Cloud intelligent gas leakage analysis and decision-making method based on lightweight discrimination

Abstract

The invention provides a cloud intelligent gas leakage analysis and decision-making method based on lightweight discrimination, and the method comprises the steps: enabling a mobile inspection platform to movably collect the gas leakage monitoring information of each monitoring point in a monitoring region, and determining a lightweight discrimination index; the mobile inspection platform determines a hard threshold discrimination result by adopting a hard threshold decision rule based on the lightweight discrimination index; if it is judged that the risk level is high, local response is immediately carried out, and a preset decision scheme is matched; if the risk level is judged to be low, edge node region decision is started; an edge node calculates a gas concentration accumulated value, a concentration fluctuation intensity degree and a sudden increase leakage amplification early warning factor, the gas concentration accumulated value, the concentration fluctuation intensity degree and the sudden increase leakage amplification early warning factor are input into a LightGBM time sequence classification model together with a lightweight discrimination index, a leakage risk level is determined according to model output and serves as a soft threshold discrimination result, and a preset decision scheme is matched; and the cloud model optimizes and collects conflict samples to carry out incremental training on the model of the edge node. According to the invention, the missed alarm rate can be reduced, and the discrimination robustness is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of gas detection, and particularly relates to a cloud intelligent gas leakage analysis and decision method based on light weight discrimination. BACKGROUND

[0002] In recent years, as one of the most important resources for social production and life, gas transmission pipelines have rapidly expanded in cities in China. The hidden location of city gas pipelines, complex connection facilities, flammable and explosive contents, and other characteristics make it bring convenience to people's life, but also mix with many safety hazards. Gas leakage accidents caused by gas pipeline aging, overload work, and other conditions occur from time to time. It is urgent to explore the causes of city underground gas pipeline leakage accidents and reverse the high incidence of accidents.

[0003] There are many researches on gas pipeline leakage accidents, and many analysis tools and methods are introduced into this field. However, the existing gas leakage detection and decision technology has the following core defects:

[0004] (1) Insufficient gas tracing accuracy: the existing vehicle-mounted detection equipment only relies on methane concentration to determine leakage, without detecting ethane concentration and methane / ethane ratio. Because there are a large number of non-pipeline methane sources in nature, such as biogas, it is impossible to distinguish natural methane from pipeline gas leakage, resulting in a false alarm rate of up to 15%-30%, causing ineffective alarm and resource waste.

[0005] (2) Risk discrimination mechanism is rigid: the traditional cloud platform uses fixed threshold rules to determine the risk level, such as gas concentration mean C t >5000ppm is judged as L3 level. However, this hard threshold scheme has two problems: one is the misjudgment of the fuzzy area, that is, when the dynamic indicator is at the threshold critical point, such as concentration change rate ΔC=18ppm / min, the hard threshold rule cannot accurately classify; the second is the high risk of missing alarm: it does not consider the spatial aggregation effect of leakage points (such as the superposition of adjacent leakage points) and the time evolution characteristics (such as the sudden increase of concentration change acceleration α). The rigid risk discrimination mechanism makes the existing scheme over-responsive to low-risk events and delayed alarm for high-risk leaks, with a missing alarm rate of >5%.

[0006] (3) Multi-source data collaboration failure: the wind speed data of the mobile inspection platform contains vehicle speed interference, and the time and space alignment error between the fixed monitoring point data and the mobile terminal is >3 minutes, which makes the time and space reference fragmented; and the local leakage cannot be associated with the global state of the pipeline, resulting in data islands. This will result in poor accuracy of source location, with a source location error of ≥50 meters.

[0007] (4) Decision response intelligence is missing: the existing platform only realizes data visualization, lacks active prediction ability, and does not use historical data to model the degradation law of network health; the closed-loop decision mechanism is missing, the alarm information is disconnected from the emergency disposal strategy, and relies on manual experience scheduling (response delay > 10 minutes). Therefore, the existing scheme has a long cycle from leakage occurrence to decision execution, which cannot curb risk diffusion. SUMMARY

[0008] Therefore, the present application provides a cloud intelligent gas leakage analysis and decision method based on lightweight discrimination, which can reduce low-risk false positives and high-risk missed judgments, reduce false alarm rates, and improve discrimination robustness.

