Fire early warning method, device, equipment and program product

Abstract

The invention provides a fire early warning method, device, equipment and program product. The method comprises the following steps: inputting smoke sensor data and environmental parameter data, which are collected in real time, in a target building into a deep learning model to obtain a smoke concentration predicted value of at least one time step in the future; the smoke concentration prediction capability of the deep learning model is trained based on a target sample, the target sample comprises historical smoke sensor data of a target building or similar buildings and environmental parameter data corresponding to the historical smoke sensor data, and the target sample is marked with a real smoke concentration value corresponding to at least one time step in the future. And based on the smoke concentration prediction value of the at least one time step, determining whether a fire early warning condition is satisfied. And if yes, spatial interpolation modeling is carried out based on the indoor structure diagram of the target building and the smoke sensor layout information in combination with the smoke concentration predicted value of the at least one time step, a smoke concentration distribution thermodynamic diagram of the target building is generated, and an evacuation path is determined and output according to the thermodynamic diagram.

Description

Technical Field

[0001] This application relates to the field of fire early warning technology, and in particular to a fire early warning method, device, equipment and program product. Background Technology

[0002] Smoke detection technology has undergone a significant evolution from mechanical detection to electronic sensing. Currently, the mainstream approach still relies on physical sensors to achieve real-time monitoring. These technologies can effectively detect the presence of smoke particles through physical principles such as photoelectric sensing, ionization detection, or laser scattering. However, they are essentially passive response monitoring and lack intelligent analysis and predictive capabilities. Summary of the Invention

[0003] This application proposes a fire early warning method, device, equipment, and program product, aiming to address the shortcomings of existing technologies in early identification of fire hazards, intelligent identification of smoke types, and control of false alarm rates in complex environments. The technical solution is as follows: Firstly, a fire early warning method is provided, including: The smoke sensor data and environmental parameter data collected in real time inside the target building are input into a deep learning model to obtain the predicted smoke concentration value of the target building at least one time step in the future; wherein, the smoke concentration prediction capability of the deep learning model is based on target samples trained on target samples, the target samples include historical smoke sensor data of the target building or similar buildings and environmental parameter data corresponding to the historical smoke sensor data, and are labeled with the actual smoke concentration value corresponding to the at least one time step in the future. Based on the predicted smoke concentration at at least one time step, determine whether the fire early warning conditions are met. If the fire warning conditions are met, then based on the indoor structural diagram of the target building and the smoke sensor layout information, combined with the predicted smoke concentration values ​​at at least one time step, spatial interpolation modeling is performed to generate a smoke concentration distribution heat map of the target building. Based on the smoke concentration distribution heat map, the evacuation route within the target building is determined and output.

[0004] Secondly, a fire early warning device is provided, comprising: The prediction module is used to input real-time collected smoke sensor data and environmental parameter data in the target building into a deep learning model to obtain the predicted smoke concentration value of the target building at least one time step in the future; wherein, the smoke concentration prediction capability of the deep learning model is based on target samples trained on target samples, the target samples include historical smoke sensor data of the target building or similar buildings and environmental parameter data corresponding to the historical smoke sensor data, and are labeled with the actual smoke concentration value corresponding to the at least one time step in the future. The judgment module is used to determine whether the fire early warning conditions are met based on the predicted smoke concentration value of the at least one time step. The mapping module is used to generate a heat map of the smoke concentration distribution of the target building by performing spatial interpolation modeling based on the indoor structure diagram of the target building and the smoke sensor layout information, combined with the predicted smoke concentration values ​​at at least one time step, if the fire early warning conditions are met. The planning module is used to determine and output the evacuation route within the target building based on the smoke concentration distribution heat map.

[0005] Thirdly, embodiments of this application provide an electronic device, including: a processor; and a memory configured to store computer-executable instructions, which, when executed, cause the processor to perform the method described in the first aspect.

[0006] Fourthly, a computer program product is provided, the computer program product including a computer-readable storage medium storing a computer program operable to cause a computer to perform the method described in the first aspect.

[0007] This application proposes a fire early warning scheme. By inputting real-time smoke sensor data and environmental parameter data collected within the target building into a specially trained deep learning model, the predicted smoke concentration for at least one future time step is determined. The deep learning model is pre-trained on target samples containing historical data from the target building or similar buildings. These target samples include not only historical smoke sensor data and corresponding environmental parameter data, but also the actual smoke concentration values ​​for at least one future time step. This endows the model with the ability to capture smoke evolution patterns from multi-dimensional time-series data, significantly improving prediction accuracy and generalization performance. After determining the predicted smoke concentration for at least one future time step using the deep learning model, it can be determined whether the fire early warning conditions are met. If so, the predicted values ​​are further spatially interpolated and modeled using the target building's interior structural diagram and smoke sensor layout information to generate a concentration distribution heat map reflecting the smoke diffusion trend. This further determines and outputs safe evacuation routes within the target building. This application's scheme achieves a fundamental shift from passive response to intelligent prediction, providing early warning before a fire occurs, effectively reducing false alarm rates, and significantly improving emergency response efficiency through visualized heat maps and dynamic evacuation routes to ensure personnel safety. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a schematic diagram of the first type of fire early warning method according to an embodiment of this application.

[0010] Figure 2 This is a schematic diagram of a second process for a fire early warning method according to an embodiment of this application.

[0011] Figure 3 This is a schematic diagram of the third process of the fire early warning method according to an embodiment of this application.

