A method and system for constructing lightweight time-series prediction models for fire situation prediction

Abstract

This invention relates to the technical field of fire prediction and protection for steel structure factory buildings, specifically to a lightweight time-series prediction model construction method and system for fire situation prediction. The method includes: firstly, defining the ignition point / scenario and constructing an expert rule base, using rules to label and enhance the risk probability of normal data; secondly, collecting real-time and historical multi-source data, preprocessing it to form a supervised sample set with risk labels; and thirdly, constructing a DecoKAN model, using multi-source data as input, learning the data and risk mapping, and outputting the fire risk probability for the next 24 hours, achieving lightweight and accurate prediction. This invention provides a comprehensive solution for active fire protection of steel structure factory buildings, integrating multi-source data fusion, high-precision time-series prediction, interpretable analysis, and intelligent strategy linkage.

Description

Technical Field

[0001] This invention relates to the technical field of fire prediction and protection for steel structure workshops, specifically a comprehensive active fire protection solution for steel structure workshops that integrates multi-source data fusion, high-precision time-series prediction, interpretable analysis, and intelligent strategy linkage. In particular, it relates to a lightweight time-series prediction model construction method and system for fire situation prediction. Background Technology

[0002] Steel-structured chemical plants, due to the frequent storage of flammable and explosive materials, experience a sharp decline in the mechanical properties of their steel structures under sustained high temperatures, resulting in extremely high fire risks and a high likelihood of rapid building collapse. Currently, fire safety management in such plants primarily relies on periodic manual inspections and fixed-threshold alarm systems. These traditional methods are significantly passive and delayed: fixed-threshold alarms only trigger when specific parameters exceed preset limits, failing to detect the early stages of a fire; manual inspections are limited by periodicity and subjectivity. Both methods are insufficient for proactively assessing the development of a fire and cannot provide a sufficient early warning window to initiate precise and effective prevention and control measures before a fire occurs. Therefore, developing an intelligent system capable of proactively predicting fire risks and dynamically assessing the situation has become an urgent need to ensure the safe operation of high-risk chemical plants and prevent major accidents.

[0003] Although time series forecasting technology provides a theoretical possibility for fire situation prediction, its practical application in the complex scenario of steel structure factory buildings faces multiple technical challenges: First, the complexity of the data and the difficulty of correlation modeling are significant. Fire risk is the result of nonlinear coupling of multi-source time-series data. Influencing factors include the spatiotemporal evolution of temperature and smoke concentration inside the factory, dynamic meteorological conditions such as external temperature, humidity, and wind, as well as the operating status of internal equipment. Traditional prediction models are unable to effectively capture and quantify the complex long-term dependencies between these multi-source, heterogeneous time-series features.

[0004] Secondly, the accuracy and interpretability of the predictions are extremely important. Fire early warning is related to major safety. The model not only needs to output the probability of fire risk in the future (such as 24 hours) with high precision, but its prediction logic should also have a certain degree of interpretability to support the formulation of targeted prevention and control strategies (such as locating high-risk areas and judging the dominant risk factors). A "black box" model that lacks interpretability is difficult to gain the full trust and effective application of safety management personnel.

[0005] Finally, the model needs to have strong adaptability and generalization ability. Different chemical plants have differences in processes, layouts, and materials, resulting in different data patterns. The model must be effectively trained on limited historical data and be able to adapt to the slow changes in data distribution with the seasons and production plans to avoid overfitting.

[0006] Therefore, it is crucial to establish a method that can systematically and quantitatively evaluate the degree of influence of various factors on workability, so as to provide precise guidance for material design and production to improve workability. Summary of the Invention

[0007] Based on the technical problems existing in the prior art, the present invention provides a lightweight time-series prediction model construction method and system for fire situation prediction.

[0008] As a first aspect of the present invention, a method for constructing a lightweight time-series prediction model for fire situation prediction is provided, comprising the following steps: S1: Data Acquisition and Preprocessing Identify fire risk sources and build an expert rule base. Then, collect real-time time-series data on temperature and smoke concentration in various areas of the factory through a network of temperature and smoke sensors. Simultaneously, access local meteorological data and factory equipment operation status signals. At the same time, integrate historical fire event records from previous years and corresponding multi-dimensional environmental data from the historical database to form a multi-source data pool. The raw time-series data in the multi-source data pool are cleaned and standardized to form a standardized supervised learning sample set; S2: Model Building and Training A lightweight time series prediction model is built based on DecoKAN. First, the data distribution offset is eliminated by reversible normalization. Then, the multi-source time series data is decomposed into approximation coefficients and detail coefficients using the Daubechies4 wavelet basis. Subsequently, the patterns of different frequency components are learned in parallel by independent KAN branches. Finally, after multi-level wavelet reconstruction and inverse normalization, the fire risk probability in the next 24 hours is output. The model is trained using a loss function combining mean squared error and cross-entropy, along with an early stopping strategy. S3: Online Prediction After the model is deployed, it receives online sensor and meteorological data streams in real time. After preprocessing, the data is input into the trained lightweight time series prediction model, which dynamically outputs the fire risk probability curve for the next 24 hours. When the probability exceeds a preset threshold, a tiered warning is automatically triggered, and the risk area and its evolution trend are displayed through a visual interface.

