Cloud-edge collaborative data processing method for gas pipeline network safety monitoring

Abstract

This invention relates to the field of gas pipeline network safety monitoring technology, and discloses a cloud-edge-device collaborative data processing method for gas pipeline network safety monitoring. This method generates a three-dimensional uncertainty map covering the entire pipeline network. The cloud monitors the values ​​of each grid point on the map in real time, generates data demand instructions, and sends them to the edge device. Upon receiving the instructions, the edge device uses a lightweight decision agent model and representativeness judgment pre-issued by the cloud to determine the target data sample, and combines a knapsack optimization algorithm to select a target uploaded data subset that maximizes total information gain and meets resource budget constraints. After receiving the data subset, the cloud performs incremental updates to the decision model and corrects the uncertainty map, calibrating the prediction deviation of the edge device's information gain pre-evaluation. This invention enables the edge device to possess intelligent filtering nodes with cloud demand perception capabilities, maximizing the actual utility of data for cloud decision-making under resource constraints, and effectively improving the timeliness of early warning and the accuracy of decision-making in gas pipeline network safety monitoring.

Description

Technical Field

[0001] This invention relates to the field of gas pipeline network safety monitoring technology, specifically to a cloud-edge-device collaborative data processing method for gas pipeline network safety monitoring. Background Technology

[0002] With the acceleration of urbanization and the adjustment of energy structure, the scale of urban gas pipeline networks is constantly expanding, and pipeline safety monitoring faces challenges such as wide coverage, numerous monitoring points, heterogeneous data types, and complex communication environments. In recent years, cloud-edge-device collaborative architecture has become the mainstream technical route in the field of gas pipeline safety monitoring. The terminal perception layer deploys pressure sensors, vibration sensors, gas concentration sensors, flow sensors, and video acquisition equipment to be responsible for the real-time acquisition of raw data; edge computing nodes are deployed at key nodes or valve chambers of the pipeline network to be responsible for data preprocessing and preliminary anomaly detection; the cloud aggregates data from the entire network and uses artificial intelligence models for global risk identification, leak location, and emergency decision-making.

[0003] However, existing cloud-edge-device collaborative solutions suffer from the following core contradictions at the data scheduling level: the disconnect between the dynamic data demands of the cloud and the static data processing methods of the edge leads to a severe imbalance between data value density and transmission decision-making utility. This is mainly reflected in the following: it is generally believed that the more comprehensive and complete the data received by the cloud, the higher the accuracy of the decision-making model; however, the communication environment along gas pipelines is complex, and 4G / 5G network signals may be unstable in remote pipe sections, underground pipe corridors, and narrow valve chambers. Edge devices are limited by the power consumption and cost of outdoor deployment, and communication bandwidth and battery power are scarce resources, forcing the edge to compress or truncate massive amounts of raw data before uploading; while existing edge data filtering strategies typically use fixed rules or static configuration distribution, aiming to ensure that the amount of data uploaded from the edge to the cloud meets preset requirements and that the uploaded data conforms to manually preset coarse-grained scenario classifications. However, in actual operation, the data requirements of cloud-based decision-making models are dynamic and highly differentiated. The information gaps of the models vary greatly in different regions, at different times, and under different weather conditions. For example, during heavy rain, vibration data of crossing sections is urgently needed to detect pipe displacement; during peak gas consumption, pressure fluctuation data near pressure regulating stations is urgently needed to predict load impacts; and during routine inspections, balanced coverage of various types of data is required to maintain the baseline accuracy of the model. These changes in requirements are minute-level and regional-level, driven by the model's cognitive state, rather than being exhaustively pre-defined by human experience through fixed rules or static configurations. Existing edge data strategies are limited to the resource constraints of the edge itself and cannot perceive the real-time cognitive state of the cloud-based decision-making model. The task objectives they face are still limited to ensuring the amount of data uploaded and coarse-grained scenario matching, which inevitably leads to the following defects: when communication deteriorates, the edge discards a large amount of seemingly redundant but actually crucial micro-fluctuation data for long-term trend prediction in the cloud, causing the cloud model to make inaccurate decisions due to a lack of key data; conversely, when communication is good, it tends to upload all raw data indiscriminately, resulting in a triple waste of bandwidth, power, and cloud computing power. Summary of the Invention

[0004] This invention aims to provide a cloud-edge-device collaborative data processing method for gas pipeline network safety monitoring. By using the uncertainty of the cloud model as a control signal to drive edge data scheduling, a complete closed loop of perception, evaluation, scheduling, and feedback is constructed to achieve the maximum amount of decision information with the least amount of data uploaded.

