Method and device for intelligently analyzing fire occurrence trend and generating report, and medium

Abstract

The invention discloses a method, a system and equipment for intelligently analyzing a fire occurrence trend and generating a report. The method comprises the following steps: collecting original data related to a fire and pre-processing; selecting features influencing a prediction target from the original data; dividing the data set into a training set and a test set; based on the selected features, using training set data to fit a logistic regression model; a cross validation technology is used for adjusting model parameters, overfitting is avoided, meanwhile, parameter information is dynamically adjusted through parameter configuration after result output, and adjusted feedback information is recorded for next adjustment till the numerical value is stable; and inputting new data into the trained model, and outputting the fire occurrence probability. The prediction result is applied to a fire early warning system, and when the prediction probability exceeds a certain threshold value, an early warning mechanism is triggered to remind related departments to take measures. And meanwhile, an analysis report is formed as a reference basis for next work arrangement to assist decision making.

Description

Technical Field

[0001] The present invention relates to the technical field of trend analysis, and in particular to a method, device and storage medium for intelligently analyzing the trend of fire occurrence and generating a report. Background Art

[0002] Trend analysis is an important data analysis method, which is widely used in many fields such as finance, marketing, and scientific research. Through trend analysis, the law and development direction of data change over time can be revealed, providing a basis for enterprise decision-making. The present invention 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. Although these methods can reveal the development trend of data to a certain extent, they have certain limitations when dealing with complex and non-linear data. Summary of the Invention

[0004] A method, device and storage medium for intelligently analyzing the trend of fire occurrence and generating a report proposed by the present invention can at least solve one of the technical problems in the background art.

[0005] To achieve the above object, the present invention adopts the following technical solutions: A method for intelligently analyzing the trend of fire occurrence and generating a report includes the following steps, including the following steps, Collect original fire-related data and perform preprocessing; Select features that have an impact on the prediction target from the original data; Divide the data set 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 construction of the model library and the design of the model scheduler are used for hybrid model training and dynamic scheduling; Perform closed-loop optimization on the dynamic parameters of the hybrid model. At the same time, dynamically adjust the parameter information through the parameter configuration after the result output, and record the adjusted feedback information for the next adjustment until the value is stable; Set up model training, and by inputting new data into the trained model, output the fire occurrence probability and generate an interpretable report.

[0006] Further, collecting original fire-related data includes, Deploy a lightweight flame recognition model based on the TinyML architecture of MobileNetV3 through the edge side, that is, an Internet of Things device, to process the camera video stream in real time, extract features such as the proportion of flame pixels and the smoke diffusion speed, and the response delay < 200ms; Receive structured data including temperature, humidity, and wind speed from the Meteorological Bureau API and unstructured data including features extracted from the edge and satellite heat maps through the cloud, and achieve real-time access to data streams through Apache Kafka.

[0007] Furthermore, data preprocessing includes, Perform standardization and missing value filling on structured data; The unstructured data is extracted with a 128-dimensional feature vector through pre-training ResNet-50, which is then concatenated with the structured data features into a unified input matrix.

[0008] Furthermore, the hybrid model training includes, The model library is constructed through the short-term warning model LSTM, which inputs the time series features, i.e. the environmental data of the past 24 hours, and outputs the probability of fire in the next 2 hours; Analyze historical fire data through the long-term trend model ARIMA to predict the quarterly risk level; The real-time classification model XGBoost is used to process structured data and output probabilistic classification results.

[0009] Furthermore, the hybrid model dynamic scheduling includes, The model scheduler is designed to be data-driven. If the input is real-time streaming data, the LSTM+XGBoost hybrid model is called first. If it is batch historical data, ARIMA is called for trend fitting. A / B test 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.

