Tank leakage prediction method based on large model multi-source data fusion

Abstract

The present application relates to the technical field of storage tank data analysis, and particularly relates to a storage tank leakage prediction method based on large model multi-source data fusion. The present application collects data transmission link parameters of multiple types of monitoring sensors, counts the signal conversion times of each type of sensing signal and records the time stamp, divides the sensing signal into several time segments on the edge side, and uploads the globally representative time segment to the cloud side for operation based on the spatial distribution of the sensor. In response to no abnormality in the preliminary operation of the cloud side, the data of the first and last time segments of each type of sensing signal is uploaded to the cloud side. The cloud side calls the tail end data of each type and the head end data of the remaining types for cross operation, and determines whether the storage tank state is abnormal according to the operation result. The present application differentiates and prioritizes the signals of different conversion links, quantifies the influence of time dislocation on fusion evaluation by cross-operation across time segments, and effectively balances the timeliness and accuracy of leakage prediction.

Description

Technical Field

[0001] This invention relates to the field of storage tank data analysis technology, and in particular to a method for predicting storage tank leaks based on the fusion of multi-source data from a large model. Background Technology

[0002] With the rapid development of the petrochemical, liquefied natural gas storage, and hazardous chemical storage industries, the safe operation monitoring of large storage tanks has become a core concern in the field of industrial safety. To ensure the structural integrity of storage tanks and prevent leakage accidents, existing technologies typically employ multi-type sensor monitoring systems, deploying various monitoring devices such as pressure sensors, acoustic emission sensors, temperature sensors, and liquid level sensors on the tank walls and key components. This allows for a comprehensive assessment of the tank's operating status by collecting multi-source heterogeneous sensor signals. In recent years, with the increasing application of large-scale modeling technology in industrial data analysis, some advanced monitoring systems have begun to upload multi-source sensor data to the cloud. Leveraging the cross-domain feature extraction and nonlinear fitting capabilities of large-scale models, they aim to comprehensively assess leakage risks, thereby improving the accuracy and robustness of predictions.

[0003] However, the differences between various sensors in terms of signal acquisition principles, analog-to-digital conversion accuracy, and sampling frequency are significant. This leads to a natural misalignment in the timestamps of various sensor signals when completing analog-to-digital conversion and generating transmittable data at the edge, resulting in application defects in actual engineering deployments.

[0004] For example, Chinese invention patent CN119984694A discloses a method for analyzing the diffusion of hydrogen leaks in high-pressure storage tanks. This method includes: constructing a physical model to accurately capture the physical properties of hydrogen and its interaction with the environment, simulating the diffusion process under different leakage scenarios; deploying sensors to monitor key parameters in real time, ensuring the real-time nature and reliability of the data, enabling rapid response to potential leakage events and reducing the likelihood and consequences of accidents; establishing a mapping relationship between real-time monitoring data and physical model parameters, allowing for dynamic adjustment of model parameters based on real-time data to optimize output results; and generating corresponding risk warning signals through quantitative assessment and risk level classification.

[0005] Existing technologies do not consider the timeliness inaccuracies caused by differences in signal conversion links of multi-source sensor signals. They cannot perform edge-side traffic splitting and priority scheduling for signals from different conversion links, nor can they quantify the impact of time misalignment on the fusion evaluation results through cross-time period cross-operation. As a result, it is difficult to balance the timeliness and accuracy of leakage prediction. Summary of the Invention

[0006] To address this, the present invention provides a tank leakage prediction method based on large-scale multi-source data fusion, which overcomes the problems of existing technologies that fail to consider the timeliness inaccuracy caused by differences in signal conversion links of multi-source sensor signals, cannot perform edge-side diversion and priority scheduling of signals from different conversion links, and cannot quantify the impact of time misalignment on the fusion evaluation results through cross-time period cross-operation, thus making it difficult to balance the timeliness and accuracy of leakage prediction.

