A fire hazard prediction system based on deep learning

Abstract

The application discloses a fire danger prediction system based on deep learning, comprising a data acquisition module for acquiring multi-source monitoring data in a fire monitoring environment and forming time series input; a feature construction module for performing feature mapping processing on the time series input and generating stage feature sequence; a stage identification module for outputting a stage identification vector corresponding to a time index; a reversible flow prediction module for generating risk distribution parameters under stage consistency monotonicity constraint; a parameter optimization module for performing update on reversible flow model parameters by adopting a covariance matrix adaptive evolution strategy with introduction of stage consistency weight; and a result output module for generating fire danger prediction results and outputting. The application realizes continuous prediction and adaptive evolution of fire danger by stage consistency constraint and distribution modeling.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and fire safety technology, and in particular to a fire hazard prediction system based on deep learning. Background Technology

[0002] With the acceleration of urbanization and the increasing complexity of building structures, energy consumption methods, and equipment types, fire safety risks are characterized by multiple sources, concealment, and stages. Fire accidents often do not occur suddenly in a short period, but rather evolve gradually under the long-term cumulative effects of various factors such as changes in environmental conditions, abnormal electrical operation, and fluctuations in equipment operating conditions. How to effectively identify and predict the evolution trend of hazards before a fire occurs has become an urgent technical problem to be solved in the field of fire safety.

[0003] Existing fire monitoring technologies primarily rely on single or limited sensor data such as temperature, smoke, and current, triggering alarms through threshold judgment or rule matching. While these methods are effective for sudden and obvious anomalies, they struggle to accurately depict the risk change process when facing long-term, cumulative, and slowly evolving fire risks, often resulting in frequent false alarms or delayed warnings. Some technical solutions incorporate machine learning or deep learning models to analyze historical data, but most methods focus on single-point prediction or classification, lacking the ability to systematically model the uncertainty and evolutionary stages of risks.

[0004] Furthermore, existing technologies typically employ models with fixed structures or parameters for training and inference, resulting in limited adaptability to different regions, operating conditions, or time periods. When the monitoring environment changes or risk evolution patterns shift, the model struggles to adjust its internal parameter structure in a timely manner, leading to insufficient stability in prediction results. Simultaneously, most methods only output a single risk score, failing to provide information on risk distribution and thus hindering refined fire safety decision-making and tiered response.

[0005] Therefore, there is an urgent need for a fire hazard prediction technology that can integrate multi-source monitoring data, characterize the evolution stages of fire hazards, model risk distribution and uncertainty, and have the ability to adaptively adjust parameters, so as to improve the foresight and reliability of fire safety early warning. Summary of the Invention

[0006] One objective of this invention is to propose a fire hazard prediction system based on deep learning. This invention fully integrates multi-source temporal feature mapping, stage identification, a reversible flow model with stage consistency monotonic constraints, and an adaptive evolutionary mechanism driven by stage consistency weights. This enables continuous modeling of fire hazards from data acquisition to risk distribution prediction, possessing the characteristics of continuous risk characterization, consistent evolution, and adaptability to environmental changes. According to an embodiment of this invention, a fire hazard prediction system based on deep learning includes: The data acquisition module is used to collect multi-source monitoring data in the fire monitoring environment and generate time series inputs; The feature construction module is used to perform feature mapping on time series inputs to generate a unified feature representation sequence and a stage feature sequence; The stage identification module is used to output a stage identifier vector based on the stage feature sequence. The reversible flow prediction module is used to receive a unified feature representation sequence and a stage identifier vector and generate risk distribution parameters. The reversible flow prediction module includes a stage consistency monotonic constraint flow layer. The parameter optimization module is used to update the parameters of the stage identification module and the reversible flow prediction module by adopting an adaptive evolution strategy of the covariance matrix that introduces stage consistency weights. The results output module is used to generate and output fire hazard prediction results based on risk distribution parameters.

[0007] Optionally, modules can be integrated using the following methods: S1. Collect multi-source monitoring data from the fire monitoring environment and arrange the multi-source monitoring data according to the time index to form multi-dimensional time series input data; S2. Perform feature mapping processing on the multidimensional time series input data, construct a unified feature representation sequence, and generate a stage feature sequence from the unified feature representation sequence; S3. Input the stage feature sequence into the stage identification module and output the stage identifier vector that corresponds one-to-one with each time index. S4. Input the unified feature representation sequence and the stage identifier vector into the reversible flow model in parallel. The reversible flow model includes a stage-consistent monotonic constraint flow layer. Within the time interval corresponding to the same stage identifier vector, the stage-consistent monotonic constraint flow layer performs monotonic constraint mapping on the transformation parameters corresponding to the preset risk-driven components and outputs risk distribution parameters. S5. Generate the risk distribution sequence corresponding to each time index based on the risk distribution parameters, and extract the risk center value sequence and uncertainty measure sequence from the risk distribution sequence; S6. Construct an evaluation vector with the stability index of the stage identifier vector and the change trend of the uncertainty measure sequence as input, and use the evaluation vector as the fitness input. Adopt an adaptive evolution strategy of covariance matrix with stage consistency weight to perform iterative updates on the parameters of the stage identification module and the parameters of the reversible flow model. S7. After the parameters are updated, generate fire hazard prediction results based on the risk distribution sequence and output the prediction results.

[0008] Optionally, the multi-source monitoring data in step S1 specifically includes: Environmental status monitoring data, electrical operation status data, equipment operation status data, space and ventilation status data, and historical hazard and event statistics.