[0010] Furthermore, the closed-loop optimization of dynamic parameters includes Bayesian optimization and sliding window mechanism; Among them, the Bayesian optimization process includes: Define the hyperparameter space, i.e., regularization coefficient λ∈[0.01,1], learning rate η∈[0.001,0.1]; After each round of training, a Gaussian process proxy model is constructed based on the validation set loss function, and the parameter combination with the largest expected improvement is selected; Record parameter adjustment trajectories and generate visual reports for review by operation and maintenance personnel; 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 KL divergence detects that the parameter distribution changes by less than 1%, the convergence mechanism is triggered and the parameter update is frozen; W new =0.8 W current 10.2 history Among them: The data of the recent 3 months has a higher weight, and the weight of historical data decays to ensure that the model adapts to the latest environment.

[0011] Furthermore, the model training is set to include MDP modeling: The state space includes seasons, namely spring / summer / autumn / winter, real-time humidity, <30% is high risk, and population density unit: people / km²; The action space includes the threshold adjustment range, namely ±5%, ±10%, ±15%; Reward function: R = fire loss reduction rate - 0.3 × number of false alarms; DQN training and decision-making: Construct a deep Q network including 3 fully connected layers, ReLU activation, input state features, and output action Q values; Combine Monte Carlo tree search to simulate the potential risks after threshold adjustment and select the optimal strategy.

[0012] Furthermore, output the fire occurrence probability and generate an interpretable report, including FireBERT model training: Include pre-training based on the BERT architecture on 100,000 fire report corpora to support the generation of "data-conclusion" causal reasoning; Sort the importance of input features and generate corresponding statements; GAN optimization of readability: Set the generator FireBERT to output the first draft; evaluate the readability of the report through the discriminator CNN + BiLSTM; Balance technical terms and popular expressions through adversarial training.

[0013] On the other hand, the present invention also discloses a computer-readable storage medium storing a computer program, which when executed by a processor causes the processor to execute the steps of the above method.

[0014] On yet another aspect, the present invention also discloses a computer device including a memory and a processor, the memory storing a computer program, which when executed by the processor causes the processor to execute the steps of the above method.

[0015] As can be seen from the above technical solutions, since traditional fire handling is all post-event processing and is arranged according to the high-incidence periods of traditional rules, although it has certain effects, it is not sufficient to support the increasingly complex fire situations. Through this invention, after using appropriate algorithm parameters, a relatively accurate prediction of the trend can be made, and at the same time, with the accumulation of data, it will become more and more accurate, thereby achieving the fundamental purpose of improving personnel efficiency and reducing workload, and further better ensuring the implementation of work.

[0016] To overcome the deficiencies of the prior art, the present invention proposes a new trend analysis method, which combines the latest advances in machine learning and statistics and can achieve accurate, flexible, and real-time trend analysis in a wider range of scenarios. Specifically, the technical solution of the present invention has the following characteristics: strong adaptability, high-precision prediction, real-time analysis ability, and flexible configuration.

[0017] The present invention applies the prediction results to a fire warning system, which can set different thresholds according to current environmental factors to trigger the warning mechanism and remind relevant departments to take measures. That is, a differential data analysis method for fire trends caused by different reasons at different times or in different regions.

[0018] The present invention applies the prediction results to a fire warning system. When the prediction probability exceeds a certain threshold, the warning mechanism is triggered to remind relevant departments to take measures. At the same time, an analysis report is formed as a reference for the next work arrangement to assist decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow block diagram of the method of the present invention; Figure 2 is a system block diagram of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention.

[0021] As Figure 1 and Figure 2 shown, the method for intelligently analyzing the fire occurrence trend and generating a report according to this embodiment includes the following steps: Collect original fire-related data and perform preprocessing; Select features that have an impact on the prediction target from the original data; Divide the data set 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; Based on the selected features, use the training set data to fit a logistic regression model; Use cross-validation techniques to adjust the model parameters to avoid overfitting. At the same time, dynamically adjust the parameter information through the parameter configuration after the result output, and record the adjusted feedback information for the next adjustment until the value is stable; By inputting new data into the trained model, output the fire occurrence probability.

[0022] The following is a specific description: Step 1: Multi-source Heterogeneous Data Collection and Preprocessing Improvement Point: Introduce Edge Node Preprocessing and Unstructured Data Feature Extraction 1. Data Source Classification and Access: Edge Side (IoT Devices): Deploy a lightweight flame recognition model (TinyML architecture based on MobileNetV3) to process camera video streams in real time, extract features such as the proportion of flame pixels and the smoke diffusion speed, and the response delay < 200ms.

