Intelligent gas alarm method and system

Abstract

The invention discloses an intelligent gas alarm method and system, and belongs to the technical field of gas safety monitoring. According to the method, multi-source sensing data of gas, flame, human body and valve states are integrated, and after missing value filling, abnormal value detection and correction, preprocessing of moving average filtering and feature extraction based on a fixed time window, the data are input into a support vector machine classification model for risk level identification and judgment; executing a multi-mode linkage control strategy according to the risk level, the valve state signal and the human body existence state signal; and anonymized feature data generated by the model and corresponding event tags are uploaded to a cloud end for incremental learning and model iterative updating based on weighted security performance indexes. According to the invention, multifunctional collaborative linkage and intelligent decision making are realized, the alarm threshold can be dynamically optimized, and the accuracy, environmental adaptability and long-term self-evolution capability of the system are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of gas safety monitoring technology, and in particular to an intelligent gas alarm method and system. Background Technology

[0002] In the field of gas safety in residential and commercial kitchens, there are various safety risks caused by factors such as gas leaks, dry burning of stoves, and personnel leaving their posts unattended. Furthermore, these application scenarios exhibit significant differences and dynamic changes in environmental temperature, humidity, and ventilation conditions. Current technical solutions for addressing these risks have the following recognized shortcomings.

[0003] The system suffers from low integration and isolated functions. Existing gas leak alarms, valve status monitors, anti-dry-burning devices, and hot work / unattended personnel monitoring systems are typically designed, deployed, and operated independently. The lack of effective data communication and collaborative control mechanisms among these devices prevents the construction of a unified and efficient safety response system. An "information gap" exists between functional modules, hindering the maximization of safety benefits.

[0004] The gas alarm lacks environmental adaptability and relies heavily on manual intervention for reliability. Changes in environmental factors directly affect the performance output of the gas alarm. However, the core decision parameters of existing devices are usually pre-stored in non-volatile memory as fixed values. Due to fluctuations in environmental parameters caused by different installation locations (such as basements and rooftop kitchens), different seasons (summer and winter), and different user habits, a fixed set of decision parameters cannot maintain optimal performance in all scenarios. This directly leads to a dilemma in actual use: setting the sensitivity too high to ensure safety will cause a surge in false alarms, while setting the sensitivity too low to reduce false alarms will bring the risk of missed alarms. Although some high-end devices allow for manual threshold correction, this is too technically challenging and difficult to implement for ordinary users, making it impossible to guarantee the long-term operational reliability of the device in most cases.

[0005] Therefore, existing gas safety devices generally suffer from low system integration and isolated functions. Functions such as gas leak alarms, valve monitoring, anti-dry burning, and remote monitoring are typically implemented by independent devices, lacking effective data collaboration and control linkage, making it difficult to form a unified safety response system. Furthermore, existing devices lack environmental adaptability; core decision parameters are mostly fixed values, unable to adapt to parameter fluctuations caused by different installation environments, seasonal changes, and user habits. This leads to a dilemma in practical applications: high false alarm rates or high risk of missed alarms. Moreover, reliance on manual threshold adjustment results in a high operational threshold, making long-term reliability difficult to guarantee. Existing technical solutions are insufficient in both systemic and intelligent aspects, failing to meet the modern kitchen's demand for highly reliable and convenient safety protection. Summary of the Invention

