Early fire risk assessment method and system based on deep learning

Abstract

The application discloses a deep learning-based early fire risk assessment method and system, relates to the technical field of fire monitoring and intelligent early warning, and comprises the following steps: collecting multi-source fire monitoring data, preprocessing and generating standard fire monitoring data; constructing a fire evolution chain and generating a fire monitoring state sample; inputting an improved TSMixer model and generating stable fire state features; predicting fire monitoring states in a future time window and generating a future fire state sequence; calculating a fire risk probability and forming a current fire development state; determining a fire risk grade and generating graded early warning information; and performing replacement recalculation and generating linkage disposal instructions. The application realizes accurate identification, graded early warning and linkage disposal of early fire risks by constructing a fire evolution chain, improving a TSMixer model and fusing future state prediction and risk contribution analysis.

Description

Technical Field

[0001] This invention relates to the field of fire monitoring and intelligent early warning technology, and in particular to a method and system for early fire risk assessment based on deep learning. Background Technology

[0002] With the continuous expansion of industrial production, electrical equipment operation, and building fire protection, fire monitoring typically uses temperature sensors, smoke sensors, gas sensors, and current sensors to collect data on ambient temperature, equipment surface temperature, smoke concentration, carbon monoxide concentration, combustible gas concentration, oxygen concentration, and equipment operating current. Fire risks are identified using methods such as fixed threshold comparison, single-indicator anomaly judgment, or statistical analysis. Some existing methods incorporate deep learning models to process multi-source monitoring data, predict fire status based on historical monitoring sequences, and generate risk levels and early warning information according to the prediction results.

[0003] However, early-stage fires typically manifest as a slow rise in temperature, a weak accumulation of smoke and gas concentrations, and sequential anomalies in different monitoring indicators. Existing methods struggle to fully extract subtle increments within a continuous time window, the sequence of indicator responses, and the transmission relationships of anomalies. Instantaneous spikes and sensor fluctuations can also interfere with the extraction of state characteristics. Current methods lack sufficient joint analysis of future fire states, risk probabilities, and development status. Early warning results lack the traceability of leading indicators and stability verification, leading to untimely identification of early fire risks and a high risk of false alarms and missed alarms.

[0004] Therefore, how to provide a method and system for early fire risk assessment based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a deep learning-based method and system for early fire risk assessment. This invention comprehensively utilizes multi-source monitoring data processing, fire evolution chain construction, temporal feature extraction, and risk prediction methods, detailing the entire process from fire data acquisition, state sample generation, feature extraction, future state prediction, risk assessment to tiered early warning and coordinated response. This invention introduces continuous incremental back-calculation, sequential calculation of indicators, and residual round-trip verification into the TSMixer model, improving the ability to identify weak fire signs and the coordinated changes of multiple indicators. It determines the dominant early warning indicators and the stability of the results through abnormal indicator replacement and recalculation. Compared with existing technologies, this invention has the advantages of accurate early identification, strong anti-interference capability, low false alarm rate, and timely coordinated response.

[0006] The deep learning-based early fire risk assessment method according to embodiments of the present invention includes:

[0007] S1. Collect multi-source fire monitoring data within the monitoring area, preprocess the multi-source fire monitoring data, and generate standard fire monitoring data;

[0008] S2. Construct a fire evolution chain based on the change continuity and anomaly transmission relationship of standard fire monitoring data, and generate fire monitoring status samples based on the fire evolution chain;

[0009] S3. Input the fire monitoring status sample into the improved TSMixer model, and generate stable fire status characteristics through continuous incremental back calculation, sequential calculation of indicators and round-trip verification of residuals.

[0010] S4. Predict the fire monitoring status within the future time window based on the characteristics of stable fire status, and generate a future fire status sequence.

[0011] S5. Calculate the fire risk probability by combining the future fire state sequence and the characteristics of stable fire states to form the current fire development status;

[0012] S6. Determine the fire risk level by analyzing the fire risk probability, the current fire development status, and the future fire status sequence, and generate graded early warning information.

[0013] S7. Perform replacement and recalculation on each abnormal monitoring indicator that triggers the graded early warning information, determine the main early warning indicator and the stability of the result, and generate linkage response instructions to send to the fire monitoring terminal.

[0014] Optionally, the generation of standard fire monitoring data includes:

[0015] S11. Sort the multi-source fire monitoring data according to the monitoring location number and collection time, unify the time format and sampling interval, and form a multi-source synchronous monitoring sequence;

[0016] S12. For the monitoring values ​​of the multi-source synchronous monitoring sequence at each acquisition time, calculate the absolute difference between the monitoring values ​​and the median of the adjacent monitoring values ​​to determine the abnormal monitoring values, perform linear interpolation to replace the abnormal monitoring values ​​and fill in the missing monitoring values.

[0017] S13. Calculate the average value of the monitored values ​​as the baseline average, correct the data offset based on the baseline average, and generate standard fire monitoring data.

[0018] Optionally, the construction of the fire evolution chain and the generation of fire monitoring status samples include:

[0019] S21. Calculate the numerical difference and duration of continuous unidirectional change for each monitoring data in the standard fire monitoring data to form a change continuity relationship;

[0020] S22. Compare the variation range of the same monitoring data at adjacent monitoring locations in the multi-source synchronous monitoring sequence at each acquisition time to determine the anomaly propagation delay and anomaly propagation direction;

[0021] S23. Take the monitoring status of each monitoring location within each time window as chain nodes, connect the chain nodes according to the change continuity relationship and the anomaly transmission relationship, and construct the fire evolution chain.

[0022] S24. Extract the monitoring data, anomaly transmission delay, and anomaly transmission direction of each node in the fire evolution chain, and combine them in chronological order to generate a fire monitoring status sample.