[0009] Optionally, S2 specifically includes: S21. Establish a field position table for the multidimensional time series input data according to the data source. The field position table records the field identifier and the fixed position of the field in the input vector. At each time index position, read the corresponding field value according to the field position table, perform missing marker writing and value normalization processing to form the original feature vector corresponding to the time index. S22. Input the original feature vector into the mapping weight matrix and the mapping bias vector, perform linear projection mapping, output fixed-dimensional feature vectors, and arrange the fixed-dimensional feature vectors according to the time index to form a unified feature representation sequence; S23. Extract continuous time window segments from the unified feature representation sequence according to the preset window length. Calculate the first-order difference vector, second-order difference vector, cumulative change vector, and volatility vector for each window segment. Then, concatenate them in the preset position order to form a stage feature vector. Arrange the stage feature vectors according to the time index to generate a stage feature sequence. At the same time, establish a risk-driven component index table and write it into the transformation parameter constraint input buffer of the reversible flow model.

[0010] Optionally, S3 specifically includes: S31. Read the stage feature vectors from the stage feature sequence in order of time index, and establish the stage feature matrix. The row index of the stage feature matrix corresponds one-to-one with the time index. S32. Perform adjacent row difference operation on the stage feature matrix according to the time index to generate the stage change matrix, and perform window aggregation operation on the stage change matrix according to the preset window length to generate the stage change intensity sequence. S33. Input the stage change intensity sequence into the stage discrimination mapping layer. The stage discrimination mapping layer contains a discrimination weight matrix and a discrimination bias vector. The stage discrimination mapping layer outputs a stage identifier vector for the stage change intensity sequence at time index. The stage identifier vector is composed of identifier components of a preset dimension. Each identifier component corresponds one-to-one with the preset stage category position in the vector position. The stage identifier vector is then output after establishing a correspondence between the stage identifier vector and the time index.

[0011] Optionally, S4 specifically includes: S41. Read the unified feature representation sequence and the stage identifier vector, write the unified feature representation sequence into the reversible flow model input buffer according to the time index, write the stage identifier vector into the stage buffer according to the time index, and read the set of dimension positions corresponding to the preset risk-driven components from the risk-driven component index table. S42. At each time index position, read the unified feature representation vector from the input buffer of the reversible flow model and input it into the parameter generation subnetwork of the reversible flow model. The parameter generation subnetwork outputs a set of transformation parameters, which includes the coupling layer scaling parameter vector and the coupling layer translation parameter vector. S43. Read the stage identifier vector from the stage buffer, perform stage interval parsing on the stage identifier vector according to the preset stage category position, and form a stage interval table. The stage interval table records the stage start time index and the stage end time index. S44. For each stage interval in the stage interval table, read the set of transformation parameters corresponding to each time index in the stage interval, and perform monotonic constraint mapping sign fixing and amplitude upper bound processing on the scaling parameter components in the transformation parameter set that correspond to the preset risk-driven components. The sign fixing process maps the scaling parameter components, the monotonic constraint mapping includes parameter values ​​that map the scaling parameter components to the same preset sign direction, and the amplitude upper bound processing trims the scaling parameter components to the preset amplitude interval. S45. Write back the set of transformation parameters that have completed the monotonic constraint mapping to the transformation parameter buffer area corresponding to the time index in the stage interval, and read the scaling parameter vector and translation parameter vector in the transformation parameter buffer area at each time index position. Perform the coupling layer invertible transformation on the unified feature representation vector. The coupling layer invertible transformation outputs the risk latent variable vector corresponding to the time index. S46. Perform probability distribution parameter mapping operation on the risk latent variable vector. The probability distribution parameter mapping operation includes inputting the risk latent variable vector into the distribution head mapping matrix and the distribution head bias vector, and outputting risk distribution parameters. The risk distribution parameters include risk center value parameters and uncertainty measurement parameters, and arrange the risk distribution parameters by time index to form a risk distribution parameter sequence.

[0012] Optionally, the probability distribution parameter mapping operation in step S46 specifically includes: At each time index position, the latent variable value corresponding to the time index is read from the risk latent variable vector output by step S45, and the risk latent variable vector is written into the distribution mapping input buffer. Input the risk latent variable vector into the risk center value mapping matrix and the risk center value bias vector, perform a linear mapping operation, and output the risk center value parameter corresponding to the time index; Input the same risk latent variable vector into the uncertainty metric mapping matrix and the uncertainty metric bias vector, perform a linear mapping operation, and output the uncertainty metric parameter corresponding to the time index; The risk center value parameter and the uncertainty measurement parameter are combined in the order of preset parameter positions to form a risk distribution parameter vector that corresponds one-to-one with the time index. The risk distribution parameter vectors are arranged in time index order to generate a risk distribution parameter sequence, which is then written into the risk distribution cache.

[0013] Optionally, S5 specifically includes: S51. Read the risk distribution parameter vector in the risk distribution parameter sequence according to the time index, write the risk distribution parameter vector into the risk distribution buffer, and perform parameter position parsing on the risk distribution parameter vector according to the preset parameter position, and read the risk center value parameter and uncertainty measurement parameter respectively. S52. At each time index position, the risk center value parameter and the uncertainty measure parameter are combined to form a distribution instance parameter group, and the distribution instance parameter group is written into the risk distribution sequence buffer in time index order to generate a risk distribution sequence arranged by time index. S53. Read the risk center value parameters from the risk distribution parameter vector in time index order and arrange them to form a risk center value sequence. Read the uncertainty measurement parameters from the risk distribution parameter vector in time index order and arrange them to form an uncertainty measurement sequence. Write the risk center value sequence and the uncertainty measurement sequence into the evaluation vector input buffer.