[0005] The basic solution provided by this invention is: a cloud-edge-device collaborative data processing method for gas pipeline network safety monitoring, including: S1, based on the decision model deployed in the cloud, determines the prediction uncertainty values ​​of each spatial segment, each data type, and each time of the gas pipeline network, and generates a three-dimensional uncertainty map covering the entire pipeline network; S2 monitors the values ​​of each grid point in the uncertainty map in real time in the cloud. When the uncertainty value of a certain area does not meet the preset conditions, a data demand instruction is generated and sent to the edge of the corresponding area through the communication network. The data demand instruction includes the required data type, uncertainty value, sampling priority weight vector and resource budget constraint parameters. S3: After receiving the data request instruction, the edge device uses a lightweight decision agent model pre-deployed in the cloud to pre-evaluate the information gain of data samples that belong to the required data type in the local cache data, and makes a representativeness determination based on a preset redundancy threshold that is dynamically adjusted based on the uncertainty value of the region to determine the target data sample. S4. At the edge, the target data samples are used as candidate sets. The knapsack optimization algorithm is used to calculate the estimated gain value based on the sampling priority weight vector and the information gain pre-evaluation result. The target upload data subset that maximizes the total information gain and meets the resource budget constraint parameter constraints is selected and uploaded to the cloud. S5 receives a subset of the target data uploaded to the cloud and performs incremental updates to the decision model and corrections to the uncertainty map. It then uses the magnitude of the change in uncertainty values ​​and the estimated gain values ​​to update the lightweight decision agent model to assess the prediction bias of the information gain pre-evaluation at the edge during the next calibration.

[0006] The working principle and advantages of this invention are as follows: This invention quantifies the cognitive gaps in cloud-based decision-making models in real time into actionable data acquisition strategies at the edge. This drives the edge to prioritize high-value data that has the highest marginal contribution to reducing uncertainty in the cloud, and achieves continuous self-optimization of the selection accuracy through closed-loop feedback. The core advantages are: This invention uses the uncertainty of the cloud model as a control signal to drive data scheduling at the edge, upgrading the criterion for judging data value from whether it exceeds a threshold to whether it can reduce the uncertainty of cloud decision-making. This enables the edge to perceive changes in cloud data demand for the first time, realizing a fundamental change in data scheduling from data-driven to decision-driven. Areas with high uncertainty in the cloud are precisely the areas that need the edge to prioritize data supply. This mechanism is highly compatible with the business characteristics of gas pipeline safety monitoring, where the cost of underreporting is extremely high.

[0007] This invention constructs a dual-layer driven edge-end intelligent filtering architecture combining data demand commands and a proxy model. Data demand commands issued from the cloud provide business guidance, while the proxy model provides value quantification capabilities. Knowledge distillation brings the cognitive capabilities of the cloud-based decision-making model down to the edge, allowing the edge to simulate cloud-based reasoning locally using the proxy model and predict the marginal contribution of each piece of data to reducing cloud uncertainty. The filtering logic has been upgraded from maximizing the amount of data transmitted to maximizing the marginal value of data for cloud-based decision-making. The synergy of these two approaches enables the edge to make globally optimal data filtering decisions even under resource constraints. This architecture can respond to sudden changes in operating conditions (such as dynamic adjustments to commands during rainstorm warnings) and independently complete high-value data filtering using the proxy model (degradation strategy) during communication interruptions.

[0008] This invention employs an adaptive deduplication mechanism driven by cloud-based cognitive state during edge data filtering. This mechanism allows the uncertainty value in the cloud to inversely adjust the redundancy tolerance at the edge. The higher the uncertainty in a region, the higher the cloud's tolerance for redundancy (the lower the threshold), allowing more superficially similar data that may contain new information to be uploaded, thus avoiding excessive deduplication and missing key data. The knapsack optimization algorithm filters the data subset that maximizes the total information gain under the dual constraints of communication bandwidth and power consumption, ensuring that the output is the optimal data combination for cloud-based decision-making under any resource state.

[0009] The closed-loop feedback correction mechanism introduced in this invention enables continuous self-optimization of the prediction accuracy at the edge. By carrying the target gain value through metadata tags, the cloud compares the actual decision contribution and generates a correction factor that is sent back to the edge, so that the prediction deviation of the edge proxy model is systematically corrected. This allows the system to maintain optimal screening accuracy without human intervention during long-term operation, which is especially suitable for gas pipeline networks, which are long-term service networks with slowly changing operating conditions. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating the cloud-edge-device collaborative data processing method for gas pipeline network safety monitoring provided in an embodiment of the present invention. Detailed Implementation

[0011] The following detailed explanation illustrates the specific implementation methods: The basic implementation examples are as follows: Figure 1 As shown: A cloud-edge-device collaborative data processing method for gas pipeline network safety monitoring, including: S1, Cloud-based pipeline uncertainty map construction steps: Based on the decision model deployed in the cloud, determine the predicted uncertainty values ​​of each spatial segment, each data type, and each time of the gas pipeline network, and generate a three-dimensional uncertainty map covering the entire pipeline network.