[0023] Optionally, the generation of stable fire state characteristics includes:

[0024] S31. Input the fire monitoring status sample into the improved TSMixer model and extract the change between adjacent monitoring locations at each acquisition time according to the time sequence of the multi-source synchronous monitoring sequence.

[0025] S32. Continuously accumulate and back-calculate the changes in each monitoring data, and generate continuous incremental features based on the difference between the back-calculation results and the original monitoring trajectory.

[0026] S33. Compare the sequential relationship of changes in different monitoring data, align the monitoring values ​​of each multi-source synchronous monitoring sequence at each acquisition time according to the response interval, and generate the coordinated change characteristics of the indicators through feature mixing processing;

[0027] S34. Mix the continuous incremental features and the index co-change features in forward and reverse order respectively, restore the reverse order results and check the residuals with the forward order results to generate stable fire state features;

[0028] S35. Calculate the training error based on the characteristics of a stable fire state, update the model parameters through backpropagation, and obtain the improved TSMixer model after training.

[0029] Optionally, the obtained improved TSMixer model after training includes:

[0030] S351. Divide the fire monitoring status samples into training samples and validation samples. Input the training samples into the improved TSMixer model to obtain the monitoring value prediction results, fire status classification results and fire risk probability.

[0031] S352. Calculate the errors between the predicted monitoring values, the fire status classification results, and the fire risk probability and the corresponding actual results, and add up each error to obtain the training error.

[0032] S353. Perform backpropagation based on the training error and update the model parameters. Use validation samples to test the updated model. Repeat the training until the validation error no longer decreases, and obtain the improved TSMixer model after training.

[0033] Optionally, generating the future fire state sequence includes:

[0034] S41. Read the stable fire state characteristics according to the monitoring location and time sequence, and input the state characteristics of each monitoring data within the current time window into the prediction output layer of the improved TSMixer model.

[0035] S42. The prediction output layer converts the stable fire state characteristics into the index changes of each future time window along the time dimension, and obtains the future state transition amount corresponding to each monitoring index.

[0036] S43. Based on the monitoring values ​​of the multi-source synchronous monitoring sequence at each acquisition time and the first future state transition quantity, the predicted values ​​of monitoring indicators for each future time window are obtained. The predicted values ​​of monitoring indicators, the direction of change, and the duration of continuous change for each future time window are combined in chronological order to generate a future fire state sequence.

[0037] Optionally, obtaining the future state transition amounts corresponding to each monitoring indicator includes:

[0038] S421. Arrange the stable fire state characteristics according to the monitoring values ​​and time sequence at each acquisition time of the multi-source synchronous monitoring sequence to obtain the time variation characteristics;

[0039] S422. Multiply the time change features by the model parameters obtained from training the improved TSMixer model and sum them up, then add the corresponding correction value to obtain the change results for each future time window.

[0040] S423. Organize the changes of each future time window in the order of future time windows to obtain the future state transition quantities.

[0041] Optionally, the formation of the current fire development status includes:

[0042] S51. Calculate the proportion of future anomalies in the future fire state sequence, count the number of time windows of the change between adjacent monitoring locations in the multi-source synchronous monitoring sequence at each acquisition time, and obtain the anomaly persistence.

[0043] S52. Combine the characteristics of stable fire state, the proportion of future anomalies, and the duration of anomalies into fire risk characteristics, and obtain the fire risk probability based on the fire risk characteristics.

[0044] S53. Compare the fire risk probability with the preset state boundary value to determine whether the current fire development state is normal, early abnormal, continuous development or high-risk.

[0045] Optionally, generating tiered early warning information includes:

[0046] S61. Read the current fire development status and future fire status sequence, and count the number of abnormal monitoring values ​​and the number of consecutive abnormal time windows.

[0047] S62. Compare the fire risk probability with the risk threshold, and determine the fire risk level by combining the current fire development status, the number of abnormal monitoring indicators and the number of consecutive abnormal time windows.

[0048] S63. Match early warning signs, early warning locations, abnormal monitoring indicators, and response time limits according to the fire risk level to generate graded early warning information.

[0049] The deep learning-based early fire risk assessment system according to an embodiment of the present invention includes the following modules:

[0050] The data preprocessing module is used to collect multi-source fire monitoring data and complete preprocessing to generate standard fire monitoring data.

[0051] The evolution chain construction module is used to analyze the continuity of changes and the transmission of anomalies in monitoring data, construct fire evolution chains, and generate fire monitoring status samples.

[0052] The state feature extraction module is used to process fire monitoring state samples by improving the TSMixer model and generate stable fire state features.

[0053] The fire status prediction module is used to predict the monitoring status within a future time window based on the characteristics of stable fire status, and generate a future fire status sequence.

[0054] The risk assessment module is used to calculate the probability of fire risk by combining the sequence of future fire states with the characteristics of stable fire states, and to determine the current fire development status.

[0055] The graded early warning module is used to determine the risk level based on the fire risk probability, the current fire development status, and the future fire status sequence, and generate graded early warning information.

[0056] The coordinated response module is used to replace and recalculate abnormal monitoring indicators, determine the leading indicators for early warning and the stability of the results, and generate coordinated response instructions.

[0057] The beneficial effects of this invention are:

[0058] This invention improves the intelligent identification capability of early fire risks by integrating multi-source monitoring data such as ambient temperature, equipment surface temperature, smoke concentration, and gas concentration. By constructing a fire evolution chain, it can extract the continuity of changes in monitoring indicators within a continuous time window and the anomaly transmission relationship between different monitoring locations. By improving the TSMixer model settings for continuous incremental back-calculation, sequential calculation of indicators, and residual round-trip verification, it can highlight weak and persistent anomalies, identify the sequential response patterns of multiple indicators, and reduce the interference of instantaneous spikes and sensor fluctuations on the assessment results, thereby improving the stability and accuracy of fire state characteristics.