[0023] Cloud: Receive structured data (temperature, humidity, wind speed from the meteorological bureau API) and unstructured data (features extracted from the edge side, satellite thermal maps), and achieve real-time access to the data stream through Apache Kafka.

[0024] 2. Data Cleaning and Fusion: Perform standardization (Z-score) and missing value filling (KNN interpolation method) on structured data; Unstructured data (such as images) extract 128-dimensional feature vectors through pre-trained ResNet-50 and splice them with structured data features into a unified input matrix.

[0025] Example: A forest area surveillance camera detects that the smoke diffusion speed is 2m / s. After the edge side extracts the features, it uploads them to the cloud and splices them with meteorological data (humidity 30%, wind speed 5m / s) into an input vector.

[0026] Step 2: Hybrid Model Training and Dynamic Scheduling Improvement Point: Multi-model Collaboration and Adaptive Weight Allocation 1. Model Library Construction: Short-term Early Warning Model (LSTM): Input time series features (environmental data in the past 24 hours) and output the fire probability in the next 2 hours; Long-term Trend Model (ARIMA): Analyze historical fire data (5-year cycle) and predict the quarterly risk level; Real-time Classification Model (XGBoost): Process structured data and output probability classification results.

[0027] 2. Model Scheduler Design: Data-driven Scheduling: If the input is real-time stream data (such as sensor readings), preferentially call the LSTM+XGBoost hybrid model; if it is batch historical data, call ARIMA for trend fitting; A / B Test Optimization: Randomly allocate 10% of the data to different model combinations every month, and dynamically adjust the model weights according to the F1-score (such as the LSTM weight increases from 0.6 to 0.7).

[0028] Technical verification: In the test set, the prediction accuracy of the hybrid model for sudden fires has been increased to 92%, which is 15% higher than that of the single model.

[0029] The specific implementation of the practical calculation process of the hybrid model is as follows: Data preparation: The test set contains 1000 samples, among which 200 are actual fire occurrences (positive class) and 800 are non-occurrences (negative class).

[0030] Prediction results of the hybrid model: Predicted as positive class: 250 (among which 180 are correctly predicted and 70 are false alarms).

[0031] Predicted as negative class: 750 (among which 730 are correctly predicted and 20 are missed alarms).

[0032] Index calculation: Precision: Precision = TP / (TP + FP) = 180 / (180 + 70) = 0.72 Recall: Recall = TP / (TP + FN) = 180 / (180 + 20) = 0.90 F1-score: F1 = 2×0.72×0.90 / (0.72 + 0.90) = 1.296 / 1.62 = 0.80 Comparison with the single model: LSTM prediction alone: F1 = 0.75 XGBoost prediction alone: F1 = 0.70 That is: The F1-score of the hybrid model (0.80) is significantly better than that of the single model, verifying the effectiveness of the integration strategy.

[0033] Step 3: Dynamic parameter closed-loop optimization (Bayesian optimization + sliding window) Improvement points: Parameter self-adaptation and stability detection 1. Bayesian optimization process: Define the hyperparameter space (regularization coefficient λ ∈ [0.01, 1], learning rate η ∈ [0.001, 0.1]); After each round of training, construct a Gaussian process surrogate model based on the validation set loss function, and select the parameter combination with the largest expected improvement (EI); Record the parameter adjustment trajectory and generate a visualization report for the operation and maintenance personnel to review.

[0034] The specific formula is: f(x) ∼ GP(m(x), k(x, x′)) m(x): Mean function (usually zero mean) k(x, x′): Kernel function (such as RBF kernel )

[0035] Among them: The relationship between the modeling hyperparameters (such as the regularization coefficient λ) and the validation set loss.

[0036] 2. Sliding window mechanism: 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 that the parameter distribution change < 1%, trigger the convergence mechanism and freeze the parameter update.

[0037] W new = 0.8·W current + 0.2·W history Among them: The data weight in the recent 3 months is higher, and the historical data weight decays to ensure that the model adapts to the latest environment.