[0012] Specifically as follows: The cloud-based decision model (main AI model) adopts a gas pipeline network leakage prediction model or corrosion rate prediction model based on a spatiotemporal graph neural network. The input of the decision model is gas pipeline network topology data and multi-source sensor time series data, and the output is the risk probability prediction value of each pipe segment. At the same time, the model uses Bayesian neural network, MC-Dropout (Monte Carlo dropout method) or deep Gaussian process to enable it to output the uncertainty measure value of each prediction result (such as prediction variance or information entropy) as the prediction uncertainty value during inference.

[0013] The multi-source sensor data of the gas pipeline network includes at least two of the following types: time-series data of gas pressure in the pipeline collected by pressure sensors, data of vibration acceleration of the pipeline wall collected by vibration sensors, data of combustible gas concentration in the pipe gallery or well collected by gas sensors, data of instantaneous gas flow in the pipeline collected by flow sensors, and image data of the surrounding environment of the pipeline network acquired by video acquisition equipment; different types of sensor data have different sampling frequencies and data type identifiers.

[0014] The dimensions of a three-dimensional uncertainty map include: spatial dimension: different sections of the pipeline (marked by latitude, longitude and mileage), temporal dimension: different times (recent / long-term), and feature dimension: different data types (pressure / vibration / gas concentration / visual).

[0015] S2, Data Request Command Generation and Issuance Steps: The cloud monitors the values ​​of each grid point in the uncertainty map in real time. When the uncertainty value of a certain area does not meet the preset conditions, a data request command is generated and issued to the edge of the corresponding area through the communication network. The data request command includes the required data type, uncertainty value, sampling priority weight vector, and resource budget constraint parameters.

[0016] Specifically as follows: The grid value H(loc, t, feat) of the uncertainty map represents the uncertainty value, characterizing the degree of uncertainty in the cloud's perception of location loc at time t and data type feat corresponding to the current network state. It is calculated by weighted sum of prediction entropy and posterior variance of model parameters. H(loc, t, feat) = -Σ p(y|x) · log p(y|x) + β · Var[θ|x] Where -Σ p(y|x) · log p(y|x) is the prediction entropy, reflecting the dispersion of the model output probability distribution; Var[θ|x] is the posterior variance of the model parameters, reflecting the confidence level of the model in its own parameters; β is the preset tradeoff coefficient.

[0017] The higher the uncertainty value, the more new data from that region the cloud needs to reduce the risk of global prediction.

[0018] When the uncertainty value of a certain area exceeds a preset threshold, it is considered not to meet the preset conditions and is marked as an information desert area, triggering a data request command to the edge. The cloud generates a data request command for the information desert area.

[0019] Gas pipeline networks are characterized by wide spatial distribution and significant differences in operating conditions across sections. Different pipeline sections (such as crossing sections, residential sections, and industrial sections) have different risk characteristics. The grid division and preset thresholds of the uncertainty map need to be matched with the pipeline network topology and monitoring point layout, and the preset thresholds need to be matched with the risk characteristics of the pipeline section where each grid point is located. For example, pipeline sections crossing rivers have a high risk of geological disasters, so their uncertainty thresholds should be set to lower values ​​to more sensitively trigger data demand commands.

[0020] Data requirement instructions include: The target area identifier is used to specify the spatial range of data collection. During application, the cloud scans an uncertainty map, performs spatial clustering, and identifies areas where the grid value H changes abruptly. If the H value of a certain pipe segment (such as a crossing segment) suddenly increases from 0.2 to 0.7, while the surrounding pipe segments remain at 0.2, its spatial gradient is calculated, and the continuous pipe segments with the largest gradient and H > a preset threshold (such as 0.5) are designated as the target area.

[0021] The required data type specifies the sensor data types to be considered. In application, multi-dimensional contribution analysis is used. Input data includes pressure (P), vibration (V), gas (G), and flow rate (F). The cloud-based decision model, trained with built-in attention or gradient attribution, calculates which input feature dimension's uncertainty is primarily caused by the absence or ambiguity of uncertainties in pipe sections where the uncertainty value consistently exceeds a preset warning threshold across multiple time windows (e.g., 30 minutes). If the area has a high risk of geological disasters, the model places a significant emphasis on the vibration (V) dimension; the absence of vibration would cause H to spike, thus the extracted data type is vibration. If it's a slow leak, the model is sensitive to pressure (P) and gas (G), so the extracted data types are pressure and gas.

[0022] The uncertainty value is used to quantify the degree of lack of current knowledge about the area in the cloud, and is used to adjust the preset redundancy threshold during the representativeness determination process. In application, the highest grid point value in the area is used.