[0059] This invention achieves dynamic, graded early warning of fire risk by predicting future fire state sequences and calculating fire risk probability and current fire development status by combining stable fire state characteristics. Through item-by-item replacement and recalculation of abnormal monitoring indicators, the dominant early warning indicators and the stability of the results can be determined, generating coordinated response instructions and sending them to the fire monitoring terminal, enhancing the traceability of early warning results and the timeliness of response. This invention overcomes the problems of untimely fixed threshold judgment, insufficient multi-indicator correlation analysis, and high false alarm and missed alarm rates, providing technical support for early fire prevention and control in industrial sites, electrical equipment areas, and building spaces. Attached Figure Description

[0060] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0061] Figure 1 This is a flowchart of the deep learning-based early fire risk assessment method proposed in this invention;

[0062] Figure 2 This is a block diagram of the improved TSMixer model for early fire risk assessment based on deep learning proposed in this invention.

[0063] Figure 3 This is a functional diagram of the deep learning-based early fire risk assessment system proposed in this invention. Detailed Implementation

[0064] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0065] refer to Figure 1 and Figure 2 Deep learning-based early fire risk assessment methods include:

[0066] S1. Collect multi-source fire monitoring data within the monitoring area, preprocess the multi-source fire monitoring data, and generate standard fire monitoring data;

[0067] S2. Construct a fire evolution chain based on the change continuity and anomaly transmission relationship of standard fire monitoring data, and generate fire monitoring status samples based on the fire evolution chain;

[0068] S3. Input the fire monitoring status sample into the improved TSMixer model, and generate stable fire status characteristics through continuous incremental back calculation, sequential calculation of indicators and round-trip verification of residuals.

[0069] S4. Predict the fire monitoring status within the future time window based on the characteristics of stable fire status, and generate a future fire status sequence.

[0070] S5. Calculate the fire risk probability by combining the future fire state sequence and the characteristics of stable fire states to form the current fire development status;

[0071] S6. Determine the fire risk level by analyzing the fire risk probability, the current fire development status, and the future fire status sequence, and generate graded early warning information.

[0072] S7. Perform replacement and recalculation on each abnormal monitoring indicator that triggers the graded early warning information, determine the main early warning indicator and the stability of the result, and generate linkage response instructions to send to the fire monitoring terminal.

[0073] In this embodiment, generating standard fire monitoring data includes:

[0074] S11. Sort the multi-source fire monitoring data according to the monitoring location number and collection time, unify the time format and sampling interval, and form a multi-source synchronous monitoring sequence. The sorting of the multi-source fire monitoring data according to the monitoring location number and collection time is as follows:

[0075] The location identifiers uploaded by each monitoring device are uniformly converted into monitoring location numbers in a fixed format. The monitoring location number is used as the first sorting field to divide the data belonging to the same monitoring location into the same data group. The collection time in different formats is uniformly converted into a time value relative to the start time. The time value is the number of seconds between the collection time and the start time. The time value is used as the second sorting field to arrange the data in each data group from earliest to latest time.

[0076] For cases where there are multiple records for the same monitoring location, the same collection time, and the same monitoring indicator, the average value of each record is calculated as the monitoring value at that time. Data from the same collection time are combined according to the order of the monitoring indicators and assigned time numbers in sequence to form a multi-source fire monitoring data sequence grouped by monitoring location and ordered by collection time.

[0077] S12. For the monitoring values ​​of the multi-source synchronous monitoring sequence at each acquisition time, calculate the absolute difference between the abnormal monitoring values ​​and the median of the preceding and following monitoring values ​​to determine the abnormal monitoring values. Perform linear interpolation to replace the abnormal monitoring values ​​and fill in the missing monitoring values. Specifically, the calculation of the absolute difference between the abnormal monitoring values ​​and the median of the preceding and following monitoring values ​​to determine the abnormal monitoring values ​​is as follows:

[0078] Taking the data point to be detected as the center, select the same number of data points before and after it to form an adjacent time window. Calculate the median value of each monitoring value in the window after arranging them from smallest to largest, and use it as the median value of the adjacent time window. Subtract the median value from the monitoring value of the data point to be detected and take the absolute value to obtain the absolute difference of the data points. Calculate the median of the absolute differences of all data points in the window. Set three times the median as the anomaly judgment threshold. When the absolute difference of the data point to be detected is greater than the anomaly judgment threshold, the data point is determined to be an abnormal data point; otherwise, it is determined to be a valid data point.

[0079] S13. Calculate the average value of the monitored values ​​as the baseline mean, correct the data offset based on the baseline mean, and generate standard fire monitoring data. Specifically, the correction of data offset based on the baseline mean involves:

[0080] Select a continuous period of time in the monitoring area where there are no fire anomalies, the equipment is operating stably, and the data is valid. Calculate the ratio of the sum of all monitoring values ​​of each monitoring location and each monitoring indicator to the number of valid data in this period to obtain the baseline mean. Calculate the real-time mean of the corresponding monitoring indicator in the current normal monitoring period according to the preset correction cycle. Subtract the baseline mean from the real-time mean to obtain the drift offset.

[0081] By uniformly sorting, synchronizing, removing anomalies, filling in missing data, correcting drift, and standardizing multi-source fire monitoring data, data deviations caused by differences in sampling frequencies, time formats, and units of measurement of different sensors can be eliminated. This reduces the interference of abnormal data and sensor drift on subsequent analysis, ensuring that each monitoring location and each monitoring indicator corresponds continuously under the same time reference. This provides a complete, accurate, and consistent data foundation for fire evolution chain construction, steady state feature extraction, and fire risk prediction.

