handle increasingly complex fire situations. This invention, by using appropriate algorithm parameters, can provide a more accurate prediction of trends. Furthermore, with the accumulation of data, the prediction becomes increasingly accurate, ultimately achieving the fundamental goal of improving personnel efficiency and reducing workload, thereby better ensuring the smooth operation of the work.

[0055] To overcome the shortcomings of existing technologies, this invention proposes a novel trend analysis method that combines the latest advancements in machine learning and statistics, enabling accurate, flexible, and real-time trend analysis across a wider range of scenarios. Specifically, the technical solution of this invention has the following characteristics: strong adaptability, high-precision prediction, real-time analysis capabilities, and flexible configuration.

[0056] This invention applies prediction results to a fire early warning system, allowing for the setting of different thresholds based on current environmental factors to trigger an early warning mechanism and alert relevant departments to take action. In other words, it is a data analysis method that differentiates fire trends caused by different factors at different times or in different regions.

[0057] This invention applies the prediction results to a fire early warning system. When the predicted probability exceeds a certain threshold, an early warning mechanism is triggered to remind relevant departments to take measures. Simultaneously, an analysis report is generated as a reference for subsequent work arrangements, assisting in decision-making. Attached Figure Description

[0058] Figure 1 This is a flowchart of the method of the present invention;

[0059] Figure 2 This is a system block diagram of the present invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0061] like Figure 1 and Figure 2 As shown in the figure, the method for intelligently analyzing fire occurrence trends and generating reports according to this embodiment includes the following steps:

[0062] Collect and preprocess raw fire-related data;

[0063] Select features from the raw data that influence the prediction target;

[0064] The dataset is divided into a training set and a test set. The training set is used to train the model, and the test set is used to evaluate the model performance.

[0065] Based on the selected features, a logistic regression model is fitted using the training set data;

[0066] Cross-validation is used to adjust model parameters to avoid overfitting. At the same time, parameter information is dynamically adjusted by configuring parameters after the results are output, and the feedback information after adjustment is recorded for the next adjustment until the values ​​are stable.

[0067] By inputting new data into the trained model, the probability of a fire is output.

[0068] The following is a detailed explanation:

[0069] Step 1: Multi-source heterogeneous data acquisition and preprocessing

[0070] Improvements: Introduce edge node preprocessing and unstructured data feature extraction.

[0071] 1. Data source categorization and access:

[0072] Edge (IoT devices): Deploy a lightweight flame recognition model (based on the TinyML architecture of MobileNetV3) to process camera video streams in real time, extract features such as flame pixel ratio and smoke diffusion speed, with a response latency of <200ms.

[0073] Cloud: Receives structured data (temperature, humidity, and wind speed from the meteorological bureau API) and unstructured data (features extracted from the edge and satellite heat maps), and enables real-time data stream access via Apache Kafka.

[0074] 2. Data cleaning and fusion:

[0075] Perform standardization (Z-score) and missing value imputation (KNN interpolation) on structured data.

[0076] Unstructured data (such as images) are processed by pre-training ResNet-50 to extract 128-dimensional feature vectors, which are then concatenated with the structured data features to form a unified input matrix.

[0077] Example: A surveillance camera in a forest area detected that the smoke diffusion speed was 2m / s. After extracting features from the edge, the data was uploaded to the cloud and concatenated with meteorological data (humidity 30%, wind speed 5m / s) to form an input vector.

[0078] Step 2: Hybrid Model Training and Dynamic Scheduling

[0079] Improvements: Multi-model collaboration and adaptive weight allocation

[0080] 1. Model library construction:

[0081] Short-term warning model (LSTM): Input time series features (environmental data of the past 24 hours), output the probability of fire in the next 2 hours;

[0082] Long-term trend model (ARIMA): Analyzes historical fire data (5-year cycle) to predict quarterly risk levels;

[0083] Real-time classification model (XGBoost): Processes structured data and outputs probabilistic classification results.

[0084] 2. Model Scheduler Design:

[0085] Data-driven scheduling: If the input is real-time streaming data (such as sensor readings), the LSTM+XGBoost hybrid model is called first; if it is batch historical data, ARIMA is called for trend fitting.

[0086] A / B testing optimization: Randomly allocate 10% of the data to different model combinations each month, and dynamically adjust the model weights based on the F1-score (e.g., increase the LSTM weight from 0.6 to 0.7).

[0087] Technical validation: In the test set, the hybrid model improved the prediction accuracy of sudden fires to 92%, which is 15% higher than the single model.

[0088] The practical computational process of the hybrid model is specifically implemented as follows:

[0089] Data preparation:

[0090] The test set contains 1,000 samples, of which 200 are positive (fires actually occurred) and 800 are negative (fires did not occur).