[0009] Based on the above scheme, step S2: model construction and training specifically includes: S201: Perform reversible normalization on the input time series data and map it to a uniform scale; S202: Iteratively decompose the time series data after normalization and transpose in step S201 using the Daubechies4 wavelet basis to generate approximation coefficients and detail coefficients; S203: Each coefficient sequence output from step S202 is processed by an independent KAN branch; S204: Synthesize the predicted coefficient sequence output from the KAN branch into the time domain using inverse multi-level wavelet transform: Generated time series data Characterize the prediction results in the normalized channel preference space; tensor Transposed to Then, a reversible normalization inverse process is performed to obtain the predicted sequence in the original data space. .

[0010] Based on the above scheme, step S202, which involves iteratively decomposing the time-series data after normalization and transpose in step S201 using the Daubechies4 wavelet basis to generate approximation coefficients and detail coefficients, specifically includes: Decomposition layer number As a hyperparameter, the default value is 1 level, which is adjusted according to the input length. According to the formula The balance between resolution and computational cost was determined and empirically verified. The DWT process employs a symmetrical fill strategy to handle boundary effects and decomposes the generated... A sequence of coefficients: in, For the first Layer approximation coefficient sequence, , This corresponds to the sequence of detail coefficients for each layer.

[0011] Based on the above scheme, in step S203: each coefficient sequence output by step S202 is processed by an independent KAN branch, and the process of each branch processing includes: ① Input Perform a reversible normalization-forward operation to generate a normalized sequence. ,in Indicates the first One KAN branch; ② The normalized sequence is processed by block partitioning, including padding and overlap calculations, to obtain... : in, Indicates the number of blocks. The block size (sliding step size is) ); After further embedding, the block results are projected onto the dimension. The formula is: Among them, tensor As the input to the first KAN mixer.

[0012] ③ The data stream passes through a time-series KAN mixer and a feature-based KAN mixer, and retains the original information through residual connections, as shown in the following formula: = ⊕ In the formula, For hybrid characterization, For the original embedded input, This is the output of the second KAN mixer.

[0013] ④ Restore the branch output to the original scale to obtain the branch prediction coefficients. .

[0014] Based on the above scheme, step S2: model construction and training also includes: Each KAN linearization layer introduces a composite regularization loss: The loss function contains two key components: the balance coefficient and the loss function. and Set to 1.0, and incorporate it into the total loss function through the global regularization intensity hyperparameter γ. : in, The mean squared error between the predicted and actual values. .

[0015] Based on the above scheme, step S2: model construction and training also includes: Input the labeled training dataset, optimize the DecoKAN parameters and early stopping strategy through backpropagation, and save the optimal model parameters.

[0016] Based on the above scheme, step S1: data acquisition and preprocessing specifically includes: S101: Based on the target steel plant's process flow diagram, equipment layout diagram, and hazardous substance safety data sheet, identify potential ignition sources and typical fire scenarios to confirm fire risk sources. Then, gather fire safety regulations, industry accident cases, and domain expert experience to build an expert rule base for data completion. S102: For cases where the factory has no fire history records but has normal sensor data, the expert rule base in step S101 is used to scan the time series segments under normal operating conditions to obtain the reverse labeling of risk probability. S103: Collect real-time data and historical data to form a multi-source data pool; S104: Clean and standardize the raw time-series data in the multi-source data pool to form a supervised learning sample set for model training.

[0017] Based on the above scheme, S102: For situations where the factory has no historical fire records but has normal sensor data, the expert rule base from step S101 is used to scan the time series segments under normal operating conditions to obtain the reverse annotation of risk probabilities. Specifically, this includes: Data segments that conform to a high-risk pattern are labeled as high-risk values, those that conform to a potential-risk pattern are labeled as medium-risk values, and those that are completely normal are labeled as low-risk values.

[0018] Based on the above scheme, in step S103: collecting real-time data and historical data to form a multi-source data pool, the specific steps are as follows: The real-time data includes time-series data of temperature and smoke concentration in various areas of the factory collected through a network of temperature and smoke sensors, and synchronously accessed local meteorological data and factory equipment operating status signals. The historical data includes records of fire events from previous years and corresponding multi-dimensional environmental data from a historical database.

[0019] As a second aspect of the present invention, a lightweight time-series prediction system for fire situation prediction is provided, including the lightweight time-series prediction model. The system collects multi-source time-series data through temperature sensors, smoke sensors, meteorological interfaces, equipment status interfaces, and historical databases. The lightweight time-series prediction model processes the multi-source time-series data and outputs the fire risk probability for the next 24 hours.

[0020] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This invention proposes an intelligent early warning and protection system that integrates multi-source heterogeneous time-series data perception with an advanced deep learning time-series prediction model. At the data level, it integrates factory sensor data, environmental meteorological data, and equipment operation data to construct a unified fire risk time-series database. At the model level, it introduces a novel Decoupled Kolmogorov-Arnold Networks (DecoKAN) architecture. This model excels in capturing complex nonlinear long-term dependencies, and its structural characteristics also contribute to improving model interpretability and aiding in the identification of key risk drivers. At the application level, the model's predicted output (future fire risk probability and situation) is directly linked to prevention and control strategies, guiding proactive safety inspections, resource pre-positioning, and emergency plan optimization, thus driving a fundamental shift in safety management from passive response to proactive early warning and precise prevention and control. Practice has shown that prevention and control strategies based on such accurate predictions can effectively reduce the probability of fire occurrence and losses.