[0082] In this embodiment, the construction of the fire evolution chain and the generation of fire monitoring status samples include:

[0083] S21. Calculate the numerical difference and duration of continuous unidirectional change for each monitoring data point in the standard fire monitoring data to establish a continuity relationship. Specifically, the formation of a continuity relationship is as follows:

[0084] Standard fire monitoring data for adjacent time windows are read according to the monitoring location and monitoring indicators. The monitoring value of the current time window is subtracted from the monitoring value of the previous time window to obtain the indicator value difference. When the value difference is greater than zero, it is marked as rising; when it is less than zero, it is marked as falling; and when it is equal to zero, it is marked as stable. The direction of change of each adjacent time window is compared sequentially from the current time window. When the direction of change is continuous and consistent, the number of windows with the same direction of change is accumulated until the opposite direction or a stable state appears, and the accumulation stops. The number of windows with the same direction of change is multiplied by the time interval of adjacent time windows to obtain the duration of continuous change in the same direction. The monitoring location, monitoring indicator, start and end time of change, direction of change, value difference, and duration of continuous change in the same direction are correlated to form a change continuity relationship.

[0085] S22. Compare the variation range of the same monitoring data at adjacent monitoring locations in the multi-source synchronous monitoring sequence at each acquisition time to determine the anomaly propagation delay and anomaly propagation direction;

[0086] S23. Using the monitoring status of each monitoring location within each time window as chain nodes, connect the chain nodes according to the change continuity relationship and the anomaly transmission relationship to construct a fire evolution chain. The construction of the fire evolution chain specifically involves:

[0087] Each monitoring location's monitoring values, change direction, and abnormal status within each time window are combined into chain nodes, arranged from earliest to latest according to the collection time. For adjacent chain nodes with continuous and consistent change directions in the same monitoring location, a time connection is established based on the continuity of change, and the numerical difference and duration of continuous change in the same direction are recorded. For chain nodes with abnormal transmission relationships in different monitoring locations, the chain node corresponding to the previous abnormal location is pointed to the chain node of the next location where the abnormality occurs after the abnormal transmission delay, and the transmission direction and transmission delay are recorded. All chain nodes are integrated according to the time connection and the transmission connection between locations, and connections with opposite time order or no change relationship are deleted to form a fire evolution chain that can characterize the continuous change of fire monitoring indicators and the abnormal spread process.

[0088] S24. Extract the monitoring data, anomaly transmission delay and anomaly transmission direction of each node in the fire evolution chain, and combine them in chronological order to generate a fire monitoring status sample.

[0089] By analyzing the numerical differences, directions of change, and durations of continuous unidirectional changes of various monitoring indicators within adjacent time windows, and combining the anomaly transmission delay and direction between different monitoring locations to construct a fire evolution chain, it is possible to simultaneously characterize the temporal duration and spatial diffusion process of fire signs. This avoids judgments based solely on a single moment or a single monitoring location, improves the ability to identify weak continuous anomalies and multi-location associated anomalies, and provides fire monitoring state samples with temporal logic and transmission relationships for improving the TSMixer model to extract stable fire state characteristics.

[0090] In this embodiment, the generation of stable fire state characteristics includes:

[0091] S31. Input the fire monitoring status sample into the improved TSMixer model and extract the change between adjacent monitoring locations at each acquisition time according to the time sequence of the multi-source synchronous monitoring sequence.

[0092] S32. Continuously accumulate and back-calculate the changes in each monitoring data point, and generate continuous incremental features based on the difference between the back-calculation results and the original monitoring trajectory. Specifically, the generation of continuous incremental features includes:

[0093] The monitoring values ​​are read in a continuous sequence according to the monitoring location, monitoring index and time window number. The monitoring value of the current time window is subtracted from the monitoring value of the previous time window to obtain the index increment corresponding to the current time window. When the index increment is greater than zero, it is determined to be a positive increment, indicating that the monitoring index is rising. When the index increment is less than zero, it is determined to be a negative increment, indicating that the monitoring index is falling. When the index increment is equal to zero, it is determined to be unchanged. If the lengths of adjacent time windows are inconsistent, the index increment is divided by the time interval between the center time of the two time windows to obtain the index increment per unit time. All adjacent time windows are calculated in chronological order to form a continuous increment feature corresponding to the original monitoring trajectory.

[0094] S33. Compare the sequential relationship of changes in different monitoring data, align the monitoring indicators according to the response interval, and generate the indicator co-change characteristics through feature mixing processing. Specifically, the generation of indicator co-change characteristics is as follows:

[0095] Select the first and second indicator increment sequences at the same monitoring location and within the same time range. Keep the time position of the first indicator increment sequence unchanged. Move the second indicator increment sequence forward and backward by one time window in turn until the maximum number of misalignment windows is reached. After each move, only the increment data of the two sequences that overlap in time are retained. Calculate the sum of the products of the two indicator increments at the overlapping positions and divide it by the square root of the product of the sum of squares of the first and second indicator increments to obtain the ratio of the corresponding misalignment. Compare the absolute values ​​of all ratios and determine the number of misalignment windows when the absolute value is the largest as the response interval. Determine the sequential response relationship of the two indicators based on the direction of movement to obtain the indicator coordinated change characteristics.

[0096] S34. The continuous incremental features and the coordinated change features of the indicators are time-mixed according to forward and reverse order respectively. The reverse-order result is restored and its residual is checked against the forward-order result to generate stable fire state features. Specifically, the generation of stable fire state features is as follows:

[0097] The continuous incremental features are input into the time mixing layer of shared parameters in both forward and reverse time order. The time mixing layer includes normalization processing, first time mapping, nonlinear transformation and second time mapping. The feature values ​​of each monitoring index within the entire time window are normalized. The feature value of each time window is multiplied by the corresponding time weight and summed. After nonlinear transformation, it is mapped again according to the time weight to obtain the processing result containing the cross-time window change relationship. The reverse processing result is restored to the original time order. Half of the sum of the forward and reverse results is calculated to obtain the bidirectional average feature. The absolute value of the difference between the two is calculated to obtain the round-trip difference. The bidirectional average feature is divided by one and summed with the round-trip difference, and then added to the original continuous incremental feature to generate the stable fire state feature.