[0012] Figure 4 This is a schematic diagram of the fire early warning device according to an embodiment of this application.

[0013] Figure 5 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0014] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0015] In the field of smart fire protection, smoke sensors, as core components of fire monitoring, trigger early warnings by collecting smoke concentration data and are widely used in environmental monitoring and home security scenarios. However, existing technologies are mainly based on photoelectric smoke detection, aspirating smoke detection, or IoT smart systems. These methods rely on comparing real-time monitoring values ​​with fixed thresholds and lack the ability to model future trends in smoke concentration. Affected by factors such as environmental noise, hardware drift, and data loss, sensor data is prone to abnormal fluctuations or omissions, leading to reduced accuracy and reliability of early warnings. This makes it difficult to capture sudden changes in the early stages of a fire, resulting in delayed alarms and failing to meet the need for early risk identification.

[0016] In view of this, this application proposes a fire early warning method, device, equipment and program product, which aims to solve the problems of existing technology in early identification of fire hazards, intelligent identification of smoke type and control of false alarm rate in complex environments, and can promote the intelligent upgrade of smoke monitoring technology towards active perception, intelligent analysis and accurate early warning.

[0017] To overcome the inherent defects of the prior art, this application proposes a fire early warning method, apparatus, equipment, and program product. The technical solutions provided by the various embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0018] One embodiment of this application provides a fire early warning method. Figure 1 This is a flowchart illustrating the fire early warning method, including: S101, input the real-time collected smoke sensor data and environmental parameter data in the target building into the deep learning model to obtain the predicted smoke concentration value of the target building at least one time step in the future; wherein, the smoke concentration prediction capability of the deep learning model is based on the target sample training, the target sample includes the historical smoke sensor data of the target building or similar buildings and the environmental parameter data corresponding to the historical smoke sensor data, and is labeled with the actual smoke concentration value corresponding to at least one time step in the future.

[0019] This embodiment leverages the powerful temporal modeling capabilities of deep learning models to predict future smoke concentration within a target building. Specifically, real-time smoke sensor data and environmental parameter data are input into a pre-trained deep learning model to directly obtain the predicted smoke concentration for the target building at least one time step in the future. The choice to predict "at least one time step" rather than "only one time step" reflects a deep understanding of the actual needs of fire early warning. The occurrence and development of a fire is a dynamic process; a prediction at a single time step only provides transient information, while predictions at multiple consecutive time steps can outline the trend of smoke concentration changes. This trend information is crucial for determining whether the fire situation is continuously worsening and identifying potential sudden risks, enabling the early warning system to provide earlier and more reliable warnings based on the assessment of future conditions, thus gaining valuable time windows for personnel evacuation and fire control.

[0020] To achieve high-precision multi-time-step predictions, deep learning models can be pre-trained in a supervised manner. Specifically, the smoke concentration prediction capability of the deep learning model is acquired through training on a set of "target samples." These target samples contain historical data collected from the target building itself or similar buildings with similar structures and risk characteristics. Each target sample consists of historical smoke sensor data and corresponding environmental parameter data, and is critically labeled with the actual smoke concentration value measured at least one future time step. By mapping the historical data to future true values, the deep learning model learns to predict smoke concentration for at least one future time step.

[0021] During training, the accuracy of deep learning model predictions is guided and optimized by a composite loss function. This loss function is quantified by the prediction error at each time step, but its unique feature is the introduction of two key perceptual penalty terms, enabling it to deeply understand the specific characteristics of fire early warning scenarios.

[0022] The first is the "threshold-aware penalty term." When the actual smoke concentration at a certain time step is higher than a preset safety threshold (first threshold), and the smoke concentration prediction given by the deep learning model underestimates this actual smoke concentration, this term significantly increases the weight of the prediction error at that time step in the training loss. Its significance lies in forcing the model to give the highest priority to concentration values ​​that are close to or exceed dangerous levels, especially avoiding underestimating the risk, thereby fundamentally ensuring the reliability of the early warning.

[0023] The second is the "change-point perception penalty term." When a drastic change is detected in the actual smoke concentration value at a certain time step compared to the nearest neighboring time step (such as the previous time step) (i.e., the absolute value of the difference is greater than the second threshold), this term will increase the weight of the prediction error at that time step in the training loss. Its significance lies in guiding the model to pay close attention to "critical moments" when smoke concentration may suddenly surge, enhancing the deep learning model's ability to capture and respond to sudden situations in the early stages of a fire, and effectively improving the timeliness of early warnings.

[0024] As an example, the loss function mainly consists of the following parts: 1) Early warning loss due to change point and threshold perception This is the core principle guiding deep learning models to focus on key risk points:

[0025] in, This represents the total number of target samples in a training batch; This represents the total number of time steps that need to be predicted. Indicates the first The target sample at the th The actual smoke concentration value at a future time step; Indicates the first The target sample at the th Predicted smoke concentration values ​​for one future time step; Dynamic weighting coefficients are used to apply the "threshold-aware penalty" and the "variable-point-aware penalty". Its formula is:

[0026] in, This indicates an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. This represents the preset fire warning concentration threshold, i.e., the first threshold mentioned above; The neighborhood range of the warning threshold is defined; Represents a hyperparameter that controls the strength of the threshold-aware penalty term. Approaching or exceeding ,and At that time, this item will increase This significantly increases the weight of the prediction error at that time step in the total training loss; It is a binary identifier, when the... The target sample at the th When a time step is detected as a concentration mutation point (e.g., ...), The second threshold mentioned above is 1 if it is set to 1, and 0 otherwise. This represents a hyperparameter that controls the strength of the change-point sensing penalty term; it increases when a sudden concentration change is detected. This guides deep learning models to pay close attention to "critical moments" when smoke concentration may suddenly spike.