[0006] To address the aforementioned problems, this invention provides an intelligent gas alarm method. It constructs a multi-module system integrating a gas sensing module, a flame monitoring module, a human body sensing module, and a valve status sensing and controller. It runs a multi-source data fusion-based alarm method encompassing data preprocessing, feature parameter extraction, support vector machine (SVM) model classification, and an optimized closed-loop system based on anonymized incremental learning and weighted safety performance index (WSPI) verification. This solves the technical problems in existing gas safety technologies, such as poor environmental adaptability, difficulty in balancing false alarm and false alarm rates, low system intelligence, and inability to continuously evolve, caused by isolated system functions and fixed decision parameters. The method achieves collaborative sensing and linkage control of multiple risk sources, dynamically optimizes alarm thresholds to significantly improve alarm accuracy, and utilizes a cloud-based continuous learning and verification mechanism to enable long-term self-optimization, achieving high reliability and adaptability.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An intelligent gas alarm method, comprising the following steps: S1: acquiring multi-source data output from a gas sensing module, a flame monitoring module, a human body sensing module, and a valve status sensor and controller. x t ; S2: Preprocess the multi-source data, including time axis alignment, missing value imputation, outlier detection and correction, and moving average filtering, and label the preprocessed data with event tags; the formula for outlier detection is: in, z Parameters for outlier monitoring; μ The mean; σ Let | be the standard deviation; if | z If |>3, then determine x t Identify outliers and correct them; S3: Extract at least the average rate of temperature change from the preprocessed data using a fixed time window. Characteristic parameters; the average rate of temperature change The formula is: in, T ( n )and T (1) are the first of the fixed time windows. n The first and second temperature values; t ( n )and t (1) These are the corresponding collection times; S4: Input the extracted feature parameters and corresponding event labels into a pre-trained machine learning classification model to obtain the risk level; S5: Based on the risk level, valve status signal and human presence status signal, execute a multimodal linkage control strategy to dynamically control the audible and visual alarm module and the valve status sensor and controller. S6: Upload the anonymized feature data and corresponding event labels generated by the model operation to the cloud for incremental learning and optimization updates of the model. A / B testing and iterative updates of the model before and after optimization are performed using the Weighted Security Performance Index (WSPI). The formula for calculating WSPI is: in, w i These are the weighting coefficients; C i For performance components; i For indexing.

[0008] Preferably, the missing value imputation in step S2 uses linear interpolation, and the formula is as follows: in, t It is discrete time; x t For the first t Multi-source data at any given time.

[0009] Preferably, the moving average filtering in step S2 has the following formula: in, k The size of the sliding window; This is the filtered data; t It is discrete time; x i For the first i Multi-source data after outlier detection and correction; i For indexing.

[0010] Preferably, the feature parameters extracted in step S3 also include the average temperature. With temperature range R : in, n The total number of temperature data points within a fixed time window; i For indexing; For the first i Temperature values ​​of each temperature data point; To all within the window Take the maximum value; To all within the window Take the minimum value.

[0011] Preferably, the machine learning classification model in step S4 is a support vector machine, and its kernel function is a radial basis function kernel: in, x i , x j The input feature vector; γ These are kernel function parameters; For feature vectors x i and x j The Euclidean distance between them.

[0012] Preferably, the multimodal linkage control strategy includes: when the flame monitoring module detects that the ambient temperature exceeds the dynamically adjusted warning threshold and the human body sensing module does not detect the presence of a human body, controlling the audible and visual alarm module to enter the warning mode; if this state continues to exceed the dynamically adjusted time threshold, controlling the audible and visual alarm module to upgrade to the alarm mode, and controlling the valve status sensor and controller to close the valve.

[0013] Preferably, the multimodal linkage control strategy further includes: when the gas sensing module detects that the target gas concentration exceeds the threshold, if the valve status sensor and controller feedback that the valve is in the open state, then immediately control it to close the valve and activate the warning mode of the audible and visual alarm module; if the valve is in the closed state, then according to the detection result of the human body sensing module, select to activate the alarm mode of the audible and visual alarm module or initiate a remote alarm notification.

[0014] Preferably, the performance component in step S6 C i The formula is: in, C 1 represents the alarm recognition accuracy component; C 2 represents the false alarm control component; C 3 is to avoid underreporting; C 4 represents the component with stable performance; TP To correctly predict the number of positive examples, FP The number of incorrectly predicted positive examples. TN To correctly predict the number of negative examples, FN This represents the number of incorrectly predicted negative examples.

[0015] Preferably, the weighting coefficient w i Specifically: w 1 = 0.4 w 2 = 0.25 w 3 = 0.2 w 4 = 0.15.

[0016] An intelligent gas alarm system includes: a main control module; a gas sensing module for monitoring gas concentration; a flame monitoring module for monitoring flames or abnormal temperatures; a human body sensing module for sensing the presence of a human body; and a valve status sensor and controller for detecting and controlling gas pipeline valves. An audible and visual alarm module for issuing audible and visual alarms; and a network transmission module for data communication; The gas sensing module, flame monitoring module, human body sensing module, valve status sensing and controller, audible and visual alarm module, and network transmission module are all connected to the main control module; the main control module is used to execute the above-mentioned intelligent gas alarm method.

[0017] By adopting the above technical solution, the present invention has the following beneficial effects.