[0098] S35. Calculate the training error based on the characteristics of a stable fire state, update the model parameters through backpropagation, and obtain the improved TSMixer model after training.

[0099] In this embodiment, obtaining the trained improved TSMixer model includes:

[0100] S351. Divide the fire monitoring status samples into training samples and validation samples. Input the training samples into the improved TSMixer model to obtain the monitoring value prediction results, fire status classification results, and fire risk probability. Specifically, the monitoring value prediction results, fire status classification results, and fire risk probability are obtained as follows:

[0101] Each stable fire state characteristic is multiplied by its corresponding training coefficient and summed to obtain the predicted value of each monitoring indicator. The scores of the four states—normal, initial abnormal, continuous development, and high risk—are calculated in the same way. The state with the highest score is taken as the fire state classification result. The stable fire state characteristics, the predicted values ​​of the monitoring indicators, and the state scores are multiplied by their corresponding training coefficients and summed. The resulting values ​​are converted to the interval between 0 and 1 to obtain the fire risk probability.

[0102] S352. Calculate the errors between the predicted monitoring values, fire state classification results, and fire risk probabilities and their corresponding actual results, respectively. Add each error together to obtain the training error. Specifically, the calculation of the errors between the predicted monitoring values, fire state classification results, and fire risk probabilities and their corresponding actual results is as follows:

[0103] The square of the difference between the predicted value and the actual monitored value of each monitoring indicator is calculated and averaged to obtain the monitoring value prediction error. The actual fire state is recorded as 1 and the other states are recorded as 0. The sum of the squares of the differences between the output values ​​of the four states and the corresponding labels is calculated to obtain the state classification error. The square of the difference between the fire risk probability and the actual risk label is calculated to obtain the risk probability error. The three errors are added together to obtain the training error.

[0104] S353. Perform backpropagation based on the training error and update the model parameters. Use validation samples to test the updated model. Repeat the training until the validation error no longer decreases, and obtain the improved TSMixer model after training.

[0105] In this embodiment, generating the future fire state sequence includes:

[0106] S41. Read the stable fire state characteristics according to the monitoring location and time sequence, and input the state characteristics of each monitoring data within the current time window into the prediction output layer of the improved TSMixer model.

[0107] S42. The prediction output layer converts the stable fire state characteristics along the time dimension into the index changes for each future time window, obtaining the future state transition values ​​corresponding to each monitoring index. Specifically, the future state transition values ​​corresponding to each monitoring index are obtained as follows:

[0108] The prediction output layer presets 12 historical time windows and 6 future time windows. For each monitoring indicator, 72 time mapping weights and 6 bias values ​​are set. Each future time window corresponds to 12 time mapping weights and 1 bias value. When calculating any future time window, the stable fire state characteristics within the 12 historical time windows are multiplied by the corresponding 12 time mapping weights. The 12 products are added together and then the corresponding bias value is added to obtain the state transition amount of the monitoring indicator within that future time window. The 6 future time windows are calculated in sequence, and all monitoring indicators are processed separately. The results are arranged in chronological order to form the future state transition amount corresponding to each monitoring indicator.

[0109] S43. Based on the monitoring values ​​at each acquisition time and the first future state transition quantity of the multi-source synchronous monitoring sequence, the predicted values ​​of monitoring indicators for each future time window are obtained. The predicted values, direction of change, and duration of continuous change of the monitoring indicators for each future time window are combined in chronological order to generate a future fire state sequence. Specifically, the predicted values ​​of monitoring indicators for each future time window are obtained as follows:

[0110] Read the value of each monitoring indicator in the current time window and its corresponding 6 future state transitions. Add the value of the monitoring indicator in the current time window to the first future state transition to obtain the predicted value of the monitoring indicator in the first future time window. Add the first predicted value to the second future state transition to obtain the second predicted value. Repeat this process of adding the previous predicted value to the current future state transition to obtain the predicted value of the monitoring indicator in the sixth future time window. Perform the above processing on all monitoring indicators and arrange them in the order of the future time windows to form the predicted values ​​of the monitoring indicators for each future time window.

[0111] By converting stable fire state characteristics into future state transition quantities along the time dimension, and generating a future fire state sequence by recursively generating it window by window starting from the current monitoring value, it can continuously predict the values, directions of change, and duration of continuous change of each monitoring indicator in the next 6 time windows. It can identify the development trends of temperature rise, smoke accumulation, increase of combustible gas, and decrease of oxygen in advance, overcome the lag in judging fire risk based solely on the current monitoring value, and provide a complete predictive basis for calculating the proportion of future anomalies, the duration of anomalies, and the probability of fire risk.

[0112] In this embodiment, obtaining the future state transition amount corresponding to each monitoring indicator includes:

[0113] S421. Arrange the stable fire state characteristics according to the monitoring values ​​and time sequence at each acquisition time of the multi-source synchronous monitoring sequence to obtain the time variation characteristics;

[0114] S422. Multiply the time-varying features by the model parameters obtained from training the improved TSMixer model and sum them up. Then add the corresponding correction value to obtain the variation results for each future time window. The specific correction value is as follows:

[0115] The corresponding correction value is initially set to 0 and trained together with the model parameters. During training, the difference between the obtained change result and the actual change of the monitored value in the future time window is calculated. The corresponding correction value is adjusted in reverse according to the difference so that the change result gradually approaches the actual change. After training, the value obtained from the last update is used as the corresponding correction value of each monitoring indicator in each future time window.

[0116] S423. Organize the changes of each future time window in the order of future time windows to obtain the future state transition quantities.