[0014] Optionally, S6 specifically includes: S61. Read each identifier component in the stage identifier vector according to the time index, calculate the non-zero count of the difference result of the stage identifier components of adjacent time indices, and form the stage identifier vector stability index by the ratio of the non-zero count to the total number of time indices. S62. Read the uncertainty measurement parameters in the uncertainty measurement sequence according to the time index, calculate the difference results of the uncertainty measurement parameters of adjacent time indices, and count the sign consistency count and the cumulative absolute value of the difference results to form an uncertainty measurement sequence change trend index. S63. The stability index of the stage identifier vector and the trend index of the uncertainty measurement sequence are concatenated in a preset position order to form an evaluation vector, and the evaluation vector is written into the fitness cache area. S64. Initialize the mean vector, covariance matrix and step size parameters of the adaptive evolution strategy. The dimension position of the mean vector corresponds one-to-one with the parameter subset of the stage identification module and the parameter subset of the reversible flow model. S65. In each iteration, a set of candidate parameter vectors is generated based on the mean vector, covariance matrix and step size parameter, and the candidate parameter vectors are split and written into the parameter cache of the stage identification module and the parameter cache of the reversible flow model. S66. For each candidate parameter vector, call S3 to S5 to generate the corresponding stage identifier vector, risk center value sequence and uncertainty measurement sequence, and generate the corresponding evaluation vector according to S61 to S63. S67. Input the evaluation vector into the weight mapping matrix and the weight mapping bias vector to generate the phase consistency weight. Perform weighted summation on the candidate parameter vector set according to the phase consistency weight to update the mean vector. Subtract the current mean vector from each candidate parameter vector in the candidate parameter vector set to generate the candidate parameter vector deviation vector set. Perform weighted extra-integral product on the candidate parameter vector deviation vector set according to the phase consistency weight to update the covariance matrix. At the same time, update the step size parameter based on the weighted norm statistic of the candidate parameter vector deviation vector set.

[0015] The beneficial effects of this invention are: (1) This invention integrates environmental status, electrical operating status, equipment condition and historical hazard information into the same prediction framework by unifying the modeling and time indexing of multi-source monitoring data, thereby realizing the continuous expression of fire risk information in the time dimension and avoiding the information fragmentation problem caused by relying solely on a single sensor signal or instantaneous threshold judgment.

[0016] (2) This invention introduces a phase feature sequence and a phase identification mechanism to characterize the evolution process of fire hazards in stages, so that the risk prediction results can reflect the differences and continuity between different evolution stages, thereby identifying long-term cumulative and slow-evolving fire risk states in advance.

[0017] (3) The present invention adopts a reversible flow model with stage consistency monotonic constraints to model the risk distribution. It not only outputs the risk center value, but also simultaneously characterizes the uncertainty measure, so that the prediction results are expressed in the form of distribution, which enhances the stability and interpretability of fire hazard prediction.

[0018] (4) This invention introduces a covariance matrix adaptive evolution strategy with stage consistency weights to dynamically update the model parameters, enabling the prediction model to continuously adjust the parameter structure as the monitoring environment and risk evolution mode change, thereby improving the system's adaptability and prediction reliability under complex operating conditions. Attached Figure Description

[0019] 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: Figure 1 This is a flowchart of a fire hazard prediction system based on deep learning proposed in this invention. Figure 2 This is a schematic diagram of the stage consistency weight-driven adaptive evolution of a deep learning-based fire hazard prediction system proposed in this invention. Figure 3 This is a schematic diagram of the reversible flow prediction based on the stage consistency monotonic constraint of a fire hazard prediction system based on deep learning proposed in this invention. Detailed Implementation

[0020] 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.

[0021] refer to Figure 1-3 A fire hazard prediction system based on deep learning includes: The data acquisition module is used to collect multi-source monitoring data in the fire monitoring environment and generate time series inputs; The feature construction module is used to perform feature mapping on time series inputs to generate a unified feature representation sequence and a stage feature sequence; The stage identification module is used to output a stage identifier vector based on the stage feature sequence. The reversible flow prediction module is used to receive a unified feature representation sequence and a stage identifier vector and generate risk distribution parameters. The reversible flow prediction module includes a stage consistency monotonic constraint flow layer. The parameter optimization module is used to update the parameters of the stage identification module and the reversible flow prediction module by adopting an adaptive evolution strategy of the covariance matrix that introduces stage consistency weights. The results output module is used to generate and output fire hazard prediction results based on risk distribution parameters.

[0022] In this embodiment, the modules are interconnected using the following method: S1. Collect multi-source monitoring data from the fire monitoring environment and arrange the multi-source monitoring data according to the time index to form multi-dimensional time series input data; S2. Perform feature mapping processing on the multidimensional time series input data, construct a unified feature representation sequence, and generate a stage feature sequence from the unified feature representation sequence; S3. Input the stage feature sequence into the stage identification module and output the stage identifier vector that corresponds one-to-one with each time index. S4. Input the unified feature representation sequence and the stage identifier vector into the reversible flow model in parallel. The reversible flow model includes a stage-consistent monotonic constraint flow layer. Within the time interval corresponding to the same stage identifier vector, the stage-consistent monotonic constraint flow layer performs monotonic constraint mapping on the transformation parameters corresponding to the preset risk-driven components and outputs risk distribution parameters. S5. Generate the risk distribution sequence corresponding to each time index based on the risk distribution parameters, and extract the risk center value sequence and uncertainty measure sequence from the risk distribution sequence; S6. Construct an evaluation vector with the stability index of the stage identifier vector and the change trend of the uncertainty measure sequence as input, and use the evaluation vector as the fitness input. Adopt an adaptive evolution strategy of covariance matrix with stage consistency weight to perform iterative updates on the parameters of the stage identification module and the parameters of the reversible flow model. S7. After the parameters are updated, generate fire hazard prediction results based on the risk distribution sequence and output the prediction results.

[0023] In this embodiment, the multi-source monitoring data in step S1 specifically includes: Environmental status monitoring data, electrical operation status data, equipment operation status data, space and ventilation status data, and historical hazard and event statistics.