[0009] To solve the above technical problems, the present application is implemented as follows.

[0010] A cloud intelligent gas leakage analysis and decision method based on lightweight discrimination, comprising:

[0011] Step 1: The mobile inspection platform moves in the monitoring area to collect gas leakage monitoring information of each monitoring point and determines lightweight discrimination indicators;

[0012] Step 2: End device hard threshold decision: the mobile inspection platform determines the hard threshold discrimination result based on the lightweight discrimination indicators using the hard threshold decision rule; if it is determined to be a high-risk level, it will immediately respond locally and match the preset decision scheme; if it is determined to be a low-risk level, the mobile inspection platform uploads the hard threshold discrimination result and the data relied on to the edge node it belongs to;

[0013] Step 3: Edge node regional decision: the edge node records the hard threshold discrimination result; determines additional gas leakage characteristics of each monitoring point based on the uploaded data, including gas concentration cumulative value, concentration fluctuation intensity and sudden increase leakage amplification warning factor; inputs the lightweight discrimination indicators and additional gas leakage characteristics into the LightGBM time series classification model to determine the leakage risk level as the soft threshold discrimination result according to the model output, and matches the preset decision scheme;

[0014] Step 4: Cloud model optimization: the hard threshold discrimination result and the soft threshold discrimination result are different, which are regarded as conflict samples and are collected to the cloud device by each edge node; the cloud device periodically uses the conflict samples to incrementally train the efficient time series classification model of the edge node to optimize the model discrimination performance.

[0015] Preferably, in step 1, the lightweight discrimination indicators include: the number of leakage points n and the average gas concentration C of the monitoring point t , the gas concentration change rate ΔC t , and the gas concentration acceleration α.

[0016] Preferably, in step 2, the hard threshold decision rule is used to determine the hard threshold discrimination result as:

[0017] The lightweight discrimination index is substituted into the following hard threshold decision rule to obtain the hard threshold discrimination result R hard :

[0018]

[0019] In the formula, L1-L4 are four leakage risk levels.

[0020] Preferably, the additional gas leakage feature is determined in the following manner:

[0021]

[0022] A = max(0, a) · ln(1 + n)

[0023] In the formula, Q is the gas concentration cumulative value; σ ΔC is the concentration fluctuation intensity; A is the sudden increase leakage amplification warning factor; C t,i is the average gas concentration of the i-th leakage point, d i is the distance from the i-th leakage point to the sensor currently calculating the Q value, n is the number of leakage points; ΔC t is the gas concentration change rate of the monitoring point at time t, N is the number of times taken before the current time t, μ ΔC is the average value of the gas concentration change rate within N times; max(,) represents the maximum value of the two in the parentheses.

[0024] Preferably, in step 3, the leakage risk level is determined according to the model output as the soft threshold discrimination result:

[0025] The probability output by the model is substituted into the following soft threshold decision rule to obtain the soft threshold discrimination result R soft :

[0026]

[0027] In the formula, P() represents the probability value of being classified into the corresponding leakage risk level determined by the efficient time series classification model.

[0028] Preferably, when training the efficient time series classification model, a loss function with a false alarm penalty is used:

[0029]

[0030] where m is the number of batch samples; w iis the class weight of sample i; i is the sample index, taking values [1, m]; FN is the false negative sample set, j is the false negative sample index, ranging from 0 to the total number of false negative samples in the batch, (AC t ) j is the gas concentration change rate data in the jth false negative sample; CE represents the cross-entropy algorithm, y i is the model output, is the sample label; λ is the amplification coefficient of the false negative penalty term.