[0018] (1) This invention breaks down the "information gap" in existing technologies by constructing a hardware system integrating a gas sensing module, a flame monitoring module, a human body sensing module, and a valve status sensing and controller, and by implementing a multimodal linkage control strategy based on the main control module. This invention enables the above-mentioned heterogeneous sensing and execution units to work collaboratively under a unified logical framework. For example, when the gas sensing module detects a leak and the valve status sensing and controller reports that the valve is open, the system immediately triggers valve closure and alarm; when the flame monitoring module detects a high temperature and the human body sensing module confirms that no one is present, the system automatically upgrades the warning and closes the valve. This close combination of hardware integration and strategy linkage solves the problem of fragmented response and realizes a closed loop of intelligent judgment and integrated handling of composite risks from single risk detection to "leak-fire-absence of personnel", significantly improving the systematicness and timeliness of safety protection.

[0019] (2) This invention introduces a time-series feature extraction method based on a fixed time window, particularly calculating dynamic statistical features such as the average rate of temperature change and temperature range, and uses these as the core input for risk judgment. This elevates the judgment basis from a static "point" to a "line" reflecting the evolution trend of the event, effectively characterizing the essential differences in dynamic patterns between normal working conditions such as cooking and environmental fluctuations and real risk events. It provides a highly discriminative and stable input for intelligent classification, jointly solving the technical problem that it is difficult to balance false alarms and false negatives caused by dynamic environmental changes in fixed thresholds, and significantly improving the accuracy and environmental adaptability of the system alarm from the data source.

[0020] (3) This invention deploys a machine learning classification model with Support Vector Machine (SVM) as the core and Radial Basis Function (RBF) as the kernel function in the main control module to perform three-level risk classification of "safety, early warning, and alarm" on the aforementioned extracted multidimensional feature vectors, replacing the traditional judgment method that relies on manually preset complex rule logic. Utilizing the inherent advantages of the SVM model in small sample, nonlinear, and high-dimensional pattern recognition, it automatically mines the deep-level mapping relationship between multi-source sensor features and complex risk levels. It solves the technical bottlenecks of manual rules being unable to exhaust all scenarios, having weak generalization ability, and being difficult to adjust, achieving high-precision and automated risk classification in variable environments, and significantly reducing the system's dependence on expert experience and frequent manual intervention.

[0021] (4) This invention constructs a complete model optimization closed loop that includes anonymized data collection, cloud-based incremental learning, A / B testing verification based on the Weighted Security Performance Index (WSPI), and OTA silent updates. This enables the system to continuously learn and evolve from the real operating environment while protecting user privacy. Anonymous features and labels generated by online inference of the local model drive incremental training in the cloud, while the WSPI index (which integrates alarm recognition accuracy, false alarm control, missed alarm avoidance, and performance stability) provides a scientific and comprehensive quantitative evaluation standard for model iteration. This solves the "aging" problem of equipment performance being fixed after leaving the factory and unable to adapt to long-term environmental changes and unknown scenarios, ensuring continuous performance optimization and reliable operation of the system throughout its entire life cycle.

[0022] (5) This invention integrates multi-source sensor data on gas, flame, human body, and valve status. After preprocessing including missing value imputation, outlier detection and correction, and moving average filtering, and feature extraction based on a fixed time window, the data is input into a support vector machine classification model for risk level identification and judgment. A multimodal linkage control strategy is executed based on the risk level, valve status signal, and human body presence signal. Anonymized feature data and corresponding event labels generated by the model are uploaded to the cloud for incremental learning and model iterative updates based on weighted safety performance indicators. This achieves multi-functional collaborative linkage and intelligent decision-making, dynamically optimizes alarm thresholds, and significantly improves the system's accuracy, environmental adaptability, and long-term self-evolution capability. Attached Figure Description

[0023] The following provides a detailed discussion of the manufacture and application of preferred embodiments of the present invention. However, it should be understood that the present invention provides many applicable inventive concepts that can be embodied in various specific environments. The specific embodiments discussed are merely illustrative of specific ways of manufacturing and using the present invention and do not limit the scope of the invention. For those skilled in the art, other drawings can be obtained from these drawings without any creative effort.

[0024] Figure 1 This is a flowchart of the present invention.

