A remote monitoring and management system and method for combustible gases based on the Internet of Things and AI

Abstract

This invention discloses a remote monitoring and management system and method for combustible gas based on the Internet of Things and AI, relating to the field of combustible gas monitoring technology. The system includes a source heterogeneous sensing module, an edge AI inference module, a network communication transmission module, an AI collaborative decision-making module, and a judgment and collaborative response module that operate sequentially. Its key technical points are: This solution adopts an AI collaborative decision-making module driven by user behavior profiles and introduces multi-dimensional feature factors to achieve dynamic perception of abnormal gas usage behavior. At the same time, it completes a fallback by configuring over-limit protection items, ensuring to a certain extent that sudden leaks and other situations are not overlooked. It overcomes the shortcomings of traditional pure data-driven models that lack interpretability and real-time performance, and integrates and binds historical behavior context with real-time physical state to complete the collaborative optimization of safety and user experience.

Description

Technical Field

[0001] This invention relates to the field of combustible gas monitoring technology, specifically to a remote monitoring and management system and method for combustible gases based on the Internet of Things and AI. Background Technology

[0002] Combustible gas monitoring refers to the process of using sensors and related detection devices to sense, collect, or analyze the concentration or related physical parameters of combustible gases in order to determine whether there is a leak or abnormal release. In the existing technology, common combustible gas monitoring methods include catalytic combustion, semiconductor, infrared absorption, and electrochemical sensor detection. By converting the gas concentration into an electrical signal output, the qualitative or quantitative identification of combustible gases is achieved. The monitoring data is transmitted to a local controller or remote management platform through wired or wireless communication modules for status display or alarm prompts.

[0003] However, traditional combustible gas monitoring technologies mostly rely on a single type of gas sensor for concentration detection, lacking an understanding of the context of gas usage behavior. This makes them prone to false alarms due to cooking fumes or sensor aging, such as triggering a leak alarm during stir-frying in the kitchen, or missing a minor leak because the concentration does not reach the threshold. Existing systems typically use fixed alarm thresholds, failing to distinguish between scenarios like late-night soup cooking and abnormal gas usage in the early morning, leading to frequent false alarms. For special cases such as ignition failure with continuous gas release, there is a certain probability that the process may be misjudged as normal ignition due to the presence of users and the small flow rate, creating a safety blind spot. Currently, there are no in-depth regulatory technology optimizations for this. In addition, most devices only upload raw data to the cloud for processing, resulting in frequent communication and high power consumption. Even if some products introduce simple timed valve shut-off logic, it is difficult to dynamically adapt to different users' gas usage habits, affecting safety and reducing user acceptance to some extent. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A remote monitoring and management system for combustible gases based on the Internet of Things and AI includes:

[0006] Multi-source heterogeneous sensing module: Deploy a sensor array in the selected area to be monitored to acquire raw data;

[0007] Edge AI Inference Module: An MCU + lightweight neural network model is set up in the monitoring terminal to perform extraction and judgment actions based on raw data, including at least feature extraction actions and multimodal fusion judgment actions;

[0008] Network communication transmission module: Based on the results of the extracted judgment actions, it runs the pre-built local private network communication and connects to the cloud public network for verification to build a data lake;

[0009] AI Collaborative Decision Module: Establishes a user behavior profile model, outputs a basic deviation index, and introduces multi-dimensional feature factors, which are then fused with the basic deviation index to obtain the final deviation index, in order to identify abnormal gas usage behavior.

[0010] The collaborative response module determines the risk level and response action based on a pre-built rule engine. When the igniter is abnormal, a dual-feedback special identification mechanism is triggered. The edge layer multi-dimensional feature coupling detection is used for the first feedback, and the cloud collaborative layer is combined to start the second feedback verification to obtain the corrected risk level and response action.

[0011] Furthermore, the sensor array includes at least an ultrasonic flow sensor, a digital pressure sensor, an infrared array sensor, and an ambient gas composite sensor; the acquired raw data includes at least: instantaneous flow rate Q, gas pressure P, stove surface temperature distribution T, human presence status human_absent, and cooking time cooking_time.

[0012] Furthermore, the feature extraction process involves extracting the flow rate fluctuation σ_Q, pressure decay rate dP / dt, and maximum surface temperature max(T) from the original data; where the flow rate fluctuation σ_Q = std(Q t Q t-1 Q t-n ), where t is the sequence number of the current sampling time, and n is the number of sampling points to be backtracked;

[0013] The process of performing multimodal fusion judgment is as follows: Boolean logic is combined with threshold comparison to form two types of criteria for target risk scenarios. The first type is the gas micro-leakage criterion, with the condition that σ_Q > the threshold Qth corresponding to the normal gas noise level and dP / dt < -0.1. The second type is the unattended dry burning criterion, with the condition that max(T) > the maximum temperature threshold Tth corresponding to the normal range of smoke point and human_absent=True and cooking_time > the duration benchmark threshold Yth. If any of the above conditions are met, the corresponding type of alarm is triggered and reported.

[0014] Furthermore, the local private domain network communication performs a status determination based on the results of the extraction and judgment actions: if there is an alarm of the corresponding type, it indicates an abnormal status and triggers a priority alarm packet; otherwise, it indicates a normal status and reports the sub-results of the feature extraction actions at preset intervals; wherein, the priority alarm packet contains at least 5 seconds of raw data segments and a predefined enumeration code for the alarm event; the sub-results of the feature extraction actions are to extract the flow fluctuation rate σ_Q, pressure decay rate dP / dt, and maximum stove surface temperature max(T) from the raw data.