[0117] In this embodiment, the formation of the current fire development state includes:

[0118] S51. Calculate the proportion of future anomalies in the future fire state sequence, and count the number of time windows for the change in adjacent monitoring locations at each acquisition time in the multi-source synchronous monitoring sequence to obtain the anomaly persistence. Specifically, the anomaly persistence is obtained as follows:

[0119] Read the predicted values ​​of each monitoring indicator in 6 future time windows in chronological order, compare them with the corresponding normal range, mark the time windows that exceed the normal range as abnormal windows, subtract the predicted value of the previous abnormal window from the predicted value of the next abnormal window, mark the difference as rising when the difference is greater than 0, mark the difference as falling when the difference is less than 0, and mark the difference as stable when the difference is equal to 0. Starting from the first abnormal window, continuously count the number of time windows with the same direction of change and always within the abnormal range. Stop counting when the direction of change changes or the predicted value returns to normal, and take the longest consecutive window number divided by 6 to obtain the abnormal persistence of the monitoring indicator.

[0120] S52. Combine the characteristics of stable fire state, the proportion of future anomalies, and the duration of anomalies into fire risk characteristics. Obtain the fire risk probability based on these characteristics. Specifically, obtaining the fire risk probability based on the fire risk characteristics involves:

[0121] The characteristics of stable fire status, the proportion of future anomalies, the duration of anomalies, and the changes in monitoring indicators for six future time windows are arranged in a fixed order. A corresponding model weight is set for each feature value, and a bias value is also set. The model weight and bias value are determined by the model training process. Each feature value is multiplied by the corresponding model weight, and all products are added together and the bias value is added to obtain the comprehensive risk value. The negative number of the comprehensive risk value is taken to calculate the exponent value. The exponent value is added to 1 and the reciprocal is taken to obtain the fire risk probability between 0 and 1. The closer the obtained value is to 1, the higher the risk of future fire occurrence and development.

[0122] S53. Compare the fire risk probability with preset state boundary values ​​to determine whether the current fire development state is normal, early abnormal, ongoing, or high-risk. Specifically, the comparison of the fire risk probability with preset state boundary values ​​is as follows:

[0123] When the fire risk probability is less than 0.25 and future monitoring indicators do not continuously change in the direction of increasing fire risk, it is determined to be a normal state. When the fire risk probability is greater than or equal to 0.25 and less than 0.50, and at least one monitoring indicator changes in the direction of increasing risk, it is determined to be an early abnormal state. When the fire risk probability is greater than or equal to 0.50 and less than 0.75, and at least two monitoring indicators change in the direction of increasing risk for three consecutive future time windows, it is determined to be a continuously developing state. When the fire risk probability is greater than or equal to 0.75, and the equipment surface temperature, smoke concentration, carbon monoxide concentration, or combustible gas concentration continuously rises, or the oxygen concentration continuously decreases, it is determined to be a high-risk state.

[0124] By integrating stable fire state characteristics, the proportion of future anomalies, the duration of anomalies, and changes in future monitoring indicators, the current state of fire signs, the scope of future anomalies, and the trend of continued development can be comprehensively reflected. Multiple risk characteristics are converted into fire risk probabilities between 0 and 1. Through three state thresholds and the direction of changes in future monitoring indicators, the current fire development state is divided into normal state, early abnormal state, continued development state, or high-risk state. This improves the quantitative assessment capability and accuracy of state determination of early and weak fire signs, and provides a reliable basis for subsequent risk classification, early warning information generation, and coordinated response.

[0125] In this embodiment, generating graded early warning information includes:

[0126] S61. Read the current fire development status and future fire status sequence, and count the number of abnormal monitoring values ​​and the number of consecutive abnormal time windows.

[0127] S62. Compare the fire risk probability with the risk threshold, and determine the fire risk level by combining the current fire development status, the number of abnormal monitoring indicators and the number of consecutive abnormal time windows, where the risk thresholds are set to 0.25, 0.50 and 0.75 respectively.

[0128] S63. Match early warning signs, early warning locations, abnormal monitoring indicators, and response time limits according to the fire risk level to generate graded early warning information.

[0129] refer to Figure 3 The deep learning-based early fire risk assessment system includes the following modules:

[0130] The data preprocessing module is used to collect multi-source fire monitoring data and complete preprocessing to generate standard fire monitoring data.

[0131] The evolution chain construction module is used to analyze the continuity of changes and the transmission of anomalies in monitoring data, construct fire evolution chains, and generate fire monitoring status samples.

[0132] The state feature extraction module is used to process fire monitoring state samples by improving the TSMixer model and generate stable fire state features.

[0133] The fire status prediction module is used to predict the monitoring status within a future time window based on the characteristics of stable fire status, and generate a future fire status sequence.

[0134] The risk assessment module is used to calculate the probability of fire risk by combining the sequence of future fire states with the characteristics of stable fire states, and to determine the current fire development status.

[0135] The graded early warning module is used to determine the risk level based on the fire risk probability, the current fire development status, and the future fire status sequence, and generate graded early warning information.

[0136] The coordinated response module is used to replace and recalculate abnormal monitoring indicators, determine the leading indicators for early warning and the stability of the results, and generate coordinated response instructions.

[0137] Example 1: To verify the feasibility of this invention in practice, this invention focuses on the operating area of ​​power distribution equipment, setting up 24 monitoring locations to collect ambient temperature, equipment surface temperature, smoke concentration, carbon monoxide concentration, combustible gas concentration, oxygen concentration, and equipment operating current, with a sampling interval of 30 seconds. 1800 training samples, 1200 validation samples, and 1500 test samples were used, each containing 12 historical time windows and 6 future time windows, covering normal states, early abnormal states, continuously developing states, and high-risk states.