[0007] To achieve the above objectives, this invention provides a method for predicting tank leakage based on large-scale model multi-source data fusion, comprising: The data transmission link parameters of various monitoring sensors in the storage tank are pre-collected at the edge, the number of signal conversions for each type of sensor signal is counted, and the completion timestamp of each signal conversion is recorded. The edge end sorts the various sensor signals according to the number of signal conversions, and performs edge-side decoupling and diversion on the various sensor signals according to the sorting order. This includes dividing a single type of sensor signal into several different time segments according to the completion timestamp, and selecting representative time segments of the whole domain based on the spatial distribution status of the monitoring sensors corresponding to the sensor signals in each time segment. The edge device prioritizes uploading data representing the entire time segment to the cloud for computation. In response to the cloud's initial assessment that the tank status is normal, the edge device filters the data from the first and last two time segments of each type of sensor signal and uploads and stores the data to the cloud. The cloud sequentially retrieves data from the tail time segment of various types of sensor signals and performs cross-fusion calculations with the data from the beginning segment of other types of sensor signals to obtain multiple sets of calculation results. The multiple sets of calculation results are then compared for consistency, and the tank status is determined based on the comparison results.

[0008] Furthermore, the process of pre-collecting data transmission link parameters of various monitoring sensors in the storage tank at the edge and counting the number of signal conversions for each type of sensor signal includes: Obtain the signal acquisition link topology of various monitoring sensors, identify the number of nodes in the signal acquisition link topology that convert analog signals to digital signals, and record the number of nodes as the number of signal conversions.

[0009] Furthermore, the monitoring sensors include a pressure sensor, an acoustic emission sensor, a temperature sensor, and a liquid level sensor; the number of signal conversions covers all signal conversion steps under the corresponding monitoring index.

[0010] Furthermore, the process of sorting the edge signals based on the number of signal conversions for various types of sensor signals includes: All categories of sensor signals are initially sorted according to the number of signal conversions from high to low. If there are multiple types of sensor signals with the same number of signal conversions, then the signals are sorted in descending order of their historical average transmission delay time to generate a sorting order.

[0011] Furthermore, the process of dividing a single type of signal into several different time segments based on the completion timestamp at the edge includes: Using the completion timestamp as the dividing boundary, the single-type sensor signal is divided into different time segments at equal time intervals according to the order of time; The earlier the sorting order, the more time segments are divided.

[0012] Furthermore, the process of determining the spatial distribution of monitoring sensors corresponding to the sensing signals within each time period at the edge includes: The entire space of the storage tank is divided into several grid units, and the number of sensors falling into each grid unit is counted. Calculate the mean and standard deviation of the number of sensors in all grid cells, and determine the spatial distribution state value by the ratio of the standard deviation to the mean.

[0013] Furthermore, the process of filtering the representative time segment of the entire domain at the edge includes: The spatial distribution status values ​​of the monitoring sensors within each time period were obtained respectively; The time segment with the smallest spatial distribution state value is selected as the representative time segment of the entire domain.

[0014] Furthermore, the process of uploading and storing data from the first and last two time segments of each type of sensor signal at the edge to the cloud includes: From several time segments of each type of sensor signal, the time segment with the earliest timestamp is selected as the first time segment, and the time segment with the latest timestamp is selected as the last time segment. Generate a beginning data packet and a end data packet for the data in the beginning time segment and the end time segment; After associating the first-end data packet and the last-end data packet with the corresponding sensor signal category identifier and timestamp index, they are uploaded to cloud storage.

[0015] Furthermore, the process of sequentially retrieving data from the tail time segments of various types of signals and the beginning time segments of other types of signals in the cloud for calculation includes: Based on the category identifier and timestamp index of the sensor signals stored in the cloud; In the i-th round of operation, the data of the tail time segment of the i-th type of sensor signal and the data of the head time segment of the other N-1 types of sensor signals are retrieved to form the i-th heterogeneous data set, i=1,2,...,N; Perform cross-fusion operation on the i-th heterogeneous data set to obtain the i-th operation result, traverse i=1 to N to obtain N operation results, and form an evaluation set.