[0015] Furthermore, the user behavior profiling model operates as follows:

[0016] Gas usage habit learning: Based on any user ID, statistically analyze the daily average gas usage period distribution H(t), historical gas flow rate and current instantaneous flow rate Q, and commonly used flow rate intervals {Q_min, Q_max};

[0017] Defining the deviation index: Step 1: Use the absolute difference between the current flow rate and the historical average gas flow rate as the numerator, and the standard deviation of the historical gas flow rate as the denominator to obtain the deviation multiple; Step 2: Count the number of days in which gas consumption events have occurred within the historical preset period, denoted as N_use(t±Δt); t represents the current time, Δt represents the time tolerance window, and dividing N_use(t±Δt) by the historical preset period gives the time period sparsity coefficient R(t); Step 3: Subtract the time period sparsity coefficient R(t) from 1 and perform a weighted calculation with the deviation multiple to obtain the basic deviation index D_base.

[0018] Furthermore, the introduced multi-dimensional feature factors include at least: flow stability factor M(S) and duration factor A(T_dur);

[0019] The derivation process of the flow stability factor M(S) is as follows: First step: Take the sampled values ​​of all instantaneous flow rates within w seconds before the current time t, and denote them as the flow rate sequence from the (t-w+1)th sampling point to the tth sampling point, and calculate the standard deviation of the flow rate sequence; Second step: Calculate the average flow rate of the w sampling points as the denominator, and the standard deviation of the flow rate sequence as the numerator. The resulting ratio is the stability index S; Third step: Introduce an amplification factor and multiply it by the stability index S, accumulate the product with 1, and invert the accumulated result to derive the flow stability factor M(S).

[0020] The derivation process of the duration factor A(T_dur) is as follows: Step 1: Perform preliminary anomaly analysis based on the basic deviation index. Compare the basic deviation index D_base with the preset traditional threshold. If D_base exceeds the traditional threshold, it indicates that there is an anomaly. Calculate the duration T_dur of the current anomaly. Step 2: Set constraints. When T_dur is less than 2 minutes, determine that A(T_dur) = 0.

[0021] When T_dur is between 2 and 5 min, determine A(T_dur) = (T_dur-2) / 3;

[0022] When T_dur exceeds 5 minutes, A(T_dur) is determined to be 1.

[0023] Furthermore, the process of merging the basic deviation index to obtain the final deviation index is as follows: The basic deviation index D_base is multiplied sequentially by the introduced multi-dimensional feature factors to obtain one result; an indicator function I(Q>Q_max_hist) is set, which returns 1 if the current instantaneous flow Q exceeds the maximum flow Q_max_hist in the user's historical records, otherwise it returns 0; the result of the indicator function I(Q>Q_max_hist) is multiplied by the set over-limit protection item to obtain two results; finally, the one result and the two results are summed to obtain the final deviation index D_fin; the basis for identifying abnormal gas usage behavior is: when D_fin exceeds the preset standard threshold and the duration exceeds the critical duration, abnormal gas usage behavior is determined, and a warning signal is issued; otherwise, no response action is taken.

[0024] Furthermore, the rules engine is configured as follows: when the current instantaneous flow rate Q is within the range of [0.05, 0.2], the risk level is determined to be L1, which is low risk, and the response actions are: APP push and voice prompt; when the current instantaneous flow rate Q exceeds 0.2 or is dry-burning, the risk level is determined to be L2, which is medium risk, and the response actions are: automatic valve closure, calling the corresponding user, and notifying the property management; when the pressure decay rate dP / dt exceeds 1 kPa / min, the risk level is determined to be L3, which is high risk, and the response actions are: immediate valve closure and power cut-off, and linkage with the fire protection platform.

[0025] Furthermore, edge layer multidimensional feature coupling detection is applied to provide feedback: the following judgment logic is added to the monitoring terminal: if the condition corresponding to the first type of gas micro-leakage criterion is met and the duration exceeds 5 seconds, a local suspected ignition failure leak warning is triggered, and the action is to start the buzzer prompt sound and simultaneously send a marked feature packet to the cloud, which includes Q and T sequences, timestamps and predefined enumeration codes.

[0026] In conjunction with the cloud-based collaborative layer, a secondary feedback verification is initiated: For this user ID, the distribution of successful ignition times over the past 7 days is statistically analyzed. If the difference between the historical average ignition time and the current ignition time exceeds a defined threshold, it is recorded as a discrepancy event. It is then determined whether the event is repeated. If the same user experiences more than one discrepancy event within 24 hours, the original risk level is upgraded by one level. The specific number of times the same user experiences more than one discrepancy event within 24 hours is counted. If the specific number is 2, the original response action is maintained. If the specific number exceeds 2, the response action corresponding to the high risk is executed.

[0027] A remote monitoring and management method for combustible gases based on the Internet of Things and AI includes the following steps:

[0028] Deploy a sensor array in the selected area to be monitored to acquire raw data;

[0029] An MCU + lightweight neural network model is set up in the monitoring terminal to perform extraction and judgment actions based on the raw data, including at least feature extraction actions and multimodal fusion judgment actions;

[0030] Based on the results of the extraction and judgment actions, the pre-built local private domain network communication is run and connected to the cloud public network for verification, thus building a data lake;

[0031] Establish a user behavior profile model, output a basic deviation index, and introduce multi-dimensional feature factors to merge with the basic deviation index to obtain the final deviation index, so as to identify abnormal gas usage behavior.

[0032] Based on a pre-built rule engine, the risk level and response action are obtained. Under the condition that the igniter is abnormal, a dual feedback special identification mechanism is triggered. The edge layer multi-dimensional feature coupling detection is used for the first feedback, and the cloud collaboration layer is combined to start the second feedback verification to obtain the corrected risk level and response action.

[0033] This invention provides a remote monitoring and management system and method for combustible gases based on the Internet of Things and AI, which has the following beneficial effects:

[0034] (1) This solution adopts a multi-source heterogeneous sensing module and combines multi-modal fusion judgment logic. It not only achieves high-sensitivity local initial screening for high-risk scenarios such as small gas leaks and unattended dry burning, but also reduces the amount or frequency of traditional communication data. It solves the problem that traditional single-type sensors are easily affected by environmental interference, resulting in false alarms / missed alarms. By utilizing the synergistic mechanism of front-end sensing and edge filtering, it ensures the accuracy of local identification while reducing invalid cloud reporting, thereby improving the overall system operation effect and user experience.

