A control method for a fire detection alarm

Abstract

The present application belongs to the technical field of fire safety, and particularly relates to a control method of a fire detection alarm. First, a distributed sensor network is deployed to collect temperature and smoke data and self-calibrate, and to eliminate abnormal data. Then, through three-stage progressive abnormal screening, i.e., first-stage primary screening to monitor mutation characteristics, second-stage characteristic verification to construct a space-time matrix for analysis by CNN, and third-stage scene adaptation to calculate probability in combination with historical data, an abnormal screening signal is output. Then, a visual camera is used to collect images to generate a visual evidence matrix, and a comprehensive confidence is calculated in combination with a sensor data vector and the abnormal screening signal, and an alarm signal is triggered if the comprehensive confidence is greater than a threshold. Finally, a one-minute period time window is slid to verify, and an alarm is determined if the alarm signal appears for greater than or equal to three periods. The method improves data accuracy, accurately identifies fire hazards, reduces false alarm rates, and ensures the timeliness and accuracy of the alarm.

Description

Technical Field

[0001] This invention belongs to the field of fire safety technology, and in particular relates to a control method for a fire detector alarm. Background Technology

[0002] Fire detectors and alarms are safety devices that monitor fire hazards in real time. They detect signals such as smoke, temperature, and flames through sensors. They can quickly identify initial fires and trigger an alarm, activating the fire suppression system. Widely used in homes, businesses, and industrial settings, they provide early fire warnings and allow for timely evacuation, reducing the risk of loss of life and property. However, existing fire detectors and alarms are susceptible to environmental interference, leading to false alarms and difficulty in distinguishing between real fires and interfering signals, resulting in a high false alarm rate. Summary of the Invention

[0003] In view of the technical problems mentioned above, the present invention proposes a control method for a fire detector alarm.

[0004] To achieve the above objectives, the technical solution adopted by the present invention includes the following steps:

[0005] S1. Deploy a distributed sensor network to collect environmental data, including temperature data and smoke data;

[0006] S2. Perform anomaly screening based on the collected data and output anomaly screening signal; the anomaly screening is a three-level progressive anomaly screening, including first-level initial screening, second-level feature verification, and third-level scene adaptation.

[0007] The first-level screening involves deploying a lightweight anomaly detection model on edge computing nodes to monitor the sudden changes in temperature change rate and smoke concentration in real time, triggering a preliminary anomaly signal A0.

[0008] The secondary feature verification involves initiating a regional collaborative detection mechanism after A0 is triggered. This mechanism retrieves temperature and smoke data from sensors within 3 meters of the anomaly point, constructs a spatiotemporal feature matrix, and verifies the existence of a gradient-ascending feature distribution using a convolutional neural network. A feature verification signal A1 is then generated. The specific steps are as follows:

[0009] First, the temperature sensor collects 60 data points in one cycle, and the smoke sensor generates 30 data points. The smoke data is upsampled using linear interpolation to generate a 60-point sequence synchronized with the temperature data, forming a synchronized time series matrix D.

[0010] A period is divided into 10 time slices. The average temperature, average smoke level, and temperature change rate are calculated for each time slice to generate a periodic feature vector. in Represents the average temperature. ΔT represents the average smoke level, and ΔT represents the rate of temperature change within the time slice.

[0011] Using outliers as the center, the DBSCAN algorithm is used to filter sensors within 3 meters to obtain a set of nearby sensors C, where each sensor contains synchronized time series data for one period.

[0012] Constructing the spatiotemporal feature matrix Where c represents the number of nearby sensors. The generation periodic feature vector of the central node;

[0013] For each sensor, calculate its Euclidean distance d from the central node. c Generate Gaussian weights And calculate the mean rate of temperature change over time. Smoke concentration gradient in spatial dimensions The weighted gradient consistency is calculated as follows: Ensure that the range of the weighted gradient consistency index is [0,1];

[0014] Then, the spatiotemporal feature matrix M is input into the CNN model, and the gradient distribution features of neighboring sensors are extracted through convolutional kernels, outputting a unique image H. A fully connected layer is then combined with Gaussian weights ω. c Calculate the weighted eigenvector V = Softmax(ω) c )·H, the feature matching degree FM = sigmoid(V·W) is output by the last neuron of the fully connected layer. out +b out ), where W out b represents the weight of the last layer. out For bias;

[0015] When the feature matching degree FM of the CNN output is greater than 0.7 and GI is greater than 0.6, a feature verification signal A1 is generated;

[0016] The three-level scene adaptation combines the building's historical fire data to establish a scene database, performs scene adaptation calibration to calculate scene conditional probabilities, and outputs an anomaly screening signal A2.