[0138] In a set of samples showing temperature increases due to loose connections, the equipment surface temperature rose from 41.2°C to 52.9°C over 12 historical time windows; smoke concentration increased from 0.011 mg / m³ to 0.079 mg / m³; carbon monoxide concentration increased from 4.6 ppm to 17.8 ppm; equipment operating current increased from 45.3 amperes to 54.6 amperes; and oxygen concentration decreased from 20.83% to 20.42%. The original data contained 17 missing values, 9 jump values, and 3 sets of time-off data. After time alignment, outlier removal, linear interpolation, and drift correction, the missing value rate was reduced to 0, and the time offset was reduced to 0 seconds.

[0139] Difference calculations were performed on adjacent time windows, revealing that the equipment operating current increased for 11 consecutive windows, the equipment surface temperature increased for 10 consecutive windows, the smoke concentration increased for 8 consecutive windows, the carbon monoxide concentration increased for 7 consecutive windows, and the oxygen concentration decreased for 6 consecutive windows. Lag calculations showed that the equipment surface temperature lagged behind the current by 1 window, the smoke concentration by 3 windows, and the carbon monoxide concentration by 4 windows. Based on this, a fire evolution chain and fire state samples were formed.

[0140] The fire state samples were input into an improved time-series hybrid network. The increments of equipment surface temperature for the most recent six windows were 0.9, 1.2, 1.3, 1.6, 1.9, and 2.2 degrees Celsius, and the increments of smoke concentration were 0.004, 0.006, 0.008, 0.011, 0.014, and 0.017 milligrams per cubic meter. The correlation strength between current and equipment surface temperature was 0.86, between equipment surface temperature and smoke concentration was 0.81, and between smoke concentration and carbon monoxide concentration was 0.78. The difference between forward and reverse processing decreased from 0.14 to 0.03, yielding stable fire state characteristics.

[0141] The predicted state transitions for the equipment surface temperature over the next six time windows are 2.4, 2.7, 3.0, 3.4, 3.7, and 4.1 degrees Celsius, corresponding to predicted values ​​of 55.3, 58.0, 61.0, 64.4, 68.1, and 72.2 degrees Celsius. The predicted smoke concentration is 0.099 to 0.311 mg / m³, the predicted carbon monoxide concentration is 21.6 to 55.8 parts per million, and the predicted oxygen concentration decreases from 20.31% to 19.45%. The average prediction error for the equipment surface temperature is 1.7 degrees Celsius, and the average error for the smoke concentration is 0.012 mg / m³.

[0142] The future anomaly percentage and duration of the equipment surface temperature, smoke concentration, and carbon monoxide concentration are all 1, the oxygen concentration is 0.67, and the fire risk probability is 0.84, classifying it as a high-risk state and level 4 risk. After replacing the equipment surface temperature, the risk probability decreased to 0.48, with a risk contribution value of 0.36; after replacing the equipment operating current, the risk probability increased to 0.59, with a contribution value of 0.25; and after replacing the smoke concentration, the risk probability decreased to 0.66, with a contribution value of 0.18. Therefore, the equipment surface temperature was determined as the primary indicator for early warning, and commands for power outage, ventilation, activation of fire-fighting facilities, and personnel evacuation were generated.

[0143] In 1500 identical test samples, the fire status identification accuracy of the traditional method and the present invention were 84.6% and 94.8%, respectively; the early anomaly identification rate was 71.2% and 91.5%, respectively; the high-risk status false alarm rate was 12.8% and 3.1%, respectively; the average early warning lead time was 2.1 minutes and 5.4 minutes, respectively; the average absolute error of risk probability was 0.118 and 0.054, respectively; and the false alarm rate of current fluctuation was 9.7% and 2.6%, respectively. This indicates that the present invention has higher identification accuracy, timely early warning, and stable results.

[0144] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based early fire risk assessment method, characterized in that, include: S1. Collect multi-source fire monitoring data within the monitoring area, preprocess the multi-source fire monitoring data, and generate standard fire monitoring data; S2. Construct a fire evolution chain based on the change continuity and anomaly transmission relationship of standard fire monitoring data, and generate fire monitoring status samples based on the fire evolution chain; S3. Input the fire monitoring status sample into the improved TSMixer model, and generate stable fire status characteristics through continuous incremental back calculation, sequential calculation of indicators and round-trip verification of residuals. S4. Predict the fire monitoring status within the future time window based on the characteristics of stable fire status, and generate a future fire status sequence. S5. Calculate the fire risk probability by combining the future fire state sequence and the characteristics of stable fire states to form the current fire development status; S6. Determine the fire risk level by analyzing the fire risk probability, the current fire development status, and the future fire status sequence, and generate graded early warning information. S7. Perform replacement and recalculation on each abnormal monitoring indicator that triggers the graded early warning information, determine the main early warning indicator and the stability of the result, and generate linkage response instructions to send to the fire monitoring terminal.

2. The deep learning-based early fire risk assessment method according to claim 1, characterized in that, The generated standard fire monitoring data includes: S11. Sort the multi-source fire monitoring data according to the monitoring location number and collection time, unify the time format and sampling interval, and form a multi-source synchronous monitoring sequence; S12. For the monitoring values ​​of the multi-source synchronous monitoring sequence at each acquisition time, calculate the absolute difference between the monitoring values ​​and the median of the adjacent monitoring values ​​to determine the abnormal monitoring values, perform linear interpolation to replace the abnormal monitoring values ​​and fill in the missing monitoring values. S13. Calculate the average value of the monitored values ​​as the baseline average, correct the data offset based on the baseline average, and generate standard fire monitoring data.