[0023] The sampling priority weight vector indicates the relative importance of different data types in resource competition. In application, the interpretability attribution module built into the decision model (such as a gradient integral-based feature attribution algorithm) is used to calculate the uncertainty sub-value Hd contributed by each input data type to the overall uncertainty H of the current target region, satisfying H = ΣHd + ε, where ε is the non-attributable residual. The Hd of each data type is normalized to obtain the attribution proportion Ad = Hd / ΣHd'. Using the effective data type with the lowest attribution proportion as a benchmark (its sensitivity coefficient is set to 1.0), the sensitivity coefficient Sd = Ad / Abase of other data types is calculated. To prevent excessive weight concentration on a single type, a non-linear compression mapping function wd = 1.0 + (Sd - 1.0) × γ is used to map the sensitivity coefficient to the sampling priority weight, where γ is a preset compression factor (between 0.5 and 0.9, for example, 0.8). For data types with a zero attribution proportion, a preset minimum weight (e.g., 0.3) is assigned, allowing them to still participate in uploading when resources are plentiful.

[0024] Resource budget constraint parameters are used to limit the upper limit of communication bandwidth and energy consumption that the edge device can use. When applied, the cloud maintains a global resource pool and allocates resources according to the regional risk level. The current available total bandwidth (e.g., 10Mbps) is divided by the number of high-risk areas (e.g., 5) and multiplied by the security factor (e.g., 0.8) to obtain the upper limit of bandwidth available to the edge device at this time (e.g., 2.0 Mbps). The cloud reads the total battery capacity (e.g., 30000mAh) and the real-time reported remaining power (e.g., 65%) from the edge device's nameplate parameters. After deducting the safety reserve power (e.g., 20%) to ensure basic standby power, the available power (15600mAh) is obtained. This available power is divided by the preset minimum remaining working time (24 hours, corresponding to 2880 30-second upload cycles) to obtain the basic energy consumption budget for a single cycle (5.42mAh). This is then multiplied by the energy consumption adjustment factor determined by the current area's risk level (1.8 in high-risk rainstorm scenarios), and finally multiplied by the execution margin coefficient (e.g., 1.2) and rounded up to obtain the final single-cycle energy consumption limit Energy_upper (12mAh) sent to the edge. This value serves as the boundary parameter for energy consumption constraints in knapsack optimization, ensuring that the edge device will not fail prematurely due to excessive power consumption during a single upload in long-term operation.

[0025] S3, Edge Data Filtering Steps: After receiving the data request instruction, the edge device uses a lightweight decision agent model pre-deployed in the cloud to pre-evaluate the information gain of data samples belonging to the required data type in the local cache data, and makes a representativeness determination based on a preset redundancy threshold that is dynamically adjusted based on the uncertainty value of the region, thus determining the target data sample.

[0026] Specifically as follows: At the edge, a local data caching mechanism is adopted to cache the multi-source sensor data reported by the terminal perception layer and perform lightweight feature extraction to obtain feature vectors for each data sample. The feature vectors include time-domain statistical features and frequency-domain energy features, such as the mean, variance, and spectral energy of pressure fluctuations. The purpose of feature extraction is to reduce the data dimensionality, reduce the computational overhead of subsequent information gain evaluation, and retain key information valuable for decision-making.

[0027] After receiving the data request instruction, the edge device uses the proxy model pre-deployed in the cloud to perform information gain pre-evaluation on the cached data samples corresponding to the required data type, obtains the basic information gain value of each sample, and uses it to filter low-gain data samples to obtain a high-gain data sample set.

[0028] The proxy model distributed from the cloud is a lightweight, compressed version of the cloud-based decision-making model (a lightweight neural network or machine learning model file with inference capabilities). It is obtained by compressing the cloud-based decision-making model using knowledge distillation, with its parameter count not exceeding 10% of the cloud-based model's total parameters. It is deployed on an edge NPU or GPU for real-time execution. Unlike conventional static thresholding schemes, this step employs a cloud-based intelligent deployment approach. The proxy model can be understood as a digital twin of the main cloud model, enabling the edge to possess the inference and computational power to simulate cloud-based thinking. When the cloud compresses the latest uncertainty states and decision logic into this lightweight model and distributes it to the edge, the edge simulates the cloud-based decision-making model's response to the input data. The cloud periodically updates the proxy model and distributes it to the edge when the uncertainty map changes significantly. In other words, the proxy model updates with the iteration of the cloud-based decision-making model (e.g., after weekly or monthly retraining and distribution), ensuring that the proxy model's parameters are not fixed but updated periodically based on the current state of the main cloud model. This guarantees that the information gain prediction at the edge always aligns with the actual needs of the cloud.