[0035] (2) This solution adopts an AI collaborative decision-making module driven by user behavior profiles and introduces multi-dimensional feature factors to achieve dynamic perception of abnormal gas usage behavior. At the same time, it also completes the fallback by configuring over-limit protection items, which to a certain extent ensures that sudden leaks and other situations are not overlooked. It overcomes the shortcomings of traditional pure data-driven models that lack interpretability and real-time performance, and integrates and binds historical behavior context with real-time physical state to complete the collaborative optimization of safety and experience.

[0036] (3) This solution uses a judgment and collaborative response module to realize primary and secondary feedback, accurately lock the special scenario of igniter failure leakage, and has a high degree of scenario specificity. It is activated only under set conditions and does not interfere with other functions. At the same time, it balances safety and user experience, reuses sensor arrays, and thus achieves seamless integration with the existing architecture. It not only solves the blind spots of some scenarios not covered in the traditional monitoring system, but also reflects the deep and effective application of the combination of IoT and AI in subdivided risk scenarios. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the system module operation in this invention;

[0038] Figure 2 This is a schematic diagram of the overall process of the method in this invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0040] Example 1:

[0041] Please see Figure 1 This embodiment provides a remote monitoring and management system for combustible gases based on the Internet of Things (IoT) and AI. It constructs a high-precision, low-false-alarm remote monitoring and management system for combustible gases in urban residential areas, catering establishments, and industrial edge scenarios. It is important to note that the core of this solution is to address the special scenario of high risk, low frequency, but serious consequences caused by the failure of traditional igniters to ignite the gas but with partial leakage. By optimizing and improving the current mainstream solutions, a dual-form feedback mechanism is introduced to construct a special protection system with scenario specificity and high sensitivity. This demonstrates the value of in-depth governance in specific scenarios through IoT + AI in this solution.

[0042] The specific details of this system are as follows:

[0043] Multi-source heterogeneous sensing module: Deploy a sensor array in the selected area to be monitored to acquire raw data;

[0044] The areas to be monitored include the kitchen, gas pipeline wells, and the vicinity of pressure regulating boxes. The areas can be selected based on actual needs or historical configuration locations. The sensor array includes ultrasonic flow sensors, digital pressure sensors, infrared array sensors, and ambient gas composite sensors.

[0045] In this embodiment, the ultrasonic flow sensor can be installed inside the ultrasonic channel of the gas safety valve. The instantaneous flow rate Q is obtained by measuring the time difference between forward and reverse propagation through transducers A and B. The explanation for obtaining the instantaneous flow rate Q is as follows: Ultrasonic waves propagate at a fixed speed in a stationary gas. However, when the gas is flowing, forward propagation is faster, meaning the time t_down required for the ultrasonic wave to travel from transducer A to B along the airflow direction is shorter. Reverse propagation is slower, meaning the time t_up required for the ultrasonic wave to travel from transducer B to A against the airflow direction is slower. The flow rate is relatively large; the gas velocity can be calculated by measuring the propagation time of the two. Specifically, the difference between 1 / t_down and 1 / t_up is first calculated. This difference is proportional to the component of the flow velocity in the direction of the sound channel. Then, it is multiplied by L / (2cosθ) to convert it into the axial flow velocity of the pipe. Here, L represents the length of the sound channel in the pipe, that is, the straight-line distance between the two transducers, and θ represents the angle between the ultrasonic sound channel and the direction of gas flow. Finally, multiplying by the calibration coefficient k gives the instantaneous flow rate Q. Here, k is obtained by experimental calibration based on the sensor structure, gas temperature, pressure, etc., and is used to correct the deviation between the theoretical model and the actual flow field. The value range is >0. This method is based on the time-of-flight difference principle, does not depend on the gas composition, and is suitable for long-term stable monitoring of small leaks.

[0046] The digital pressure sensor can be a high-precision digital pressure sensor, which collects the inlet / outlet pressure of combustible gas in real time, and records them as P_in and P_out respectively. It is used to detect abnormal pressure drop events such as hose detachment or loose connection of combustible gas, and records the gas pressure value P after the valve is closed. The infrared array sensor is installed in the kitchen ceiling, but it is ensured that it is not directly under the stove to avoid oil contamination. It acquires the stove surface temperature distribution T at N points in real time, and also has a built-in low-power human presence detection submodule using infrared pyroelectric or millimeter-wave radar to determine whether there is anyone in the kitchen. Therefore, the entire infrared array sensor can be used to detect risks such as dry burning and unattended cooking. The value of N is a positive integer greater than 0. In this embodiment, the ambient gas composite sensor can be a low-power MEMS CO / CH4 sensor. It is only used as an auxiliary verification in this embodiment and is not used as the main criterion to avoid misjudgment due to lifespan decay. At the same time, the ambient gas composite sensor has an embedded timer to record the duration of combustible gas use, cooking_time.

[0047] Therefore, the raw data obtained includes: instantaneous flow rate Q, air pressure value P, stove surface temperature distribution T, human presence status human_absent, and cooking time, etc.

[0048] Edge AI Inference Module: An MCU + lightweight neural network model is set up in the monitoring terminal to perform extraction and judgment actions based on raw data, including feature extraction actions and multimodal fusion judgment actions;

[0049] In this embodiment, a monitoring terminal or a gas safety valve can be used, whichever is more appropriate to the specific requirements.

[0050] The process of performing feature extraction is as follows: The flow fluctuation rate σ_Q, pressure decay rate dP / dt, and maximum surface temperature max(T) are extracted from the raw data; where the flow fluctuation rate σ_Q = std(Q t Q t-1 Q t-n ); where t represents the sequence number of the current sampling time, and n represents the number of sampling points to be backtracked.

