Fire risk dynamic assessment and early warning method based on deep reinforcement learning

Abstract

The invention relates to the technical field of fire safety monitoring and early warning, in particular to a fire risk dynamic assessment and early warning method based on deep reinforcement learning, comprising the following steps: acquiring multi-source sensing data of a target area; a dynamically updated fire risk assessment state space is constructed based on multi-source sensing data, and a real-time state in the state space is processed and a risk assessment decision is generated through a deep reinforcement learning agent optimized based on an accumulated award function weighted by an accuracy award, a false alarm punishment and a missing alarm punishment; the deep reinforcement learning agent adopts a deep deterministic strategy gradient network structure; mapping the continuous output of the risk assessment decision into a discrete dynamic fire risk level; and triggering an early warning signal matched with the dynamic fire risk level according to the dynamic fire risk level. Through multi-source data standardization, dynamic state space, reinforcement learning optimization and grading early warning binding, accurate dynamic evaluation, reliable early warning and efficient emergency response are realized.

Description

Technical Field

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

[0002] Fire risk assessment, as a core support component of the fire emergency response system, has undergone three key technological iterations. Early stages were primarily manual, relying on on-site inspections by firefighters, paper-based hazard records, and qualitative assessments using fixed fire safety checklists. This approach was not only inefficient and limited in coverage but also susceptible to the experience and subjectivity of assessors, making it difficult to capture hidden or dynamically changing risks. With the widespread adoption of IoT technology, the system entered a multi-source sensing and fusion stage. Devices such as temperature, smoke, CO concentration, and flame detectors achieve real-time data collection and preliminary integration through edge computing gateways. Some solutions introduce simple threshold judgments to trigger early warnings, but this still remains at the level of passively responding to anomalies, lacking proactive analysis of risk evolution trends. In recent years, the deep integration of artificial intelligence technology has driven the intelligent transformation of the assessment system. Deep learning models such as convolutional neural networks and recurrent neural networks are used for accurate extraction of fire characteristics. Some solutions combine time series analysis to predict fire development or integrate building information models and digital twin technology to simulate fire scenarios, significantly improving assessment accuracy and scenario coverage. However, shortcomings still exist in the details of technology implementation.

[0003] Although existing technologies have made a leap from manual qualitative analysis to intelligent quantitative analysis, they still have significant limitations in complex real-world scenarios. Most intelligent assessment models adopt a static architecture with fixed state-space parameters, making it impossible to adapt in real time to continuous dynamic variables such as temperature gradient changes and nonlinear growth in smoke spread during fire evolution. This makes it difficult to accurately reflect the real-time development of the fire. Secondly, the assessment objectives lack scenario-specific design. General optimization objectives, such as classification accuracy, do not fully consider the differentiated hazards of missed and false alarms in fire scenarios. Missed alarms may lead to the neglect of key risks and cause the fire to get out of control, while false alarms will consume a lot of emergency resources and disrupt normal production and life. Furthermore, there are defects in the connection between data processing and result output: the standardization method for multi-source data is simplistic and does not dynamically adjust for the differences in the range of different sensors and environmental interference such as dust and water vapor. The state vector update frequency does not match the actual risk change rhythm. The rules for converting continuous assessment results to discrete warning levels are vague and do not refine the thresholds based on scenario factors such as building type and personnel density, resulting in less intuitive warning signal transmission and affecting the rapid decision-making of managers. In addition, some models lack robustness. When faced with extreme weather such as high temperature, strong wind or temporary equipment failure, the assessment results are prone to deviation, reducing the reliability of emergency response. Summary of the Invention

[0004] The purpose of this invention is to overcome the above-mentioned problems and provide a dynamic fire risk assessment and early warning method based on deep reinforcement learning. To achieve the above objective, this invention adopts the following technical solution:

[0005] A method for dynamic assessment and early warning of fire risk based on deep reinforcement learning includes the following steps:

[0006] Step S1: Acquire multi-source sensing data for the target area;

[0007] Step S2: Construct a dynamically updated fire risk assessment state space based on multi-source sensing data. The state space is dynamically adjusted through a weighted rolling update mechanism and contains at least five sensing data features.

[0008] Step S3: A deep reinforcement learning agent optimized by a cumulative reward function weighted by accuracy reward, false positive penalty, and false negative penalty processes the real-time state in the state space and generates risk assessment decisions; the deep reinforcement learning agent adopts a deep deterministic policy gradient network structure.

[0009] Step S4: Map the continuous output of the risk assessment decision to discrete dynamic fire risk levels;

[0010] Step S5: Trigger the corresponding early warning signal based on the dynamic fire risk level.