[0031] Preferably, the method further comprises: dynamically adjusting the amplification coefficient λ of the false negative penalty term according to historical false negative events, controlling the sensitivity of the model to false negatives; if a missed detection event occurs, increasing the amplification coefficient λ to strengthen the penalty; if no missed detection occurs for a set time period, reducing the amplification coefficient λ to reduce the conservatism.

[0032] Preferably, the leakage risk level has four levels, and the L1 and L2 levels are low risk levels, and the L3 and L4 levels are high risk levels.

[0033] Preferably, the matching preset decision scheme is:

[0034] The response strategy corresponding to the leakage risk level L1 is to record the monitoring data.

[0035] The response strategy corresponding to the leakage risk level L2 is early warning notification + 24-hour inspection.

[0036] The response strategy corresponding to the leakage risk level L3 is depressurization operation + 30-minute on-site verification.

[0037] The response strategy corresponding to the leakage risk level L4 is immediate gas stop + evacuation radius of 500 meters.

[0038] Beneficial effects:

[0039] (1) The existing hard threshold method cannot accurately handle the critical transition zone of multi-dimensional indicators such as concentration change rate (AC) and number of leakage points (n), resulting in false positives for low risk or missed judgments for high risk. The present application designs a soft and hard collaborative dual-track discrimination mechanism, a parallel architecture design of hard rule base and LightGBM time series classification model, to achieve the balance between extreme risk second-level response and accurate classification of fuzzy areas, and solves the problem of poor adaptability of traditional threshold rules to dynamic leakage scenarios.

[0040] (2) In the design of the LightGBM time series classification model, considering that the prior art lacks modeling capability for enhanced features such as the superposition effect of adjacent leakage points, the present application introduces features in the aspect of time series evolution analysis, including gas concentration cumulative value, concentration fluctuation intensity, and sudden increase leakage amplification warning factor. The gas concentration cumulative value represents the superposition effect of adjacent leakage points, and the concentration fluctuation intensity and sudden increase leakage amplification warning factor represent the time series evolution characteristics. Through the consideration of these features, the robustness of the soft threshold discrimination is improved.

[0041] (3) The present application establishes a three-level decision fusion architecture of "end (single point)-edge (region)-cloud (global)". The end (single point) uses a hard threshold discrimination scheme based on lightweight indicators to make risk discrimination, and responds immediately when a high-risk situation is identified, avoiding risk diffusion. Moreover, the use of lightweight indicators is conducive to the lightweight discrimination of end devices. The edge (region) is responsible for soft threshold risk discrimination, solving the problem of accurate identification of fuzzy areas. The cloud (global) accumulates conflict samples in the global range, incrementally optimizes the edge (region) model, and thus continuously optimizes the model discrimination performance. The three-level decision fusion solves the contradiction between lightweight demand and high-precision discrimination, and through the dual-track mechanism, it simultaneously realizes extreme risk second-level response and fuzzy area accurate classification, significantly reducing the false alarm rate. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 The flowchart of the cloud-end intelligent gas leakage analysis and decision method based on lightweight discrimination in the embodiments of the present application.

[0043] Figure 2 The schematic diagram of false alarm penalty self-adaption. DETAILED DESCRIPTION

[0044] The present application will be described in detail below with reference to the accompanying drawings and embodiments.

[0045] The present application provides a cloud-end intelligent gas leakage analysis and decision method based on lightweight discrimination, as shown in Figure 1 The method comprises the following steps:

[0046] Step 1: The mobile inspection platform moves in the monitoring area, collects gas leakage monitoring information of each monitoring point, and determines the lightweight discrimination indicators.