[0025] Figure 2 This is a schematic diagram of the structure of the present invention. Detailed Implementation

[0026] The following provides a detailed discussion of the manufacture and application of preferred embodiments of the present invention. However, it should be understood that the present invention provides many applicable inventive concepts that can be embodied in various specific environments. The specific embodiments discussed are merely illustrative of specific ways of manufacturing and using the invention and do not limit the scope of the invention.

[0027] like Figure 1 The method shown is an intelligent gas alarm method. Through continuous learning and self-optimization, this method can provide accurately matched alarm strategies for different environmental conditions, user habits, and installation locations. This intelligent gas alarm method includes the following steps: S1: Acquiring multi-source data output from the gas sensing module, flame monitoring module, human body sensing module, valve status sensor, and controller. x t The multi-source data x t It includes various real-world scenarios (such as different cooking modes, ambient temperature, and sensor installation locations), collecting raw, multi-source data throughout the entire process of safety, early warning, and alarm, forming a benchmark dataset for building the initial model.

[0028] S2: Preprocess the multi-source data, including time axis alignment, missing value imputation, outlier detection and correction, and moving average filtering, and label the preprocessed data with event tags.

[0029] The missing value imputation uses linear interpolation, and the formula is as follows: in, t It is discrete time; x t For the first t Multi-source data at different times. After data imputation, outlier detection is performed on the complete data sequence. This invention employs a statistical threshold-based method for identification, using predefined outlier monitoring parameters. z To determine multi-source data x t Is it abnormal? The formula for abnormal value detection is: in, z Parameters for outlier monitoring; μ The mean; σLet | be the standard deviation; if | z If |>3, then determine x t For outliers, continue to use linear interpolation to fill in the gaps and perform corrections and data recovery.

[0030] Then, a moving average filtering process is performed. The formula for the moving average filtering is: in, k The size of the sliding window; This is the filtered data; t It is discrete time; x i For the first i Multi-source data after outlier detection and correction; i For indexing. The time of safety status, warning status, and alarm status is manually recorded. The corresponding sensor data samples (including temperature, human presence status, etc.) are found by time alignment. Each data sample is labeled with an event tag of "safe status", "warning status", or "alarm status".

[0031] S3: Extract at least the average rate of temperature change from the preprocessed data using a fixed time window. Characteristic parameters; the average rate of temperature change The formula is: in, T ( n )and T (1) are the first of the fixed time windows. n The first and second temperature values; t ( n )and t (1) These represent the corresponding data collection times. The average temperature change rate is... This feature, by capturing the overall temperature change trend within a time window, can more fundamentally distinguish between the slow heating of normal cooking and the rapid heating patterns that pose risks such as dry burning, compared to simple instantaneous temperature values ​​or fixed thresholds. Experimental data shows that in a simulated kitchen environment, introducing... After being used as a feature, the model's accuracy in distinguishing between "normal cooking" and "abnormal dry burning" increased from 82% when only temperature thresholds were used to 95%, and the number of false alarms decreased by about 30%.

[0032] The characteristic parameters also include the average temperature. With temperature range R : in, n The total number of temperature data points within a fixed time window; i For indexing; For the first i Temperature values ​​of each temperature data point; To all within the window Take the maximum value; To all within the window Take the minimum value.

[0033] The characteristic parameters also include the temperature change rate. Maximum temperature change rate v max Duration of the absence event t : in, For the first i Temperature values ​​of each temperature data point; i For indexing; For the first i The acquisition time corresponding to each temperature value; maxv ( i () represents the rate of change of all temperatures within the window. Take the maximum value.

[0034] Ultimately, this step transforms the raw, unstructured sensor data into a well-structured, fixed-dimensional feature vector. x i , y i ), i ∈[1, n ], x i For example, the eigenvector can be represented as the eigenvalue of the eigenvector. i The average rate of change of temperature, the first i Features such as average temperature; y i ∈[-1, 1], where 0 represents a safe state, 1 represents a warning state, and -1 represents an alarm state, thus laying the foundation for subsequent model training and event classification.

[0035] S4: Input the extracted feature parameters and corresponding event labels into a pre-trained machine learning classification model to obtain the risk level. The machine learning classification model is a Support Vector Machine (SVM). SVM is favored for its advantages in solving small-sample, nonlinear, and high-dimensional pattern recognition problems. Other feasible candidate models include, but are not limited to, decision trees, random forests, or neural network models.