[0011] Furthermore, the multi-source sensing data includes temperature data, smoke concentration data, flame characteristic data, building structure parameter data, and personnel density data for the target area. Temperature data is collected by an array of temperature sensors evenly distributed on the ceiling or walls of the target area. Smoke concentration data is collected by smoke sensors installed on the top of the target area. Flame characteristic data is extracted from images captured by surveillance cameras in the target area using a flame detection algorithm. Building structure parameter data includes the building floor height, fire compartment area, and evacuation route width data for the target area. Personnel density data is calculated from infrared sensors or video images in the target area using a human detection algorithm.

[0012] Furthermore, in step S2, each type of perceived data is processed using the min-max normalization method, with the following formula:

[0013]

[0014] Where, x i,t X is the standardized value of the i-th type of sensing data at time t; i,t X represents the original value of the i-th type of sensory data at time t; i,min X is the minimum value of the i-th type of sensory data; i,maxt represents the maximum value of the i-th type of sensing data; i is the index of the sensing data type, i∈{1,2,…,n}, t is the current time; n is the number of types of sensing data and n≥5;

[0015] The standardized values ​​of all sensed data are combined into a state vector, and the expression for the state vector is:

[0016] S t =[x 1,t ,x 2,t ,...,x n,t ];

[0017] Among them, S t Let x be the fire risk assessment state vector at time t; i,t Let be the standardized value of the i-th type of sensing data at time t; i is the index of the sensing data type, i∈{1,2,…,n}, t is the current time; n is the number of types of sensing data and n≥5.

[0018] 4. The fire risk dynamic assessment and early warning method based on deep reinforcement learning according to claim 3, characterized in that, in step S3, during the training process of the deep reinforcement learning agent, a reward value is calculated after each decision step, and the expression of the reward function is:

[0019] R t =α·r 1,t +β·r 2,t +γ·r 3,t ;

[0020] Among them, R t Let be the reward value at time t; α be the weighting coefficient of the accuracy reward term; r 1,t β is the reward item corresponding to the accuracy of risk assessment at time t; β is the weight coefficient of the false alarm penalty item; r 2,t Let be the penalty term corresponding to the false alarm rate at time t; γ be the weight coefficient of the false alarm penalty term; t be the current time; the weight coefficients satisfy α>β>γ and α+β+γ=1;

[0021] The expression for the reward item corresponding to the accuracy reward item is:

[0022]

[0023] Where, r 1,t C represents the reward item corresponding to the accuracy of risk assessment at time t. t T represents the number of times the assessed fire severity level matches the actual fire severity level within time t; t The total number of evaluations within time t; t is the current time.

[0024] The expression for the penalty term corresponding to the false positive rate is:

[0025]

[0026] Where, r 2,t F is the penalty term corresponding to the false alarm rate at time t; t T represents the number of times the assessed fire severity level is higher than the actual fire severity level within time t; t The total number of evaluations within time t; t is the current time.

[0027] The expression for the penalty term corresponding to the false negative rate is:

[0028]

[0029] Where, r 3,t M is the penalty term corresponding to the false negative rate at time t; t T represents the number of times the assessed fire severity level is lower than the actual fire severity level within time t; t Let t be the total number of evaluations within time t; t is the current time.

[0030] Furthermore, in step S3, the formula for updating the model parameters of the deep reinforcement learning agent is:

[0031]

[0032] Where, θ t+1 Here are the policy network parameters at time t+1; θ t Let be the policy network parameters at time t; η be the learning rate. For the objective function J with respect to θ t The gradient; t is the current time;

[0033] The expression for the objective function is:

[0034]

[0035] Where J(θ) t ) represents the expected cumulative reward from time t to time t+N; E is the mathematical expectation operator; k is the time step index; N is the time step size; γ is the discount factor; R t+k Let t be the reward value at time t+k; t is the current time.

[0036] Furthermore, in step S4, the continuous output values ​​of the deep reinforcement learning agent's policy output layer are compared one by one with the intervals corresponding to the discrete risk levels obtained from statistical analysis of its training data; each interval uniquely corresponds to a dynamic risk level; and the intervals do not overlap.

[0037] Furthermore, in step S5, when the dynamic risk level is low, a solid green light is triggered; when the dynamic risk level is medium, a flashing yellow light is triggered; when the dynamic risk level is high, a flashing red light and a buzzer alarm are triggered as an audible and visual warning signal; when the dynamic risk level is extremely high, a flashing red light and a buzzer alarm are triggered as an audible and visual warning signal, along with a text warning signal containing the target area name, the current risk level, and specific guidance instructions for evacuation along the nearest evacuation route. The text warning signal is sent to the fire control room monitoring terminal and the management personnel's mobile terminal.

[0038] Furthermore, in step S2, the state space is dynamically updated over time, and the update interval is consistent with the acquisition frequency of the multi-source sensing data.