[0047] In this step, the mobile inspection platform serves as an end device, each of which collects data and calculates lightweight discrimination indicators. The lightweight discrimination indicators are part of the gas leakage features. In this embodiment, the lightweight discrimination indicators include:

[0048] n: number of leak points; the number of leak points can be determined by prior art, for example, assuming a number of sources K. Define inversion parameters for each source: including location coordinates (x_i, y_i, z_i), leak rate Q_i; construct a forward model based on the Gaussian plume model: use the currently assumed parameters of the K sources to calculate their predicted concentration values at all sensor locations, all sampling times. Calculate the difference between predicted and measured values: define an objective function to measure this difference (such as least squares error). Then optimize source parameters: adjust all parameters of the K sources to minimize the objective function. Evaluate the fitting results of different K values: compare the minimum objective function values and the rationality of model fitting under different assumptions of K = 1, 2, 3,..., select the optimal K as the number of leak points n.

[0049] C t : current concentration mean of the current monitoring point at the current time t (unit: ppm);

[0050] ΔC t : concentration change rate of the current monitoring point (unit: ppm / min), calculation formula:

[0051] ΔC t = C t -C t-1 ΔtΔC t = ΔtC t -C t-1 ;

[0052] α: concentration acceleration of the current monitoring point (unit: ppm / min 2 ), calculation formula:

[0053] α = ΔC t - ΔC t-1 Δtα = ΔtΔC t - ΔC t-1

[0054] Step 2: End device hard threshold decision:

[0055] The mobile inspection platform determines the hard threshold discrimination result based on the lightweight discrimination index and uses the hard threshold decision rule. If it is determined to be a high risk level, it will immediately respond locally and match the preset decision scheme. If it is determined to be a low risk level, the mobile inspection platform will upload the hard threshold discrimination result and the data relied on for determination to the edge node to which it belongs.

[0056] This step is the processing at the end device granularity.

[0057] The mobile inspection platform determines the hard threshold discrimination result R hard using the following hard threshold decision rule:

[0058]

[0059] In the formula, L1-L4 are four leakage risk levels.

[0060] In this embodiment, four leakage risk levels are set, and the risks are L1-L4 from low to high; wherein it is further provided that L1 and L2 levels are low risk levels, and L3 and L4 levels are high risk levels. When the end device determines the high risk level, an extreme risk second-level response is realized to prevent the danger from further spreading. If the end device determines the low risk level, a soft discrimination model based on a machine learning algorithm of the edge device needs to be started.

[0061] Step 3: Edge node area decision:

[0062] The edge node records the hard threshold discrimination result; the additional gas leakage features of each monitoring point are determined according to the uploaded data, including the gas concentration cumulative value, the concentration fluctuation intensity and the sudden increase leakage amplification warning factor; the lightweight discrimination index and the additional gas leakage features are combined to form a feature vector input into a LightGBM time series classification model, the leakage risk level is determined according to the model output as a soft threshold discrimination result, and a preset decision scheme is matched.

[0063] In order to solve the accurate classification of the fuzzy area between the hard thresholds (such as ΔC = 18-22 ppm / min), enhance the adaptability and robustness of the model, the application introduces a soft discrimination model based on a machine learning algorithm, and sets it in the edge node. Considering the ability of machine learning, on the basis of the lightweight discrimination index, additional gas leakage features are further added. Considering the superimposed influence of adjacent leakage points, the fluctuation intensity, and the sudden increase leakage amplification effect, three additional features are set: the gas concentration cumulative value Q, the concentration fluctuation intensity σ ΔC and the sudden increase leakage amplification warning factor A:

[0064]

[0065] A = max (0, a) · ln (1 + n)

[0066] In the formula, Q is the gas concentration cumulative value; σ ΔC is the concentration fluctuation intensity; A is the sudden increase leakage amplification warning factor; C t,i is the average value of the gas concentration of the i-th leakage point, d i is the distance from the i-th leakage point to the sensor currently calculating the Q value, n is the number of leakage points; ΔC t is the gas concentration change rate of the monitoring point at t, N is N time points before the current time t, μ ΔC is the average value of the gas concentration change rate within N time points; max (,) represents the maximum value of the two in the parentheses.