[0038] Case: After 3 rounds of optimization of a certain urban area model, the false alarm rate decreased from 18% to 12%, and the parameter adjustment time was reduced by 40%.

[0039] Step 4: Model training Improvement points: Scenario-based threshold generation and real-time policy evaluation 1. MDP modeling: State space: Season (spring / summer / autumn / winter), real-time humidity (< 30% is high risk), population density (people / km²); Action space: Threshold adjustment range (±5%, ±10%, ±15%); Reward function: R = Fire loss reduction rate - 0.3 × number of false alarms `.

[0040] 2. DQN training and decision-making: Build a deep Q network (3-layer fully connected, ReLU activation), input the state features, and output the action Q value; Combine Monte Carlo tree search (MCTS) to simulate the potential risks after threshold adjustment and select the optimal strategy.

[0041] Application effect: The threshold in the rainy season is automatically lowered to 0.65 (original 0.75), and the false alarm rate is reduced by 28%.

[0042] Step 5: Automatic interpretable report generation Improvements: Multimodal NLG and user-customized output 1. FireBERT model training: Pre-trained on 100,000 fire report corpora based on the BERT architecture, supporting causal reasoning generation of "data - conclusion"; Sort the importance of input features (such as SHAP values) and generate statements like "Wind speed > 8m / s leads to a 40% increase in risk".

[0043] 2. GAN optimization of readability: Generator: The initial draft is output by FireBERT; Discriminator (CNN + BiLSTM) evaluates the comprehensibility of the report; Balance technical terms and popular expressions through adversarial training.

[0044] Output example: Management report: The heat map shows high-risk areas (such as the risk value in the East Area is 0.82), and it is recommended to deploy 3 additional patrol teams; Front-line staff report: Generate the inspection path (sequence of GPS coordinates) and the equipment inspection list (fire extinguisher pressure detection).

[0045] Step 6: Edge-cloud collaborative computing and federated learning Improvements: Hierarchical processing and privacy protection 1. Lightweight inference at the edge: Deploy the quantized Tiny-XGBoost model (model size < 2MB) to achieve real-time warning of sudden temperature rise (Δ > 5℃ / min); Only upload high-confidence alarms (probability > 0.7) to the cloud to reduce bandwidth occupancy.

[0046] 2. Cloud federated learning: Each edge node locally trains the model and encrypts and uploads the parameter gradients to the aggregation server; Use the FedAvg algorithm to update the global model and synchronize it to the edge side monthly.

[0047] Performance comparison: The cloud load is reduced by 60%, and the false alarm rate at the edge is < 0.8%.

[0048] Step 7: Effect evaluation and system iteration Improvements: Multidimensional evaluation metrics and closed-loop iteration 1. Evaluation system: Prediction performance: F1-score, AUC-ROC curve; Business metrics: Early warning response time, year-on-year decline in fire losses; User experience: Report readability score (1 - 5 points), operation interface response speed.

[0049] 2. Iterative mechanism: Retrain the model quarterly based on new data and update the threshold strategy; Optimize the report template according to user feedback (such as adding multilingual support).

[0050] Verification of implementation effect (data of a pilot city):

[0051] As Figure 2 shown, the system corresponding to the method of the present invention is as follows: Two types of data are accessed through the system. The structured data includes: pre-incident alarm information, in-incident disposal information, and post-incident report information. The unstructured data includes: connected camera images and device-uploaded photo information.

[0052] The data is processed through steps such as data cleaning, transformation, and processing by the big data capability platform to form a data standard format that can be used by the model library from numerous data with inconsistent standards.

[0053] Model library processing. A large number of different types of algorithms are preset in the model library here, such as: autoregressive integrated moving average model, Holt-Winters exponential smoothing method, logistic regression, etc. In the system of the present invention, the logistic regression algorithm is mainly adopted. The probability of fire occurrence can be predicted based on multiple factors. Its advantage lies in its simple logic, fast speed, and being conducive to urban fire analysis. If a large amount of historical data is analyzed, the autoregressive integrated moving average model is adopted in the present invention, and its trend analysis of such data is more stable and accurate. After calling different model library methods according to different types of data, the system of the present invention also aggregates a large amount of other system and configuration attribute information aggregated after the use of this system, that is, the secondary adjustment of the analysis results of the model library. The aim is to make the output result more in line with user usage.