[0021] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0022] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0023] Figure 1 This is a flowchart illustrating a method for constructing a lightweight time-series prediction model for fire situation prediction, according to an exemplary embodiment. Figure 2 This is an overall architecture diagram of the DecoKAN model shown according to an exemplary embodiment; Figure 3 This is a structural diagram of a KAN branch according to an exemplary embodiment; Figure 4 This is a structural diagram of a KAN mixer according to an exemplary embodiment. Detailed Implementation

[0024] The following description and accompanying drawings fully illustrate specific embodiments of this application to enable those skilled in the art to practice them. Some parts and features of some embodiments may be included in or replace parts and features of other embodiments.

[0025] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0026] like Figure 1As shown, this application provides a lightweight time-series prediction model construction method for fire situation prediction, including the following steps: S1: Data Acquisition and Preprocessing The system collects real-time time-series data on temperature and smoke concentration in various areas of the factory through a network of temperature and smoke sensors, and simultaneously accesses local meteorological data and factory equipment operation status signals. It also integrates historical fire event records from previous years and corresponding multi-dimensional environmental data from the database to form a multi-source data pool. The raw time-series data in the multi-source data pool is cleaned and standardized to form a standardized supervised learning sample set. Specifically, S101: Based on the target steel plant's process flow diagram, equipment layout diagram, and hazardous substance safety data sheet, identify potential ignition sources and typical fire scenarios to confirm fire risk sources, and then compile fire safety regulations, industry accident cases, and domain expert experience to build an expert rule base for data completion. S102: For situations where the factory has no historical fire records but has normal sensor data, the expert rule base from step S101 is used to scan time-series segments under normal operating conditions (temperature and smoke concentration fluctuating within safe thresholds) to obtain inverse labeling of risk probabilities, specifically: Segments that meet high-risk patterns such as risk accumulation but do not trigger alarms are marked as potential risks; Those that fully conform to normal patterns are marked as low risk; S103: Two types of core data are collected to form a multi-source data pool: Real-time data: The temperature and smoke concentration data of various areas of the factory are collected through a network of temperature and smoke sensors, and local meteorological data (temperature, humidity, wind) and factory equipment operation status signals are accessed simultaneously. Historical data: Integrates records of fire events from previous years in the historical database (including time, scale, and duration) and corresponding multi-dimensional environmental data at the time.

[0027] S104: Cleaning and standardizing the raw time-series data in the multi-source data pool: The original time-series data is processed by filling missing values, correcting outliers, and denoising and smoothing. The multi-source environmental data are aligned with a unified timestamp to form a standardized supervised learning sample set for model training.

[0028] Each sample in the supervised learning sample set (a segment of historical multi-source time-series data) corresponds to a fire risk label, and the fire risk label is generated based on the following factors: If there are historical fire events: data within a specific time window before the event (such as temperature, smoke, and meteorological data 1 hour before the fire) are marked as high risk (e.g., 1.0). If there is no fire record: Scan normal data using an expert database. Data that meets the high-risk pattern is marked as higher risk (e.g., 0.7), potential risk is marked as medium risk (e.g., 0.3), and completely normal is marked as low risk (e.g., 0.05), forming a continuous monitoring signal from normal to abnormal.

[0029] S2: Model Building and Training A lightweight time series prediction model is constructed based on DecoKAN. First, reversible normalization is used to eliminate data distribution bias. Then, the Daubechies4 wavelet basis is used to decompose multi-source time series data into approximation coefficients and detail coefficients. Subsequently, independent KAN branches learn patterns of different frequency components in parallel. Finally, after multi-level wavelet reconstruction and inverse normalization, the fire risk probability for the next 24 hours is output. The model is trained using a loss function combining mean squared error and cross-entropy and an early stop strategy.

[0030] Specifically, S201: Perform reversible normalization on the input time series data and map it to a uniform scale; S202: Iteratively decompose the time series data after normalization and transpose in step S201 using the Daubechies4 wavelet basis to generate approximation coefficients and detail coefficients: Decomposition layer number As a hyperparameter (default value is 1 layer), it depends on the input length. According to the formula The balance between resolution and computational cost was determined and empirically verified. The DWT process employs a symmetrical fill strategy to handle boundary effects and decomposes the generated... A sequence of coefficients: in, For the first Layer approximation coefficient sequence, This corresponds to the sequence of detail coefficients for each layer.

[0031] S203: Each coefficient sequence output from step S202 is processed by an independent KAN branch. The process for each branch includes: ① Input Perform a reversible normalization-forward operation to generate a normalized sequence. ,in Indicates the first One KAN branch; ② The normalized sequence is processed by block partitioning, including padding and overlap calculations, to obtain... ;in ,in Indicates the number of blocks. The block size (sliding step size is) ); After further embedding, the block results are projected onto the dimension. The formula is: tensor As the input to the first KAN mixer.

[0032] ③ The data stream passes through a time-series KAN mixer and a feature-based KAN mixer, and retains the original information through residual connections, as shown in the following formula: = ⊕ In the formula, For hybrid characterization, For the original embedded input, This is the output of the second KAN mixer.