[0067] Lightweight discriminant indicators (n, C t , ΔC t , ) and additional gas leakage features (Q, σ ΔC , A) form a feature vector, input into the LightGBM time series classification model, and determine the leakage risk level according to the model output as the soft threshold discriminant result, and match the pre-set decision scheme.

[0068] The present application uses a discriminant model of LightGBM (high-efficiency time series classification) + Bayesian optimization of hyperparameters, adopts a LightGBM time series classification model as a core prediction engine, introduces time window sliding features when necessary, uses Bayesian hyperparameter optimization to maximize recall rate (Recall) while controlling false negative rate (FN), sets a loss function with false negative penalty and cost-sensitive learning, improves the priority of high-risk early warning through class weight, and finally adopts an output strategy of probability calibration and rule fusion.

[0069] The loss function with false negative penalty is expressed as:

[0070]

[0071] Where m is the number of batch samples; w is the class weight of sample i; i is the sample index, taking value [1, m]; FN is the false negative sample set; j is the false negative sample index, ranging from [0, the total number of false negative samples in the batch], (ΔC t ) j is the gas concentration change rate data in the jth false negative sample; CE represents the cross-entropy algorithm, y i is the model output, is the sample label; λ is the amplification coefficient of the false negative penalty term, controlling the sensitivity of the model to false negatives; the larger λ is, the heavier the false negative penalty is, making the model more inclined to high-risk early warning, and the smaller λ is, the lighter the false negative penalty is, making the model more inclined to reduce false positives.

[0072] The trained model processes input features and can output the probability of 4 risk levels, denoted as P(L1), P(L2), P(L3), P(L4), and the probability output by the model is substituted into the following soft threshold decision rule to obtain the soft threshold discriminant result R soft :

[0073]

[0074] In the formula, P() represents the probability value of the high-efficiency time series classification model determined by the classification into the corresponding leakage risk level.

[0075] Step 4: Cloud model optimization:

[0076] The present application designs two dynamic iterative optimization mechanisms:

[0077] One is conflict sample automatic training: each edge node takes the case that the hard threshold discrimination result and the soft threshold discrimination result are different as the conflict sample (hard ≠ soft), and collects it to the cloud device; the cloud device periodically uses the conflict sample to perform incremental training on the high-efficiency time sequence classification model of the edge node, and continuously optimizes the model discrimination performance. The period can be selected to perform incremental training once every quarter.

[0078] Two is the self-adaptive punishment of missed alarm: according to the historical missed alarm event, the amplification coefficient λ of the missed alarm punishment term is dynamically adjusted, and the high-risk scene recognition is strengthened. As shown in the formula (2), if the missed detection event occurs, the amplification coefficient λ is increased to strengthen the punishment; if there is no missed detection for a set time, the amplification coefficient λ is reduced to reduce the conservatism. After updating λ, the loss function is updated, and the model is retrained. Figure 2

[0079] Thus, the process ends.

[0080] In the above embodiment, steps 2 and 3 can use the following decision setting scheme according to the risk level matching preset decision scheme:

[0081] Risk level Response strategy example L4 Immediate gas shut-off + evacuation radius 500 meters L3 Depressurization run + 30 minutes on-site verification L2 Early warning notification + 24 hours patrol L1 Record monitoring data

[0082] The above specific embodiments only describe the design principles of the present application, and the shapes and names of the components in the description can be different and are not limited. Therefore, the person skilled in the art of the present application can modify or equivalently replace the technical solutions recorded in the foregoing embodiments; and these modifications and replacements do not deviate from the purpose and technical solutions of the present application, and should all belong to the protection scope of the present application.​