[0016] Furthermore, the process of determining whether there is an anomaly in the tank's condition based on the comparison of several calculation results includes: Calculate the dispersion index of N sets of operation results in the evaluation set, the dispersion index includes the range value and the standard deviation of the operation result; compare the range value with the preset range reference value, and compare the standard deviation of the operation result with the preset standard deviation reference value; If the range value is greater than the range reference value, or the standard deviation of the calculation result is greater than the standard deviation reference value, the cloud determines that the storage tank status is abnormal and triggers an early warning signal. If the range value is less than or equal to the range reference value, and the standard deviation of the calculation result is less than or equal to the standard deviation reference value, then the cloud determines that the storage tank is in normal condition.

[0017] The beneficial effects of the technical solution presented in this application include: by pre-statistically counting the number of signal conversions of various sensor signals and sorting them according to conversion complexity, the edge side achieves differential decoupling and diversion of sensor signals from different links, enabling signals with multiple conversion links and large delays to obtain more granular time segment divisions. This allows for the priority uploading of key segment data while ensuring data representativeness, effectively alleviating the cloud time asynchrony problem caused by differences in analog-to-digital conversion links in multi-source heterogeneous data; by selecting the most uniformly distributed sensor spatial distribution across the entire domain representative time segment for the first round of calculation, the uploaded data ensures spatial coverage and representativeness of the overall state of the storage tank, avoiding interference from local sensor data deviations on the fusion evaluation; after initially determining that there are no anomalies, by cross-calculating the first and last time segments and comparing risk scores, the time misalignment between different types of signals is quantified into a measurable evaluation deviation, enabling the large model to identify anomaly precursors caused by time asynchrony, thereby significantly enhancing the accuracy and robustness of multi-source data fusion while improving the timeliness of prediction.

[0018] Furthermore, by pre-collecting the signal acquisition link topology of various monitoring sensors and identifying the number of nodes that convert analog signals to digital signals as the number of signal conversions, the system can quantify the time cost caused by differences in hardware architecture among different types of sensors.

[0019] Furthermore, this invention divides a single type of sensor signal into several time segments at equal time intervals based on the completion timestamp, and the earlier the sorting order, the more time segments are divided, thus realizing time segmentation of high-latency signal data streams. By segmenting high-latency signals into finer-grained time segments, the system can filter and upload segments that have been converted before all data is ready, exchanging incremental computing at the edge for improved timeliness of cloud-based fusion computing, thereby optimizing the concurrency and real-time performance of data transmission.

[0020] Furthermore, this invention employs a tiered triggering strategy: first, upload representative time segments across the entire domain for preliminary calculations; then, after an initial assessment confirms no anomalies, upload the first and last time segments. This achieves a dynamic balance between communication bandwidth, cloud computing resources, and prediction accuracy. In the first stage, only selected representative segment data is uploaded, resulting in a small data volume and high regional coverage, enabling a preliminary assessment of the tank's status in the shortest possible time and optimizing early warning timeliness. In the second stage, the first and last segments retain key information with the largest time span for cross-calculation, satisfying the need for refined prediction while controlling the total data volume. This achieves a balance between timeliness and prediction accuracy.