[0033] ④ Restore the branch output to the original scale to obtain the branch prediction coefficients. .

[0034] S204: Synthesize the predicted coefficient sequence output from the KAN branch into the time domain using inverse multi-level wavelet transform: Generated time series data Characterize the prediction results in the normalized channel preference space; tensor Transposed to Then, a reversible normalization inverse process is performed to obtain the predicted sequence in the original data space. .

[0035] S205: Each KAN linearization layer introduces a composite regularization loss: The loss function contains two key components (balance coefficient). and Set to 1.0), and through global regularization of the intensity hyperparameter γ (such as... Included in the total loss function : in The mean squared error (MSE) is the difference between the predicted and actual values.

[0036] S206: Input labeled training dataset, optimize DecoKAN parameters and early stopping strategy through backpropagation, and save the optimal model parameters.

[0037] S3: Online Prediction After the model is deployed, the system receives online sensor and meteorological data streams in real time. After preprocessing, the data is input into the trained DecoKAN model, which dynamically outputs the fire risk probability curve for the next 24 hours. When the probability exceeds the preset threshold, the system automatically triggers a graded warning and displays the risk area and evolution trend through a visual interface.

[0038] Operators can use early warning information to locate high-risk areas and conduct equipment maintenance, ventilation and cooling, or manual inspections in advance, thus shifting from passive response to proactive prevention.

[0039] Specifically, S301: receives sensor, meteorological, and equipment data in real time, preprocesses it, inputs it into a trained lightweight time series prediction model, and dynamically outputs the fire risk probability curve for the next 24 hours. S302: Three-level response based on risk probability: (1) High risk (>70%): Trigger sound and light / large screen / mobile terminal alarms and implement 3 types of measures: ①Precise resource pre-positioning: dispatch mobile fire-fighting equipment (fire-fighting robots) to high-risk areas locked by the model, and notify fire stations to be on standby; ② Proactive environmental intervention: The linkage automation system activates enhanced ventilation, spray cooling, or inert gas protection in the target area; ③ Production control: It is recommended to suspend non-essential operations in high-risk areas and generate evacuation routes and emergency plans.

[0040] (2) Medium risk (30%-70%): Generate preventive inspection work orders, guide key inspections of specific equipment / lines, and mark key attention periods.

[0041] (3) Low risk (<30%): Maintain routine monitoring, no additional measures.

[0042] S303: The system adaptively optimizes the model by continuously learning from new data.

[0043] This invention provides a comprehensive proactive fire protection solution for steel structure factory buildings, integrating multi-source data fusion, high-precision time-series prediction, interpretable analysis, and intelligent strategy linkage. It overcomes the lag limitations of traditional methods by using the advanced DecoKAN model to achieve forward-looking and quantitative assessment of complex fire risks, directly transforming predictive information into actionable and precise prevention and control measures. This shifts the fire safety management paradigm from passive response to proactive early warning and precise prevention and control, possessing significant practical value and widespread applicability for ensuring safe production in high-risk industries such as chemical manufacturing.

[0044] Example 2 Based on the lightweight time-series prediction model construction method for fire situation prediction in Example 1, this application provides a specific example as follows: Figure 1 The workflow of the lightweight time-series prediction model construction method for fire situation prediction in this application is illustrated. The method includes the following steps: S1: Data Acquisition and Preprocessing Ignition point and hazardous scenario definition: Based on the process flow diagram, equipment layout diagram and hazardous material safety data sheet of the target steel plant, identify potential ignition sources (such as high temperature reactors, electrical cabinets, flammable material storage areas) and typical fire scenarios (such as pool fire, jet fire, equipment overheating, etc.).

[0045] a. Expert Rule Base Construction: This system gathers fire safety regulations, industry accident case reports, and expert experience to extract qualitative and semi-quantitative rules governing the relationship between fire risk and multi-source data. For example: a sudden increase in regional temperature exceeding 50°C within 10 minutes, accompanied by a gradual increase in smoke concentration, indicates a high risk level; continuous equipment overload operation with ambient humidity below 30% increases the cumulative fire risk coefficient.

[0046] Labeling and Enhancement of Limited Real-World Data: Even without historical fire records, normal sensor operating data from the plant (temperature and smoke concentration typically fluctuate within safe thresholds) can still be obtained. Using the aforementioned expert rules, risk probability is inversely labeled for time-series segments under normal operating conditions. For example, segments that conform to the "risk accumulation" rule pattern but do not trigger an alarm are identified and labeled as "low-risk" or "potential-risk" samples. Simultaneously, data augmentation techniques such as noise addition, time warping, and amplitude scaling are used to expand the sample size of normal and rule-defined abnormal patterns.

[0047] b. The system collects real-time time-series data on temperature and smoke concentration in various areas of the factory through a network of temperature and smoke sensors, and simultaneously accesses local meteorological data (temperature, humidity, wind speed) and factory equipment operating status signals. The historical database integrates records of fire events over the years (time, scale, duration) and corresponding multi-dimensional environmental data. Missing value imputation, outlier correction, and noise reduction and smoothing are performed on the raw time-series data. The multi-source data are aligned according to a unified timestamp to form a standardized supervised learning sample set for model training.