[0039] The state vector expression at time t+1 is:

[0040] S t+1 =λ·S t +(1-λ)·[x 1,t+1 ,x 2,t+1 ,...,x n,t+1 ];

[0041] Among them, S t+1 Let S be the fire risk assessment state vector at time t+1; λ is the state update coefficient; S t Let x be the fire risk assessment state vector at time t; i,t+1 Let be the standardized value of the i-th type of sensing data at time t+1; i∈{1,2,…,n}; t is the current time; n is the number of types of sensing data and n≥5.

[0042] Furthermore, in step S3, the deep deterministic policy gradient network structure includes a state input layer, a hidden layer, and a policy output layer; the state input layer receives the standardized state vector S. t The number of neurons is equal to the state vector S. t The dimension is n (n≥5); the hidden layer consists of at least two fully connected layers, each with 2n neurons; the policy output layer has 1 neuron, and the risk assessment decision output value is obtained by scaling after applying the tanh activation function. The expression for the output value is:

[0043]

[0044] Among them, a t Let be the risk assessment decision output value at time t; tanh is the hyperbolic tangent activation function; o t t represents the output value of the strategy output layer when it is not activated; t represents the current time.

[0045] The advantages of this invention are:

[0046] 1. This invention introduces a deep deterministic strategy gradient agent to process the continuous state space of fire risk assessment and adopts a weighted rolling update mechanism to dynamically adjust the state vector. This enables the accurate capture of real-time changes in parameters such as temperature, smoke concentration, and personnel density during the fire evolution process, and can flexibly respond to the dynamic changes in the risk situation, avoiding delays in emergency response caused by assessment lag.

[0047] 2. This invention designs a weighted cumulative reward function for fire scenarios, prioritizing accuracy rewards, followed by penalties for missed reports, and supplemented by penalties for false alarms. This guides the reinforcement learning agent to focus more on reducing the risk of missed reports during training, while reasonably controlling the probability of false alarms. This achieves a high degree of alignment between the assessment results and actual fire protection needs, effectively improving the reliability of risk judgment and reducing safety hazards caused by assessment bias.

[0048] 3. This invention compares the continuous output values ​​of reinforcement learning with the discrete risk level intervals obtained from training data statistics one by one, and binds them with the corresponding graded early warning signals, thus achieving a seamless connection from fine assessment to actual early warning. This not only preserves the assessment accuracy of intelligent algorithms, but also simplifies the engineering implementation process and improves the convenience of emergency operations.

[0049] 4. This invention uses min-max standardization to process multi-source sensing data and combines it with a dynamically updated state space to ensure the consistency and effectiveness of different types of data such as temperature, smoke, and building structural parameters in the evaluation. This enables the system to operate stably in complex environments, reduces evaluation bias caused by data fluctuations, and enhances the robustness and adaptability of the model.

[0050] 5. This invention directly binds dynamic risk levels with differentiated early warning signals, enabling intuitive transmission of early warning information, helping managers and on-site personnel to quickly understand the degree of risk and take corresponding measures, thereby improving the efficiency and accuracy of emergency response. Attached Figure Description

[0051] The accompanying drawings, which form part of this application, are used to provide a further understanding of the application and to make other features, objects, and advantages of the application more apparent. The illustrative embodiments and descriptions of this application are used to explain the application and do not constitute an undue limitation of the application.

[0052] In the attached diagram:

[0053] Figure 1 This is an overall flowchart of the fire risk dynamic assessment and early warning method based on deep reinforcement learning in Example 1.

[0054] Figure 2 This is the data acquisition and processing framework for the fire risk dynamic assessment and early warning method based on deep reinforcement learning in Example 1.

[0055] Figure 3 This is a flowchart of the deep reinforcement learning agent decision-making process for the fire risk dynamic assessment and early warning method based on deep reinforcement learning in Example 1.

[0056] Figure 4 This is a flowchart illustrating the risk level mapping of the fire risk dynamic assessment and early warning method based on deep reinforcement learning in Example 1.

[0057] Figure 5 This is a flowchart of the early warning signal triggering process for the fire risk dynamic assessment and early warning method based on deep reinforcement learning in Example 1.

[0058] Figure 6 This is a schematic diagram of the state space dynamic update of the fire risk dynamic assessment and early warning method based on deep reinforcement learning in Example 1. Detailed Implementation

[0059] The present invention will now be described in detail and specifically through specific embodiments to enable a better understanding of the invention. However, the following embodiments do not limit the scope of protection of the present invention.

[0060] Example 1

[0061] like Figure 1 As shown, the fire risk dynamic assessment and early warning method based on deep reinforcement learning includes the following steps:

[0062] Step S1: Acquire multi-source sensing data for the target area;

[0063] Step S2: Construct a dynamically updated fire risk assessment state space based on multi-source sensing data. The state space is dynamically adjusted through a weighted rolling update mechanism and contains at least five sensing data features.