[0024] In this embodiment, S2 specifically includes: S21. Establish a field position table for the multidimensional time series input data according to the data source. The field position table records the field identifier and the fixed position of the field in the input vector. At each time index position, read the corresponding field value according to the field position table, perform missing marker writing and value normalization processing to form the original feature vector corresponding to the time index. S22. The original feature vectors are input to the mapping weight matrix and the mapping bias vector to perform linear projection mapping, outputting fixed-dimensional feature vectors, and the fixed-dimensional feature vectors are arranged by time index to form a unified feature representation sequence; the linear projection mapping is completed by matrix multiplication and vector addition, and is used to map the original feature vectors from different sources to a unified dimensional space.

[0025] S23. Extract continuous time window segments from the unified feature representation sequence according to the preset window length. Calculate the first-order difference vector, second-order difference vector, cumulative change vector, and volatility vector for each window segment. Then, concatenate them in the preset position order to form a stage feature vector. Arrange the stage feature vectors according to the time index to generate a stage feature sequence. At the same time, establish a risk-driven component index table and write it into the transformation parameter constraint input buffer of the reversible flow model.

[0026] In this embodiment, the cumulative transformation vector in step S2 specifically includes: Within each window segment, the system first reads the unified feature representation vector corresponding to the window's start time index and uses it as the base feature vector for that window segment. Subsequently, the system sequentially reads the unified feature representation vectors corresponding to each time index within the window segment and performs a one-dimensional difference operation between each feature vector and the base feature vector. The system then accumulates and sums all the difference results within the window segment along the time dimension, forming a cumulative change vector with the same dimension as the unified feature representation vector. The unified feature representation vector is a fixed-dimensional numerical vector formed by aligning, normalizing, and linearly mapping multi-source monitoring data at the same time index position, used to represent the comprehensive status information corresponding to that time index.

[0027] In this embodiment, the volatility vector in step S2 specifically includes: Within each window segment, the uniform feature representation vectors are read in chronological order, and the mean vector of the feature vectors within the window segment is calculated along the time dimension. Next, the dimension-wise deviations between the uniform feature representation vectors corresponding to each time index within the window segment and the mean vector are statistically analyzed. The deviation results for each dimension are then processed by squaring, averaging, and square rooting along the time dimension to obtain a volatility vector reflecting the dispersion of feature changes within the window segment. The dimension of the volatility vector is consistent with that of the uniform feature representation vector.

[0028] In this embodiment, S3 specifically includes: S31. Read the stage feature vectors from the stage feature sequence in order of time index, and establish the stage feature matrix. The row index of the stage feature matrix corresponds one-to-one with the time index. S32. Perform adjacent row difference operation on the stage feature matrix according to the time index to generate the stage change matrix, and perform window aggregation operation on the stage change matrix according to the preset window length to generate the stage change intensity sequence. S33. Input the stage change intensity sequence into the stage discrimination mapping layer. The stage discrimination mapping layer contains a discrimination weight matrix and a discrimination bias vector. The stage discrimination mapping layer outputs a stage identifier vector for the stage change intensity sequence at time index. The stage identifier vector is composed of identifier components of a preset dimension. Each identifier component corresponds one-to-one with the preset stage category position in the vector position. The stage identifier vector is then output after establishing a correspondence between the stage identifier vector and the time index.

[0029] In this embodiment, S4 specifically includes: S41. Read the unified feature representation sequence and the stage identifier vector. Write the unified feature representation sequence into the input buffer of the reversible flow model according to the time index. Write the stage identifier vector into the stage buffer according to the time index. Read the set of dimension positions corresponding to the preset risk driving components from the risk driving component index table. Pre-configure the risk distribution parameter mapping substructure in the reversible flow model. The risk distribution parameter mapping substructure includes the risk center value mapping matrix, the risk center value bias vector, the uncertainty metric mapping matrix, and the uncertainty metric bias vector.

[0030] S42. At each time index position, read the unified feature representation vector from the input buffer of the reversible flow model and input it into the parameter generation sub-network of the reversible flow model. The parameter generation sub-network outputs a set of transformation parameters, which includes the coupling layer scaling parameter vector and the coupling layer translation parameter vector. S43. Read the stage identifier vector from the stage buffer, perform stage interval parsing on the stage identifier vector according to the preset stage category position, and form a stage interval table. The stage interval table records the stage start time index and the stage end time index. S44. For each stage interval in the stage interval table, read the transformation parameter set corresponding to each time index within the stage interval. The transformation parameter set includes the coupling layer scaling parameter vector and the coupling layer translation parameter vector. Perform monotonic constraint mapping sign fixing processing and amplitude upper bound processing on the scaling parameter components in the transformation parameter set that correspond to the preset risk-driven components. The sign fixing processing maps the scaling parameter components. The monotonic constraint mapping includes parameter values ​​that map the scaling parameter components to the same preset sign direction. The amplitude upper bound processing trims the scaling parameter components to the preset amplitude interval. S45. Write back the set of transformation parameters that have completed the monotonic constraint mapping to the transformation parameter buffer area corresponding to the time index in the stage interval, and read the scaling parameter vector and translation parameter vector in the transformation parameter buffer area at each time index position. Perform the coupling layer invertible transformation on the unified feature representation vector. The coupling layer invertible transformation outputs the risk latent variable vector corresponding to the time index. S46. In the reversible flow model, a risk distribution parameter mapping substructure is pre-configured. This substructure includes a distribution head mapping matrix and a distribution head bias vector. A probability distribution parameter mapping operation is performed on the risk latent variable vector. This operation involves inputting the risk latent variable vector into the distribution head mapping matrix and the distribution head bias vector, and outputting risk distribution parameters. These parameters include a risk center value parameter and an uncertainty metric parameter, and are arranged by time index to form a risk distribution parameter sequence.

[0031] In this embodiment, the reversible change of the coupling layer in step S45 specifically includes: At each time index, the system reads the coupling layer scaling parameter vector and coupling layer translation parameter vector corresponding to the current time index from the transformation parameter buffer. The system reads the unified feature representation vector corresponding to the current time index from the unified feature representation sequence and uses the unified feature representation vector as the input vector for the invertible transformation of the coupling layer.