[0048] Each sample in this dataset (a segment of historical multi-source time-series data) is assigned a corresponding "fire risk label." This label is generated based on: Historical fire records: If there are historical fire events, the data within a specific time window before the event is marked as "high risk" or a specific risk probability value (e.g., 1.0).

[0049] Expert rule derivation: For "normal" data with no fire records, a built-in expert rule base (e.g., the rule for "sudden temperature rise accompanied by slow smoke rise") is used for scanning. Data segments that match the high-risk pattern are labeled with a higher risk value (e.g., 0.7), those that match the potential risk pattern are labeled with a medium value (e.g., 0.3), and those that are completely normal are labeled with a low value (e.g., 0.05). This provides the model with a large number of supervisory signals that continuously transition from normal to abnormal.

[0050] Therefore, the fire risk label is provided to the model as a known answer during the training phase. The model's task is to learn how to deduce this answer from the input time-series data [temperature, smoke concentration, weather, equipment status, etc.].

[0051] Initial training set construction: The data labeled with the above rules are mixed with the limited real normal operating condition data collected to form an initial supervised learning sample set covering multiple states from normal to abnormal. Although this dataset is not entirely from real fires, its internal logic conforms to fire science, providing the model with a basic pattern for learning fire precursors and multivariate coupling relationships.

[0052] S2: Model Building and Training A time-series forecasting model based on DecoKAN is constructed, whose mechanism can capture the long-term dependencies between multi-dimensional sensor data and meteorological data. Input features include historical temperature, smoke concentration, meteorological parameters, and equipment status, and the output is the probability of fire risk in the next 24 hours. Model training uses a loss function combining mean squared error and cross-entropy, and parameters are optimized through backpropagation. An early stopping strategy is set to prevent overfitting. When the training loss converges to a stable value or reaches 1000 iterations, the optimal model parameters are saved, and the model is solidified.

[0053] During training, only real-time data such as temperature and smoke concentration are available, but how to establish a corresponding relationship between this data and fire risk, and how fire risk is derived, remains to be seen.

[0054] The final output layer of the model is designed as a classification head, which maps the high-level, abstract feature representations extracted by the DecoKAN network to a scalar value between 0 and 1, namely the fire risk probability.

[0055] The learned mapping: This output probability value is essentially a comprehensive quantitative assessment of the similarity between the system state represented by the current input sequence and the "dangerous state," based on all the patterns it has learned from massive amounts of labeled data (these patterns are associated with scenarios that have been historically or rule-defined as "dangerous"). It is not judged by a single threshold, but is the result of a weighted matching of all learned dangerous patterns.

[0056] Interpretability Verification: Thanks to the interpretability of KAN, the input variables that contribute most to the final risk probability and their interaction function forms can be traced. For example, the analysis might show that this high-risk warning is mainly driven by the quadratic growth characteristic of temperature in region X over the past 10 minutes and the first derivative characteristic of smoke concentration in region Y with respect to temperature. This reveals how the data relates to risk at both mathematical and physical levels, making the model's decision-making logic transparent rather than a black box.

[0057] S3: Online Forecasting and Operational Guidance During deployment, the system receives real-time data from online sensors and meteorological data streams. After preprocessing, this data is input into a trained DecoKAN model, dynamically outputting a fire risk probability curve for the next 24 hours. When the probability exceeds a preset threshold, the system automatically triggers a tiered early warning and displays the risk area and its evolution trend through a visual interface. Operators can use the early warning information to locate high-risk areas and proactively conduct equipment maintenance, ventilation, or manual inspections, shifting from reactive response to proactive prevention.

[0058] This process integrates multi-dimensional real-time data with historical patterns to achieve accurate and proactive prediction of fire risks in steel structure workshops, significantly improving the timeliness of safety hazard detection and handling, and providing intelligent decision support for safe chemical production.

[0059] The following is a further explanation of a lightweight time-series prediction model construction method for fire situation prediction in this embodiment.

[0060] Let the multivariate time series data of historical factory fires be... ,in This indicates the total length of the sequence.

[0061] Each Indicates time The observation vector contains Several variables (e.g., sensor temperature, smoke concentration, device status, etc.). Given a length of... Historical observation window sequence , The goal is to learn a mapping function. To predict future length sequence , .

[0062] The prediction results characterize the evolution trend of key indicators and are crucial for applications ranging from fire early warning to emergency decision-making. This prediction task can be formally represented as: ( ) in This is the predicted future sequence.

[0063] 1. DecoKAN model architecture To effectively model the inherent heterogeneous dynamic characteristics in fire trend prediction, the DecoKAN model proposed in this application adopts a decomposition-fusion-reconstruction paradigm architecture.

[0064] Figure 2 The end-to-end overall architecture diagram of the DecoKAN model is shown.

[0065] The core innovation of DecoKAN lies in replacing the traditional implicit mixer based on multilayer perceptron with Kolmogorov-Arnold networks (KANs) that have stronger expressive power and higher interpretability, and completing all core fusion operations in the wavelet domain.

[0066] like Figure 2 As shown, the model processing flow is as follows: input time series Performing reversible normalization-forward operation, we obtain Then the transpose operation converts it to Subsequently, it is decomposed into multiple approximate coefficient sequences and detail coefficient sequences through multi-level wavelet transform, as shown below. This decomposition mechanism enables customized feature extraction of the decoupled components.