[0036] The Support Vector Machine (SVM) is used to solve a subconvex optimization problem: Among them, constraints ; w These are weight parameters; b For bias; For kernel function mapping; These are slack variables; C This is the penalty parameter.

[0037] The kernel function of the Support Vector Machine (SVM) is the radial basis function kernel: in, x i , x j The input feature vector; γ Kernel function parameters; eigenvectors x i and x j The Euclidean distance between them.

[0038] The Sequence Minimum Optimization (SMO) algorithm was used to obtain... w and b Decision function used to predict event labels for new samples f ( x This can be represented as: in, w These are weight parameters; b For bias; For kernel function mapping; sign To make linear functions The numerical output is a discrete time stamp.

[0039] The extracted feature parameters and corresponding event labels are then input into a pre-trained machine learning classification model for supervised offline training. Through multiple iterations, the internal parameters are optimized to obtain the risk level, ultimately enabling the model to reliably and accurately determine the event type from the feature input.

[0040] S5: Based on the risk level, valve status signal, and human presence status signal, execute a multimodal linkage control strategy to dynamically control the audible and visual alarm module and the valve status sensor and controller.

[0041] The multimodal linkage control strategy includes: when the flame monitoring module detects that the ambient temperature exceeds the dynamically adjusted warning threshold and the human body sensing module does not detect the presence of a human body, the audible and visual alarm module is controlled to enter the warning mode; if this state continues to exceed the dynamically adjusted time threshold, the preset time threshold in this embodiment is 120 seconds, then the audible and visual alarm module is controlled to upgrade to the alarm mode, and the valve status sensor and controller are controlled to close the valve.

[0042] The multimodal linkage control strategy further includes: when the gas sensing module detects that the target gas concentration exceeds a threshold, the threshold for the target gas concentration in this embodiment is 15% LEL. If the valve status sensor and controller feedback indicates that the valve is open, it immediately controls the valve to close and activates the warning mode of the audible and visual alarm module; if the valve is closed, based on the detection result of the human body sensing module, it selects to activate the alarm mode of the audible and visual alarm module or initiates a remote alarm notification.

[0043] S6: Upload the anonymized feature data and corresponding event labels generated by the model to the cloud for incremental learning and optimization updates of the model. A / B testing and iterative updates of the model before and after optimization are performed using the weighted security performance index WSPI.

[0044] This step aims to endow the model with self-evolution capabilities. First, after completing online inference, the model removes personally identifiable information (such as device ID and location), retaining only anonymized feature data and event labels. The anonymized feature data and corresponding event labels can still be used for incremental learning, updating the model to adapt to new scenarios and improve generalization ability, without infringing on privacy. Then, the cloud utilizes this incremental data from the real environment to periodically retrain and optimize the existing model.

[0045] Subsequently, to ensure the reliability and effectiveness of algorithm iteration, the cloud uses anonymized feature data and incremental data of corresponding event labels to optimize and train the original model, generating a new version of the model. Then, by deploying the new version of the model on a portion of devices and randomly dividing the terminal devices into two groups, Group A uses the original model and Group B uses the new model, prediction results and real event data are collected within the same time period, and the weighted security performance index WSPI value is calculated. The model with the higher WSPI value is determined to be the better model.

[0046] The formula for calculating WSPI is as follows: in, w i The weighting coefficients are the weighting factors. w i Specifically: w 1 = 0.4w 2 = 0.25 w 3 = 0.2 w 4 = 0.15; C i For performance components; i For indexing.

[0047] The performance components C i The formula is: in, C 1 represents the alarm recognition accuracy component; C 2 represents the false alarm control component; C 3 is to avoid underreporting; C 4 represents the component with stable performance; TP To correctly predict the number of positive examples, FP The number of incorrectly predicted positive examples. TN To correctly predict the number of negative examples, FN This represents the number of false negative predictions. In the actual iterative testing of this embodiment, the WSPI value of the new model reached 0.91, an improvement of 4.6% compared to the original model's 0.87, where the false positive control component... C The improvements are particularly significant. The new model, confirmed to have superior performance, was ultimately deployed silently to all terminal devices via OTA (Over-The-Air).

[0048] Based on the results of A / B testing, the model with the higher WSPI value is determined to be the optimized model with significant and stable performance advantages. The system then initiates a silent update across the entire network via OTA, officially deploying the optimized model to all devices, thus forming a complete algorithm loop that continuously improves itself and realizes iterative updates of the model.