[0032] According to the preset vector component partitioning rules, the input vector is divided into a first component subset that participates in the transformation and a second component subset that maintains pass-through. For the first component subset that participates in the transformation, the system performs numerical mapping processing on each component based on the coupling layer scaling parameter component and coupling layer translation parameter component at the corresponding position to generate the transformed component values; for the second component subset, the system keeps its original values ​​unchanged and directly passes them to the output.

[0033] After mapping the first subset, the transformed first subset and the unchanged second subset are merged according to their original component order to form an output vector with the same dimension as the input vector. The output vector is the risk latent variable vector corresponding to the current time index position.

[0034] In this embodiment, the execution phase interval parsing in step S43 specifically includes: After obtaining the stage identifier vectors arranged by time index, the stage identifier vectors are first parsed time-by-time index according to the preset stage category positions. Each stage identifier vector consists of several identifier components, and the position of each identifier component in the vector corresponds one-to-one with the preset stage category. The value of the identifier component is used to indicate whether the corresponding stage category is in a valid state at that time index position. The time-by-time index parsing reads the data or vector corresponding to each time index position in the time series in chronological order, and after completing the preset parsing operation at the current time index position, it proceeds to the processing of the next time index position.

[0035] Following the order of the time indices, starting from the beginning of the stage identifier vector sequence, the stage identifier vectors corresponding to each time index are read sequentially. At each time index position, each identifier component in the stage identifier vector is traversed according to the preset stage category position. When the value of a certain identifier component meets the valid determination condition of the stage, the time index is recorded as the candidate start time index of the corresponding stage category.

[0036] Within the time index position, the identifier component corresponding to the same stage category is continuously read, and it is determined whether the identifier component continuously meets the stage validity determination condition. The stage validity determination condition is used to determine whether the identifier component of the corresponding stage category in the stage identifier vector is valid at the current time index position, serving as the basis for determining the start and end of the stage interval. When it is detected that the identifier component corresponding to the stage category changes from meeting the stage validity determination condition to not meeting the stage validity determination condition at a certain time index position, the previous time index is recorded as the stage end time index of that stage category, and the recorded stage start time index and stage end time index are combined to form a complete stage interval.

[0037] For multiple consecutive valid intervals of the same stage category appearing in the stage identifier vector sequence, the corresponding stage start time index and stage end time index are recorded respectively in the manner described above, and the multiple stage intervals are sequentially written into the stage interval table. Each record in the stage interval table contains a stage category identifier, a stage start time index, and a stage end time index, and the record order in the stage interval table is consistent with the time index order.

[0038] After the traversal of the stage identifier vector sequence is completed, the end index completion process is performed on the unclosed stage intervals. The time index of the last time that meets the valid determination condition of the stage is recorded as the stage end time index, thus completing the construction of the stage interval table.

[0039] In this embodiment, the probability distribution parameter mapping operation in step S46 specifically includes: At each time index position, the latent variable value corresponding to the time index is read from the risk latent variable vector output by step S45, and the risk latent variable vector is written into the distribution mapping input buffer. Input the risk latent variable vector into the risk center value mapping matrix and the risk center value bias vector, perform a linear mapping operation, and output the risk center value parameter corresponding to the time index; Input the same risk latent variable vector into the uncertainty metric mapping matrix and the uncertainty metric bias vector, perform a linear mapping operation, and output the uncertainty metric parameter corresponding to the time index; The risk center value parameter and the uncertainty measurement parameter are combined in the order of preset parameter positions to form a risk distribution parameter vector that corresponds one-to-one with the time index. The risk distribution parameter vectors are arranged in time index order to generate a risk distribution parameter sequence, which is then written into the risk distribution cache.

[0040] In this embodiment, S5 specifically includes: S51. Read the risk distribution parameter vector in the risk distribution parameter sequence according to the time index, write the risk distribution parameter vector into the risk distribution buffer, and perform parameter position parsing on the risk distribution parameter vector according to the preset parameter position, and read the risk center value parameter and uncertainty measurement parameter respectively. S52. At each time index position, the risk center value parameter and the uncertainty measure parameter are combined to form a distribution instance parameter group, and the distribution instance parameter group is written into the risk distribution sequence buffer in time index order to generate a risk distribution sequence arranged by time index. S53. Read the risk center value parameters from the risk distribution parameter vector in time index order and arrange them to form a risk center value sequence. Read the uncertainty measurement parameters from the risk distribution parameter vector in time index order and arrange them to form an uncertainty measurement sequence. Write the risk center value sequence and the uncertainty measurement sequence into the evaluation vector input buffer.

[0041] In this embodiment, the parameter position parsing in step S51 specifically includes: At each time index position, the risk distribution parameter vector corresponding to the current time index is first read from the risk distribution parameter sequence and written to the risk distribution buffer. Based on preset parameter position rules, position parsing is performed on each parameter component in the risk distribution parameter vector. Parameter position parsing involves reading and splitting parameter components located at different fixed positions in the risk distribution parameter vector according to preset parameter position rules, distinguishing and outputting parameters with different meanings from the same parameter vector.

[0042] The preset parameter position rules are set during the model initialization phase and are used to clarify the positional correspondence of each parameter component in the risk distribution parameter vector. The rules include at least the fixed set of positions of the risk center value parameter and the fixed set of positions of the uncertainty measure parameter in the parameter vector.

[0043] When performing parameter location parsing, according to the preset parameter location rules, the components located in the risk center value parameter location set are first read from the risk distribution parameter vector, and the read components are combined to form the risk center value parameter corresponding to the current time index; the components located in the uncertainty measurement parameter location set are read from the risk distribution parameter vector, and the read components are combined to form the uncertainty measurement parameter corresponding to the current time index.