Claims

1. A cloud-based intelligent gas leakage analysis and decision-making method based on lightweight discrimination, characterized in that, Step 1: The mobile inspection platform moves in the monitoring area to collect gas leakage monitoring information of each monitoring point and determines the lightweight discrimination index; Step 2: End device hard threshold decision: Based on the lightweight discrimination index, the mobile inspection platform adopts a hard threshold decision rule to determine the hard threshold discrimination result; If it is determined to be a high-risk level, a local response is immediately matched with the preset decision scheme; if it is determined to be a low-risk level, the mobile inspection platform uploads the hard threshold discrimination result and the data relied on for determination to the edge node to which it belongs; Step 3: Edge node regional decision: The edge node records the hard threshold discrimination result; determines additional gas leakage characteristics of each monitoring point, including gas concentration cumulative value, concentration fluctuation intensity, and sudden increase leakage amplification warning factor, according to the uploaded data; inputs the lightweight discrimination index and the additional gas leakage characteristics into a LightGBM time series classification model to determine the leakage risk level according to the model output as a soft threshold discrimination result, and matches the preset decision scheme; Step 4: Cloud model optimization: The hard threshold discrimination result and the soft threshold discrimination result are different, which are collected as conflict samples to the cloud device; The cloud device periodically uses the conflict samples to perform incremental training on the efficient time series classification model of the edge node to optimize the model discrimination performance.

2. The method of claim 1, wherein, In step 1, the light-weight discrimination index includes: the number of leakage points n and the average gas concentration C of the monitoring points t , the gas concentration change rate ΔC t , and the gas concentration acceleration α.

3. The method of claim 1 or 2, wherein, In step 2, the hard threshold decision rule is adopted to determine the hard threshold discrimination result as: The lightweight discrimination index is substituted into the following hard threshold decision rule to obtain a hard threshold discrimination result R hard : In the formula, L1-L4 are four leakage risk levels.

4. The method of claim 1, wherein, The determination method of the additional gas leakage characteristics is: A = max(0, α)·ln(1 + n) In the formula, Q is the cumulative value of gas concentration; σ ΔC is the degree of concentration fluctuation; A is the amplification warning factor of sudden increase leakage; C t,i is the average value of gas concentration of the ith leakage point, d i is the distance from the ith leakage point to the sensor currently calculating the Q value, n is the number of leakage points; ΔC t is the gas concentration change rate of the monitoring point at time t, N is the N time points before the current time t, μ ΔC is the average value of gas concentration change rate in N time points; max(,) represents the maximum value of the two in the parentheses.

5. The method of claim 1, wherein, In step 3, the leakage risk level is determined according to the model output as a soft threshold discrimination result: The probability output by the model is substituted into the following soft threshold decision rule to obtain a soft threshold discrimination result R soft : In the formula, P() represents the probability value of being classified into the corresponding leakage risk level by the efficient time series classification model.

6. The method of claim 1, wherein, When training the efficient time series classification model, a loss function with a false alarm penalty is introduced as: wherein m is the number of batch samples; w i is the class weight of sample i; i is the sample index, taking values [1, m]; FN is the false alarm sample set, j is the false alarm sample index, ranging from [0, the total number of false alarm samples in the batch], (AC t ) j is the gas concentration change rate data in the jth false alarm sample; CE represents the cross-entropy algorithm, y i is the model output, is the sample label; λ is the amplification coefficient of the false alarm penalty term.

7. The method of claim 6, wherein, The method further includes: dynamically adjusting the amplification coefficient λ of the false alarm penalty term according to historical false alarm events to control the sensitivity of the model to false alarms; if a false detection event occurs, the amplification coefficient λ is increased to strengthen the penalty; if no false detection occurs for a set period of time, the amplification coefficient λ is reduced to reduce the conservatism.

8. The method of claim 1, wherein, There are four levels of leakage risk levels, L1 and L2 are low-risk levels, and L3 and L4 are high-risk levels.

9. The method of claim 8, wherein, The matching of the preset decision scheme is: The response strategy corresponding to the leakage risk level L1 is: recording monitoring data; The response strategy corresponding to the leakage risk level L2 is: early warning notification + 24-hour inspection; The response strategy corresponding to the leakage risk level L3 is: pressure reduction operation + 30-minute on-site verification; The response strategy corresponding to the leakage risk level L4 is: immediate gas stop + evacuation radius of 500 meters.

Images