[0027] 2) Mean absolute percentage error loss This is used to balance the prediction accuracy at different concentration levels.

[0028] in, It is a very small positive number (e.g., 0.0001) used to prevent when When the denominator is zero, the fraction is zero.

[0029] 3) Trend consistency loss This is used to ensure that the predicted value and the actual value remain consistent in the direction of change:

[0030] in, This indicates the actual smoke concentration value from the time step. arrive The change in; This indicates that the predicted smoke concentration value is from the time step. arrive The change in quantity.

[0031] Furthermore, at the model architecture level, this embodiment innovatively employs a hybrid neural network as its core encoding structure to comprehensively capture the complex temporal patterns in the smoke concentration sequence. This hybrid neural network consists of three cooperating modules: 1) The short-term linear prediction module is configured to: fit the short-term trend of the input data based on the autoregressive algorithm to obtain short-term trend features; and concatenate the short-term trend features with the original features in the input data to obtain enhanced time-series features. The significance of this is that it provides a clear and easy-to-understand short-term change prior for the subsequent complex nonlinear network, which helps the model to capture recent concentration change trends more stably and quickly.

[0032] As an example, the formula for calculating short-term trend fitting is:

[0033] in, Indicates at time step The obtained short-term trend characteristics; Indicates the order of the autoregressive algorithm; This indicates that the autoregressive algorithm is at the 1st... The coefficients corresponding to the order; Indicates historical time step The actual smoke concentration value on the screen; Indicates time step The noise term; it should be noted that the text will Used as a serial number, it does not have any special parameter meaning.

[0034] 2) The dual-channel temporal neural network module is configured as follows: one channel extracts multi-scale local features from the enhanced temporal features using dilated convolution; the other channel extracts forward-temporal state features from the enhanced temporal features using a forward long short-term memory network, and extracts reverse-temporal state features using a backward long short-term memory network. The forward-temporal and reverse-temporal state features are then fused to obtain global bidirectional context-dependent features, enabling the understanding of long-term trends and cyclical patterns. Finally, the multi-scale local features and global bidirectional context-dependent features are fused to obtain comprehensive temporal features. The significance of this approach is that it allows the deep learning model to simultaneously consider and integrate the short-term abrupt changes and long-term evolution patterns of smoke concentration, providing a comprehensive and rich feature representation for accurate prediction.

[0035] 3) The change-point awareness attention mechanism module is configured as follows: First, it determines the change-point prior weights for each time step in the comprehensive temporal features. These change-point prior weights characterize the probability of a sudden change in smoke concentration at the corresponding time step. Then, it fuses the change-point prior weights with the self-attention mechanism weights to obtain a change-point awareness attention weight matrix that can dynamically focus on key abrupt changes. Finally, it weights the comprehensive temporal features based on the change-point awareness attention weight matrix to obtain the target features for smoke concentration prediction. The significance of this approach lies in simulating the behavior of human experts who pay particular attention to outliers when analyzing data, giving the deep learning model "insight" into potential risk moments and significantly improving its sensitivity to sudden fires.

[0036] In this embodiment, for each time step in the comprehensive time series features, the feature statistics (such as mean and variance) within the forward and backward time windows centered on that time step are determined. Based on the difference between the statistics in the forward and backward time windows, the change-point prior weight for that time step is determined. As an example, the formula for calculating the change-point prior weight is:

[0037]

[0038] in, Indicates the prior weights of the variable points; Indicates time step The corresponding variable point fraction; Indicates time step The characteristic statistics within the forward and backward time windows centered on the subject; The covariance matrix represents the forward time window; Represents the covariance matrix of the backward time window; It is a hyperparameter that balances the changes in mean and variance; Correspondingly, this change-point prior weight is fused with the standard self-attention mechanism weight to generate a "change-point perception attention weight matrix," the calculation formula of which is as follows:

[0039] in, Indicates the weights of the self-attention mechanism; It is a hyperparameter that controls the degree of influence of the variable point prior on the final attention weight distribution.

[0040] In practical implementation, the target building in this embodiment can be any type of place with corresponding sensing and computing facilities deployed, such as office buildings, shopping malls, hotels, factories, or residences. Its internal facilities typically include multiple smoke sensors distributed in different areas for real-time monitoring of smoke particle concentration in the air; and environmental parameter sensors for synchronously collecting key environmental variables that may affect smoke diffusion, such as temperature, humidity, and wind speed. These sensors transmit data in real-time to a central processing unit or edge computing node via wired or wireless networks. Furthermore, before inputting the real-time collected smoke sensor data and environmental parameter data into the deep learning model, a series of data preprocessing operations must be performed to ensure data quality and consistency with the model input. These preprocessing steps include, but are not limited to, at least one of the following: 1) To interpolate missing values ​​in real-time smoke sensor data and environmental parameter data to address data incompleteness caused by sensor communication interruption or data packet loss. 2) Perform data cleaning on outliers and noise in the real-time collected smoke sensor data and environmental parameter data to remove invalid or distorted data points caused by equipment interference or momentary false alarms. 3) Standardize the real-time collected smoke sensor data and environmental parameter data to eliminate the differences in the dimensions and numerical ranges of different physical quantities (such as concentration, temperature, and humidity) and accelerate model convergence; 4) Convert the real-time collected smoke sensor data and environmental parameter data into the input format required by the deep learning model, such as converting the cleaned data into a specific tensor format and time window structure required by the deep learning model.