[0051] The process of performing multimodal fusion judgment is as follows: Boolean logic is combined with threshold comparison to form two types of criteria for target risk scenarios; the first type is the gas micro-leakage criterion, with the condition that σ_Q>Qth and dP / dt<-0.1, in kPa / min; the second type is the unattended dry burning criterion, with the condition that max(T)>Tth and human_absent=True and cooking_time>Yth; as long as any of the above conditions are met, the corresponding type of alarm is triggered and reported.

[0052] Wherein, Qth represents the threshold corresponding to the normal gas noise level, which can be obtained by averaging historical data; Tth represents the maximum temperature threshold corresponding to the normal range of smoke point. For example, when the liquid in the pot boils dry, the temperature at the bottom of the pot will quickly rise to over 300℃. Therefore, in this embodiment, 250℃ can be set as the maximum temperature threshold to cover most of the initial stage of dry burning, while avoiding misjudgment under normal frying conditions of about 180℃ to 220℃; Yth represents the duration threshold, with a value greater than 0. In this embodiment, a value of 30 minutes can be used. When the conditions corresponding to the first type are met, a leak alarm is triggered; when the conditions corresponding to the second type are met, a dry burning alarm is triggered; if similar alarms are generated subsequently, the subsequent judgment process shall prevail.

[0053] Specifically, the execution code for the multimodal fusion determination action is as follows:

[0054] if (σ_Q>tQth) and (dP / dt<-0.1 kPa / min):

[0055] trigger_leak_alert()

[0056] elif (max(T)>250℃) and (human_absent=True) and (cooking_time>30min):

[0057] trigger_dry_burn_alert();

[0058] It is worth noting that the two types of actions performed in this solution employ a combination of rule engines and conditions, unlike traditional complex models. For example, traditional deep neural networks typically require hundreds of KB of parameters, far exceeding the available Flash memory of an MCU. This solution, however, only needs to store a few thresholds and a small number of state variables reflecting the original data, resulting in a significantly lower total memory footprint than the traditional hundreds of KB. Furthermore, all comparison operations can be performed using integers or Q-format fixed-point numbers, eliminating the need for a floating-point coprocessor and reducing hardware costs. The entire judgment process involves only related operations, and their execution time on the MCU meets the low-latency requirement. In summary, the system running this solution can achieve over 98% local initial screening accuracy even under extreme resource constraints, reducing invalid cloud reporting, saving communication costs, and preventing users from being frequently bothered by false alarms.

[0059] This solution employs a multi-source heterogeneous sensing module, combined with multi-modal fusion judgment logic. On one hand, it achieves highly sensitive local initial screening for high-risk scenarios such as minor gas leaks and unattended dry burning. On the other hand, it reduces the amount or frequency of communication data by uploading only feature summaries or alarm fragments. To a certain extent, this solves the problem of false alarms / missed alarms caused by environmental interference, which is easily affected by traditional single-type sensors. It avoids the bottleneck of complex deep models being unable to be deployed on resource-constrained MCUs. By utilizing the synergistic mechanism of front-end sensing and edge filtering, it can ensure local recognition accuracy while reducing invalid cloud reporting, thereby improving the overall system performance and user experience.

[0060] Network communication transmission module: Based on the results of the extracted judgment actions, it runs the pre-built local private network communication and connects to the cloud public network for verification to build a data lake;

[0061] In this embodiment, the local private domain network communication performs a status determination based on the results of the extraction and judgment action: if a corresponding type of alarm exists, it indicates an abnormal status and triggers a priority alarm packet; otherwise, it indicates a normal status and reports the sub-results of the feature extraction action at a preset interval. In this embodiment, the local private domain network communication can use LoRa / NB-IoT, and the triggered priority alarm packet contains at least 5 seconds of raw data segments and a predefined enumeration code for the alarm event, such as: 0x01 = leakage alarm, 0x02 = dry burning alarm. Under normal status conditions, the preset interval duration can be configured to 5 minutes in this embodiment. The sub-results of the feature extraction action are the extraction of the flow fluctuation rate σ_Q, pressure decay rate dP / dt, and maximum stove surface temperature max(T) from the raw data.

[0062] Access to the cloud public network for verification corresponds to a two-level confirmation mechanism, and the data lake it constructs includes:

[0063] The system includes a multi-tenant data isolation set, a time-series database storage set, and a terminal APP integration set. The multi-tenant data isolation set is divided into data partitions by community / building. The time-series database storage set retains high-frequency raw data for one week and stores aggregated data permanently. High frequency refers to raw data that exceeds a preset frequency threshold. The terminal APP integration set supports WeChat mini programs and native iOS / Android applications.

[0064] AI Collaborative Decision Module: Establishes a user behavior profile model, outputs a basic deviation index, and introduces multi-dimensional feature factors, which are then fused with the basic deviation index to obtain the final deviation index, in order to identify abnormal gas usage behavior.

[0065] Step 1, the user behavior profiling model operates as follows:

[0066] Gas usage habit learning: Based on any user ID, statistically analyze the daily average gas usage period distribution H(t), historical gas flow rate and current instantaneous flow rate Q, and commonly used flow rate intervals {Q_min, Q_max};

[0067] Defining the deviation index: Step 1 (same as the traditional method): Use the absolute difference between the current flow rate and the historical average gas flow rate as the numerator, and the standard deviation of the historical gas flow rate as the denominator to obtain the dimensionless deviation multiple; Step 2 (replacing the second step of the traditional method): Count how many days in the historical preset period have gas consumption events occurring within the time period (t±Δt), denoted as N_use(t±Δt); t represents the current time, Δt represents the time tolerance window, which can be set to 15 minutes in this embodiment. Divide N_use(t±Δt) by the historical preset period to obtain the time period sparsity coefficient R(t); Step 3: Subtract the time period sparsity coefficient R(t) from 1 and perform a weighted calculation with the deviation multiple to obtain the basic deviation index D_base; In actual use, a very small positive number, such as 0.001, is added to the denominator in Step 1 to prevent the denominator from being 0.