[0017] S3. Use the visual camera of the fire detector to collect images of the scene, generate a visual evidence matrix and retrieve real-time sensor data to construct a sensor data vector. Combine the visual evidence matrix, sensor data vector and anomaly screening signal to calculate the comprehensive confidence level. When the confidence level is greater than the threshold, trigger the alarm signal.

[0018] S4. Finally, a time window sliding verification is performed. The time window sliding verification selects one minute as a cycle. If the alarm signal appears for more than or equal to 3 cycles, an alarm is confirmed.

[0019] Preferably, when using a sensor network to collect data in real time, it is necessary to perform a self-calibration operation on the sensor data, remove abnormal sensor data with excessive deviations, and fill in new data. Specifically, this is implemented as follows:

[0020] Build a sensor data prediction model, input real-time environmental data, and obtain predicted environmental data;

[0021] For each sensor's real-time environmental data, calculate and predict the residual of the environmental data, and set a residual threshold. When the residual is greater than the residual threshold, it is marked as the first abnormal data.

[0022] The first marked abnormal data is subjected to second abnormal monitoring. The second abnormal detection is an abnormal detection that introduces physical constraints, including temperature thresholds between adjacent sensors and smoke diffusion direction constraints. Data points marked with physical constraint abnormalities are recorded as second abnormal data by using Wasserstein distance metric deviation.

[0023] The second abnormal data is removed and new data is added. The data is filled with the weighted average of the neighboring sensors to obtain sensor data without abnormalities.

[0024] Preferably, in step S2, the lightweight anomaly detection model in the primary screening performs real-time monitoring of temperature mutation rate and sudden changes in smoke concentration to trigger the preliminary anomaly signal A0. The specific implementation of this is as follows:

[0025] First, asynchronous sampling is used for the temperature sensor and the smoke sensor, and a synchronous data sequence is generated through a timestamp alignment algorithm to calculate the real-time change rate of temperature and smoke within one cycle.

[0026] Set mutation thresholds, respectively Forward mutation threshold and Negative mutation threshold, smoke jump threshold is Where μ S σ represents the mean smoke level. S μ represents the standard deviation of smoke. ΔT Represents the average temperature, σ ΔT Represents the standard deviation of temperature;

[0027] when or Furthermore, when the current temperature exceeds 30 degrees Celsius, a temperature anomaly signal is triggered. Furthermore, a smoke anomaly signal is triggered when ΔS(t) > 0.5%LEL / s;

[0028] If the temperature abnormality signal and the smoke abnormality signal are triggered successively within 10 seconds, the initial abnormality signal A0 will be triggered.

[0029] Preferably, the three-level scene adaptation obtains the anomaly screening signal A2 by performing scene adaptation calibration to calculate the scene conditional probability, and the calculation formula is as follows: Where Scene represents the scene label of the region where the anomaly is located; when P(Fire|T,S,Scene)>0.6, the anomaly filtering signal A2 is output.

[0030] Preferably, the visual evidence matrix construction process in step S3 involves optical flow processing of the camera video stream to obtain dynamic and static features of the flame. The dynamic features of the flame include the flicker frequency f. flame and diffusion velocity v smoke The static characteristics of the flame include the percentage of pixels with hue ∈ [0,30]° ∩ saturation > 0.3 in the RGB-HSV color space.