3. The deep learning-based early fire risk assessment method according to claim 1, characterized in that, The construction of the fire evolution chain and the generation of fire monitoring status samples include: S21. Calculate the numerical difference and duration of continuous unidirectional change for each monitoring data in the standard fire monitoring data to form a change continuity relationship; S22. Compare the variation range of the same monitoring data at adjacent monitoring locations in the multi-source synchronous monitoring sequence at each acquisition time to determine the anomaly propagation delay and anomaly propagation direction; S23. Take the monitoring status of each monitoring location within each time window as chain nodes, connect the chain nodes according to the change continuity relationship and the anomaly transmission relationship, and construct the fire evolution chain. S24. Extract the monitoring data, anomaly transmission delay and anomaly transmission direction of each node in the fire evolution chain, and combine them in chronological order to generate a fire monitoring status sample.

4. The deep learning-based early fire risk assessment method according to claim 1, characterized in that, The characteristics for generating a stable fire state include: S31. Input the fire monitoring status sample into the improved TSMixer model and extract the change between adjacent monitoring locations at each acquisition time according to the time sequence of the multi-source synchronous monitoring sequence. S32. Continuously accumulate and back-calculate the changes in each monitoring data, and generate continuous incremental features based on the difference between the back-calculation results and the original monitoring trajectory. S33. Compare the sequential relationship of changes in different monitoring data, align the monitoring values ​​of each multi-source synchronous monitoring sequence at each acquisition time according to the response interval, and generate the coordinated change characteristics of the indicators through feature mixing processing; S34. Mix the continuous incremental features and the index co-change features in forward and reverse order respectively, restore the reverse order results and check the residuals with the forward order results to generate stable fire state features; S35. Calculate the training error based on the characteristics of a stable fire state, update the model parameters through backpropagation, and obtain the improved TSMixer model after training.

5. The deep learning-based early fire risk assessment method according to claim 4, characterized in that, The improved TSMixer model that has been trained includes: S351. Divide the fire monitoring status samples into training samples and validation samples. Input the training samples into the improved TSMixer model to obtain the monitoring value prediction results, fire status classification results and fire risk probability. S352. Calculate the errors between the predicted monitoring values, the fire status classification results, and the fire risk probability and the corresponding actual results, and add up each error to obtain the training error. S353. Perform backpropagation based on the training error and update the model parameters. Use validation samples to test the updated model. Repeat the training until the validation error no longer decreases, and obtain the improved TSMixer model after training.

6. The deep learning-based early fire risk assessment method according to claim 1, characterized in that, The generation of the future fire state sequence includes: S41. Read the stable fire state characteristics according to the monitoring location and time sequence, and input the state characteristics of each monitoring data within the current time window into the prediction output layer of the improved TSMixer model. S42. The prediction output layer converts the stable fire state characteristics into the index changes of each future time window along the time dimension, and obtains the future state transition amount corresponding to each monitoring index. S43. Based on the monitoring values ​​of the multi-source synchronous monitoring sequence at each acquisition time and the first future state transition quantity, the predicted values ​​of monitoring indicators for each future time window are obtained. The predicted values ​​of monitoring indicators, the direction of change, and the duration of continuous change for each future time window are combined in chronological order to generate a future fire state sequence.

7. The deep learning-based early fire risk assessment method according to claim 6, characterized in that, The process of obtaining the future state transition values ​​corresponding to each monitoring indicator includes: S421. Arrange the stable fire state characteristics according to the monitoring values ​​and time sequence at each acquisition time of the multi-source synchronous monitoring sequence to obtain the time variation characteristics; S422. Multiply the time change features by the model parameters obtained from training the improved TSMixer model and sum them up, then add the corresponding correction value to obtain the change results for each future time window. S423. Organize the changes of each future time window in the order of future time windows to obtain the future state transition quantities.

8. The deep learning-based early fire risk assessment method according to claim 1, characterized in that, The current state of the fire includes: S51. Calculate the proportion of future anomalies in the future fire state sequence, count the number of time windows of the change between adjacent monitoring locations in the multi-source synchronous monitoring sequence at each acquisition time, and obtain the anomaly persistence. S52. Combine the characteristics of stable fire state, the proportion of future anomalies, and the duration of anomalies into fire risk characteristics, and obtain the fire risk probability based on the fire risk characteristics. S53. Compare the fire risk probability with the preset state boundary value to determine whether the current fire development state is normal, early abnormal, continuous development or high-risk.

9. The deep learning-based early fire risk assessment method according to claim 1, characterized in that, The generation of tiered early warning information includes: S61. Read the current fire development status and future fire status sequence, and count the number of abnormal monitoring values ​​and the number of consecutive abnormal time windows. S62. Compare the fire risk probability with the risk threshold, and determine the fire risk level by combining the current fire development status, the number of abnormal monitoring indicators and the number of consecutive abnormal time windows. S63. Match early warning signs, early warning locations, abnormal monitoring indicators, and response time limits according to the fire risk level to generate graded early warning information.

10. A deep learning-based early fire risk assessment system, executing the deep learning-based early fire risk assessment method according to any one of claims 1 to 9, characterized in that, Includes the following modules: The data preprocessing module is used to collect multi-source fire monitoring data and complete preprocessing to generate standard fire monitoring data. The evolution chain construction module is used to analyze the continuity of changes and the transmission of anomalies in monitoring data, construct fire evolution chains, and generate fire monitoring status samples. The state feature extraction module is used to process fire monitoring state samples by improving the TSMixer model and generate stable fire state features. The fire status prediction module is used to predict the monitoring status within a future time window based on the characteristics of stable fire status, and generate a future fire status sequence. The risk assessment module is used to calculate the probability of fire risk by combining the sequence of future fire states with the characteristics of stable fire states, and to determine the current fire development status. The graded early warning module is used to determine the risk level based on the fire risk probability, the current fire development status, and the future fire status sequence, and generate graded early warning information. The coordinated response module is used to replace and recalculate abnormal monitoring indicators, determine the leading indicators for early warning and the stability of the results, and generate coordinated response instructions.

Images