[0054] Three types of outputs are generated according to the data after the model library calculation. Trend data: used for system display. Analysis report: Based on the trend data, through the report template, combined with natural language processing methods, the trend data is processed twice to finally complete a text information report based on the user's daily and next work arrangements. Configuration attributes: It is a summary of the post-evaluation information according to the trend data and the analysis report. By page embedding, the usage degree of each data is analyzed to judge the data effect, which is used to adjust the configuration attributes.

[0055] The following is a specific example: Data analysis logic Combined with historical fire data (such as 310 fire fighting operations in the first three quarters of a certain district 24 years ago, 52% of which were caused by outdoor burning of waste), dynamically adjust the weights of the early warning model: Statistical model (ARIMA): Capturing seasonal patterns (e.g., the weight of the high-incidence period of burning stubble in autumn and winter is increased to 0.4).

[0056] Machine learning model (XGBoost): Analyzing the correlation features between electrical faults (accounting for 14.2% of the fire causes) and careless use of fire by residents (23.9%).

[0057] Deep learning model (LSTM): Predicting high-risk periods (e.g., the warning sensitivity during the high-incidence period of fires from 12:00 to 15:00 is increased by 35%).

[0058] Effect verification False alarm rate optimization: The false alarm rate of the hybrid model (12.1%) is significantly lower than that of single models (ARIMA 27.9%, XGBoost 19.2%) [citation: user example data].

[0059] Capturing high-risk events: The early warning of residential fires is advanced by 2.8 hours, and the fatality risk is reduced by 36% (compared with a 50% year-on-year decrease in the number of deaths from residential fires in a certain district in 24 years).

[0060] In summary, the present invention applies the prediction results to the fire warning system. When the prediction probability exceeds a certain threshold, the warning mechanism is triggered to remind relevant departments to take measures. At the same time, an analysis report is formed as a reference for the next work arrangement to assist in decision-making.

[0061] Traditional fire handling is all post-event handling and is arranged according to the high-incidence periods of traditional patterns. Although it has certain effects, it is not sufficient to support the increasingly complex fire situations. Through this invention, after using appropriate algorithm parameters, there will be a relatively accurate prediction of the trend, and it will become more and more accurate with the accumulation of data, so as to achieve the fundamental purpose of improving personnel efficiency and reducing workload, and further better ensure the implementation of work.

[0062] On the other hand, the present invention also discloses a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor is caused to execute the steps of the above method.

[0063] On yet another aspect, the present invention also discloses a computer device including a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor is caused to execute the steps of the above method.

[0064] In another embodiment provided by the present application, there is also provided a computer program product containing instructions. When it runs on a computer, the computer is caused to execute the method of any of the above embodiments for intelligently analyzing the fire occurrence trend and generating a report.

[0065] It is understandable 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. For the explanations, examples, and beneficial effects of the relevant content, reference can be made to the corresponding parts in the above methods.

[0066] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part 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, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a dedicated computer, a computer network, or other programmable devices. 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 a website, computer, server, or data center to another website, computer, server, or data center by wire (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that the computer can access or a data storage device such as a server or data center that includes one or more integrated available media. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a DVD), or a semiconductor medium (for example, a solid-state disk (SSD)).

[0067] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including", or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article, or device that includes a series of elements includes not only those elements but also other elements that are not explicitly listed, or further includes elements that are inherent to such process, method, article, or device. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of additional identical elements in the process, method, article, or device that includes the element.

[0068] Each embodiment in this specification is described in a related manner. For the same or similar parts among the embodiments, reference can be made to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and for the related parts, reference can be made to the partial description of the method embodiment.

[0069] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements 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: The following steps are included: Collect fire-related raw data and do preprocessing; Select features that have an impact on the prediction target from the original data; Divide the data set 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 construction of the model library and the model scheduler are designed for hybrid model training and dynamic scheduling; Closed-loop optimization of dynamic parameters of hybrid model is performed. At the same time, parameter information is dynamically adjusted through parameter configuration after result output, and 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.