[0068] The traditional definition of the deviation index is as follows:

[0069] Step 1: Use the absolute difference between the current gas flow rate and the historical average gas flow rate as the numerator, and the standard deviation of the historical gas flow rate as the denominator to obtain the dimensionless deviation multiple. Step 2: Introduce a judgment function: if the current time t is not within the set H_active of the user's historically frequent gas usage time periods, return 1; otherwise, return 0. H_active includes, for example, 7 AM to 9 AM and 6 PM to 8 PM. Step 3: Weight the deviation multiple with the result of the judgment function to obtain the original deviation index D_orig.

[0070] Traditional gas usage identification schemes only consider deriving the original deviation index D_orig and comparing it with the corresponding threshold to identify or trigger an alarm signal for abnormal gas usage. This traditional approach has limitations. It only considers instantaneous flow deviation, ignoring whether the flow is stable (e.g., a slight leak is a continuous low flow, while stir-frying is a sudden high flow); it also doesn't consider the duration of the event (e.g., a brief malfunction during stove adjustment shouldn't trigger an alarm); and it doesn't utilize the sparsity of historical events (if gas has never been used during a certain period, even if the flow is normal, it should be considered high-risk). Therefore, to address these limitations, this embodiment further designs a subsequent scheme to solve and optimize these issues.

[0071] Step 2, introduce multi-dimensional feature factors including: flow stability factor M(S) and duration factor A(T_dur);

[0072] The derivation of the flow stability factor M(S) is as follows:

[0073] Step 1: Take the sampled values ​​of all instantaneous flow rates within w seconds prior to the current time t, denoted as the flow rate sequence from the (t-w+1)th sampling point to the tth sampling point, and calculate the standard deviation of this flow rate sequence; where, the standard deviation is a statistical measure of the deviation between the sampled values ​​of the instantaneous flow rate and its average value, and w is usually taken as 1 min, i.e., 60 s in this embodiment; Step 2: Calculate the average flow rate of the w sampling points as the denominator, and the standard deviation of the flow rate sequence as the numerator. The resulting ratio is the dimensionless stability index S; if S is close to 0, such as small If S is less than 0.1, it indicates that the flow rate is almost constant, which may be a minor leak; however, if S is greater than 0.5, it indicates that the flow rate changes frequently, which is more likely to be normal cooking behavior; the third step is to introduce an amplification factor and multiply it by the stability index S, accumulate the product with 1, and invert the result to derive the flow stability factor M(S); where, in this embodiment, the amplification factor takes a value greater than 0; its mechanism is: when S is very small, M(S) is close to 1, which does not weaken the abnormal score; when S is very large, M(S) is significantly less than 1, which greatly reduces the abnormal score and avoids false alarms.

[0074] The derivation of the duration factor A(T_dur) is as follows:

[0075] Step 1: Perform preliminary anomaly analysis based on the baseline deviation index. Following the traditional method, compare the baseline deviation index D_base (equivalent to the original deviation index D_orig) with a preset traditional threshold. If D_base exceeds the traditional threshold, an anomaly is indicated. This initial anomaly determination is used to calculate the duration T_dur (in minutes) of the current anomaly. Step 2: Set constraints. If the duration T_dur of the anomaly is less than 2 minutes, then determine A(T_dur) as an anomaly. If the duration of the anomaly is between 2 and 5 minutes, then A(T_dur) = (T_dur - 2) / 3, using a linear growth form. For example, A(T_dur) = 1 / 3 after 3 minutes and A(T_dur) = 2 / 3 after 4 minutes. If the duration of the anomaly is more than 5 minutes, then A(T_dur) = 1, indicating full activation. The logic is based on using a piecewise linear function to ensure that short-term anomalies are not triggered.

[0076] Specifically, this solution employs an AI collaborative decision-making module driven by user behavior profiles and introduces multi-dimensional feature factors. This not only enables dynamic perception of abnormal gas usage behavior but also provides a safety net by configuring over-limit protection items. To a certain extent, it ensures that sudden leaks are not overlooked, overcoming the shortcomings of traditional pure data-driven models that lack interpretability and real-time performance. By fusing and binding historical behavioral context with real-time physical states, the overall false alarm rate at night is significantly reduced, and the micro-leak detection rate is improved, achieving a collaborative optimization of safety and user experience.

[0077] Step 3, the process of merging the basic deviation index to obtain the final deviation index is based on the following:

[0078] First, the basic deviation index D_base obtained in Step 1 is multiplied sequentially by the introduced multi-dimensional feature factors, i.e., D_base, M(S), and A(T_dur) are multiplied to obtain a result. Second, an indicator function I(Q>Q_max_hist) is set. If the current instantaneous flow Q exceeds the maximum flow Q_max_hist in the user's historical records, it returns 1; otherwise, it returns 0. The result of the indicator function I(Q>Q_max_hist) is then multiplied by the set over-limit protection item to obtain two results. This over-limit... In this embodiment, the protection item can be set to 1; finally, the result of one item and the result of two items are summed to obtain the final deviation index D_fin; this design realizes a dual protection mechanism, namely, routine anomaly detection: finely identify small and continuous anomalies, such as micro-leakage or gas usage at night, through one result; emergency risk backup: once the flow rate far exceeds the historical level, even if other conditions are not met, a high-risk alarm can be triggered immediately through the over-limit item according to the actual situation to avoid missed reports. By adopting this dual protection mechanism in the form of a fusion strategy, sensitivity and reliability can be effectively balanced.