[0044] After completing the above reading and combination operations, the parsed risk center value parameter and uncertainty measurement parameter are written into the corresponding cache area or intermediate storage structure, respectively, and their one-to-one correspondence with the current time index is maintained.

[0045] In this embodiment, S6 specifically includes: S61. Read each identifier component in the stage identifier vector according to the time index, calculate the non-zero count of the difference result of the stage identifier components of adjacent time indices, and form the stage identifier vector stability index by the ratio of the non-zero count to the total number of time indices. S62. Read the uncertainty measurement parameters in the uncertainty measurement sequence according to the time index, calculate the difference results of the uncertainty measurement parameters of adjacent time indices, and count the sign consistency count and the cumulative absolute value of the difference results to form an uncertainty measurement sequence change trend index. S63. The stability index of the stage identifier vector and the trend index of the uncertainty measurement sequence are concatenated in a preset position order to form an evaluation vector, and the evaluation vector is written into the fitness cache area. S64. Initialize the mean vector, covariance matrix and step size parameters of the adaptive evolution strategy. The dimension position of the mean vector corresponds one-to-one with the parameter subset of the stage identification module and the parameter subset of the reversible flow model. S65. In each iteration, a set of candidate parameter vectors is generated based on the mean vector, covariance matrix and step size parameter, and the candidate parameter vectors are split and written into the parameter cache of the stage identification module and the parameter cache of the reversible flow model. S66. For each candidate parameter vector, call S3 to S5 to generate the corresponding stage identifier vector, risk center value sequence and uncertainty measurement sequence, and generate the corresponding evaluation vector according to S61 to S63. S67. Input the evaluation vector into the weight mapping matrix and the weight mapping bias vector to generate the phase consistency weight. Perform weighted summation on the candidate parameter vector set according to the phase consistency weight to update the mean vector. Subtract the current mean vector from each candidate parameter vector in the candidate parameter vector set to generate the candidate parameter vector deviation vector set. Perform weighted extra-integral product on the candidate parameter vector deviation vector set according to the phase consistency weight to update the covariance matrix. At the same time, update the step size parameter based on the weighted norm statistic of the candidate parameter vector deviation vector set.

[0046] Example 1: To verify the feasibility and effectiveness of this invention in practical applications, it was applied to a typical fire monitoring scenario. This scenario involves various electrical devices operating continuously for extended periods, environmental conditions changing with operational status, and historical instances of fire hazards arising from the combined effects of electrical anomalies and environmental changes. Traditional fire protection systems primarily rely on fixed thresholds or single signals for alarms, making it difficult to identify the evolution of risks in advance, often resulting in delayed alarms or false alarms. To address these issues, this invention constructs a deep learning-based fire hazard prediction system that continuously models and predicts the evolution of fire risks. In practical applications, the system continuously collects environmental status data, electrical operating status data, equipment operating status data, and historical hazard statistics, aligning these data according to time indices to form a multi-dimensional time series input. A feature construction module maps the multi-source time series into a unified feature representation and further generates a stage feature sequence reflecting the rhythm and cumulative trend of risk changes. A stage identification module parses the stage feature sequence, outputting a time-varying stage identifier vector to characterize the evolution of fire hazards from a stable state to a state of risk accumulation. In the risk prediction phase, the system simultaneously inputs the unified feature representation sequence and the stage identifier vector into the reversible flow prediction module. Under the influence of stage consistency monotonic constraints, the model constrains the direction of change of key risk-driving components within the same stage interval, causing the risk distribution to exhibit continuous evolution characteristics over time. The model output is no longer a single numerical value, but a sequence of risk distribution parameters containing risk center values ​​and uncertainty measures, thus reflecting the risk level and its fluctuation range. During parameter updates, the system constructs an evaluation vector based on the stability of the stage identifier vector and the changing trend of the uncertainty measure sequence, and introduces stage consistency weights to adjust the adaptive evolution strategy of the covariance matrix. Through multiple rounds of iterative updates, the model parameters gradually converge towards a direction that can stably characterize the risk evolution pattern, enabling the system to maintain good prediction consistency under different operating conditions. During continuous operation, the prediction results of this invention are compared and analyzed with traditional threshold-based fire early warning methods. The results show that this invention can provide risk escalation signals earlier than obvious anomalies, and the prediction results change smoothly over time, avoiding frequent fluctuations in alarms. In multiple risk accumulation processes, the system's accuracy in identifying high-risk stages is significantly higher than the comparative methods, and the lead time for risk changes is more stable.

[0047] Table 1: Comparison of Fire Hazard Prediction Effectiveness Data

[0048] Table 1 presents a comparison of the fire hazard prediction performance between the method of this invention and traditional thresholding methods and conventional machine learning methods. All indicators in the table are based on operational data collected within the same statistical period. By quantitatively comparing the prediction results of different methods under the same data conditions, the differences in risk identification timing, stage judgment stability, and result fluctuation control of different technical solutions can be intuitively reflected.

[0049] The indicator of early risk identification time reveals that traditional thresholding methods typically require 2 to 4 time indexing periods to reflect risk changes in the prediction results after risk begins to accumulate; conventional machine learning methods, under the same conditions, gradually reflect risk changes after 5 to 8 time indexing periods; while the method of this invention can output the risk upward trend 10 to 14 time indexing periods when the risk is still in its early evolution stage. This result demonstrates that this invention can reflect risk changes earlier in the time indexing dimension, reserving a more sufficient response period for subsequent handling.

[0050] Regarding the accuracy of high-risk stage identification, the traditional thresholding method achieves an accuracy of 0.70, the conventional machine learning method achieves an accuracy of 0.82, while the method of this invention achieves an accuracy of 0.93. Numerical results demonstrate that this invention, through stage feature construction and stage identification mechanisms, provides a more stable characterization of the risk evolution process, and maintains consistency in stage identification results across continuous time indices, thereby reducing the occurrence of stage misjudgments.