[0064] Step S3: A deep reinforcement learning agent optimized by a cumulative reward function weighted by accuracy reward, false positive penalty, and false negative penalty processes the real-time state in the state space and generates risk assessment decisions; the deep reinforcement learning agent adopts a deep deterministic policy gradient network structure.

[0065] Step S4: Map the continuous output of the risk assessment decision to discrete dynamic fire risk levels;

[0066] Step S5: Trigger the corresponding early warning signal based on the dynamic fire risk level.

[0067] In a specific embodiment, this solution revolves around the dynamic assessment and early warning of fire risks, achieving precise prevention and control through five closely linked steps. First, multi-source sensing data of the target area is acquired, forming the foundation of the assessment. Multi-source data comprehensively captures various fire-related information, avoiding the limitations of a single data dimension. Next, a dynamically updated fire risk assessment state space is constructed. A weighted rolling update mechanism allows the state space to respond to data changes in real time, ensuring the assessment always aligns with the actual situation. At least five sensing data features provide ample information support for the assessment, enhancing its comprehensiveness. Then, a reinforcement learning agent with a deep deterministic policy gradient network structure processes the real-time state. This network structure performs excellently in continuous decision-making scenarios. Combined with a weighted cumulative reward function that includes accuracy rewards, false alarm penalties, and missed alarm penalties, it guides the agent to continuously optimize decisions and reduce erroneous judgments. Continuous outputs are mapped to discrete dynamic fire risk levels, making the risk status more intuitive and easier to understand, and facilitating the execution of early warning operations. Finally, matching early warning signals are triggered based on the risk level, ensuring that different risk levels receive corresponding responses, achieving a complete closed loop from risk perception to early warning execution, effectively improving the timeliness and accuracy of fire early warnings.

[0068] Furthermore, such as Figure 2 As shown, the multi-source sensing data includes temperature data, smoke concentration data, flame characteristic data, building structure parameter data, and personnel density data of the target area. Temperature data is collected by an array of temperature sensors evenly distributed on the ceiling or walls of the target area. Smoke concentration data is collected by smoke sensors installed on the top of the target area. Flame characteristic data is extracted from images captured by surveillance cameras in the target area using a flame detection algorithm. Building structure parameter data includes the building floor height, fire compartment area, and evacuation route width of the target area. Personnel density data is calculated from infrared sensors or video images in the target area using a human detection algorithm.

[0069] In a specific embodiment, the multi-source sensing data are defined as temperature data, smoke concentration data, flame characteristic data, building structural parameter data, and personnel density data. These five types of data approach the issue from three key dimensions: fire physical characteristics, environmental conditions, and personnel situation, comprehensively covering the core influencing factors of fire assessment. Temperature data is collected through an array of temperature sensors evenly distributed on the ceiling or walls. This even distribution avoids local data deviations, captures overall temperature changes in the target area, and promptly identifies areas of abnormal temperature rise. Smoke concentration data is collected by smoke sensors installed at the top of the target area. Combined with the physical characteristics of rising smoke, top-mounted sensors can capture initial smoke signals more quickly and accurately, increasing the likelihood of early warning. Flame characteristic data is extracted from images captured by surveillance cameras using flame detection algorithms, directly capturing the most direct visual features of a fire and making the assessment more targeted. Building structural parameter data includes building floor height, fire compartment area, and evacuation route width. These parameters directly affect the speed of fire spread and the efficiency of personnel evacuation, and are important bases for assessing risk levels and formulating early warning responses. Personnel density data is calculated using human detection algorithms through infrared sensors or video images. The level of personnel density is related to the difficulty of evacuation and the potential risk of casualties. Including it in the assessment can make the early warning response more in line with the actual rescue needs.

[0070] Furthermore, in step S2, each type of perceived data is processed using the min-max normalization method, with the following formula:

[0071]

[0072] Where, x i,t X is the standardized value of the i-th type of sensing data at time t; i,t X represents the original value of the i-th type of sensory data at time t; i,min X is the minimum value of the i-th type of sensory data; i,max t represents the maximum value of the i-th type of sensing data; i is the index of the sensing data type, i∈{1,2,…,n}, t is the current time; n is the number of types of sensing data and n≥5;

[0073] The standardized values ​​of all sensed data are combined into a state vector, and the expression for the state vector is:

[0074] S t =[x 1,t ,x 2,t ,...,x n,t ];

[0075] Among them, S t Let x be the fire risk assessment state vector at time t; i,tLet be the standardized value of the i-th type of sensing data at time t; i is the index of the sensing data type, i∈{1,2,…,n}, t is the current time; n is the number of types of sensing data and n≥5.