[0041] As an example, refer to Figure 2 As shown, the preprocessing flow in this embodiment includes: Step 1: Data Collection By deploying smoke sensors and related environmental parameter sensors, smoke sensor data (such as smoke concentration) and environmental parameter data (such as temperature, humidity, and wind speed) are synchronously collected within the target building at a set sampling frequency (e.g., once every 5 seconds) to form the raw input dataset. This dataset includes smoke concentration sequences and corresponding environmental characteristics at each time step, ensuring the real-time nature and multi-dimensionality of the data source.

[0042] Step 2: Imputing Missing Values To address missing values ​​caused by sensor packet loss, communication delays, and other factors during the data acquisition process, a hierarchical imputation strategy is employed: when the missing percentage is below a set threshold (e.g., 5%), linear interpolation is used to imput the missing values ​​using adjacent time steps; when the missing percentage exceeds the threshold, an interpolation method based on K-Nearest Neighbors (KNN) is used to find similar samples in the feature space based on the multidimensional features of the missing points for estimation, thereby generating a complete time series to ensure data continuity.

[0043] Step 3: Outlier Detection and Noise Cleaning Outlier detection and denoising are performed on the complete time series generated in step two. First, the 3σ principle is used to detect and mark obvious outliers; then, mean filtering is applied to each feature sequence to remove instantaneous spike noise; further, wavelet denoising method is used to decompose the signal into low-frequency (trend) and high-frequency (noise) components, retaining only the low-frequency and key high-frequency components, and finally obtaining the denoised time series, thus improving data quality.

[0044] Step 4: Standardization Processing To eliminate differences in numerical range and units among various physical quantities and improve the stability and convergence speed of model training, the time series after denoising in step three is standardized. The Z-score standardization method is used to convert each feature into a standard normal distribution with a mean of 0 and a standard deviation of 1. The calculation formula is as follows: ,in This represents the mean of the feature. This represents the standard deviation of the feature. Ultimately, a standardized time series is formed, ensuring a balanced distribution of gradient weights across different features in the model.

[0045] Step 5: Time windowing construction and dataset partitioning The standardized time series generated in step four is converted into the input format required by the deep learning model. Time windowing is used to construct the model, with a window length of T = 12 time steps (corresponding to 60 seconds of historical data). Samples are generated using a sliding window approach, with each sliding step being 1 time step (5 seconds). Each sample contains features from T consecutive time steps as model input. The prediction target is the smoke concentration value for K = 6 time steps (corresponding to 30 seconds) after the end of the window. The sample tensor dimension is... (in (The number of features is, for example, 4). Finally, the samples can be divided into training and test sets in a ratio of 8:2 for model training and testing evaluation.

[0046] S102, based on the predicted smoke concentration at at least one time step, determine whether the fire early warning conditions are met.

[0047] In this embodiment, the setting of fire early warning conditions directly determines the sensitivity and reliability of the early warning system. In specific implementation, there are several feasible judgment logics. For example, it can check whether the predicted value at a specific future time exceeds a high emergency threshold, suitable for rapid response to sudden and extremely intense fires; it can also calculate the average or integral of the predicted values ​​over a future period, and trigger an alarm if the statistic exceeds a threshold, a method more sensitive to persistent and slowly developing fires; or, it can introduce more complex environmental parameter coupling judgments, such as when the predicted smoke concentration increases while the predicted temperature also rises abnormally, thus enhancing the reliability of the early warning.

[0048] As a preferred embodiment of this application, this embodiment employs a judgment strategy that balances response speed and anti-interference capability: if the predicted smoke concentration values ​​for a consecutive preset number of time steps are all higher than a preset concentration threshold, then the fire early warning condition is determined to be met. A key design constraint in this strategy is that the required preset number is not greater than the total number of time steps (i.e., the total number of "at least one time step" output in S101). This means that the continuous exceeding trend upon which the early warning judgment relies must be entirely within the model's prediction scope.

[0049] This design has significant practical implications. First, by requiring "continuous" exceedances over multiple time steps, rather than just "a single moment," it effectively filters false alarms caused by transient sensor interference or random fluctuations in model predictions, significantly improving the reliability of early warnings. Second, it essentially identifies the early development trend of a fire. Exceeding the limit in a single time step may be accidental, but continuous exceedances over multiple time steps strongly suggest a sustained, rather than transient, combustion process, thus capturing the key characteristic of early, continuous deterioration before the fire truly becomes uncontrollable. Finally, by flexibly configuring preset quantities and concentration thresholds, system administrators can customize different early warning sensitivities based on the risk levels of different areas (such as warehouses, offices, and corridors), achieving customizable and refined early warning strategies. For example, in a high-risk area, an alarm can be triggered if the predicted value exceeds the threshold for two consecutive time steps (i.e., the next 10 seconds); while in an area with lower tolerance for false alarms, it may require four consecutive time steps (i.e., the next 20 seconds) of predicted value exceeding the limit to trigger an alarm, thus achieving an optimal balance between safety and efficiency.

[0050] S103, if the fire early warning conditions are met, then based on the indoor structural diagram of the target building and the smoke sensor layout information, combined with the predicted smoke concentration at at least one time step, spatial interpolation modeling is performed to generate a smoke concentration distribution heat map of the target building.