[0031] Preferably, the specific implementation steps for calculating the comprehensive confidence level by combining the visual evidence matrix, sensor data vector, and anomaly screening signal in step S3 are as follows:

[0032] First, the flame and smoke regions are extracted using the YOLOv8 object detection model, and the feature matching degree is calculated as follows: Where ρ h ρ represents the percentage of flame pixels. z f represents the percentage of noisy pixels. freq_match The degree of matching between the flicker frequency and the characteristic frequency of a fire is represented by v. smoke_dir_match The degree of matching between the direction of smoke diffusion and the direction of gravity;

[0033] Then, the Mahalanobis distance between the sensor parameters and the fire sample space is calculated as follows: Where Σ is the covariance matrix of the sensor parameters;

[0034] The anomaly filtering signal A2 output by the third-level scene adaptation is converted into probability m by fusing the scene prior probability and the filtering result. A ;

[0035] Finally, the overall confidence level is calculated, and the calculation method is as follows: Where δ represents the conflict factor, γ i Represents weighting factors. This represents the non-fire confidence level, where ε is a very small constant to prevent the denominator from being 0. When the overall confidence level is greater than the set threshold, an alarm signal is triggered.

[0036] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0037] 1. Construct a distributed sensor network to collect data and perform self-calibration operations to remove abnormal data and fill in new data, thereby improving data accuracy and reliability and overcoming the problems of traditional single-point sensors being susceptible to interference and having limited monitoring range.

[0038] 2. A three-level progressive anomaly screening mechanism: the first level monitors temperature and smoke mutation characteristics; the second level verifies features using spatiotemporal matrix and CNN analysis; and the third level adapts scenarios by combining historical data to calculate probabilities. This progressive approach accurately identifies fire hazards and reduces false alarm rates.

[0039] 3. By integrating visual evidence matrix, sensor data vectors, and anomaly screening signals to calculate comprehensive confidence, fires can be judged from multiple dimensions, which improves the reliability of fire judgment compared to relying solely on sensor data.

[0040] 4. The time window sliding verification mechanism uses a one-minute cycle and only confirms an alarm after the alarm signal has appeared for more than three cycles, avoiding false alarms and missed alarms and ensuring the timeliness and accuracy of alarms. Attached Figure Description

[0041] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is a flowchart of a control method for a fire detector alarm. Detailed Implementation

[0043] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0044] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.

[0045] In this embodiment, to address the problem that existing fire detectors are susceptible to environmental interference and prone to false alarms, making it difficult to distinguish between real fires and interference signals, resulting in a high false alarm rate, this invention proposes a control method for fire detectors, specifically implemented as follows: Figure 1 As shown.

[0046] Firstly, considering that traditional single-point sensors are susceptible to environmental interference and have limited monitoring range, making it difficult to acquire comprehensive and accurate environmental data, resulting in insufficient detection of fire hazards, this invention constructs a distributed sensor network to achieve comprehensive and high-precision acquisition of environmental data. The acquired environmental data includes temperature data and smoke data.

[0047] After acquiring the data, a self-calibration process is performed on the sensor data to remove abnormal sensor data with excessive deviations and to fill in new data. Specifically, a Long Short-Term Memory (LSTM) network is selected to build the sensor data prediction model. Environmental data over a period of time is collected and organized according to time series. The data is normalized to improve the model training effect. The processed data is divided into training, validation, and test sets according to a certain ratio. The LSTM model is trained using the training set, and the model parameters are continuously adjusted to achieve good prediction performance on the validation set. Finally, the accuracy and stability of the model are evaluated using the test set to ensure that the model can accurately predict environmental data. The real-time collected environmental data is input into the trained prediction model to obtain the predicted environmental data. For the real-time environmental data of each sensor, the residual between the actual data and the corresponding predicted environmental data is calculated. A residual threshold is set based on historical data and actual conditions. When the residual of a sensor exceeds the residual threshold, this data is marked as the first abnormal data. For the marked first abnormal data, physical constraints are introduced for second anomaly monitoring. Temperature thresholds and smoke diffusion direction constraints are established between adjacent sensors. The deviation of each first outlier data point from the physical constraints is measured using Wasserstein distance. When the deviation exceeds a certain range, the data point is marked as a second outlier. These marked second outliers are then removed from the original data. For the removed data points, a weighted average of the neighboring sensors is used to fill the gaps. The weights are determined based on the distance between the sensors, with closer sensors receiving a larger weight. This method yields sensor data free of outliers.