[0079] Step 4: The basis for identifying abnormal gas usage behavior is as follows: when the final deviation index D_fin exceeds the preset standard threshold and the duration exceeds the critical duration, abnormal gas usage behavior is determined, and an early warning signal is issued; otherwise, no response is taken. Regarding the standard threshold, it is dynamically set through an adaptive threshold adjustment algorithm: the system statistically analyzes the distribution of D_fin based on the user's historical normal gas usage data, such as the 95th percentile, and uses this as the initial threshold; subsequently, combined with false alarm / missed alarm feedback, the threshold is fine-tuned online using the exponentially weighted moving average (EWMA), achieving personalized adaptation while ensuring safety. The setting of this standard threshold will not be elaborated further here. The critical duration in this embodiment can be set to 3 minutes, which can be adjusted according to actual needs.

[0080] The following table illustrates the effects of the AI ​​collaborative decision-making module before optimization (compared to the explanation of the traditional definition of deviation index in Step 1 above) and after optimization (compared to the solution proposed for the AI ​​collaborative decision-making module in this embodiment):

[0081] Table 1: Comparison of feedback results for some indicators before and after optimization:

[0082]

[0083] As can be seen from Table 1 above, the technical solution in this embodiment significantly improves the accuracy, robustness, and scenario adaptability of abnormal gas usage identification while maintaining lightweight computation.

[0084] Determine the collaborative response module: Based on a pre-built rule engine, obtain the risk level and response action. Under the condition that there is an abnormality in the flammable gas igniter, trigger the dual feedback special identification mechanism. Apply edge layer multi-dimensional feature coupling detection to perform the first feedback, and combine the cloud collaborative layer to start the second feedback verification to obtain the corrected risk level and response action.

[0085] The settings in the rules engine are explained below:

[0086] When the current instantaneous flow rate Q is within the range of [0.05, 0.2], the risk level is determined to be L1, which is low risk. The response actions are: APP push and voice prompt, with the prompt content being: "It is recommended to check the hose." When the current instantaneous flow rate Q exceeds 0.2 or is dry-burning, the risk level is determined to be L2, which is medium risk. The response actions are: automatic valve closure, calling the corresponding user, and notifying property management, etc. When the pressure decay rate dP / dt exceeds 1 kPa / min, the risk level is determined to be L3, which is high risk. The response actions are: immediately closing the valve and cutting off the power, and linking the fire protection platform.

[0087] The following conditions indicate an abnormality in the equipment equipped with a combustible gas igniter:

[0088] Ignition malfunction causing continuous ignition failure but gas leakage occurs in the following scenario: The user presses the stove knob and rotates it to the ignition position. Due to aging, moisture, or electrode carbon buildup, the igniter repeatedly fails to ignite, such as making a clicking sound for more than 5 seconds. However, the gas valve is open, resulting in continuous gas release without a flame. At this time, the ultrasonic flow sensor detects a stable low flow rate. Because there is no combustion, the infrared temperature does not rise significantly, and a human body is present. Therefore, traditional systems may easily misinterpret this as a normal ignition process or low-flame gas usage, resulting in a very high risk of false alarm.

[0089] A special dual-feedback identification mechanism was developed to address the unique circumstances in this scenario.

[0090] Therefore, the triggered dual-feedback special identification mechanism process is based on the following: First, a feedback is performed using multi-dimensional feature coupling detection at the edge layer. The following judgment logic is added to the monitoring terminal: If the condition corresponding to the first type of minor gas leak criterion is met, and the duration exceeds 5 seconds, a suspected ignition failure leak warning is triggered locally. The actions performed are: activating the buzzer sound and simultaneously sending a tagged feature packet to the cloud, containing Q and T sequences (corresponding to the stove surface temperature distribution T, timestamp, and predefined enumeration code); Second, a secondary feedback verification is initiated in conjunction with the cloud collaboration layer: First, the historical ignition success rate is queried, and for this user ID, statistics are... The distribution of successful ignition times over the past 7 days is analyzed. If the difference between the historical average ignition time and the current ignition time exceeds a defined threshold, it is recorded as a discrepancy event. Next, it is determined whether the event is recurring. If the same user experiences more than one discrepancy event within 24 hours, the original risk level is upgraded by one level; for example, if the original risk level is L1, it is upgraded to L2. Finally, the specific number of times the same user experiences more than one discrepancy event within 24 hours is counted. If the specific number is 2, the original response action is maintained; if the specific number exceeds 2, the response action corresponding to the high risk is executed. The above constitutes the corrected risk level and response action.

[0091] This solution uses primary and secondary feedback to accurately pinpoint the specific scenario of igniter malfunction-induced leakage, which occurs when the igniter equipped with the combustible gas is malfunctioning. It exhibits high scenario specificity, activating only under set conditions without interfering with other functions. At the same time, it balances safety and user experience by reusing the sensor array, thus achieving seamless integration with the existing architecture. It not only solves some blind spots not covered by traditional monitoring systems, but also demonstrates the deep and effective application of the combination of IoT and AI in segmented risk scenarios.

[0092] Example 2:

[0093] Based on Example 1, referring to Figure 2 As shown, this embodiment also provides a remote monitoring and management method for combustible gases based on the Internet of Things and AI, including the following steps:

[0094] Deploy a sensor array in the selected area to be monitored to acquire raw data;

[0095] An MCU + lightweight neural network model is set up in the monitoring terminal to perform extraction and judgment actions based on the raw data, including at least feature extraction actions and multimodal fusion judgment actions;

[0096] Based on the results of the extraction and judgment actions, the pre-built local private domain network communication is run and connected to the cloud public network for verification, thus building a data lake;

[0097] Establish a user behavior profile model, output a basic deviation index, and introduce multi-dimensional feature factors to merge with the basic deviation index to obtain the final deviation index, so as to identify abnormal gas usage behavior.

[0098] Based on a pre-built rule engine, the risk level and response action are obtained. Under the condition that the igniter is abnormal, a dual feedback special identification mechanism is triggered. The edge layer multi-dimensional feature coupling detection is used for the first feedback, and the cloud collaboration layer is combined to start the second feedback verification to obtain the corrected risk level and response action.