[0021] Furthermore, this invention completes multi-round cross-fusion by rotating and combining the first and last time segments of various sensor signals. Based on the range and standard deviation of the evaluation set, it quantifies the dispersion of leakage risk results of each group and compares them with a preset steady-state reference threshold. The range can capture extreme evaluation offsets caused by time lag of a single type of signal, and the standard deviation can measure the global coordination consistency of all monitoring indicators. The dual-indicator linkage judgment can accurately distinguish anomalies with asynchronous time sequences. Combined with the hierarchical transmission strategy of prioritizing the uploading of representative segments of the entire domain at the edge for rapid initial judgment, it significantly shortens the early warning response latency, improves the predictive timeliness, reduces the probability of false alarms caused by time sequence misalignment, and enhances the adaptability of large-scale model multi-source data fusion evaluation under various working conditions. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the steps of the tank leakage prediction method based on large-model multi-source data fusion in an embodiment of the present invention. Figure 2 This is a diagram illustrating the steps of uploading and storing data from the first and last two time segments to the cloud at the edge of an embodiment of the present invention. Figure 3 This diagram illustrates the steps of retrieving data from the beginning and end time segments in the cloud for computation, as described in an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0024] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0025] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this disclosure should have the ordinary meaning understood by one of ordinary skill in the art to which this disclosure pertains. The terms "first," "second," and similar terms used in the embodiments of this disclosure do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0026] Please see Figure 1 The diagram illustrates the steps of a tank leakage prediction method based on large-model multi-source data fusion according to an embodiment of the present invention. The tank leakage prediction method based on large-model multi-source data fusion of the present invention includes: It is important to note that in the actual monitoring deployment of large storage tanks, there are significant topological differences in the signal acquisition links of different types of monitoring sensors. Taking pressure sensors as an example, they typically employ strain gauge or piezoresistive principles, outputting weak millivolt-level analog signals. These signals must pass through a preamplifier circuit, a low-pass filter circuit, a signal isolation circuit, and an analog-to-digital converter (ADC) before being converted into a transmittable digital signal. If the storage tank monitoring scale is large and a multiplexed acquisition architecture is used, analog switching is also required, further increasing the conversion stage. Acoustic emission sensors are used to capture high-frequency stress waves generated by material crack propagation. Their signal conditioning links are even more complex, requiring multiple stages of processing, including a preamplifier, a bandpass filter, an envelope detector, and a high-speed ADC. In contrast, some temperature or level sensors may only require one or a few signal conversions to complete the data digitization. Each additional analog-to-digital conversion node introduces not only hardware processing delays but also cumulative jitter due to asynchronous clock sources at different nodes. This causes the timestamp for the generation of transmittable data packets at the edge of this type of sensor signal to lag significantly behind other types. In this invention, the number of signal conversions quantifies the process from physical quantity acquisition to data readiness for this type of sensor signal. By pre-acquiring the signal acquisition link topology and identifying the number of analog-to-digital conversion nodes, the system can accurately assess the time readiness differences of various signals before data fusion, providing an objective quantitative basis for differentiated processing of high-latency signals and preventing implicit time misalignments caused by link differences from being directly carried into cloud fusion computation.

[0027] Step S100: The edge end pre-collects the data transmission link parameters of multiple monitoring sensors of the storage tank, counts the number of signal conversions for each type of sensor signal, and records the completion timestamp of each signal conversion; Step S200: The edge end sorts the various types of sensor signals according to the number of signal conversions, and performs edge-side decoupling and diversion on the various types of sensor signals according to the sorting order. This includes dividing a single type of sensor signal into several different time segments according to the completion timestamp, and selecting a representative time segment of the whole domain based on the spatial distribution status of the monitoring sensors corresponding to the sensor signals in each time segment. In step S300, the edge device prioritizes uploading data representing the entire time segment to the cloud for computation. In response to the cloud's preliminary assessment that there are no abnormalities in the tank status, the edge device filters the data from the first and last two time segments of each type of sensor signal and uploads and stores the data to the cloud. In step S400, the cloud sequentially retrieves the data from the tail time segment of various types of sensor signals and performs cross-fusion calculations with the data from the head time segment of other types of sensor signals to obtain multiple sets of calculation results. The multiple sets of calculation results are then compared for consistency, and the tank status is determined based on the comparison results.