[0051] In this embodiment, the smoke concentration distribution heat map is a two-dimensional or three-dimensional graphic that uses color depth to intuitively represent the smoke concentration in different areas. Its purpose is to overcome the limitations of the physical location of sensors and to estimate the concentration in areas not directly monitored. This provides an indispensable spatial risk map for a comprehensive understanding of the fire situation, especially for subsequent intelligent evacuation route planning, upgrading emergency decision-making from "point-based perception" to "global insight".

[0052] The key inputs for generating the smoke concentration distribution heatmap include the interior structural diagram of the target building and the smoke sensor layout information. The interior structural diagram defines the physical space of the building, including room layouts, corridors, walls, and exits; these structural elements determine the possible diffusion paths and obstructions of smoke. The sensor layout information precisely provides the spatial coordinates of each smoke sensor within the building's plane. Based on this information, a spatial interpolation modeling algorithm can be used to extrapolate the predicted smoke concentration value of each sensor at a specific future time step to every location within the building's plane.

[0053] Specifically, this embodiment can use the Inverse Distance Weighting (IDW) method for spatial interpolation, and its core calculation formula is as follows:

[0054] in, This represents the location point inside the target building that needs to be solved. Estimated smoke concentration; Indicates the total number of smoke sensors; Indicates the first The coordinates of the smoke sensor; Indicates the first The smoke sensor in the future The predicted smoke concentration at each time step (obtained from step S101) is used as a known data point in the interpolation calculation. This is the weight decay factor (usually set to 2), used to control the degree to which distance affects the weights. The larger the value, the stronger the influence of the closer sensor, and the more the interpolation result highlights local details.

[0055] The significance of the above formula is that the smoke concentration value at an unknown point is a weighted average of all known smoke sensor predictions. Its weight is inversely proportional to the distance from the point to be measured to the smoke sensor; that is, the closer a location is to the smoke sensor, the greater the influence of the sensor's reading on its smoke concentration value. Using this method, discrete smoke sensor data can be "spread" across the entire building plane, generating a continuous and smooth heat map of smoke concentration distribution that reflects the spatial distribution of predicted concentration at a future moment.

[0056] S104, based on the smoke concentration distribution heat map, determines and outputs the evacuation route within the target building.

[0057] This embodiment utilizes a generated smoke concentration distribution heatmap to plan and output a dynamic evacuation route that maximizes personnel safety. The evacuation route is planned starting from the personnel's location and ending at a safety exit within the target building. The personnel's location can be determined in various ways: for fixed workstations or rooms, a preset fixed location can be used; for public areas or moving personnel, real-time data can be obtained through deployed cameras combined with visual analysis technology. The goal of the planning algorithm is to search for an optimal evacuation route from the starting point to the ending point. The key is to define an objective function that aims to minimize the overall risk exposure of the area traversed by the evacuation route, which is the cumulative predicted smoke concentration value of the area traversed by the evacuation route. Accordingly, this embodiment searches for an evacuation route by minimizing the objective function.

[0058] As an example, the objective function is expressed as follows:

[0059] in, Indicates starting from the origin To the finish line Among all possible evacuation paths, find the optimal path that minimizes the objective function. ; Indicates the evacuation route The cost of each spatial unit (or node) traversed is summed. This represents the base distance cost for each step in the evacuation path, usually set to 1, to characterize the physical length of the path; Indicates the weight of smoke concentration; Representing distance weights; this embodiment uses... Set to higher (For example The value is 0.7. (Taking a value of 0.3), the objective function explicitly prioritizes avoiding high-concentration areas, guiding the algorithm to generate a "low smoke concentration evacuation path" that is longer but safer, rather than simply the shortest path.

[0060] It should be understood that in practical applications, the planned optimal evacuation routes will be output to multiple terminals and platforms in real time through system interfaces to ensure timely information delivery. First, the routes will be sent to the building's public address system and intelligent evacuation signs, dynamically updating escape direction instructions. Second, route information can be pushed to the command and dispatch screens of security personnel, providing decision support for their emergency command. Most importantly, routes can be directly distributed to the mobile terminals of people within the building via mobile apps, SMS, or internal communication systems, achieving personalized and precise guidance. To ensure the routes remain effective, the system will periodically (e.g., every 5 seconds) recalculate the routes based on the latest predicted concentration distribution. This dynamically updated mechanism can cope with the rapidly changing fire and smoke spread, continuously providing personnel with the safest and most reliable evacuation guidance at any given moment, maximizing the protection of lives.

[0061] Among them, reference Figure 3 As shown, the method in this embodiment realizes a complete closed-loop process from data acquisition to intelligent evacuation: First, the system inputs real-time smoke sensor data (such as smoke concentration) and environmental parameter data (such as temperature and humidity) collected in the target building into a pre-trained deep learning model. This model learns the smoke evolution pattern based on historical data and outputs predicted smoke concentration values ​​for multiple future time steps. Subsequently, the system determines whether the fire early warning conditions are met based on these predicted values ​​(e.g., the predicted values ​​for consecutive preset time steps all exceed a preset concentration threshold). If the conditions are met, the system combines the indoor structural diagram of the target building and the smoke sensor layout information to generate a concentration distribution heat map reflecting the smoke diffusion trend through spatial interpolation modeling (such as the inverse distance weighting method). Finally, the system divides the area into safe and dangerous zones based on the heat map, and uses the personnel location as the starting point and the safety exit as the ending point to dynamically plan and output the optimal evacuation path by minimizing the smoke concentration accumulation function on the path, thereby realizing integrated prevention and control of early warning, accurate perception and intelligent evacuation.