[0051] The continuity index of the prediction results shows that the traditional thresholding method has a continuity index of 0.61, the conventional machine learning method has a continuity index of 0.74, while the method of this invention reaches 0.88. This index reflects the magnitude of change of the prediction results between adjacent time indices. The risk center value output by this invention changes more smoothly with the time index, which is beneficial for continuous observation of risk evolution trends.

[0052] Regarding the frequency of false alarms, the traditional threshold method recorded 18 false alarms within the statistical period, the conventional machine learning method recorded 9 false alarms, while the method of this invention recorded 3 false alarms. Numerical comparison results show that the present invention demonstrates a more stable response to short-term abnormal disturbances during long-term operation, and effectively controls the number of false alarms.

[0053] Furthermore, the uncertainty fluctuation index shows that the traditional threshold method has an uncertainty fluctuation of 0.42, the conventional machine learning method has an uncertainty fluctuation of 0.28, while the method of this invention has an uncertainty fluctuation of 0.15. The uncertainty measure shows a convergence trend with time index, indicating that the present invention has a higher stability in risk assessment results. A comprehensive review of the data in Table 1 shows that the present invention demonstrates reliable technical performance in terms of early risk identification, stage identification accuracy, prediction continuity, and false alarm control.

[0054] 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 fire hazard prediction system based on deep learning, characterized in that, include: The data acquisition module is used to collect multi-source monitoring data in the fire monitoring environment and generate time series inputs; The feature construction module is used to perform feature mapping on time series inputs to generate a unified feature representation sequence and a stage feature sequence; The stage identification module is used to output a stage identifier vector based on the stage feature sequence. A reversible flow prediction module is used to receive a unified feature representation sequence and a stage identifier vector and generate risk distribution parameters. The reversible flow prediction module includes a reversible flow model, which includes a stage-consistent monotonic constraint flow layer. The parameter optimization module is used to update the parameters of the stage identification module and the reversible flow prediction module by adopting an adaptive evolution strategy of the covariance matrix that introduces stage consistency weights. The results output module is used to generate and output fire hazard prediction results based on risk distribution parameters.

2. The fire hazard prediction system based on deep learning according to claim 1, characterized in that, The modules are connected in the following way: S1. Collect multi-source monitoring data from the fire monitoring environment and arrange the multi-source monitoring data according to the time index to form multi-dimensional time series input data; S2. Perform feature mapping processing on the multidimensional time series input data, construct a unified feature representation sequence, and generate a stage feature sequence from the unified feature representation sequence; S3. Input the stage feature sequence into the stage identification module and output the stage identifier vector that corresponds one-to-one with each time index. S4. Input the unified feature representation sequence and the stage identifier vector into the reversible flow model in parallel. The reversible flow model includes a stage-consistent monotonic constraint flow layer. Within the time interval corresponding to the same stage identifier vector, the stage-consistent monotonic constraint flow layer performs monotonic constraint mapping on the transformation parameters corresponding to the preset risk-driven components and outputs risk distribution parameters. S5. Generate the risk distribution sequence corresponding to each time index based on the risk distribution parameters, and extract the risk center value sequence and uncertainty measure sequence from the risk distribution sequence; S6. Construct an evaluation vector with the stability index of the stage identifier vector and the change trend of the uncertainty measure sequence as input, and use the evaluation vector as the fitness input. Adopt an adaptive evolution strategy of covariance matrix with stage consistency weight to perform iterative updates on the parameters of the stage identification module and the parameters of the reversible flow model. S7. After the parameters are updated, generate fire hazard prediction results based on the risk distribution sequence and output the prediction results.

3. The fire hazard prediction system based on deep learning according to claim 2, characterized in that, The multi-source monitoring data in step S1 specifically includes: Environmental status monitoring data, electrical operation status data, equipment operation status data, space and ventilation status data, and historical hazard and event statistics.

4. The fire hazard prediction system based on deep learning according to claim 3, characterized in that, S2 specifically includes: S21. Establish a field position table for the multidimensional time series input data according to the data source. The field position table records the field identifier and the fixed position of the field in the input vector. At each time index position, read the corresponding field value according to the field position table, perform missing marker writing and value normalization processing to form the original feature vector corresponding to the time index. S22. Input the original feature vector into the mapping weight matrix and the mapping bias vector, perform linear projection mapping, output fixed-dimensional feature vectors, and arrange the fixed-dimensional feature vectors according to the time index to form a unified feature representation sequence; S23. Extract continuous time window segments from the unified feature representation sequence according to the preset window length. Calculate the first-order difference vector, second-order difference vector, cumulative change vector, and volatility vector for each window segment. Then, concatenate them in the preset position order to form a stage feature vector. Arrange the stage feature vectors according to the time index to generate a stage feature sequence. At the same time, establish a risk-driven component index table and write it into the transformation parameter constraint input buffer of the reversible flow model.

5. A fire hazard prediction system based on deep learning according to claim 4, characterized in that, S3 specifically includes: S31. Read the stage feature vectors from the stage feature sequence in order of time index, and establish the stage feature matrix. The row index of the stage feature matrix corresponds one-to-one with the time index. S32. Perform adjacent row difference operation on the stage feature matrix according to the time index to generate the stage change matrix, and perform window aggregation operation on the stage change matrix according to the preset window length to generate the stage change intensity sequence. S33. Input the stage change intensity sequence into the stage discrimination mapping layer. The stage discrimination mapping layer contains a discrimination weight matrix and a discrimination bias vector. The stage discrimination mapping layer outputs a stage identifier vector for the stage change intensity sequence at time index. The stage identifier vector is composed of identifier components of a preset dimension. Each identifier component corresponds one-to-one with the preset stage category position in the vector position. The stage identifier vector is then output after establishing a correspondence between the stage identifier vector and the time index.