[0028] In this embodiment of the invention, when the cloud performs a preliminary assessment of the tank status based on data representing the entire time period, the following directly implementable rule engine is used for rapid threshold determination: the time-domain statistical characteristics of each type of sensor signal are compared item by item with the preset normal baseline range. The arithmetic mean of pressure must be within ±5% of the tank's design working pressure; the arithmetic mean of temperature must be within ±10°C of the ambient temperature; the effective value of acoustic emission must be lower than the preset threshold of 0.5V; and the arithmetic mean of liquid level must be within 20% to 90% of the tank's safe liquid level. If all dimensions are within the corresponding baseline range, the cloud determines that the tank status is normal and sends a "normal" status to the edge, triggering the upload of the first and last segments of the second stage; if any dimension exceeds the baseline range, it is immediately determined to be abnormal, and an early warning is directly triggered.

[0029] In the online tank monitoring process of this invention, communication bandwidth and cloud computing resources are limited. Uploading data from all time segments of all sensors without differentiation would not only cause network congestion but also put immense pressure on the cloud-based large model to process massive amounts of redundant data. Therefore, in the first stage, only data from a selected representative time segment of the entire domain is uploaded to the cloud for preliminary calculation. This segment has a small data volume and high regional coverage, enabling a preliminary judgment of the tank's status in the shortest possible time. If the preliminary calculation determines an anomaly, the system can immediately trigger an early warning, thereby maximizing the timeliness of the warning. The second stage only starts when the preliminary calculation determines no anomalies, uploading data from the beginning and end time segments of each type of sensor signal. This retains information on the maximum time span used for cross-calculation while controlling the total amount of data, achieving a balance between timeliness and predictive accuracy in engineering.

[0030] Specifically, the process of pre-collecting data transmission link parameters of various monitoring sensors in the storage tank at the edge and counting the number of signal conversions for each type of sensor signal includes: Obtain the signal acquisition link topology of various monitoring sensors, identify the number of nodes in the signal acquisition link topology that convert analog signals to digital signals, and record the number of nodes as the number of signal conversions.

[0031] This document provides an example of counting the number of signal conversions for each type of sensor signal. A link topology scanning module can be built into the edge protocol parsing gateway. The configuration registers of each sensor and its corresponding acquisition module can be read through standard industrial protocols to obtain the number of channels, filter order, amplifier gain bits, A / D conversion chip model, and channel enable status. The gateway has a pre-built link topology knowledge base that stores standard link templates for typical sensors. For example, for a strain gauge pressure sensor, the data transmission link is: preamplifier, second-order low-pass filter, photoelectric signal isolation, analog-to-digital converter; the number of nodes is 4, and the number of signal conversions is 4. For the acoustic emission sensor, the data transmission link is as follows: preamplifier, eighth-order bandpass filter, envelope detector, high-speed analog-to-digital converter; the number of nodes is 3, and the number of signal conversions is 3. For a temperature sensor, the data transmission link consists of: signal conditioning and single analog-to-digital conversion; the number of signal conversions is 1. For a liquid level sensor, the data transmission link consists of: signal shaping and single analog-to-digital conversion; the number of signal conversions is 1.

[0032] Specifically, the monitoring sensors include pressure sensors, acoustic emission sensors, temperature sensors, and liquid level sensors; the number of signal conversions covers all signal conversion steps under the corresponding monitoring index.

[0033] Specifically, the process of sorting the edge signals based on the number of signal conversions for various types of sensor signals includes: All categories of sensor signals are initially sorted according to the number of signal conversions from high to low. If there are multiple types of sensor signals with the same number of signal conversions, then the signals are sorted in descending order of their historical average transmission delay time to generate a sorting order.

[0034] In this invention, the edge-end protocol parsing gateway records the transmission delay of each type of sensor signal in each continuous monitoring cycle. The time difference is calculated from the moment the edge sends a sampling trigger command to the corresponding sensor or acquisition module to the moment the sensor signal passes through all link conversion nodes, forms a complete and valid data frame, and is received and parsed by the gateway. The time difference is continuously recorded for the most recent 50 monitoring cycles, and the arithmetic mean is taken as the historical average transmission delay time of that type of sensor signal. This is stored in the edge-end local configuration table and updated periodically.