[0029] The information gain pre-evaluation process includes: for the feature vector xi of the i-th data sample, calculating the estimated reduction in cloud uncertainty by the sample, Gain. i = Hcurrent - E[Hafter(xi)], serving as the basic information gain value for this sample, where Hcurrent is the current uncertainty value of the region in the cloud (obtained from the most recent data request instruction or uncertainty map synchronization information), and E[Hafter(xi)] is the expected posterior uncertainty value calculated by Monte Carlo dropout or a single Bayesian forward propagation after inputting the feature vector xi into the surrogate model. Low gain data (Gain) iEven if a data sample is deemed novel (completely different from historical data) in the representativeness assessment (with a threshold of 0.10), it is not worth uploading because it does not carry sufficient information. Therefore, such a data sample is directly filtered out and does not enter the subsequent processing flow. (The last sentence appears to be incomplete and possibly refers to a different data point.) i Data samples exceeding the threshold are considered high-gain data and further evaluated for representativeness. Information gain pre-evaluation avoids the high communication costs associated with uploading and evaluating the original data.

[0030] The representativeness determination of high-gain data samples retained through information gain pre-evaluation includes: calculating the similarity distance (Mahaviran distance or cosine similarity distance) between the feature vector xi of the current data sample corresponding to the required data type and the feature vectors of the N most recent uploaded data samples in the local cache; if the similarity distance is less than the preset redundancy threshold, it means that the distance is too small (i.e., highly redundant with the uploaded data), and the data sample is marked as redundant data and discarded; if the similarity distance does not exceed the preset redundancy threshold, it is retained as the target data sample.

[0031] The preset redundancy threshold is inversely correlated with the uncertainty value; that is, the higher the uncertainty, the lower the preset redundancy threshold. This means the cloud tolerates lower redundancy in areas with high uncertainty, allowing more superficially similar but potentially new data samples to enter the candidate set. The preset redundancy threshold can be dynamically determined by multiplying the basic redundancy threshold by (1 - current uncertainty value / maximum uncertainty value).

[0032] Among them, the distance calculation for representativeness determination adopts the Local Sensitive Hash (LSH) approximation algorithm, which maps high-dimensional feature vectors into low-dimensional hash codes, greatly reducing computational complexity.

[0033] The technical significance of representativeness determination lies in the fact that, under normal operating conditions of gas pipeline networks, time-series data collected by the same sensor exhibit strong temporal correlation, and a large number of continuous data samples are highly similar in the feature space, resulting in extremely low marginal contribution to cloud-based decision-making. Representativeness determination can effectively filter out this repetitive data, avoiding the waste of communication resources.

[0034] Information gain pre-assessment and representativeness determination are two independent, sequential, and indispensable screening steps that together constitute a dual-guarantee mechanism for edge data screening. By performing gain assessment first, approximately 40% to 50% of low-value data can be filtered out before entering the high-overhead representativeness determination stage, significantly reducing the computational load on the edge and demonstrating the engineering rationality on resource-constrained edge devices.

[0035] Information gain pre-evaluation determines whether a single data point is useful, while representativeness determination determines whether the data point is duplicated. After quality screening by information gain pre-evaluation and redundancy screening by representativeness determination, the target data sample is finally obtained.

[0036] S4, Knapsack Optimized Upload Steps: At the edge, the target data samples are used as candidate sets. The knapsack optimization algorithm is used to calculate the estimated gain value based on the sampling priority weight vector and the information gain pre-evaluation result. The target upload data subset that maximizes the total information gain and meets the resource budget constraint parameter constraints is selected and uploaded to the cloud.

[0037] Specifically as follows: Resource budget constraints include communication bandwidth availability and remaining power. The edge device obtains the currently available communication bandwidth availability and remaining power, and under these constraints, solves the following knapsack optimization problem: maxΣ(Gain i ×wi) stΣSize i ≤Bandwidth, ΣCost i ≤ Energy Among them, Gain i The basic information gain value of the i-th data sample is calculated for the information gain pre-evaluation; wi is the weight value corresponding to this data type in the sampling priority weight vector issued by the cloud; Size i Cost is the amount of uploaded data (after compression) for the i-th data sample. i The estimated energy consumption for uploading this data sample is defined as follows: Bandwidth is the current available communication bandwidth, and Energy is the current remaining power.

[0038] Estimated energy cost i The energy consumption model is calculated based on the current communication standard, signal strength, and amount of data to be uploaded by the edge device; the communication standard includes at least one of NB-IoT, Cat1, 4G, and 5G.

[0039] When the current battery level of the edge device is lower than the first battery threshold, the weight coefficient of the battery constraint term is increased in the knapsack optimization problem, and data samples with high gain and low power consumption are uploaded first to extend the continuous working time of the device in the low battery state.

[0040] Data demand instructions and the proxy model work together to form the basis for data scheduling at the edge. The proxy model is used to objectively quantify at the edge the marginal reduction capability (i.e., the base gain value) of a single data sample in terms of the uncertainty of cloud-based decision-making. iThe data requirement instruction is used to subjectively adjust the priority coefficients (i.e., weights wi and uncertainty levels) of different data types and regions in resource competition. When performing data filtering at the edge, the inference output of the proxy model is used as the base gain, and the weights in the data requirement instruction are used as the utility multipliers to jointly construct the objective function of the knapsack optimization problem, thereby achieving unified optimization of business needs and data value.