[0048] Next, considering that traditional fire detectors struggle to distinguish between real fires and interference signals, leading to a high false alarm rate, this solution employs a three-level progressive anomaly screening mechanism to gradually eliminate interference factors, accurately identify fire hazards, and reduce the false alarm rate. Anomaly screening is performed on collected data, outputting anomaly screening signals. This three-level progressive anomaly screening includes a primary screening, a secondary feature verification, and a tertiary scene adaptation. The primary screening deploys a lightweight anomaly detection model on edge computing nodes to monitor sudden temperature changes and smoke concentration jumps in real time, triggering a preliminary anomaly signal A0. The secondary feature verification, upon triggering A0, initiates a regional collaborative detection mechanism, retrieving temperature and smoke data from sensors within 3 meters of the anomaly point to construct a spatiotemporal feature matrix. A convolutional neural network is then used to verify the existence of a gradient-ascending feature distribution, generating a feature verification signal A1. The tertiary scene adaptation combines historical fire data from the building to establish a scene database, performs scene adaptation calibration, calculates scene conditional probabilities, and outputs anomaly screening signal A2.

[0049] The lightweight anomaly detection model in the primary screening process monitors the temperature mutation rate and sudden changes in smoke concentration in real time, triggering the initial anomaly signal A0. Specifically, it first uses asynchronous sampling for the temperature and smoke sensors, and generates a synchronous data sequence using a timestamp alignment algorithm. It then calculates the real-time change rate of temperature and smoke within one period; and sets mutation thresholds, respectively... Forward mutation threshold and Negative mutation threshold, smoke jump threshold is Where μ S σ represents the mean smoke level. S μ represents the standard deviation of smoke. ΔT Represents the average temperature, σ ΔT Represents the standard deviation of temperature; when or Furthermore, when the current temperature exceeds 30 degrees Celsius, a temperature anomaly signal is triggered. The smoke anomaly signal is triggered when ΔS(t) > 0.5%LEL / s; if the temperature anomaly signal and the smoke anomaly signal are triggered successively within 10 seconds, the preliminary anomaly signal A0 is triggered.

[0050] The specific steps of the secondary feature verification are as follows: First, the temperature sensor collects 60 data points within one cycle, and the smoke sensor generates 30 data points. The smoke data is upsampled using linear interpolation to generate a 60-point sequence synchronized with the temperature data, forming a synchronized time series matrix D. One cycle is divided into 10 time slices, and the average temperature, average smoke level, and temperature change rate are calculated for each time slice to generate a cycle feature vector. in Represents the average temperature. The average smoke value is represented by ΔT, and the temperature change rate within the time slice is represented by ΔT. Using outliers as the center, sensors within a 3-meter radius are filtered using the DBSCAN algorithm to obtain a set of nearby sensors C, where each sensor contains synchronized time-series data for one period. A spatiotemporal feature matrix is ​​then constructed. Where c represents the number of nearby sensors. Let be the generation periodic feature vector of the central node; for each sensor, calculate its Euclidean distance d from the central node. c Generate Gaussian weights And calculate the mean rate of temperature change over time. Smoke concentration gradient in spatial dimensions The weighted gradient consistency is calculated as follows: Ensure the weighted gradient consistency index is within the range of [0,1]; then input the spatiotemporal feature matrix M into the CNN model, extract the gradient distribution features of neighboring sensors through convolutional kernels, and output a unique image H. The fully connected layer is then combined with Gaussian weights ω. c Calculate the weighted eigenvector V = Softmax(ω) c )·H, the feature matching degree FM = sigmoid(V·W) is output by the last neuron of the fully connected layer. out +b out ), where W out b represents the weight of the last layer. out As a bias, when the feature matching degree FM of the CNN output is greater than 0.7 and GI is greater than 0.6, the feature verification signal A1 is generated.