6. A fire hazard prediction system based on deep learning according to claim 5, characterized in that, S4 specifically includes: S41. Read the unified feature representation sequence and the stage identifier vector, write the unified feature representation sequence into the reversible flow model input buffer according to the time index, write the stage identifier vector into the stage buffer according to the time index, and read the set of dimension positions corresponding to the preset risk-driven components from the risk-driven component index table. S42. At each time index position, read the unified feature representation vector from the input buffer of the reversible flow model and input it into the parameter generation subnetwork of the reversible flow model. The parameter generation subnetwork outputs a set of transformation parameters, which includes the coupling layer scaling parameter vector and the coupling layer translation parameter vector. S43. Read the stage identifier vector from the stage buffer, perform stage interval parsing on the stage identifier vector according to the preset stage category position, and form a stage interval table. The stage interval table records the stage start time index and the stage end time index. S44. For each stage interval in the stage interval table, read the set of transformation parameters corresponding to each time index in the stage interval, and perform monotonic constraint mapping sign fixing and amplitude upper bound processing on the scaling parameter components in the transformation parameter set that correspond to the preset risk-driven components. The sign fixing process maps the scaling parameter components, the monotonic constraint mapping includes parameter values ​​that map the scaling parameter components to the same preset sign direction, and the amplitude upper bound processing trims the scaling parameter components to the preset amplitude interval. S45. Write back the set of transformation parameters that have completed the monotonic constraint mapping to the transformation parameter buffer area corresponding to the time index in the stage interval, and read the scaling parameter vector and translation parameter vector in the transformation parameter buffer area at each time index position. Perform the coupling layer invertible transformation on the unified feature representation vector. The coupling layer invertible transformation outputs the risk latent variable vector corresponding to the time index. S46. Perform probability distribution parameter mapping operation on the risk latent variable vector. The probability distribution parameter mapping operation includes inputting the risk latent variable vector into the distribution head mapping matrix and the distribution head bias vector, and outputting risk distribution parameters. The risk distribution parameters include risk center value parameters and uncertainty measurement parameters, and arrange the risk distribution parameters by time index to form a risk distribution parameter sequence.

7. A fire hazard prediction system based on deep learning according to claim 6, characterized in that, The probability distribution parameter mapping operation in step S46 specifically includes: At each time index position, the latent variable value corresponding to the time index is read from the risk latent variable vector output by step S45, and the risk latent variable vector is written into the distribution mapping input buffer. Input the risk latent variable vector into the risk center value mapping matrix and the risk center value bias vector, perform a linear mapping operation, and output the risk center value parameter corresponding to the time index; Input the same risk latent variable vector into the uncertainty metric mapping matrix and the uncertainty metric bias vector, perform a linear mapping operation, and output the uncertainty metric parameter corresponding to the time index; The risk center value parameter and the uncertainty measurement parameter are combined in the order of preset parameter positions to form a risk distribution parameter vector that corresponds one-to-one with the time index. The risk distribution parameter vectors are arranged in time index order to generate a risk distribution parameter sequence, which is then written into the risk distribution cache.

8. A fire hazard prediction system based on deep learning according to claim 7, characterized in that, S5 specifically includes: S51. Read the risk distribution parameter vector in the risk distribution parameter sequence according to the time index, write the risk distribution parameter vector into the risk distribution buffer, and perform parameter position parsing on the risk distribution parameter vector according to the preset parameter position, and read the risk center value parameter and uncertainty measurement parameter respectively. S52. At each time index position, the risk center value parameter and the uncertainty measure parameter are combined to form a distribution instance parameter group, and the distribution instance parameter group is written into the risk distribution sequence buffer in time index order to generate a risk distribution sequence arranged by time index. S53. Read the risk center value parameters from the risk distribution parameter vector in time index order and arrange them to form a risk center value sequence. Read the uncertainty measurement parameters from the risk distribution parameter vector in time index order and arrange them to form an uncertainty measurement sequence. Write the risk center value sequence and the uncertainty measurement sequence into the evaluation vector input buffer.

9. A fire hazard prediction system based on deep learning according to claim 8, characterized in that, S6 specifically includes: S61. Read each identifier component in the stage identifier vector according to the time index, calculate the non-zero count of the difference result of the stage identifier components of adjacent time indices, and form the stage identifier vector stability index by the ratio of the non-zero count to the total number of time indices. S62. Read the uncertainty measurement parameters in the uncertainty measurement sequence according to the time index, calculate the difference results of the uncertainty measurement parameters of adjacent time indices, and count the sign consistency count and the cumulative absolute value of the difference results to form an uncertainty measurement sequence change trend index. S63. The stability index of the stage identifier vector and the trend index of the uncertainty measurement sequence are concatenated in a preset position order to form an evaluation vector, and the evaluation vector is written into the fitness cache area. S64. Initialize the mean vector, covariance matrix and step size parameters of the adaptive evolution strategy. The dimension position of the mean vector corresponds one-to-one with the parameter subset of the stage identification module and the parameter subset of the reversible flow model. S65. In each iteration, a set of candidate parameter vectors is generated based on the mean vector, covariance matrix and step size parameter, and the candidate parameter vectors are split and written into the parameter cache of the stage identification module and the parameter cache of the reversible flow model. S66. For each candidate parameter vector, call S3 to S5 to generate the corresponding stage identifier vector, risk center value sequence and uncertainty measurement sequence, and generate the corresponding evaluation vector according to S61 to S63. S67. Input the evaluation vector into the weight mapping matrix and the weight mapping bias vector to generate the phase consistency weight. Perform weighted summation on the candidate parameter vector set according to the phase consistency weight to update the mean vector. Subtract the current mean vector from each candidate parameter vector in the candidate parameter vector set to generate the candidate parameter vector deviation vector set. Perform weighted extra-integral product on the candidate parameter vector deviation vector set according to the phase consistency weight to update the covariance matrix. At the same time, update the step size parameter based on the weighted norm statistic of the candidate parameter vector deviation vector set.

Images