[0067] Each generated coefficient sequence is processed by an independent KAN branch, which are represented in the figure as KAN branch 1 to KAN branch 2. This branching structure supports parallel and independent processing of approximate and detail components, minimizing spectral interference while enabling frequency-specific pattern learning.

[0068] The results of all branch processing are represented as follows .

[0069] Finally, the prediction coefficients of all branches are synthesized into a coherent time-domain signal through multi-level wavelet reconstruction. After transposing, we get Finally, the final prediction result is generated through reversible normalization and inverse processing. .

[0070] Of the above parameters, This indicates the length of the historical observation window, where m represents the branch number (1 to m). ), ,.... , Indicates the subscript number of the branch. , , ...... This indicates that the wavelet transform is used to decompose the wavelet into a sequence of multiple approximate coefficients and a sequence of detail coefficients. The results of the KAN branch processing are shown in Table 1.

[0071] 2. Reversible normalization Given the drastic fluctuations and non-stationarity of fire prediction data, mitigating distribution bias is crucial for improving model generalization ability. Therefore, a reversible instance normalization technique (Reversible Instance NormalizationRevIN) is employed. This method maintains consistency with the overall model implementation and does not introduce learnable affine parameters. This reversible normalization consists of two operations: reversible normalization-forward and reversible normalization-inverse. For the input sequence... Reversible normalization-forward operation will affect the input sequence Perform mean subtraction and standard deviation division operations along the channel dimension, and save the corresponding statistics. To ensure numerical stability, a minimum value is included in the variance calculation. And set a lower bound cutoff value for the standard deviation ( Reversible normalization-forward operation maps raw data of different dimensions, ranges, and distributions to a unified standard scale (usually [0, 1] or a range with a mean of 0 and a variance of 1).

[0072] Reversible normalization-inverse operation is the opposite computational method, the purpose of which is to restore the data and convert the model's output back to its original physical scale, so that it has practical and interpretable meaning.

[0073] 3. Multi-level wavelet decomposition The multi-level wavelet decomposition module uses the multi-level discrete wavelet transform method, which aims to analyze the composite signals in fire trend prediction data and separate potential low-frequency trends from high-frequency noise and fluctuations.

[0074] This method applies to normalized transposed time series data We decompose the data by iteratively applying high-pass and low-pass filters generated from the selected wavelet basis. We chose the Daubechies 4 wavelet basis (db4), denoted as Ψ. This wavelet basis achieves a good balance between smoothness and tight support, making it suitable for jointly capturing transient and trend components in non-stationary time series.

[0075] Decomposition layer number As a hyperparameter (default value is 1 level), it is usually determined based on the input length. According to the formula The balance between resolution and computational cost was determined and empirically verified. A symmetric padding strategy was employed to handle boundary effects during the DWT process. The final decomposition yields... A sequence of coefficients: in, For the first Layer approximation coefficient sequence, This corresponds to the sequence of detail coefficients for each layer.

[0076] make express Any sequence in the coefficient sequence (e.g.) Its corresponding input length Each dimension is coefficient sequence This will be handled by a separate KAN branch.

[0077] 3. KAN branch like Figure 3As shown, each KAN branch first checks its input. Perform a reversible normalization-forward operation to generate a normalized sequence. ,in Indicates the first One KAN branch.

[0078] The normalized sequence is then subjected to block processing involving padding and overlap calculations to obtain... .in ,in Indicates the number of blocks. The block size (sliding step size is) ).

[0079] Then, after embedding, the block results are projected onto the dimension. The formula is: This tensor This serves as the input to the first KAN mixer (KANMixer-1). Data flows sequentially through two KAN mixer modules, KANMixer-1 and KANMixer-2: The output of the second KAN mixer, KANMixer-2 and original embedded input Perform residual connections to generate hybrid characterizations. The formula is: = ⊕ The obtained tensor It is passed to the header module for processing, and then... Finally, after reversible normalization inverse operation, we obtain... .

[0080] 4. KAN Mixer Module The structure of each KAN mixer module is as follows: Figure 4As shown, a hierarchical structure based on KAN is used to model the interaction between temporal and feature dimensions sequentially. This module consists of a temporal KAN mixer and a feature KAN mixer connected together. Taking KAN mixer 1 (denoted as KAN Mixer-1) as an example, the structure of the temporal KAN mixer and feature KAN mixer is as follows: 1) Timing KAN Mixer enter Layer normalization process is performed sequentially. ) and dimensional permutation ( Shape conversion Then, after two consecutive KAN linearizations ( Acting on Dimension processing, and finally inverse dimension replacement ( Its output is added to the original input via a residual connection with Dropout, as shown in the formula: ) in, This indicates Dropout calculation. This represents the final output.

[0081] 2) Featured KAN Mixer Output of the timing KAN mixer Layer normalization ( After that, we obtained Then, after two consecutive KAN linearizations ( The process involves adding the output to the original input via a residual connection with Dropout, as shown in the formula: tensor This is the final output of the feature KAN mixer d.

[0082] The core advantage of the KAN mixer lies in its fundamental building block, the KAN linearization layer. Each connection in this layer uses a learnable univariate activation function. (Default grid size = 5, spline order = 3), through grid parameterization Spline function implementation typically involves superimposing basis functions. (The SiLU function is used in this implementation): in, For B-spline basis functions, , , For learnable parameters, This represents the basis function (the SiLU function is used in this implementation).