[0062] In summary, the method of this embodiment inputs real-time smoke sensor data and environmental parameter data collected within the target building into a specially trained deep learning model to determine the predicted smoke concentration value for at least one future time step. The deep learning model is pre-trained on target samples containing historical operating conditions of the target building or similar buildings. These target samples include not only historical smoke sensor data and corresponding environmental parameter data, but also the actual smoke concentration values ​​for at least one future time step, thereby endowing the model with the ability to capture smoke evolution patterns from multi-dimensional time-series data, significantly improving the accuracy and generalization performance of the prediction. After determining the predicted smoke concentration value for at least one future time step through the deep learning model, it can be determined whether the fire warning conditions are met. If so, the predicted value is further spatially interpolated and modeled using the target building's interior structural diagram and smoke sensor layout information to generate a concentration distribution heat map reflecting the smoke diffusion trend, thereby further determining and outputting safe evacuation routes within the target building. The solution of this embodiment achieves a fundamental shift from passive response to intelligent prediction, providing early warning before a fire occurs, effectively reducing the false alarm rate, and significantly improving emergency response efficiency through visualized heat maps and dynamic evacuation routes to ensure personnel safety.

[0063] In addition, corresponding to Figure 1 In addition to the method shown in this embodiment, another embodiment of this invention also provides a fire early warning method. Figure 5 This is a structural diagram of the fire early warning method 400, including: The prediction module 410 is used to input real-time collected smoke sensor data and environmental parameter data in the target building into a deep learning model to obtain a predicted smoke concentration value for the target building at least one time step in the future; wherein, the smoke concentration prediction capability of the deep learning model is based on target samples trained on target samples, the target samples include historical smoke sensor data of the target building or similar buildings and environmental parameter data corresponding to the historical smoke sensor data, and are labeled with the actual smoke concentration value corresponding to the at least one time step in the future.

[0064] The judgment module 420 is used to determine whether the fire early warning conditions are met based on the predicted smoke concentration value of the at least one time step.

[0065] The mapping module 430 is used to generate a heat map of the smoke concentration distribution of the target building by performing spatial interpolation modeling based on the indoor structural diagram of the target building and the smoke sensor layout information, combined with the predicted smoke concentration values ​​at at least one time step, if the fire early warning conditions are met.

[0066] The planning module 440 is used to determine and output the evacuation route within the target building based on the smoke concentration distribution heat map.

[0067] Optionally, the deep learning model employs a hybrid neural network as its encoding structure, the hybrid neural network comprising: The short-term linear prediction module is configured to: fit the short-term trend of the input data based on the autoregressive algorithm to obtain short-term trend features; and concatenate the short-term trend features with the original features in the input data to obtain enhanced time-series features. The dual-channel temporal neural network module is configured to: extract multi-scale local features from the enhanced temporal features based on dilated convolution; extract forward temporal state features of the enhanced temporal features based on a forward long short-term memory network, and extract reverse temporal state features of the enhanced temporal features based on a backward long short-term memory network; fuse the forward temporal state features and the reverse temporal state features to obtain global bidirectional context-dependent features; and fuse the multi-scale local features with the global bidirectional context-dependent features to obtain comprehensive temporal features. The change point perception attention mechanism module is configured to: determine the change point prior weights for each time step in the comprehensive temporal features, wherein the change point prior weights characterize the probability of a sudden change in smoke concentration at the corresponding time step; fuse the change point prior weights with the self-attention mechanism weights to obtain a change point perception attention weight matrix; and weight the comprehensive temporal features based on the change point perception attention weight matrix to obtain target features for smoke concentration prediction.

[0068] Optionally, the change point awareness attention mechanism module determines the change point prior weights for each time step in the comprehensive temporal features, including: for each time step in the comprehensive temporal features, determining the feature statistics within the forward time window and the backward time window centered on it, and determining the change point prior weights for that time step based on the difference between the statistics of the forward time window and the statistics of the backward time window.

[0069] Optionally, the training loss of the deep learning model is quantified by the prediction error at each time step, and the loss function used to calculate the training loss is configured as follows: A threshold-aware penalty term corresponding to the prediction error at each time step is used to increase the weight of the prediction error at that time step in the training loss when the actual smoke concentration value at that time step is higher than the first threshold and at the same time higher than the corresponding predicted smoke concentration value. A change point perception penalty term corresponding to the prediction error at each time step is used to increase the weight of the prediction error at that time step in the training loss when the absolute value of the difference between the true smoke concentration value at that time step and the true smoke concentration value at the nearest neighboring time step is greater than a second threshold. Optionally, the planning module 440 determines and outputs evacuation routes within the target building based on the smoke concentration distribution heat map, including: dividing the area represented by the smoke concentration distribution heat map into a safe zone and a danger zone according to a preset smoke concentration safety standard; searching for evacuation routes within the safe zone with the location of personnel within the target building as the starting point of the route, the safety exit within the target building as the ending point of the route, and minimizing the objective function as the objective; wherein the objective function is a cumulative function of the predicted smoke concentration values ​​of the areas traversed by the evacuation route.