[0049] like Figure 2 The illustrated intelligent gas alarm system includes a main control module, which serves as the system's control center, responsible for connecting and coordinating all other modules, performing data aggregation, logical judgment, and issuing control commands. A gas sensing module is used to monitor the concentration of specific gases (such as methane and carbon monoxide) in the environment in real time and transmit the concentration signal to the main control module. Its implementation methods include, but are not limited to, laser gas sensing, electrochemical sensing, or catalytic combustion sensing. A flame monitoring module is responsible for detecting open flames or abnormally high temperature signals in the environment, serving as an important criterion for fire risk assessment. Technologies that can be used include, but are not limited to, pyroelectric sensors, infrared array detectors, or thermistor sensors.

[0050] The human body sensing module uses infrared, sound, or millimeter wave technologies to detect the presence of human bodies within the vicinity of the alarm, providing a basis for alarm strategy formulation. Specific sensor types include, but are not limited to, infrared sensors, sound sensors, or millimeter wave sensors. A valve status sensor and controller is used to detect and control gas pipeline valves. This module integrates status detection and execution functions, detecting the open / closed state of gas pipeline valves and automatically closing the valve upon receiving instructions from the main control module during an alarm. An audible and visual alarm module can emit a high-decibel alarm sound and a strong flashing light according to instructions from the main control module, achieving on-site audible and visual warning. A network transmission module is responsible for wireless communication, enabling the alarm to connect to an IoT platform for remote alarm information push and user interaction functions.

[0051] The gas sensing module, flame monitoring module, human body sensing module, valve status sensing and controller, audible and visual alarm module, and network transmission module are all connected to the main control module. The main control module is used to execute the aforementioned intelligent gas alarm method. When the valve status sensing and controller detects that the valve is not closed, and the flame monitoring module confirms that there is no flame, the main control module immediately closes the gas valve and triggers the audible and visual alarm. If the valve is frequently not closed, the system will send a reminder to the user's pre-registered contact information via the network transmission module, suggesting that the valve be checked.

[0052] When the gas sensing module detects the target gas, the main control module initiates the linkage control logic: First, it checks the valve status. If the valve is not fully closed, it immediately closes the valve and activates the audible and visual alarm module to enter the warning mode. If the valve is closed, it further judges based on the detection results of the human body sensing module. If a person is detected, it activates the alarm mode to prompt the personnel on site to evacuate. If no one is present, it sends alarm information to the user through the network sending module and simultaneously performs remote alarm.

[0053] When the flame monitoring module detects that the ambient temperature exceeds the preset warning temperature, and the human body sensing module does not detect any human signal, the main control module controls the audible and visual alarm module to activate the warning mode. If the above-mentioned "high temperature and no one in the room" condition continues for more than a preset time threshold (in this embodiment, the preset time threshold is 120 seconds), the main control module upgrades the audible and visual alarm from the warning mode to the alarm mode and activates the valve status sensor and controller to close the gas valve. Once the flame monitoring module detects that the ambient temperature exceeds the preset alarm temperature (which is higher than the warning temperature), regardless of whether any personnel are detected, the main control module immediately performs an emergency operation: closing the gas valve and triggering the audible and visual alarm module to enter the alarm mode.

[0054] Although the specification has provided a detailed description, it should be understood that various changes, substitutions, and modifications can be made without departing from the spirit and scope of the invention as defined by the appended claims. Furthermore, the specific embodiments described are not intended to limit the scope of the invention, and those skilled in the art will readily understand based on this invention that existing or future-developed processes, machines, manufactures, compositions of matter, means, methods, or steps can perform substantially the same functions or achieve substantially the same results as the embodiments of the invention. Therefore, the appended claims are intended to include such processes, machines, manufactures, compositions of matter, means, methods, or steps within their scope.