[0041] The metadata tags of the data subset to be uploaded shall include at least: data acquisition timestamp, data source sensor identifier, location information of the pipeline section to which the data belongs, and the basic gain value Gaini of the data sample.

[0042] S5 introduces a cloud-based feedback correction step with a feedback-driven convergence mechanism: After receiving a subset of the target uploaded data, the cloud performs incremental updates to the decision model and corrections to the uncertainty map, and uses the change magnitude of the uncertainty value and the estimated gain value to update the lightweight decision agent model to adjust the prediction bias of the information gain pre-evaluation at the edge in the next calibration.

[0043] Specifically as follows: Incremental model updates involve receiving data uploaded from the edge device in the cloud, incorporating the uploaded data into the online learning process of the decision model (such as using online stochastic gradient descent or Bayesian filtering), and updating the model parameters to reduce the uncertainty in that region.

[0044] Uncertainty map correction includes: the cloud recalculates the uncertainty value (predicted entropy Hnew) of the corresponding area based on the updated model and updates the map; if the recalculated uncertainty value decreases by no less than the preset effective threshold compared to the original value, it indicates that the uncertainty value of the area has decreased significantly and the data collection is valid; if the recalculated uncertainty value decreases by less than the preset effective threshold compared to the original value, it indicates that the decrease is insufficient and there is a deviation in the gain prediction at the edge. The cloud will introduce correction factors (such as adjusting the parameters of the gain prediction model) in the next proxy model to gradually correct the filtering logic at the edge.

[0045] The specific method for generating the correction factor in the cloud is as follows: Calculate the ratio η = ΔHactual / ΔHestimated, which is the actual decrease in uncertainty ΔHactual to the estimated decrease ΔHestimated. If η is less than a preset effective threshold, then generate the correction factor α = α. old ×(1+λ×(1-η)), where λ is the preset learning rate. The correction factor is embedded in the output layer of the surrogate model. The original gain prediction value output by the surrogate model is multiplied by the correction factor to obtain the corrected gain prediction value, which is used to correct the gain prediction value output by the surrogate model. This allows the prediction accuracy at the edge to continuously converge with the increase of iterations under the actual feedback drive, forming a closed-loop optimization.

[0046] This method is illustrated using the data scheduling application of gas pipeline crossing sections as an example: A city's gas pipeline network includes a section of pipeline that crosses a river. This section is approximately 500 meters long and is equipped with two pressure sensors (sampling frequency 500Hz), four vibration sensors (sampling frequency 2kHz), two gas concentration sensors (sampling frequency 5Hz), and one edge computing node. It communicates with the cloud via a 4G network.

[0047] Under normal operating conditions, the uncertainty map grid value in this area is low, no special data requirement instructions are triggered in the cloud, and the edge device executes data scheduling according to the default strategy.

[0048] After the meteorological station issued a rainstorm warning, the uncertainty of the output of the cloud-based geological disaster risk assessment sub-model for the crossing section increased significantly. The grid value of the area in the uncertainty map exceeded the preset threshold. The cloud generated a data request instruction: requiring the edge terminal to increase the sampling priority weight of the vibration sensor to 1.5 times (the default weight is 1.0) within the next 72 hours, keep the sampling priority weight of the pressure sensor at 1.0, and reduce the sampling priority weight of the gas concentration sensor to 0.5.

[0049] Upon receiving the instruction, the edge device dynamically adjusts the weights of the objective function of the knapsack optimization problem, giving vibration data a higher upload priority under resource-constrained conditions. When the communication signal attenuation due to heavy rain reduces the bandwidth margin to 60% of the normal value, the edge device automatically discards repetitive gas concentration data with low gain by solving the knapsack optimization problem, prioritizing the upload of vibration spectrum data. This ensures that critical information about pipeline stress anomalies induced by geological disasters is not overlooked.

[0050] This embodiment provides a cloud-edge-device collaborative data processing method for gas pipeline network safety monitoring. Example 2 Unlike Example 1, a data value prediction network is constructed in the cloud. This network includes accumulating historical uploaded data and their corresponding actual decision contributions, and training an independent data value prediction network to predict the long-term marginal value of different regions and types of data. The long-term marginal value and the uncertainty map serve as input signals for the cloud to generate data demand instructions. Using the long-term marginal value as an auxiliary signal to assist in the generation of the uncertainty map avoids the short-term error problem caused by relying solely on short-term uncertainty.