[0083] 5. Multi-level wavelet reconstruction like Figure 3 As shown, the input for multi-level wavelet reconstruction is That is, the above A series of prediction coefficients from the output of each KAN branch.

[0084] Its meaning is: Approximation coefficient Detail factor In the multi-level wavelet reconstruction stage, the predicted coefficient sequence output by the KAN branch is synthesized into the time domain through inverse multi-level wavelet transform: Generated time series data This tensor represents the prediction results in the normalized channel preference space. Transposed to Then, a reversible normalization inverse process is performed to finally obtain the predicted sequence in the original data space. .

[0085] 6. Loss Function In time series deep learning models, the loss function plays multiple roles, acting as a navigator and evaluator. During training, it guides the model to adjust in the direction of reducing prediction error at each epoch. Choosing an appropriate loss function is equivalent to equipping the model with a professional guide to understand time dynamics, directly affecting whether the model can capture complex trends, periodicities, and sudden event patterns in time series.

[0086] In the training of this model, to improve interpretability and sparsity, a composite regularization loss is introduced for each KAN linearization layer: The loss function contains two key components (balance coefficient). and Typically set to 1.0), and through global regularization of the intensity hyperparameter γ (such as... Included in the total loss function: in The mean squared error (MSE) is the difference between the predicted and actual values.

[0087] The specific composition of the regularization component: ①L1 norm term The average amplitude s of the spline components promotes function sparsity. ② Entropy term Based on the normalized spline coefficient amplitude calculate: This entropy constraint promotes structural sparsity at the neuron hierarchy level; The dual regularization mechanism, controlled by parameter γ, guides the model to obtain a simple and interpretable solution during training.

[0088] After generating an advanced fire risk prediction, this invention implements a graded and dynamic intelligent prevention and control strategy based on the prediction results. Specifically, the system divides the predicted risk curve for the next 24 hours into three levels: low risk (<30%), medium risk (30%-70%), and high risk (>70%), and automatically triggers a response plan that matches the risk level.

[0089] For high-risk warnings, the system not only pushes emergency alerts through sound and light, the control center's large screen, and mobile terminals, but also automatically executes or suggests the following prevention and control measures: Precise resource pre-positioning: Based on the high-risk areas identified by the model's interpretability analysis, mobile fire-fighting facilities (such as fire-fighting robots and fire monitors) are automatically dispatched for pre-deployment, and the nearest fire station is notified to enter an enhanced standby state; Proactive environmental intervention: In conjunction with the factory's automation system, activate enhanced ventilation, spray cooling, or inert gas protection in the target area; Production activity control: It sends recommendations to the production management system to suspend or reduce non-essential operations in high-risk areas and automatically generates optimal evacuation routes and emergency plans for management to make decisions.

[0090] For low- to medium-risk warnings, the system generates preventative inspection work orders, instructing personnel to conduct focused inspections of specific equipment or lines, and marking the period as a key period requiring enhanced monitoring.

[0091] In actual deployment and verification, this system has demonstrated significant application value. First, it advances the detection of fire hazards from the traditional "post-event alarm" or "threshold trigger" stage to the early stages of fire development, with an average warning lead time of several hours to over ten hours, providing a crucial window for proactive measures. Second, through precise risk positioning and attribution analysis, it changes the previous crude emergency response model of blanket plant-wide shutdowns, improving the efficiency of prevention and control resource allocation by over 60% and greatly reducing production interruption losses caused by false alarms or over-response. Finally, by continuously learning from new operational data and constantly adaptively optimizing the predictive model, the system forms a closed loop of prediction-intervention-feedback-optimization safety management, sustainably improving the inherent safety level of the plant.

[0092] Finally, it should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0093] The above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them; although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can still be made to the specific implementation of the present invention or equivalent substitutions can be made to some technical features without departing from the spirit of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the scope of the technical solutions claimed in the present invention.

Claims

1. A method for constructing a lightweight time-series prediction model for fire situation prediction, characterized in that, Includes the following steps: S1: Data Acquisition and Preprocessing Identify fire risk sources and build an expert rule base. Then, collect real-time time-series data on temperature and smoke concentration in various areas of the factory through a network of temperature and smoke sensors. Simultaneously, access local meteorological data and factory equipment operation status signals. At the same time, integrate historical fire event records from previous years and corresponding multi-dimensional environmental data from the historical database to form a multi-source data pool. The raw time-series data in the multi-source data pool are cleaned and standardized to form a supervised learning sample set; S2: Model Building and Training A lightweight time series prediction model is built based on DecoKAN. First, the data distribution offset is eliminated by reversible normalization. Then, the multi-source time series data is decomposed into approximation coefficients and detail coefficients using the Daubechies4 wavelet basis. Subsequently, the patterns of different frequency components are learned in parallel by independent KAN branches. Finally, after multi-level wavelet reconstruction and inverse normalization, the fire risk probability in the next 24 hours is output. The model is trained using a loss function combining mean squared error and cross-entropy, along with an early stopping strategy. S3: Online Prediction After the model is deployed, it receives online sensor and meteorological data streams in real time. After preprocessing, the data is input into the trained lightweight time series prediction model, which dynamically outputs the fire risk probability curve for the next 24 hours. When the probability exceeds a preset threshold, a tiered warning is automatically triggered, and the risk area and its evolution trend are displayed through a visual interface.