[0099] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0100] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0101] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A remote monitoring and management system for combustible gases based on the Internet of Things and AI, characterized in that, include: Multi-source heterogeneous sensing module: Deploy a sensor array in the selected area to be monitored to acquire raw data; Edge AI Inference Module: An MCU + lightweight neural network model is set up in the monitoring terminal to perform extraction and judgment actions based on raw data, including at least feature extraction actions and multimodal fusion judgment actions; Network communication transmission module: Based on the results of the extracted judgment actions, it runs the pre-built local private network communication and connects to the cloud public network for verification to build a data lake; AI Collaborative Decision Module: Establishes a user behavior profile model, outputs a basic deviation index, and introduces multi-dimensional feature factors, which are then fused with the basic deviation index to obtain the final deviation index, in order to identify abnormal gas usage behavior. The user behavior profiling model operates as follows: Gas usage habit learning: Based on any user ID, statistically analyze the daily average gas usage period distribution H(t), historical gas flow rate and current instantaneous flow rate Q, and commonly used flow rate intervals {Q_min, Q_max}; Defining the deviation index: Step 1: Take the absolute difference between the current flow rate and the historical average gas flow rate as the numerator and the standard deviation of the historical gas flow rate as the denominator to obtain the deviation multiple; Step 2: Count the number of days that gas usage events occurred within the preset historical period, and record it as N_use(t±Δt); t represents the current time, Δt represents the time tolerance window, and the time period sparsity coefficient R(t) can be obtained by dividing N_use(t±Δt) by the historical preset period. Step 3: Subtract the sparsity coefficient R(t) from 1 and calculate the weighted average of the deviation multiple to obtain the basic deviation index D_base; The introduced multi-dimensional feature factors include at least: flow stability factor M(S) and duration factor A(T_dur); The derivation process of the flow stability factor M(S) is as follows: First step: Take the sampled values ​​of all instantaneous flow rates within w seconds before the current time t, and denote them as the flow rate sequence from the (t-w+1)th sampling point to the tth sampling point, and calculate the standard deviation of the flow rate sequence; Second step: Calculate the average flow rate of the w sampling points as the denominator, and the standard deviation of the flow rate sequence as the numerator. The resulting ratio is the stability index S; Third step: Introduce an amplification factor and multiply it by the stability index S, accumulate the product with 1, and invert the accumulated result to derive the flow stability factor M(S). The derivation process of the duration factor A(T_dur) is as follows: Step 1: Perform preliminary anomaly analysis based on the basic deviation index. Compare the basic deviation index D_base with the preset traditional threshold. If D_base exceeds the traditional threshold, it indicates that there is an anomaly. Calculate the duration T_dur of the current anomaly. Step 2: Set constraints. When T_dur is less than 2 minutes, determine that A(T_dur) = 0. When T_dur is between 2 and 5 min, determine A(T_dur) = (T_dur-2) / 3; When T_dur exceeds 5 minutes, A(T_dur) is determined to be 1. The process of merging the basic deviation index to obtain the final deviation index is as follows: The basic deviation index D_base is multiplied sequentially by the introduced multi-dimensional feature factors to obtain one result; an indicator function I(Q>Q_max_hist) is set, which returns 1 if the current instantaneous flow Q exceeds the maximum flow Q_max_hist in the user's historical records, otherwise returns 0; the result of the indicator function I(Q>Q_max_hist) is multiplied by the set over-limit protection item to obtain two results; finally, the one result and the two results are summed to obtain the final deviation index D_fin; the basis for identifying abnormal gas usage behavior is: when D_fin exceeds the preset standard threshold and the duration exceeds the critical duration, abnormal gas usage behavior is determined, and a warning signal is issued; otherwise, no response is taken. The collaborative response module determines the risk level and response action based on a pre-built rule engine. When the igniter is abnormal, a dual-feedback special identification mechanism is triggered. The edge layer multi-dimensional feature coupling detection is used for the first feedback, and the cloud collaborative layer is combined to start the second feedback verification to obtain the corrected risk level and response action.

2. The combustible gas remote monitoring and management system based on the Internet of Things and AI according to claim 1, characterized in that: The sensor array includes at least an ultrasonic flow sensor, a digital pressure sensor, an infrared array sensor, and an ambient gas composite sensor; the acquired raw data includes at least: instantaneous flow rate Q, gas pressure P, stove surface temperature distribution T, human presence status human_absent, and cooking time cooking_time.

3. The combustible gas remote monitoring and management system based on the Internet of Things and AI according to claim 2, characterized in that: The process of performing feature extraction is as follows: extract the flow fluctuation rate σ_Q, pressure decay rate dP / dt, and maximum temperature of the stove surface max(T) from the original data; where the flow fluctuation rate σ_Q=std(Qt, Qt-1, ..., Qt-n), t is the sequence number of the current sampling time, and n is the number of backtracking sampling points; The process of performing multimodal fusion judgment is as follows: Boolean logic is combined with threshold comparison to form two types of criteria for target risk scenarios. The first type is the gas micro-leakage criterion, with the condition that σ_Q > the threshold Qth corresponding to the normal gas noise level and dP / dt < -0.

1. The second type is the unattended dry burning criterion, with the condition that max(T) > the maximum temperature threshold Tth corresponding to the normal range of smoke point and human_absent=True and cooking_time > the duration benchmark threshold Yth. If any of the above conditions are met, the corresponding type of alarm is triggered and reported.

4. The combustible gas remote monitoring and management system based on the Internet of Things and AI according to claim 3, characterized in that: After the results of the extraction and judgment actions are obtained, the local private domain network communication performs a status judgment: if there is an alarm of the corresponding type, it indicates that the status is abnormal and a priority alarm packet is triggered. Conversely, if the status is normal, the sub-results of the feature extraction action will be reported at preset intervals; among them, the priority alarm package contains at least 5 seconds of raw data segments and predefined enumeration codes for alarm events; the sub-results of the feature extraction action are the extraction of the flow fluctuation rate σ_Q, pressure decay rate dP / dt, and maximum stove surface temperature max(T) from the raw data.