[0051] The implementation of the three-level scene adaptation to obtain the anomaly screening signal A2 involves calculating the scene conditional probability through scene adaptation calibration, using the following formula: Where Scene represents the scene label of the region where the anomaly is located; when P(Fire|T,S,Scene)>0.6, the anomaly filtering signal A2 is output.

[0052] The aforementioned three-tiered progressive anomaly screening significantly improves the accuracy of fire detection. The first-level initial screening, as a preliminary detection, uses a lightweight anomaly detection model on edge computing nodes to monitor obvious sudden features such as temperature fluctuations and smoke concentration jumps in real time. This step quickly identifies potential fire signs, providing a foundation for subsequent in-depth detection. The second-level feature verification is initiated after the first-level initial screening triggers a preliminary anomaly signal. It constructs a spatiotemporal feature matrix by retrieving sensor data within 3 meters of the anomaly point and uses a convolutional neural network (CNN) to verify the existence of a gradient-ascending feature distribution. This step comprehensively analyzes from both spatial and temporal dimensions, further confirming the possibility of a fire and eliminating some false alarms. The third-level scene adaptation builds a scene database based on historical fire data of the building and calculates the conditional probability of the scene. Based on the scene label of the area where the anomaly point is located, an anomaly screening signal is output. This step considers the probability of fire occurrence under different scenarios, making the judgment more realistic and reducing false alarms. This progressive screening mechanism, with each level building upon the results of the previous level for deeper analysis and gradually eliminating interfering factors, greatly improves the accuracy of fire detection. From simple threshold judgment to complex feature analysis, and then to combining scenario probability, fire hazards are investigated from multiple dimensions and angles, effectively reducing the false alarm rate.

[0053] Following this, considering the potential for false alarms even when relying solely on sensor data for judgment, the present invention further improves the reliability of fire detection by fusing a visual evidence matrix, sensor data vectors, and anomaly screening signals to calculate a comprehensive confidence level. Specifically, the visual camera of the fire detector alarm captures images of the scene, generates a visual evidence matrix, retrieves real-time sensor data, constructs a sensor data vector, and combines the visual evidence matrix, sensor data vector, and anomaly screening signals to calculate the comprehensive confidence level. An alarm signal is triggered when the confidence level exceeds a threshold. The construction of the visual evidence matrix involves optical flow processing of the camera video stream to obtain dynamic and static flame characteristics. The dynamic flame characteristics include the flicker frequency f. flame and diffusion velocity v smoke The static characteristics of the flame include the percentage of pixels with hue ∈ [0,30]° ∩ saturation > 0.3 in the RGB-HSV color space.

[0054] Then, the specific steps for calculating the overall confidence score by combining the visual evidence matrix, sensor data vectors, and anomaly screening signals are as follows: First, extract the flame and smoke regions using the YOLOv8 object detection model, and calculate the feature matching degree. The calculation method is as follows: Where ρ h ρ represents the percentage of flame pixels. z f represents the percentage of noisy pixels. freq_match The degree of matching between the flicker frequency and the characteristic frequency of a fire is represented by v.smoke_dir_match The degree of matching between the smoke diffusion direction and the gravity direction is calculated; then, the Mahalanobis distance between the sensor parameters and the fire sample space is calculated as follows: Where Σ is the covariance matrix of the sensor parameters; the anomaly filtering signal A2 output by the three-level scene adaptation is converted into probability m by fusing the scene prior probability and the filtering result. A Finally, the overall confidence level is calculated using the following method: Where δ represents the conflict factor, γ i Represents weighting factors. This represents the non-fire confidence level, where ε is a very small constant to prevent the denominator from being 0. When the overall confidence level is greater than the set threshold, an alarm signal is triggered.