[0070] Optionally, before inputting the real-time collected smoke sensor data and environmental parameter data from the target building into the deep learning model, the prediction module 410 is further configured to: preprocess the real-time collected smoke sensor data and environmental parameter data; wherein the preprocessing includes at least one of the following: Missing values ​​in real-time collected smoke sensor data and environmental parameter data are interpolated; Data cleaning is performed on outliers and noise in real-time smoke sensor data and environmental parameter data. Standardize the real-time collected smoke sensor data and environmental parameter data; The real-time collected smoke sensor data and environmental parameter data are converted into the input format required by the deep learning model.

[0071] Optionally, the judgment module 420 determines whether the fire warning conditions are met based on the predicted smoke concentration values ​​of the at least one time step, including: if the predicted smoke concentration values ​​of a consecutive preset number of time steps are all higher than a preset concentration threshold, then the fire warning conditions are met; wherein, the preset number is not greater than the total number of time steps of the at least one time step.

[0072] It should be noted that the fire early warning device in this embodiment can be used as... Figure 1 The execution body of the method shown is therefore able to achieve... Figure 1 The steps and functions of the method shown are illustrated.

[0073] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Please refer to it. Figure 5 At the hardware level, the electronic device includes a processor, and optionally also includes an internal bus, a network interface, and memory. The memory may include main memory, such as high-speed random-access memory (RAM), or non-volatile memory, such as at least one disk drive. Of course, the electronic device may also include other hardware required for other business operations.

[0074] The processor, network interface, and memory can be interconnected via an internal bus, which can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. This bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 5 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0075] Memory is used to store computer programs. Specifically, a computer program may include program code, which includes computer operation instructions. Memory may include main memory and non-volatile memory, and provides the computer program to the processor.

[0076] Specifically, the processor reads the corresponding computer program from non-volatile memory into memory and then runs it, forming the above-mentioned logical structure. Figure 4 The fire warning device is shown. Correspondingly, the processor executes the program stored in the memory and specifically performs the following operations: The smoke sensor data and environmental parameter data collected in real time inside the target building are input into a deep learning model to obtain the predicted smoke concentration value of the target building at least one time step in the future. The smoke concentration prediction capability of the deep learning model is based on target samples, which include historical smoke sensor data of the target building or similar buildings and environmental parameter data corresponding to the historical smoke sensor data, and are labeled with the actual smoke concentration value corresponding to the at least one time step in the future.

[0077] Based on the predicted smoke concentration at at least one time step, determine whether the fire early warning conditions are met.

[0078] If the fire warning conditions are met, then based on the indoor structural diagram of the target building and the smoke sensor layout information, combined with the predicted smoke concentration values ​​at at least one time step, spatial interpolation modeling is performed to generate a smoke concentration distribution heat map of the target building.

[0079] Based on the smoke concentration distribution heat map, the evacuation route within the target building is determined and output.

[0080] The above is as described in this instruction manual. Figure 1The fire early warning method disclosed in the illustrated embodiment can be applied to a processor and implemented by the processor. The processor may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the integrated logic circuit in the processor or by instructions in the form of software. The processor mentioned above can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the embodiments of this application can be directly embodied as being executed by a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the field, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method.

[0081] Of course, in addition to software implementation, the electronic device described in this specification does not exclude other implementation methods, such as logic devices or a combination of hardware and software. In other words, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0082] Furthermore, embodiments of this application also propose a computer program product, including a computer-readable storage medium storing one or more computer programs, the one or more computer programs including instructions.

[0083] When the aforementioned instructions are executed by a portable electronic device that includes multiple applications, they enable the portable electronic device to perform... Figure 1 The steps in the method shown include: The smoke sensor data and environmental parameter data collected in real time inside the target building are input into a deep learning model to obtain the predicted smoke concentration value of the target building at least one time step in the future. The smoke concentration prediction capability of the deep learning model is based on target samples, which include historical smoke sensor data of the target building or similar buildings and environmental parameter data corresponding to the historical smoke sensor data, and are labeled with the actual smoke concentration value corresponding to the at least one time step in the future.

[0084] Based on the predicted smoke concentration at at least one time step, determine whether the fire early warning conditions are met.

[0085] If the fire warning conditions are met, then based on the indoor structural diagram of the target building and the smoke sensor layout information, combined with the predicted smoke concentration values ​​at at least one time step, spatial interpolation modeling is performed to generate a smoke concentration distribution heat map of the target building.

[0086] Based on the smoke concentration distribution heat map, the evacuation route within the target building is determined and output.

[0087] Those skilled in the art will understand that the embodiments of this specification can be provided as methods, systems, or computer program products. Therefore, this specification may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this specification may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0088] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0089] The above are merely embodiments of this specification and are not intended to limit the scope of this specification. Various modifications and variations can be made to this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this specification should be included within the scope of the claims of this specification. Furthermore, all other embodiments obtained by those skilled in the art without inventive effort should fall within the protection scope of this document.

Claims

1. A fire early warning method, characterized in that, include: The smoke sensor data and environmental parameter data collected in real time inside the target building are input into a deep learning model to obtain the predicted smoke concentration value of the target building at least one time step in the future; wherein, the smoke concentration prediction capability of the deep learning model is based on target samples trained on target samples, the target samples include historical smoke sensor data of the target building or similar buildings and environmental parameter data corresponding to the historical smoke sensor data, and are labeled with the actual smoke concentration value corresponding to the at least one time step in the future. Based on the predicted smoke concentration at at least one time step, determine whether the fire early warning conditions are met. If the fire warning conditions are met, then based on the indoor structural diagram of the target building and the smoke sensor layout information, combined with the predicted smoke concentration values ​​at at least one time step, spatial interpolation modeling is performed to generate a smoke concentration distribution heat map of the target building. Based on the smoke concentration distribution heat map, the evacuation route within the target building is determined and output.