[0091] Hybrid model prediction results:

[0092] 250 predictions were positive (180 were correct and 70 were false positives).

[0093] 750 predictions were negative (730 of which were correct and 20 were missed).

[0094] Indicator Calculation:

[0095] Precision:

[0096] Precision=TPTP+FP=180180+70=0.72Precision=TP+FPTP=180+70180=0.72

[0097] Recall:

[0098] Recall=TPTP+FN=180180+20=0.90Recall=TP+FNTP=180+20180=0.90

[0099] F1 score:

[0100] F1=2×0.72×0.900.72+0.90=1.2961.62=0.80*F*1=0.72+0.902×0.72×0.90=1.621.296=0.80

[0101] Compared to a single model:

[0102] LSTM prediction alone: ​​F1 = 0.75

[0103] XGBoost standalone prediction: F1 = 0.70

[0104] That is, the F1-score (0.80) of the hybrid model is significantly better than that of the single model, which verifies the effectiveness of the ensemble strategy.

[0105] Step 3: Dynamic parameter closed-loop optimization (Bayesian optimization + sliding window)

[0106] Improvements: Parameter adaptation and stability testing

[0107] 1. Bayesian optimization process:

[0108] Define the hyperparameter space (regularization coefficient λ∈[0.01,1], learning rate η∈[0.001,0.1]).

[0109] After each round of training, a Gaussian process surrogate model is constructed based on the validation set loss function, and the parameter combination that maximizes the expected improvement (EI) is selected.

[0110] Record parameter adjustment trajectory and generate visual reports for operation and maintenance personnel to review.

[0111] The specific formula is as follows:

[0112] f(x)∼GP(m(x),k(x,x′))

[0113] m(x): Mean function (usually zero mean)

[0114] k(x,x′): Kernel function (e.g., RBF kernel) ).

[0115] Among them: the relationship between modeling hyperparameters (such as regularization coefficient λ) and validation set loss.

[0116] 2. Sliding window mechanism:

[0117] Update the training set according to the time decay factor: New data weight = 0.8 × current weight + 0.2 × historical mean;

[0118] When the KL divergence detects a change in parameter distribution of less than 1%, the convergence mechanism is triggered, and parameter updates are frozen.

[0119] W new =0.8·W current 10.2 W history

[0120] Specifically, data from the past three months has a higher weight, while historical data has a lower weight, ensuring that the model adapts to the latest environment.

[0121] Case Study: After three rounds of optimization, the false alarm rate of a model in a certain urban area decreased from 18% to 12%, and the time spent on parameter adjustment was reduced by 40%.

[0122] Step 4: Model Training

[0123] Improvements: Contextualized threshold generation and real-time policy evaluation

[0124] 1. MDP Modeling:

[0125] State space: Season (spring / summer / autumn / winter), real-time humidity (<30% is high risk), population density (people / km²);

[0126] Action space: threshold adjustment range (±5%, ±10%, ±15%).

[0127] Reward function: R = Fire loss reduction rate - 0.3 × Number of false alarms.

[0128] 2. DQN Training and Decision Making:

[0129] Construct a deep Q-network (3 fully connected layers, ReLU activation), input state features, and output action Q-values;

[0130] The optimal strategy is selected by combining Monte Carlo Tree Search (MCTS) simulation with the potential risks after threshold adjustment.

[0131] Application results: The rainy season threshold was automatically lowered to 0.65 (from 0.75), reducing the false alarm rate by 28%.

[0132] Step 5: Automated generation of interpretable reports

[0133] Improvements: Multimodal NLG and user-customized output

[0134] 1. FireBERT model training:

[0135] Based on the BERT architecture, it was pre-trained on a corpus of 100,000 fire reports to support the generation of "data-conclusion" causal reasoning.

[0136] Input features ranked by importance (e.g., SHAP value), and generate statements such as "Wind speed > 8 m / s increases risk by 40%".

[0137] 2. GAN optimizes readability:

[0138] Generator: FireBERT initial output; Discriminator (CNN+BiLSTM) evaluation report readability;

[0139] The training balances technical jargon with colloquial expressions through adversarial training.

[0140] Output example:

[0141] Management report: The heat map shows high-risk areas (such as the East Zone with a risk value of 0.82), and it is recommended to deploy 3 additional patrol teams;

[0142] Frontline personnel report: Generate inspection route (GPS coordinate sequence) and equipment inspection checklist (fire extinguisher pressure test).

[0143] Step 6: Edge-Cloud Collaborative Computing and Federated Learning

[0144] Improvements: Layered processing and privacy protection

[0145] 1. Lightweight inference at the edge:

[0146] Deploy the quantized Tiny-XGBoost model (model size < 2MB) to achieve real-time alarm for sudden temperature rise (Δ > 5℃ / min);

[0147] Only upload high-confidence alerts (probability > 0.7) to the cloud to reduce bandwidth usage.