[0076] In a specific embodiment, step S2 processes each type of sensing data using a min-max standardization method. Different types of sensing data have different dimensions and numerical ranges; directly involving them in the evaluation can amplify or diminish the impact of some data. Standardization can map all data to the same scale, eliminating dimensional interference and ensuring equal weight for all types of data in the evaluation, thus improving the fairness and accuracy of the evaluation results. Where x... i,t As the standardized value of the i-th type of sensing data at time t, it is the core foundation for state construction, and its value can objectively reflect the relative level of this type of data within the overall range. i,t It is the raw data, retaining the true information about how the data was collected. X i,min and X i,max These are the extreme values ​​of this type of data. Standardization is achieved through operations with the original data, allowing x to... i,t The values ​​should be within a reasonable range to avoid outliers interfering with the evaluation. All standardized values ​​are combined into a state vector S. t It can integrate scattered multi-source data into a unified input format, which facilitates centralized processing by deep reinforcement learning agents. The setting of n≥5 ensures that the data dimensions of the state vector are sufficient to fully carry fire risk-related information.

[0077] Furthermore, such as Figure 3 As shown in step S3, during the training process of the deep reinforcement learning agent, the reward value is calculated after each decision step. The expression for the reward function is:

[0078] R t =α·r 1,t +β·r 2,t +γ·r 3,t ;

[0079] Among them, R t Let be the reward value at time t; α be the weighting coefficient of the accuracy reward term; r 1,t β is the reward item corresponding to the accuracy of risk assessment at time t; β is the weight coefficient of the false alarm penalty item; r 2,t Let be the penalty term corresponding to the false alarm rate at time t; γ be the weight coefficient of the false alarm penalty term; t be the current time; the weight coefficients satisfy α>β>γ and α+β+γ=1;

[0080] The expression for the reward item corresponding to the accuracy reward item is:

[0081]

[0082] Where, r1,t C represents the reward item corresponding to the accuracy of risk assessment at time t. t T represents the number of times the assessed fire severity level matches the actual fire severity level within time t; t The total number of evaluations within time t; t is the current time.

[0083] The expression for the penalty term corresponding to the false positive rate is:

[0084]

[0085] Where, r 2,t F is the penalty term corresponding to the false alarm rate at time t; t T represents the number of times the assessed fire severity level is higher than the actual fire severity level within time t; t The total number of evaluations within time t; t is the current time.

[0086] The expression for the penalty term corresponding to the false negative rate is:

[0087]

[0088] Where, r 3,t M is the penalty term corresponding to the false negative rate at time t; t T represents the number of times the assessed fire severity level is lower than the actual fire severity level within time t; t Let t be the total number of evaluations within time t; t is the current time.

[0089] In a specific embodiment, during the deep reinforcement learning agent training process in step S3, a reward value is calculated after each decision, and the reward function R... t The accuracy bonus term r is weighted by three coefficients: α, β, and γ. 1,t False alarm penalty item r 2,t Penalty for underreporting 3,t By using weighted combinations, this design can guide the agent to focus on core performance metrics during training.

[0090] α, as the weighting coefficient for the accuracy bonus item, has the largest value and satisfies α>β>γ, while α+β+γ=1. This reflects the core position of accuracy in the evaluation system. Accurate risk assessment is a prerequisite for effective early warning and can provide a reliable basis for response measures. 1,t C The number of times the assessed fire level matches the actual fire level t Total number of evaluations T t The ratio directly reflects the degree to which the agent's decision matches the actual situation; the higher the value, the stronger the reliability of the agent's decision.

[0091] r 2,t and r 3,t By the number of false alarms F t and the number of missed reports Mi Total number of evaluations T t The ratio is negative. False alarms can cause unnecessary panic and waste of resources, while false alarms can lead to the spread of fire and serious losses. The setting of the penalty term can constrain the agent's wrong decision-making, prompting the model to continuously reduce false alarms and false alarms during training, and ultimately improve the practicality and credibility of the entire evaluation system.

[0092] Furthermore, in step S3, the formula for updating the model parameters of the deep reinforcement learning agent is:

[0093]

[0094] Where, θ t+1 Here are the policy network parameters at time t+1; θ t Let be the policy network parameters at time t; η be the learning rate. For the objective function J with respect to θ t The gradient; t is the current time;

[0095] The expression for the objective function is:

[0096]

[0097] Where J(θ) t ) represents the expected cumulative reward from time t to time t+N; E is the mathematical expectation operator; k is the time step index; N is the time step size; γ is the discount factor; R t+k Let t be the reward value at time t+k; t is the current time.

[0098] In a specific embodiment, in the deep reinforcement learning agent model parameter update formula of step S3, θ t+1 These are the policy network parameters at time t+1, θ t These are the parameters at time t, obtained through the learning rate η and the gradient of the objective function. The parameters are updated using the product of the learning rate and the objective function J(θ). The learning rate η controls the step size of the parameter update. A reasonable step size can avoid oscillations or slow convergence during model training, allowing the model to learn the optimal policy stably and efficiently. t ) is the expected cumulative reward from time t to time t+N. The mathematical expectation operator E can reduce the influence of random factors on reward calculation and make the objective function more stable.