Claims

1. A smart gas alarm method characterized by: The method comprises the following steps: S1: Acquiring multi-source data output by a gas sensing module, a flame monitoring module, a human sensing module, and a valve state sensing and controller x t ; S2: preprocessing the multi-source data, including time axis alignment, missing value filling, outlier detection and correction, and moving average filtering, and labeling the event label for the preprocessed data; the formula for outlier detection is: wherein, z is an outlier monitoring parameter; μ is a mean value; σ is a standard deviation; if z |>3, then it is determined that x t is an outlier and is corrected; S3: extracting at least a characteristic parameter comprising a temperature average change rate from the pre-processed data through a fixed time window; the temperature average change rate is calculated by the formula: ​ wherein, T ( n ) and T (1) are the first and the last temperature values of the fixed time window, respectively; n t ( n ) and t (1) are the corresponding acquisition times, respectively.​ S4: inputting the extracted feature parameters and the corresponding event label into a pre-trained machine learning classification model to obtain a risk level; S5: according to the risk level, the valve state signal and the human existence state signal, a multi-modal linkage control strategy is executed to dynamically control the sound and light alarm module, the valve state sensor and the controller; S6: uploading the anonymized feature data generated by the model running and the corresponding event label to the cloud for incremental learning and optimization update of the model, and verifying and updating the model through A / B testing of the weighted safety performance index WSPI; the calculation formula of WSPI is: wherein, w i is a weight coefficient; C i is a performance component; i is an index.

2. A smart gas alarming method as claimed in claim 1, wherein: The missing value filling in the S2 step adopts a linear interpolation method, and the formula is: wherein, t is a discrete time; x t is a first t moment of multi-source data.

3. A smart gas alarming method as claimed in claim 1, wherein: The moving average filtering in the S2 step has the formula: wherein, k is the size of the sliding window; is the filtered data; t is the discrete time; x i is the i-th multi-source data after outlier detection and correction; i is the i-th multi-source data after outlier detection and correction; i is the index.

4. A smart gas alarming method as claimed in claim 1, wherein: The feature parameters extracted in the S3 step also include temperature mean value With temperature range R : in, n The total number of temperature data points within a fixed time window; i For indexing; For the first i Temperature values ​​of each temperature data point; To all within the window T ( i Take the maximum value; To all within the window Take the minimum value.

5. A smart gas alarming method as claimed in claim 1, wherein: The machine learning classification model in the S4 step is a support vector machine, and the kernel function is a radial basis function kernel: wherein, x i , x j is the input feature vector; γ is the kernel function parameter; is the feature vector x i and x j the Euclidean distance between 6. A smart gas alarming method as claimed in claim 1, wherein: The multi-modal linkage control strategy includes: when the flame monitoring module detects that the environmental temperature exceeds the dynamically adjusted warning threshold and the human body perception module does not detect the presence of human body, the sound and light alarm module is controlled to enter the warning mode; if this state lasts for more than a dynamically adjusted time threshold, the sound and light alarm module is controlled to upgrade to the alarm mode, and the valve state sensor and controller are controlled to close the valve.

7. A smart gas alarm method as claimed in claim 6, wherein: The multi-modal linkage control strategy also includes: when the gas sensing module detects that the target gas concentration exceeds the threshold, if the valve state sensor and controller feedback that the valve is in an open state, the valve is immediately controlled to be closed and the warning mode of the sound and light alarm module is started; if the valve is in a closed state, according to the detection result of the human body perception module, the alarm mode of the sound and light alarm module is selected to be started or a remote alarm notification is initiated.

8. A smart gas alarming method as claimed in claim 1, wherein: The performance component in the S6 step C i The formula is: wherein, C 1 is an alarm recognition accuracy component; C 2 is a false alarm control component; C 3 is a false negative avoidance component; C 4 is a performance stability component; TP is the number of correct predictions as positive examples, FP is the number of incorrect predictions as positive examples, TN is the number of correct predictions as negative examples, FN is the number of incorrect predictions as negative examples.

9. A smart gas alarming method as claimed in claim 1, wherein: The weight coefficient w i Specifically: w 1 = 0.4, w 2 = 0.25, w 3 = 0.2, w 4 = 0.

15.

10. An intelligent gas alarm system characterized in that, It comprises: a master control module; a gas sensing module for monitoring gas concentration; a flame monitoring module for monitoring flame or abnormal temperature; a human body perception module for sensing human existence state; a valve state sensor and controller for detecting and controlling gas pipeline valve; a sound and light alarm module for issuing sound and light alarm; and a network sending module for data communication; The gas sensing module, flame monitoring module, human body perception module, valve state sensor and controller, sound and light alarm module and network sending module are connected with the master control module; the master control module is used to execute the intelligent gas alarm method as claimed in any one of claims 1 to 9.

Images