2. The method according to claim 1, characterized in that, The deep learning model uses a hybrid neural network as its encoding structure, and the hybrid neural network includes: The short-term linear prediction module is configured to: fit the short-term trend of the input data based on the autoregressive algorithm to obtain short-term trend features; and concatenate the short-term trend features with the original features in the input data to obtain enhanced time-series features. The dual-channel temporal neural network module is configured to: extract multi-scale local features from the enhanced temporal features based on dilated convolution; extract forward temporal state features of the enhanced temporal features based on a forward long short-term memory network, and extract reverse temporal state features of the enhanced temporal features based on a backward long short-term memory network; fuse the forward temporal state features and the reverse temporal state features to obtain global bidirectional context-dependent features; and fuse the multi-scale local features with the global bidirectional context-dependent features to obtain comprehensive temporal features. The change point perception attention mechanism module is configured to: determine the change point prior weights for each time step in the comprehensive temporal features, wherein the change point prior weights characterize the probability of a sudden change in smoke concentration at the corresponding time step; fuse the change point prior weights with the self-attention mechanism weights to obtain a change point perception attention weight matrix; and weight the comprehensive temporal features based on the change point perception attention weight matrix to obtain target features for smoke concentration prediction.

3. The method according to claim 2, characterized in that, The determination of the prior weights of the change points at each time step in the comprehensive time series features includes: For each time step in the comprehensive time series features, the feature statistics within the forward time window and the backward time window centered on it are determined, and the change point prior weight of that time step is determined based on the difference between the statistics of the forward time window and the statistics of the backward time window.

4. The method according to claim 1, characterized in that, The training loss of the deep learning model is quantified by the prediction error at each time step, and the loss function used to calculate the training loss is configured as follows: A threshold-aware penalty term corresponding to the prediction error at each time step is used to increase the weight of the prediction error at that time step in the training loss when the actual smoke concentration value at that time step is higher than the first threshold and at the same time higher than the corresponding predicted smoke concentration value. A change-point perception penalty term corresponding to the prediction error at each time step is used to increase the weight of the prediction error at that time step in the training loss when the absolute value of the difference between the true smoke concentration value at that time step and the true smoke concentration value at the nearest neighboring time step is greater than a second threshold.

5. The method according to any one of claims 1 to 4, characterized in that, Based on the smoke concentration distribution heatmap, the evacuation routes within the target building are determined and output, including: According to the preset smoke concentration safety standard, the area represented by the smoke concentration distribution heat map is divided into a safe zone and a danger zone; Using the location of personnel within the target building as the starting point of the path, the safety exit within the target building as the ending point of the path, and minimizing the objective function as the objective function, an evacuation path is searched within the safe zone; wherein, the objective function is a cumulative function of the predicted smoke concentration values ​​of the areas traversed by the evacuation path.

6. The method according to any one of claims 1 to 4, characterized in that, Before inputting the real-time smoke sensor data and environmental parameter data collected inside the target building into the deep learning model, the method further includes: The real-time collected smoke sensor data and environmental parameter data are preprocessed; wherein the preprocessing includes at least one of the following: Missing values ​​in real-time collected smoke sensor data and environmental parameter data are interpolated; Data cleaning is performed on outliers and noise in real-time smoke sensor data and environmental parameter data. Standardize the real-time collected smoke sensor data and environmental parameter data; The real-time collected smoke sensor data and environmental parameter data are converted into the input format required by the deep learning model.

7. The method according to any one of claims 1 to 4, characterized in that, Based on the predicted smoke concentration values ​​at at least one time step, determine whether the fire early warning conditions are met, including: If the predicted smoke concentration values ​​for a consecutive preset number of time steps are all higher than a preset concentration threshold, then the fire warning conditions are met; wherein, the preset number is not greater than the total number of time steps of the at least one time step.

8. A fire early warning device, characterized in that, include: The prediction module is used to input real-time collected smoke sensor data and environmental parameter data in the target building into a deep learning model to obtain the predicted smoke concentration value of the target building at least one time step in the future; wherein, the smoke concentration prediction capability of the deep learning model is based on target samples trained on target samples, the target samples include historical smoke sensor data of the target building or similar buildings and environmental parameter data corresponding to the historical smoke sensor data, and are labeled with the actual smoke concentration value corresponding to the at least one time step in the future. The judgment module is used to determine whether the fire early warning conditions are met based on the predicted smoke concentration value of the at least one time step. The mapping module is used to generate a heat map of the smoke concentration distribution of the target building by performing spatial interpolation modeling based on the indoor structure diagram of the target building and the smoke sensor layout information, combined with the predicted smoke concentration values ​​at at least one time step, if the fire early warning conditions are met. The planning module is used to determine and output the evacuation route within the target building based on the smoke concentration distribution heat map.

9. An electronic device, comprising: processor; And a memory arranged to store computer-executable instructions, characterized in that, when executed, the executable instructions cause the processor to perform the method as described in any one of claims 1 to 7.

10. A computer program product, the computer program product comprising a computer-readable storage medium storing a computer program, characterized in that, The computer program is operable to cause the computer to perform the method as described in any one of claims 1 to 7.

Images