[0099] k is the time step index, N is the time step size, and γ is the discount factor. The introduction of the discount factor balances the importance of current and future rewards, guiding the agent to not only focus on the effect of immediate decisions but also consider long-term evaluation performance, avoiding short-sighted decisions. Cumulative reward R t+kThe summation operation can comprehensively reflect the model's decision-making performance over a period of time, providing a comprehensive basis for parameter updates. Through continuous parameter updates, the agent's decision-making ability will be continuously improved, ultimately achieving accurate fire risk assessment.

[0100] Furthermore, such as Figure 4 As shown, in step S4, the continuous output values ​​of the deep reinforcement learning agent's policy output layer are compared one by one with the intervals corresponding to the discrete risk levels obtained from statistical analysis of its training data; each interval uniquely corresponds to a dynamic risk level; the intervals do not overlap.

[0101] In a specific embodiment, step S4 compares the continuous output values ​​of the deep reinforcement learning agent's policy output layer with the corresponding intervals of discrete risk levels. The range of continuous output values ​​is quite wide, making direct use for early warning insufficiently intuitive and hindering the development of targeted response measures. The corresponding intervals obtained through statistical analysis of training data accurately reflect the actual risk level corresponding to different continuous output values. Each interval uniquely corresponds to a dynamic risk level, and the intervals do not overlap, ensuring the uniqueness and certainty of risk level determination and avoiding ambiguous assessment results. This mapping method transforms abstract continuous decision results into clear and easily understandable discrete levels, enabling relevant personnel to quickly grasp the fire risk status of the target area, providing clear guidance for triggering early warning signals and implementing response measures, and improving the efficiency of early warning response.

[0102] Furthermore, such as Figure 5 As shown, in step S5, when the dynamic risk level is low, a solid green light is triggered; when the dynamic risk level is medium, a flashing yellow light is triggered; when the dynamic risk level is high, a flashing red light and a buzzer alarm are triggered as an audible and visual warning signal; when the dynamic risk level is extremely high, a flashing red light and a buzzer alarm are triggered as an audible and visual warning signal, along with a text warning signal containing the target area name, the current risk level, and specific guidance instructions for evacuating along the nearest evacuation route. The text warning signal is sent to the fire control room monitoring terminal and the mobile terminal of the management personnel.

[0103] In a specific embodiment, step S5 designs differentiated warning signals based on different dynamic risk levels. This tiered warning approach makes the warnings more targeted and practical. Low-risk levels trigger a solid green light, clearly conveying the information that the area is safe without disturbing the surrounding environment, maintaining normal production and living order. Medium-risk levels trigger a flashing yellow light, gently alerting relevant personnel to potential risks and prompting them to prepare for prevention. High-risk levels trigger a combination of flashing red light and a buzzer alarm, quickly attracting attention through visual and auditory stimulation and prompting initial response measures, such as checking for hazards and preparing for evacuation. Extremely high-risk levels, in addition to the visual and auditory warnings, add a text warning signal containing the target area name, current risk level, and specific evacuation guidance instructions. This text information provides clear action guidance and is simultaneously sent to the fire control room monitoring terminal and management personnel's mobile terminals, allowing fire departments and relevant management personnel to obtain crucial information promptly, quickly organize evacuation and rescue work, and minimize casualties and property losses caused by fire.

[0104] Furthermore, such as Figure 6 As shown, in step S2, the state space is dynamically updated over time, and the update interval is consistent with the acquisition frequency of the multi-source sensing data.

[0105] The state vector expression at time t+1 is:

[0106] S t+1 =λ·S t +(1-λ)·[x 1,t+1 ,x 2,t+1 ,...,x n,t+1 ];

[0107] Among them, S t+1 Let S be the fire risk assessment state vector at time t+1; λ is the state update coefficient; S t Let x be the fire risk assessment state vector at time t; i,t+1 Let be the standardized value of the i-th type of sensing data at time t+1; i∈{1,2,…,n}; t is the current time; n is the number of types of sensing data and n≥5.

[0108] In a specific embodiment, the state space in step S2 is dynamically updated over time, and the update interval is consistent with the collection frequency of multi-source sensing data. This allows the state space to incorporate the latest sensing data in a timely manner, accurately reflect the real-time changes in the fire risk of the target area, and avoid the assessment results from being out of touch with the actual situation due to data lag.