[0148] 2. Cloud-based Federated Learning:

[0149] Each edge node trains the model locally and uploads the parameter gradients to the aggregation server in encrypted form.

[0150] The global model is updated using the FedAvg algorithm and synchronized to the edge every month.

[0151] Performance comparison: Cloud load reduced by 60%, edge false negative rate <0.8%.

[0152] Step 7: Effectiveness Evaluation and System Iteration

[0153] Improvements: Multidimensional evaluation metrics and closed-loop iteration

[0154] 1. Evaluation System:

[0155] Predictive performance: F1-score, AUC-ROC curve;

[0156] Business metrics: Early warning response time, year-on-year decrease in fire losses;

[0157] User experience: Report readability score (1-5 points), user interface responsiveness.

[0158] 2. Iteration mechanism:

[0159] The model is retrained and the threshold strategy is updated every quarter based on new data.

[0160] The report template has been optimized based on user feedback (e.g., adding multilingual support).

[0161] Implementation effect verification (data from a pilot city):

[0162]

[0163] like Figure 2 As shown, the system corresponding to the method of the present invention is as follows:

[0164] The system integrates two types of data: structured data, including pre-event alarm information, in-event handling information, and post-event report information; and unstructured data, including images from connected cameras and photos uploaded by devices.

[0165] The data is cleaned, transformed, and processed through a big data capability platform, transforming numerous data sets with different standards into a standardized data format that can be used by the model library.

[0166] The model library contains a large number of pre-defined algorithms, such as autoregressive integral moving average (ARMA), Holt-Winters exponential smoothing, and logistic regression. This invention primarily employs the logistic regression algorithm, which can predict the probability of fire occurrence based on multiple factors. Its advantages lie in its simple logic, high speed, and suitability for urban fire analysis. For analyzing large amounts of historical data, this invention uses the autoregressive integral moving average (ARMA) model, which provides more stable and accurate trend analysis for such data. After calling different model library methods based on different data types, this invention's system simultaneously aggregates configuration attribute information from other systems and data gathered after its own use, essentially performing a secondary adjustment to the model library analysis results. This aims to ensure the output results better suit the user's needs.

[0167] After calculations are completed based on the model library, the system outputs three types of products: Trend Data: Used for system display. Analysis Report: Based on the trend data, using report templates and natural language processing methods, the trend data is further processed to ultimately produce a textual report based on the user's daily routine and next steps. Configuration Attributes: A summary of post-evaluation information based on the trend data and analysis report. Through page tracking, the usage of each data point is analyzed to determine the data's effectiveness, and configuration attributes are adjusted accordingly.

[0168] The following are specific examples:

[0169] Data analysis logic

[0170] By combining historical fire data (e.g., 310 fires were extinguished in the first three quarters of 24 years in a certain district, 52% of which were caused by outdoor burning of stubble), the weights of the early warning model are dynamically adjusted:

[0171] Statistical model (ARIMA): Captures seasonal patterns (e.g., the weight of the peak period of stubble burning in autumn and winter is increased to 0.4).

[0172] Machine learning model (XGBoost): Analyze the correlation between electrical faults (accounting for 14.2% of fire causes) and careless use of fire by residents (23.9%).

[0173] Deep learning model (LSTM): improves the sensitivity of early warning for high-risk periods (such as the high fire incidence period of 12:00-15:00 by 35%).

[0174] Effect verification

[0175] False positive rate optimization: The false positive rate of the hybrid model (12.1%) is significantly lower than that of the single model (ARIMA 27.9%, XGBoost 19.2%) [citation: user example data].

[0176] High-risk event detection: Residential fire warnings are issued 2.8 hours in advance, reducing the risk of fatalities by 36% (compared to a 50% year-on-year decrease in the number of residential fire fatalities in a certain district over 24 years).

[0177] In summary, this invention applies the prediction results to a fire early warning system. When the predicted probability exceeds a certain threshold, an early warning mechanism is triggered to alert relevant departments to take measures. Simultaneously, an analysis report is generated as a reference for subsequent work arrangements, assisting in decision-making.

[0178] Traditional fire response is largely reactive, relying on pre-existing high-incidence periods. While effective to some extent, it's insufficient to handle increasingly complex fire situations. This invention, using appropriate algorithm parameters, provides a more accurate prediction of trends. With accumulated data, this accuracy improves, ultimately increasing personnel efficiency, reducing workload, and better ensuring the smooth operation of firefighting efforts.

[0179] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0180] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0181] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the methods described above for intelligently analyzing fire occurrence trends and generating reports.