[0051] Specifically, actual decision contribution (training label): After historical uploaded data is absorbed by the cloud model, the actual reduction in uncertainty for that region is the difference between the uncertainty before the upload and the uncertainty after the upload and model update. The larger the difference, the more useful the data is. Long-term marginal value (predicted output): The data value prediction network (LSTM / Transformer) uses upload records from a historical estimated time period (e.g., 30 days) as a training set to learn "what kind of data can achieve a higher actual decision contribution under what spatiotemporal and environmental conditions," and then predicts the expected average decision contribution of each type of data in each region within the next 24 hours, as the long-term marginal value prediction value. In application, the cloud performs a weighted fusion of the long-term marginal value prediction value and the grid values ​​of the current uncertainty map to generate a comprehensive data demand score. When the comprehensive data demand score exceeds a preset fusion threshold, a data demand instruction is triggered.

[0052] Example 3 Unlike Examples 1 and 2, the communication environment along gas pipelines is complex, and communication blind spots or temporary interruptions may occur in remote pipeline sections. Therefore, a local independent scheduling degradation strategy is introduced at the edge end in the event of a communication interruption: When the edge device detects a continuous interruption in the communication link with the cloud (when the interruption duration exceeds a preset interruption threshold), the edge device switches to local independent scheduling mode and performs local independent scheduling. This includes performing a round of independent scheduling based on the most recently received cloud data request instruction, independently executing data caching, feature extraction, information gain pre-evaluation, redundancy filtering for representativeness determination, and resource constraint optimization upload steps. High-value data is cached locally (and local cached data is sorted from high to low value gain, with the lowest gain data being discarded first when local storage capacity reaches a preset limit). After the communication link is restored, the cached data is uploaded to the cloud in batches from high to low gain, with the upload data packet carrying the collection timestamp and cache duration tag. When the cloud merges the data, it applies an exponential decay weighting to the data value based on the cache duration. The decay coefficient is related to the time sensitivity of the data type. For example, leakage alarm data has high time sensitivity and a large decay coefficient, while long-term trend analysis data has low time sensitivity and a small decay coefficient.

[0053] This method is illustrated using a degradation handling application in a scenario where communication is interrupted at the edge as an example: An edge node in a remote pipeline section experienced a 31-minute communication outage with the cloud due to a base station malfunction. Upon detecting that the communication outage had persisted for more than a preset 30-minute threshold, the edge node switched to a local independent scheduling mode. The edge node then identified high-value data generated by the anomaly according to the rules in the most recent cloud data request instruction.

[0054] During the outage, an unusual event occurred in the area where third-party construction machinery approached the pipeline. Vibration sensors at the edge captured the abnormal vibration signal, and the information gain evaluator determined the gain of this data sample. i The value is eight times the daily average. The edge device stores this data sample, along with the vibration waveform data 30 seconds before and after it, in a local high-value cache.

[0055] After communication is restored, the edge device prioritizes uploading the high-value data packet. Upon receiving it, the cloud immediately triggers a leakage warning verification process. After the spatiotemporal graph neural network model confirms that there is no leakage, the event is recorded as a high-risk proximity event in the historical database for use in optimizing subsequent risk warning models.

[0056] The above descriptions are merely embodiments of the present invention. Commonly known structures and characteristics of the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent.

Claims

1. A cloud-edge-device collaborative data processing method for gas pipeline network safety monitoring, characterized in that, include: S1, based on the decision model deployed in the cloud, determines the prediction uncertainty values ​​of each spatial segment, each data type, and each time of the gas pipeline network, and generates a three-dimensional uncertainty map covering the entire pipeline network; S2: The cloud monitors the values ​​of each grid point in the uncertainty map in real time. When the uncertainty value of a certain area does not meet the preset conditions, a data demand instruction is generated and sent to the edge of the corresponding area through the communication network. The data requirement instruction includes the required data type, uncertainty value, sampling priority weight vector, and resource budget constraint parameters; S3: After receiving the data request instruction, the edge device uses a lightweight decision agent model pre-deployed in the cloud to pre-evaluate the information gain of data samples that belong to the required data type in the local cache data, and makes a representativeness determination based on a preset redundancy threshold that is dynamically adjusted based on the uncertainty value of the region to determine the target data sample. S4. At the edge, the target data samples are used as candidate sets. The knapsack optimization algorithm is used to calculate the estimated gain value based on the sampling priority weight vector and the information gain pre-evaluation result. The target upload data subset that maximizes the total information gain and meets the resource budget constraint parameter constraints is selected and uploaded to the cloud. S5 receives a subset of the target data uploaded to the cloud and performs incremental updates to the decision model and corrections to the uncertainty map. It then uses the magnitude of the change in uncertainty values ​​and the estimated gain values ​​to update the lightweight decision agent model to assess the prediction bias of the information gain pre-evaluation at the edge during the next calibration.