[0109] The state vector S at time t+1 t+1 The state vector S at time t is updated using the state update coefficient λ. tCombined with the new standardized data at time t+1, the presence of input balances the influence of historical states and real-time data. It retains the effective information accumulated in previous evaluation processes, ensuring the continuity of evaluation, while also responding promptly to risk changes reflected in new data, giving the state vector dynamic adaptability. This weighted rolling update method allows the state space to continuously conform to the actual situation of the target region, providing accurate input for the deep reinforcement learning agent, thereby improving the dynamic response capability and evaluation accuracy of the entire evaluation system.

[0110] Furthermore, in step S3, the deep deterministic policy gradient network structure includes a state input layer, a hidden layer, and a policy output layer; the state input layer receives the standardized state vector S. t The number of neurons is equal to the state vector S. t The dimension is n (n≥5); the hidden layer consists of at least two fully connected layers, each with 2n neurons; the policy output layer has 1 neuron, and the risk assessment decision output value is obtained by scaling after applying the tanh activation function. The expression for the output value is:

[0111]

[0112] Among them, a t Let be the risk assessment decision output value at time t; tanh is the hyperbolic tangent activation function; o t t represents the output value of the strategy output layer when it is not activated; t represents the current time.

[0113] In a specific embodiment, the deep deterministic policy gradient network structure in step S3 adopts a hierarchical design, with the state input layer receiving the standardized state vector S. t The number of neurons is kept consistent with the dimension n of the state vector, ensuring that the input layer receives all information from the state vector without missing any data related to fire risk. The hidden layers consist of at least two fully connected layers, each with 2n neurons. This sufficient number of neurons and multi-layered network structure effectively mines deep features in the state vector, enhancing the model's ability to process and express complex data, and better capturing the intrinsic correlation between various perceived data and fire risk. The policy output layer has only one neuron, and the risk assessment decision output value 'a' is obtained by scaling after applying the tanh activation function. t The tanh activation function maps the network output to the interval [-1, 1], and then transforms it into an a value in the interval [0, 1] through scaling operations. t The output values ​​in this interval are easily compared with the corresponding intervals of discrete risk levels, simplifying the mapping process. Simultaneously, the nonlinear characteristics of the tanh function enhance the model's fitting ability, enabling the agent to more accurately output decision results that match the actual risk, thus improving the accuracy of the assessment.

[0114] The specific embodiments of the present invention have been described in detail above, but they are merely examples, and the present invention is not equivalent to the specific embodiments described above. For those skilled in the art, any equivalent modifications and substitutions to the present invention are also within the scope of the present invention. Therefore, all equivalent transformations and modifications made without departing from the spirit and scope of the present invention should be covered within the scope of the present invention.

Claims

1. A method for dynamic assessment and early warning of fire risk based on deep reinforcement learning, characterized in that, Includes the following steps: Step S1: Acquire multi-source sensing data for the target area; Step S2: Construct a dynamically updated fire risk assessment state space based on multi-source sensing data. The state space is dynamically adjusted through a weighted rolling update mechanism and contains at least five sensing data features. Step S3: A deep reinforcement learning agent optimized based on a cumulative reward function weighted by accuracy reward, false positive penalty, and false negative penalty processes the real-time state in the state space and generates a risk assessment decision; the deep reinforcement learning agent adopts a deep deterministic policy gradient network structure. Step S4: Map the continuous output of the risk assessment decision to discrete dynamic fire risk levels; Step S5: Trigger the corresponding early warning signal based on the dynamic fire risk level.

2. The fire risk dynamic assessment and early warning method based on deep reinforcement learning according to claim 1, characterized in that, The multi-source sensing data includes temperature data, smoke concentration data, flame characteristic data, building structure parameter data, and personnel density data of the target area; the temperature data is collected by an array of temperature sensors evenly distributed on the ceiling or walls of the target area; the smoke concentration data is collected by smoke sensors installed on the top of the target area; the flame characteristic data is extracted from images captured by surveillance cameras in the target area using a flame detection algorithm; and the building structure parameter data includes the building floor height, fire compartment area, and evacuation route width of the target area. The personnel density data is calculated by infrared sensors or video images within the target area using a human detection algorithm.

3. The fire risk dynamic assessment and early warning method based on deep reinforcement learning according to claim 2, characterized in that, In step S2, each type of sensing data is processed using the min-max normalization method, and the processing formula is as follows: Where, x i,t X is the standardized value of the i-th type of sensing data at time t; i,t X represents the original value of the i-th type of sensory data at time t; i,min X is the minimum value of the i-th type of sensory data; i,max t represents the maximum value of the i-th type of sensing data; i is the index of the sensing data type, i∈{1,2,…,n}, t is the current time; n is the number of types of sensing data and n≥5; The standardized values ​​of all sensed data are combined into a state vector, the expression of which is: S t =[x 1,t ,x 2,t ,...,x n,t ]; Among them, S t Let x be the fire risk assessment state vector at time t; i,t Let be the standardized value of the i-th type of sensing data at time t; i is the index of the sensing data type, i∈{1,2,…,n}, t is the current time; n is the number of types of sensing data and n≥5.