[0055] Finally, considering that in practical applications, occasional abnormal signals may not be actual fires, and immediate alarms could easily lead to false alarms; while delayed alarms could allow the fire to spread. This solution uses a time-window sliding verification mechanism to ensure timely alarms while avoiding false alarms and missed alarms. The time-window sliding verification selects one minute as a cycle, and if an alarm signal appears for more than three cycles, an alarm is confirmed. Specifically, in actual operation, alarm signals are continuously monitored in one-minute cycles. When the overall confidence level is greater than a set threshold, the alarm signal is triggered. If an alarm is triggered in the first cycle, the system does not immediately confirm a fire, but starts timing and recording the alarm. If an alarm is triggered again in the next one-minute cycle, the alarm count is incremented by one. When the alarm signal is triggered for three consecutive one-minute cycles, that is, the number of alarms is greater than or equal to three, the system determines that a fire has occurred, activates fire-fighting linkage equipment, and issues audible and visual alarms to notify personnel to evacuate. If the number of alarm signal triggers is less than three in three cycles, the counting restarts, and this cycle continues to ensure that there are no false alarms due to brief interference, nor is the optimal time for fire suppression delayed due to delayed judgment.

[0056] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A control method for a fire detector alarm, characterized in that, Includes the following steps: S1. Deploy a distributed sensor network to collect environmental data, including temperature data and smoke data; S2. Perform anomaly screening based on the collected data and output anomaly screening signal; the anomaly screening is a three-level progressive anomaly screening, including first-level initial screening, second-level feature verification, and third-level scene adaptation. The first-level screening involves deploying a lightweight anomaly detection model on edge computing nodes to monitor the sudden changes in temperature change rate and smoke concentration in real time, triggering a preliminary anomaly signal A0. The secondary feature verification involves initiating a regional collaborative detection mechanism after A0 is triggered. This mechanism retrieves temperature and smoke data from sensors within 3 meters of the anomaly point, constructs a spatiotemporal feature matrix, and verifies the existence of a gradient-ascending feature distribution using a convolutional neural network. A feature verification signal A1 is then generated. The specific steps are as follows: First, the temperature sensor collects 60 data points in one cycle, and the smoke sensor generates 30 data points. The smoke data is upsampled using linear interpolation to generate a 60-point sequence synchronized with the temperature data, forming a synchronized time series matrix D. A period is divided into 10 time slices. The average temperature, average smoke level, and temperature change rate are calculated for each time slice to generate a periodic feature vector. in Represents the average temperature. ΔT represents the average smoke level, and ΔT represents the rate of temperature change within the time slice. Using outliers as the center, the DBSCAN algorithm is used to filter sensors within 3 meters to obtain a set of nearby sensors C, where each sensor contains synchronized time series data for one period. Constructing the spatiotemporal feature matrix Where c represents the number of nearby sensors. The generation periodic feature vector of the central node; For each sensor, calculate its Euclidean distance d from the central node. c Generate Gaussian weights And calculate the mean rate of temperature change over time. Smoke concentration gradient in spatial dimensions The weighted gradient consistency is calculated as follows: Ensure that the range of the weighted gradient consistency index is [0,1]; Then, the spatiotemporal feature matrix M is input into the CNN model, and the gradient distribution features of neighboring sensors are extracted through convolutional kernels, outputting a unique image H. A fully connected layer is then combined with Gaussian weights ω. c Calculate the weighted eigenvector V = Softmax(ω) c )·H, the feature matching degree FM = sigmoid(V·W) is output by the last neuron of the fully connected layer. out +b out ), where W out b represents the weight of the last layer. out For bias; When the feature matching degree FM of the CNN output is greater than 0.7 and GI is greater than 0.6, a feature verification signal A1 is generated; The three-level scene adaptation combines the building's historical fire data to establish a scene database, performs scene adaptation calibration to calculate scene conditional probabilities, and outputs an anomaly screening signal A2. S3. Use the visual camera of the fire detector to collect images of the scene, generate a visual evidence matrix and retrieve real-time sensor data to construct a sensor data vector. Combine the visual evidence matrix, sensor data vector and anomaly screening signal to calculate the comprehensive confidence level. When the confidence level is greater than the threshold, trigger the alarm signal. S4. Finally, a time window sliding verification is performed. The time window sliding verification selects one minute as a cycle. If the alarm signal appears for more than or equal to 3 cycles, an alarm is confirmed.