2. The cloud-edge-device collaborative data processing method for gas pipeline network safety monitoring according to claim 1, characterized in that, In S1, the cloud-based decision model adopts a gas pipeline network leakage prediction model or corrosion rate prediction model based on a spatiotemporal graph neural network. The input of the decision model is gas pipeline network topology data and multi-source sensor time series data, and the output is the risk probability prediction value of each pipe segment. At the same time, the model uses a Bayesian neural network, Monte Carlo dropout method or deep Gaussian process to enable it to output the uncertainty measure value of each prediction result during inference, as the prediction uncertainty value.

3. The cloud-edge-device collaborative data processing method for gas pipeline network safety monitoring according to claim 1, characterized in that, In S2, the uncertainty map grid value H(loc, t, feat) is used as the uncertainty value, which represents the degree of uncertainty in the cloud's perception of the current network state corresponding to the location loc at time t and the data type feat. It is calculated by the weighted sum of the prediction entropy and the posterior variance of the model parameters. The higher the uncertainty value, the more the cloud needs new data in this area to reduce the global prediction risk.

4. The cloud-edge-device collaborative data processing method for gas pipeline network safety monitoring according to claim 1, characterized in that, In S3, a lightweight decision agent model pre-deployed from the cloud is used to perform information gain pre-evaluation on data samples belonging to the required data type in the locally cached data to filter out low-gain data samples and obtain a high-gain data sample set. The information gain pre-evaluation includes extracting the feature vector xi of the i-th data sample, inputting the feature vector xi into the agent model, and calculating the expected posterior uncertainty value through Monte Carlo dropout or a first Bayesian forward propagation. The difference between the expected posterior uncertainty value and the current uncertainty value of the region in the cloud is used to define the reduction of the uncertainty of the cloud by the sample, which is used as the basic information gain value of the sample.

5. The cloud-edge-device collaborative data processing method for gas pipeline network safety monitoring according to claim 1, characterized in that, In S3, the surrogate model is obtained by compressing the cloud decision model through knowledge distillation. Its number of parameters does not exceed the preset ratio of the number of parameters in the cloud decision model. It is deployed on the neural grid processor or graphics processor at the edge and runs in real time. The cloud updates the surrogate model periodically or when the uncertainty map changes significantly and then distributes it to the edge.

6. The cloud-edge-device collaborative data processing method for gas pipeline network safety monitoring according to claim 1, characterized in that, In S3, the representativeness of the data samples that have been pre-evaluated and screened by information gain is determined, including: calculating the similarity distance between the feature vector xi of each data sample and the feature vectors of the N most recent uploaded data samples of the same type in the local cache; if the similarity distance is less than the preset redundancy threshold, the data sample is marked as redundant data and discarded, so as to filter the redundant data and determine the target data sample; the preset redundancy threshold is inversely correlated with the uncertainty value.

7. The cloud-edge-device collaborative data processing method for gas pipeline network safety monitoring according to claim 1, characterized in that, In S4, the resource budget constraint parameters include communication bandwidth margin and remaining power information. The edge device obtains the currently available communication bandwidth margin and remaining power information. Under the constraints of communication bandwidth and remaining power, the following knapsack optimization problem is solved: maxΣ(Gain i ×wi) s.t.ΣSize i ≤Bandwidth,ΣCost i ≤ Energy Among them, Gain i The basic information gain value of the i-th data sample is calculated for the information gain pre-evaluation; wi is the weight value of the data type corresponding to the i-th data sample in the sampling priority weight vector issued by the cloud; Size i Cost represents the amount of data uploaded for the i-th data sample. i The estimated energy consumption for uploading the i-th data sample is defined as follows: Bandwidth is the current available communication bandwidth, and Energy is the current remaining power.

8. The cloud-edge-device collaborative data processing method for gas pipeline network safety monitoring according to claim 1, characterized in that, In S5, the cloud recalculates the uncertainty value of the corresponding area based on the updated model and updates the map accordingly. If the recalculated uncertainty value decreases less than the preset effective threshold compared to the original value, it indicates that the decrease is insufficient and there is a deviation in the gain prediction at the edge. The cloud will introduce a correction factor in the next proxy model to correct the gain prediction result output by the proxy model.

9. The cloud-edge-device collaborative data processing method for gas pipeline network safety monitoring according to claim 1, characterized in that, By leveraging historical uploaded data accumulated in the cloud and their corresponding actual decision-making contributions, an independent data value prediction network is trained to predict the long-term marginal value of different regions and different types of data in the future. The long-term marginal value and the uncertainty map together serve as input signals for data demand instructions generated in the cloud.

10. The cloud-edge-device collaborative data processing method for gas pipeline network safety monitoring according to claim 1, characterized in that, In the event of a communication interruption, a local independent scheduling degradation strategy is introduced for the edge device. When the edge device detects a continuous interruption in the communication link with the cloud, it switches to the local independent scheduling mode to perform local independent scheduling. This includes independently executing S3 and S4 according to the most recently received cloud data request instruction, and caching the selected target upload data subset to local storage. After the communication link is restored, the data is uploaded to the cloud in a preset manner.

Images