2. The method for intelligently analyzing fire occurrence trends and generating reports according to claim 1 is characterized in that: Collecting fire-related raw data includes: Through the edge, that is, the IoT device, a lightweight flame recognition model is deployed based on the TinyML architecture of MobileNetV3 to process the camera video stream in real time, extract features such as the percentage of flame pixels and the speed of smoke diffusion, and the response delay is less than 200ms. Receive structured data including temperature, humidity, and wind speed from the Meteorological Bureau API and unstructured data including features extracted from the edge and satellite heat maps through the cloud, and achieve real-time access to data streams through Apache Kafka.

3. The method for intelligently analyzing fire occurrence trends and generating reports according to claim 2 is characterized in that: Data preprocessing includes: Perform standardization and missing value filling on structured data; The unstructured data is extracted with a 128-dimensional feature vector through pre-training ResNet-50, which is then concatenated with the structured data features into a unified input matrix.

4. The method for intelligently analyzing fire occurrence trends and generating reports according to claim 1 is characterized in that: Hybrid model training includes, The model library is constructed through the short-term warning model LSTM, which inputs the time series features, i.e. the environmental data of the past 24 hours, and outputs the probability of fire in the next 2 hours; Analyze historical fire data through the long-term trend model ARIMA to predict the quarterly risk level; The real-time classification model XGBoost is used to process structured data and output probabilistic classification results.

5. The method for intelligently analyzing fire occurrence trends and generating reports according to claim 4 is characterized in that: Hybrid model dynamic scheduling includes, The model scheduler is designed to be data-driven. If the input is real-time streaming data, the LSTM+XGBoost hybrid model is called first. If it is batch historical data, ARIMA is called for trend fitting. A / B test 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.

6. The method for intelligently analyzing fire occurrence trends and generating reports according to claim 5 is characterized in that: Dynamic parameter closed-loop optimization includes Bayesian optimization and sliding window mechanism; Among them, the Bayesian optimization process includes: Define the hyperparameter space, i.e., regularization coefficient λ∈[0.01,1], learning rate η∈[0.001,0.1]; After each round of training, a Gaussian process proxy model is constructed based on the validation set loss function, and the parameter combination with the largest expected improvement is selected; Record parameter adjustment trajectories and generate visual reports for review by operation and maintenance personnel; 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 KL divergence detects that the parameter distribution changes by less than 1%, the convergence mechanism is triggered and the parameter update is frozen; W new =0.8 W current 10.2 history Among them: the data in the past three months has a higher weight, and the weight of historical data is attenuated to ensure that the model adapts to the latest environment.

7. The method for intelligently analyzing fire occurrence trends and generating reports according to claim 6 is characterized in that: Setting up model training includes, MDP Modeling: The state space includes seasons, i.e. spring / summer / autumn / winter, real-time humidity, <30% is considered high risk, and population density unit: people / km²; The action space includes the threshold adjustment range, namely ±5%, ±10%, and ±15%; Reward function: R = fire loss reduction rate - 0.3 × number of false alarms; DQN training and decision making: Constructing a deep Q network includes 3 layers of full connection, ReLU activation, input state features, and output action Q value; Combined with Monte Carlo tree search, the potential risk after threshold adjustment is simulated to select the optimal strategy.

8. The method for intelligently analyzing fire occurrence trends and generating reports according to claim 7 is characterized in that: Output fire probability and generate interpretable reports, including, FireBERT model training: Including pre-training on 100,000 fire report corpora based on the BERT architecture to support "data-conclusion" causal reasoning generation; Input feature importance ranking and generate corresponding statements; GAN optimizes readability: Set up the generator FireBERT to output the first draft; evaluate the report's understandability through the discriminator CNN+BiLSTM; Balance professional terminology with popular expressions through adversarial training.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the processor is caused to perform the steps of the method according to claims 1 to 8.

10. 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 is caused to perform the steps of the method according to claims 1 to 8.

Images