2. The method for constructing a lightweight time-series prediction model for fire situation prediction according to claim 1, characterized in that, The steps of S2: model construction and training specifically include: S201: Perform reversible normalization on the input time series data and map it to a uniform scale; S202: Iteratively decompose the time series data after normalization and transpose in step S201 using the Daubechies4 wavelet basis to generate approximation coefficients and detail coefficients; S203: Each coefficient sequence output from step S202 is processed by an independent KAN branch; S204: Synthesize the predicted coefficient sequence output from the KAN branch into the time domain using inverse multi-level wavelet transform: Generated time series data Characterize the prediction results in the normalized channel preference space; tensor Transposed to Then, a reversible normalization inverse process is performed to obtain the predicted sequence in the original data space. .

3. The method for constructing a lightweight time-series prediction model for fire situation prediction according to claim 2, characterized in that, S202: The step of iteratively decomposing the time series data after normalization and transpose in step S201 using the Daubechies4 wavelet basis to generate approximation coefficients and detail coefficients specifically includes: Decomposition layer number As a hyperparameter, the default value is 1 level, which is adjusted according to the input length. According to the formula The balance between resolution and computational cost was determined and empirically verified. The DWT process employs a symmetrical fill strategy to handle boundary effects and decomposes the generated... A sequence of coefficients: in, For the first Layer approximation coefficient sequence, This corresponds to the sequence of detail coefficients for each layer.

4. The method for constructing a lightweight time-series prediction model for fire situation prediction according to claim 2, characterized in that, In step S203: Each coefficient sequence output from step S202 is processed by an independent KAN branch. The process for each branch includes: ① Input Perform a reversible normalization-forward operation to generate a normalized sequence. ,in Indicates the first One KAN branch; ② The normalized sequence is processed by block partitioning, including padding and overlap calculations, to obtain... : in, Indicates the number of blocks. The block size (sliding step size is) ); After further embedding, the block results are projected onto the dimension. The formula is: Among them, tensor As the input to the first KAN mixer; ③ The data stream passes through a time-series KAN mixer and a feature-based KAN mixer, and retains the original information through residual connections, as shown in the following formula: = ⊕ In the formula, For hybrid characterization, For the original embedded input, This is the output of the second KAN mixer; ④ Restore the branch output to the original scale to obtain the branch prediction coefficients. .

5. A method for constructing a lightweight time-series prediction model for fire situation prediction according to claim 2, characterized in that, The S2 step, model construction and training, also includes: Each KAN linearization layer introduces a composite regularization loss: The loss function contains two key components: the balance coefficient and the equilibrium coefficient. and Set to 1.0, and incorporate it into the total loss function through the global regularization intensity hyperparameter γ. : in, The mean squared error between the predicted and actual values. .

6. A method for constructing a lightweight time-series prediction model for fire situation prediction according to claim 2, characterized in that, The S2 step, model construction and training, also includes: Input the labeled training dataset, optimize the DecoKAN parameters and early stopping strategy through backpropagation, and save the optimal model parameters.

7. The method for constructing a lightweight time-series prediction model for fire situation prediction according to claim 1, characterized in that, S1: The data acquisition and preprocessing steps specifically include: S101: Based on the target steel plant's process flow diagram, equipment layout diagram, and hazardous substance safety data sheet, identify potential ignition sources and typical fire scenarios, confirm fire risk sources, and then compile fire safety regulations, industry accident cases, and domain expert experience to build an expert rule base for data completion. S102: For cases where the factory has no fire history records but has normal sensor data, the expert rule base in step S101 is used to scan the time series segments under normal operating conditions to obtain the reverse labeling of risk probability. S103: Collect real-time data and historical data to form a multi-source data pool; S104: Clean and standardize the raw time-series data in the multi-source data pool to form a supervised learning sample set for model training.

8. A method for constructing a lightweight time-series prediction model for fire situation prediction according to claim 7, characterized in that, S102: For situations where the factory has no historical fire records but has normal sensor data, the expert rule base from step S101 is used to scan time series segments under normal operating conditions to obtain reverse labeling of risk probabilities. This specifically includes: Data segments that conform to a high-risk pattern are labeled as high-risk values, those that conform to a potential-risk pattern are labeled as medium-risk values, and those that are completely normal are labeled as low-risk values.

9. A method for constructing a lightweight time-series prediction model for fire situation prediction according to claim 7, characterized in that, S103: The steps for collecting real-time data and historical data to form a multi-source data pool are as follows: The real-time data includes time-series data of temperature and smoke concentration in various areas of the factory collected through a network of temperature and smoke sensors, and synchronously accessed local meteorological data and factory equipment operating status signals. The historical data includes records of fire events from previous years and corresponding multi-dimensional environmental data from a historical database.

10. A lightweight time-series prediction system for fire situation prediction, characterized in that, The system includes the lightweight time-series prediction model according to any one of claims 1-9, wherein the system collects multi-source time-series data through temperature sensors, smoke sensors, meteorological interfaces, equipment status interfaces and historical databases, and the lightweight time-series prediction model processes the multi-source time-series data and outputs the fire risk probability for the next 24 hours.

Images