4. The fire risk dynamic assessment and early warning method based on deep reinforcement learning according to claim 3, characterized in that, In step S3, during the training process of the deep reinforcement learning agent, a reward value is calculated after each decision step. The expression for the reward function is: R t =α·r 1,t +b·c 2,t +γ·r 3,t ; Among them, R t Let be the reward value at time t; α be the weighting coefficient of the accuracy reward term; r 1,t β is the reward item corresponding to the accuracy of risk assessment at time t; β is the weight coefficient of the false alarm penalty item; r 2,t Let be the penalty term corresponding to the false alarm rate at time t; γ be the weight coefficient of the false alarm penalty term; t be the current time; the weight coefficients satisfy α>β>γ and α+β+γ=1; The expression for the reward item corresponding to the accuracy reward item is: Where, r 1,t C represents the reward item corresponding to the accuracy of risk assessment at time t. t T represents the number of times the assessed fire severity level matches the actual fire severity level within time t; t The total number of evaluations within time t; t is the current time. The expression for the penalty term corresponding to the false alarm rate is: Where, r 2,t F is the penalty term corresponding to the false alarm rate at time t; t T represents the number of times the assessed fire severity level is higher than the actual fire severity level within time t; t The total number of evaluations within time t; t is the current time. The expression for the penalty term corresponding to the false negative rate is: Where, r 3,t M is the penalty term corresponding to the false negative rate at time t; t T represents the number of times the assessed fire severity level is lower than the actual fire severity level within time t; t Let t be the total number of evaluations within time t; t is the current time.

5. The fire risk dynamic assessment and early warning method based on deep reinforcement learning according to claim 4, characterized in that, In step S3, the formula for updating the model parameters of the deep reinforcement learning agent is: Where, θ t+1 Here are the policy network parameters at time t+1; θ t Let be the policy network parameters at time t; η be the learning rate. For the objective function J with respect to θ t The gradient; t is the current time; The expression for the objective function is: Wherein, J(θ) t ) represents the expected cumulative reward from time t to time t+N; E is the mathematical expectation operator; k is the time step index; N is the time step size; γ is the discount factor; R t+k Let t be the reward value at time t+k; t is the current time.

6. The fire risk dynamic assessment and early warning method based on deep reinforcement learning according to claim 5, characterized in that, In step S4, the continuous output values ​​of the deep reinforcement learning agent's policy output layer are compared one by one with the intervals corresponding to discrete risk levels obtained from statistical analysis of its training data; each interval uniquely corresponds to a dynamic risk level; the intervals do not overlap with each other.

7. The fire risk dynamic assessment and early warning method based on deep reinforcement learning according to claim 6, characterized in that, In step S5, when the dynamic risk level is low, a solid green light is triggered; when the dynamic risk level is medium, a flashing yellow light is triggered; when the dynamic risk level is high, a flashing red light and a buzzer alarm are triggered as an audible and visual warning signal; when the dynamic risk level is extremely high, a flashing red light and a buzzer alarm are triggered as an audible and visual warning signal, along with a text warning signal containing the target area name, the current risk level, and specific guidance instructions for evacuating along the nearest evacuation route. The text warning signal is sent to the fire control room monitoring terminal and the management personnel's mobile terminal.

8. The fire risk dynamic assessment and early warning method based on deep reinforcement learning according to claim 7, characterized in that, In step S2, the state space is dynamically updated over time, and the update interval is consistent with the acquisition frequency of the multi-source sensing data. The state vector expression at time t+1 is: S t+1 =λ·S t +(1-λ)·[x 1,t+1 ,x 2,t+1 ,...,x n,t+1 ]; Among them, S t+1 Let S be the fire risk assessment state vector at time t+1; λ is the state update coefficient; S t Let x be the fire risk assessment state vector at time t; i,t+1 Let be the standardized value of the i-th type of sensing data at time t+1; i∈{1,2,…,n}; t is the current time; n is the number of types of sensing data and n≥5.

9. The fire risk dynamic assessment and early warning method based on deep reinforcement learning according to claim 8, characterized in that, In step S3, the deep deterministic policy gradient network structure includes a state input layer, a hidden layer, and a policy output layer; the state input layer receives the standardized state vector S. t The number of neurons is equal to the state vector S. t The dimension n (n≥5); the hidden layer includes at least two fully connected layers, each with 2n neurons; the policy output layer has 1 neuron, and the risk assessment decision output value is obtained by scaling after applying the tanh activation function, and the expression for the output value is: Among them, a t Let be the risk assessment decision output value at time t; tanh is the hyperbolic tangent activation function; o t t represents the output value of the strategy output layer when it is not activated; t represents the current time.

Images