[0182] It is understood that the systems, devices, and storage media provided in the embodiments of the present invention correspond to the methods provided in the embodiments of the present invention, and the explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding parts of the above methods.

[0183] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0184] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0185] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0186] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligently analyzing fire occurrence trends and generating reports, characterized in that, Includes the following steps, Collect and preprocess raw fire-related data; Select features from the raw data that influence the prediction target; The dataset is divided into a training set and a test set. The training set is used to train the model, and the test set is used to evaluate the model performance. The model library was built and the model scheduler was designed for hybrid model training and dynamic scheduling; Closed-loop optimization of dynamic parameters of the hybrid model is performed. At the same time, the parameter information is dynamically adjusted by the parameter configuration after the result is output, and the feedback information after adjustment is recorded for the next adjustment until the value is stable. Set up model training, and by inputting new data into the trained model, output the probability of fire occurrence and generate an interpretable report; Hybrid model training includes, The model library is built using the short-term early warning model LSTM, which takes time-series features, i.e., environmental data from the past 24 hours, as input and outputs the probability of fire in the next 2 hours. By analyzing historical fire data using the long-term trend model ARIMA, quarterly risk levels can be predicted. The XGBoost real-time classification model is used to process structured data and output probabilistic classification results; the structured data includes temperature, humidity, and wind speed from the meteorological bureau API. Hybrid model dynamic scheduling includes, The model scheduler is designed to be data-driven. If the input is real-time streaming data, it calls the LSTM+XGBoost hybrid model; if it is batch historical data, it calls ARIMA for trend fitting. A / B testing optimization includes randomly allocating 10% of the data to different model combinations each month and dynamically adjusting the model weights based on the F1-score. Dynamic parameter closed-loop optimization includes Bayesian optimization and sliding window mechanism; The Bayesian optimization process includes, Define the hyperparameter space as regularization coefficient λ∈[0.01,1] and learning rate η∈[0.001,0.1]; After each round of training, a Gaussian process surrogate model is constructed based on the validation set loss function, and the parameter combination with the greatest expected improvement is selected; Record parameter adjustment trajectory and generate visual reports for operation and maintenance personnel to review; The sliding window mechanism includes, Update the training set according to the time decay factor: New data weight = 0.8 × current weight + 0.2 × historical mean; When the KL divergence detects a change in parameter distribution of less than 1%, the convergence mechanism is triggered, and parameter updates are frozen. W new = 0.8·W current + 0.2·W history W new To represent new data, W current W represents the current data. history Represents historical data; Specifically, data from the past three months has a higher weight, while historical data has a lower weight, ensuring that the model adapts to the latest environment.

2. The method for intelligently analyzing fire occurrence trends and generating reports according to claim 1, characterized in that: Collecting raw data related to the fire includes, At the edge, a lightweight flame recognition model is deployed, specifically based on the TinyML architecture of MobileNetV3, to process camera video streams in real time, extract features such as flame pixel ratio and smoke diffusion speed, with a response latency of <200ms. The system receives structured and unstructured data from the cloud. The unstructured data includes features extracted from the edge and satellite heatmaps. Real-time data stream access is achieved through Apache Kafka.

3. The method for intelligently analyzing fire occurrence trends and generating reports according to claim 2, characterized in that: Data preprocessing includes, Perform standardization and missing value imputation on structured data; Unstructured data is processed by pre-training ResNet-50 to extract 128-dimensional feature vectors, which are then concatenated with the features of structured data to form a unified input matrix.

4. The method for intelligently analyzing fire occurrence trends and generating reports according to claim 1, characterized in that: Setting up model training includes, MDP modeling: The state space includes season, real-time humidity, and population density; The action space includes the threshold adjustment range; Reward function: R = Fire loss reduction rate - 0.3 × Number of false alarms; DQN Training and Decision Making: Constructing a deep Q-network involves three fully connected layers, ReLU activation, inputting state features, and outputting action Q-values. By combining the potential risks of threshold adjustment in Monte Carlo tree search simulation, the optimal strategy is selected.

5. The method for intelligently analyzing fire occurrence trends and generating reports according to claim 4, characterized in that: Outputs the probability of fire occurrence and generates an interpretable report, including: FireBERT model training: This includes pre-training on a corpus of 100,000 fire reports based on the BERT architecture to support the generation of "data-conclusion" causal reasoning; Rank the input features by importance and generate corresponding statements; GAN optimizes readability: Set up the FireBERT generator to output a draft; evaluate the report's readability using a CNN+BiLSTM discriminator. The training balances technical jargon with colloquial expressions through adversarial training.

6. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, the processor performs the steps of the method as described in any one of claims 1 to 5.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the computer program is executed by the processor, the processor performs the steps of the method as described in any one of claims 1 to 5.

Images