5. The combustible gas remote monitoring and management system based on the Internet of Things and AI according to claim 3, characterized in that: The rules engine is configured as follows: when the current instantaneous flow rate Q is within the range of [0.05, 0.2], the risk level is determined to be L1, which is low risk, and the response actions are: APP push and voice prompt; when the current instantaneous flow rate Q exceeds 0.2 or is dry-burning, the risk level is determined to be L2, which is medium risk, and the response actions are: automatic valve closure, calling the corresponding user, and notifying the property management; when the pressure decay rate dP / dt exceeds 1 kPa / min, the risk level is determined to be L3, which is high risk, and the response actions are: immediate valve closure and power cut-off, and linkage with the fire protection platform.

6. The combustible gas remote monitoring and management system based on the Internet of Things and AI according to claim 5, characterized in that: The edge layer multidimensional feature coupling detection is used to provide feedback: the following judgment logic is added to the monitoring terminal: if the condition corresponding to the first type of gas micro-leakage criterion is met and the duration exceeds 5 seconds, a suspected ignition failure leak warning is triggered locally. The action is to start the buzzer prompt sound and simultaneously send a marked feature packet to the cloud, which includes Q and T sequences, timestamps and predefined enumeration codes. Combined with the cloud collaboration layer to initiate secondary feedback verification: For this user ID, statistically analyze the distribution of successful ignition time over the past 7 days. If the difference between the historical average ignition time and the current ignition time exceeds the defined threshold, it is recorded as a difference event. To determine whether the event is repeated, if the same user experiences more than one different event within 24 hours, the original risk level is increased by one level. The specific number of times the same user experiences more than one different event within 24 hours is counted. If the specific number is 2, the original response action is maintained. If the specific number exceeds 2, the response action corresponding to the high risk is executed.

7. A method for remote monitoring and management of combustible gases based on the Internet of Things and AI, characterized in that, The steps include: Deploy a sensor array in the selected area to be monitored to acquire raw data; An MCU + lightweight neural network model is set up in the monitoring terminal to perform extraction and judgment actions based on the raw data, including at least feature extraction actions and multimodal fusion judgment actions; Based on the results of the extraction and judgment actions, the pre-built local private domain network communication is run and connected to the cloud public network for verification, thus building a data lake; Establish a user behavior profile model, output a basic deviation index, and introduce multi-dimensional feature factors to merge with the basic deviation index to obtain the final deviation index, so as to identify abnormal gas usage behavior. The user behavior profiling model operates as follows: Gas usage habit learning: Based on any user ID, statistically analyze the daily average gas usage period distribution H(t), historical gas flow rate and current instantaneous flow rate Q, and commonly used flow rate intervals {Q_min, Q_max}; Defining the deviation index: Step 1: Take the absolute difference between the current flow rate and the historical average gas flow rate as the numerator and the standard deviation of the historical gas flow rate as the denominator to obtain the deviation multiple; Step 2: Count the number of days that gas usage events occurred within the preset historical period, and record it as N_use(t±Δt); t represents the current time, Δt represents the time tolerance window, and the time period sparsity coefficient R(t) can be obtained by dividing N_use(t±Δt) by the historical preset period. Step 3: Subtract the sparsity coefficient R(t) from 1 and calculate the weighted average of the deviation multiple to obtain the basic deviation index D_base; The introduced multi-dimensional feature factors include at least: flow stability factor M(S) and duration factor A(T_dur); The derivation process of the flow stability factor M(S) is as follows: First step: Take the sampled values ​​of all instantaneous flow rates within w seconds before the current time t, and denote them as the flow rate sequence from the (t-w+1)th sampling point to the tth sampling point, and calculate the standard deviation of the flow rate sequence; Second step: Calculate the average flow rate of the w sampling points as the denominator, and the standard deviation of the flow rate sequence as the numerator. The resulting ratio is the stability index S; Third step: Introduce an amplification factor and multiply it by the stability index S, accumulate the product with 1, and invert the accumulated result to derive the flow stability factor M(S). The derivation process of the duration factor A(T_dur) is as follows: Step 1: Perform preliminary anomaly analysis based on the basic deviation index. Compare the basic deviation index D_base with the preset traditional threshold. If D_base exceeds the traditional threshold, it indicates that there is an anomaly. Calculate the duration T_dur of the current anomaly. Step 2: Set constraints. When T_dur is less than 2 minutes, determine that A(T_dur) = 0. When T_dur is between 2 and 5 min, determine A(T_dur) = (T_dur-2) / 3; When T_dur exceeds 5 minutes, A(T_dur) is determined to be 1. The process of merging the basic deviation index to obtain the final deviation index is as follows: The basic deviation index D_base is multiplied sequentially by the introduced multi-dimensional feature factors to obtain one result; an indicator function I(Q>Q_max_hist) is set, which returns 1 if the current instantaneous flow Q exceeds the maximum flow Q_max_hist in the user's historical records, otherwise returns 0; the result of the indicator function I(Q>Q_max_hist) is multiplied by the set over-limit protection item to obtain two results; finally, the one result and the two results are summed to obtain the final deviation index D_fin; the basis for identifying abnormal gas usage behavior is: when D_fin exceeds the preset standard threshold and the duration exceeds the critical duration, abnormal gas usage behavior is determined, and a warning signal is issued; otherwise, no response is taken. Based on a pre-built rule engine, the risk level and response action are obtained. Under the condition that the igniter is abnormal, a dual feedback special identification mechanism is triggered. The edge layer multi-dimensional feature coupling detection is used for the first feedback, and the cloud collaboration layer is combined to start the second feedback verification to obtain the corrected risk level and response action.

Images