2. The control method for a fire detector alarm according to claim 1, characterized in that, When using a sensor network to collect data in real time, it is necessary to perform a self-calibration operation on the sensor data, remove abnormal sensor data with excessive deviations, and fill in new data. The specific implementation is as follows: Build a sensor data prediction model, input real-time environmental data, and obtain predicted environmental data; For each sensor's real-time environmental data, calculate and predict the residual of the environmental data, and set a residual threshold. When the residual is greater than the residual threshold, it is marked as the first abnormal data. The first marked abnormal data is subjected to second abnormal monitoring. The second abnormal detection is an abnormal detection that introduces physical constraints, including temperature thresholds between adjacent sensors and smoke diffusion direction constraints. Data points marked with physical constraint abnormalities are recorded as second abnormal data by using Wasserstein distance metric deviation. The second abnormal data is removed and new data is added. The data is filled with the weighted average of the neighboring sensors to obtain sensor data without abnormalities.

3. The control method for a fire detector alarm according to claim 1, characterized in that, In step S2, the lightweight anomaly detection model in the primary screening stage monitors the temperature mutation rate and sudden changes in smoke concentration in real time, and triggers the preliminary anomaly signal A0. The specific implementation of this is as follows: First, asynchronous sampling is used for the temperature sensor and the smoke sensor, and a synchronous data sequence is generated through a timestamp alignment algorithm to calculate the real-time change rate of temperature and smoke within one cycle. Set mutation thresholds, respectively Forward mutation threshold and Negative mutation threshold, smoke jump threshold is Where μ S σ represents the mean smoke level. S μ represents the standard deviation of smoke. ΔT Represents the average temperature, σ ΔT Represents the standard deviation of temperature; when or Furthermore, when the current temperature exceeds 30 degrees Celsius, a temperature anomaly signal is triggered. Furthermore, a smoke anomaly signal is triggered when ΔS(t) > 0.5%LEL / s; If the temperature abnormality signal and the smoke abnormality signal are triggered successively within 10 seconds, the initial abnormality signal A0 will be triggered.

4. The control method for a fire detector alarm according to claim 1, characterized in that, The implementation of the three-level scene adaptation to obtain the anomaly screening signal A2 involves calculating the scene conditional probability through scene adaptation calibration, using the following formula: Where Scene represents the scene label of the region where the anomaly is located; when P(Fire|T,S,Scene)>0.6, the anomaly filtering signal A2 is output.

5. The control method for a fire detector alarm according to claim 1, characterized in that, The process of constructing the visual evidence matrix in step S3 involves processing the camera video stream using optical flow to obtain dynamic and static features of the flame. The dynamic features of the flame include the flicker frequency f. flame and diffusion velocity v smoke The static characteristics of the flame include the percentage of pixels with hue ∈ [0,30]° ∩ saturation > 0.3 in the RGB-HSV color space.

6. The control method for a fire detector alarm according to claim 1, characterized in that, The specific steps for calculating the overall confidence level by combining the visual evidence matrix, sensor data vector, and anomaly screening signal in step S3 are as follows: First, the flame and smoke regions are extracted using the YOLOv8 object detection model, and the feature matching degree is calculated as follows: Where ρ h ρ represents the percentage of flame pixels. z f represents the percentage of noisy pixels. freq_match The degree of matching between the flicker frequency and the characteristic frequency of a fire is represented by v. smoke_dir_match The degree of matching between the direction of smoke diffusion and the direction of gravity; Then, the Mahalanobis distance between the sensor parameters and the fire sample space is calculated as follows: Where Σ is the covariance matrix of the sensor parameters; The anomaly filtering signal A2 output by the third-level scene adaptation is converted into probability m by fusing the scene prior probability and the filtering result. A ; Finally, the overall confidence level is calculated, and the calculation method is as follows: Where δ represents the conflict factor, γ i Represents weighting factors. This represents the non-fire confidence level, where ε is a very small constant to prevent the denominator from being 0. When the overall confidence level is greater than the set threshold, an alarm signal is triggered.

Images