[0035] Those skilled in the art should understand that sorting signals by the number of conversions from highest to lowest essentially establishes a mapping relationship between signal link complexity and time delay levels. For example, in engineering scenarios for monitoring tank leaks, acoustic emission signals are extremely sensitive to crack propagation, but their high-frequency and high-dynamic characteristics determine that their signal conditioning links are complex and have many conversion nodes; pressure signals also involve multi-stage amplification and filtering, resulting in a large link depth. If the edge device does not prioritize the identification and processing of these high-latency signals, but instead waits for all types of data to arrive synchronously before uploading them uniformly, the waiting time for cloud-based fusion computing will be determined by the slowest link, affecting the timeliness of leak warnings. In addition, the physical response speed and transmission protocol of sensor signals may also differ: for example, temperature changes are extremely slow in a tank under steady-state conditions, and although their historical average transmission delay is long, the real-time requirements are low; while acoustic emission signals change rapidly, and although their historical delay is similar to that of temperature signals, they require faster processing and response. Therefore, introducing historical average transmission delay for secondary sorting can more accurately reflect the real time asynchrony of various signals in the actual deployment environment, ensuring that edge computing and transmission resources are tilted towards the most lagging signals, thereby eliminating structural delay differences caused by link differences before multi-source heterogeneous data are incorporated into the cloud.

[0036] Specifically, the process of dividing a single type of signal into several different time segments based on the completion timestamp at the edge includes: Using the completion timestamp as the dividing boundary, the single-type sensor signal is divided into different time segments at equal time intervals according to the order of time; The earlier the sorting order, the more time segments are divided.

[0037] This provides an example of time segmentation. If a storage tank has N types of monitoring sensor signals deployed, after the edge device completes the sorting, it allocates M time segments to the sensor signal with sorting sequence number k. k Where k = 1, 2, ..., N, k = 1 indicates that the signal has the most transitions; The quantization mapping relationship is: M k =max(2, N-k+2); For example, when N=4, the acoustic emission signal ranked 1st is divided into M1=max(2,4-1+2)=5 time segments; the pressure signal ranked 2nd is divided into M2=4 time segments; the temperature signal ranked 3rd is divided into M3=3 time segments; and the liquid level signal ranked 4th is divided into M4=2 time segments. Each segment is divided according to the time interval of the completion timestamp, and the segment length Δt k =T / M k Where T is the total effective data duration of this type of sensor signal within the current monitoring period, which is uniformly set to 60 seconds in the implementation of this invention.

[0038] Understandably, dividing a single type of sensor signal into several time segments based on its completion timestamp is essentially time-segmenting a continuous data stream. For sensor signals ranked high, if a batch upload strategy is adopted, it is necessary to wait for all signals of that type to be converted before they can be packaged together, causing the cloud fusion computation time to be limited by the slowest link. By dividing these signals into more time segments with equal time intervals, the system can filter and upload segments that have already been converted before all data is ready, achieving simultaneous conversion, filtering, and transmission. The number of segments is negatively correlated with the sorting order; that is, signals with more conversions are segmented more finely. The significance lies in using incremental computation at the edge to improve the timeliness of cloud fusion computation. This processing strategy ensures that high-latency signals do not block the entire multi-source fusion process, prioritizing the push of completed and representative segments to the cloud, maximizing the concurrency and timeliness of data transmission.

[0039] Specifically, the process of determining the spatial distribution of monitoring sensors corresponding to sensing signals within each time interval at the edge includes: The entire space of the storage tank is divided into several grid units, and the number of sensors falling into each grid unit is counted. Calculate the mean and standard deviation of the number of sensors in all grid cells, and determine the spatial distribution state value by the ratio of the standard deviation to the mean.

[0040] This provides an embodiment of dividing the overall space of a storage tank into several grid cells, dividing the total height H of the storage tank into H equal parts along the tank's axial direction. n The system consists of four horizontal layers; the circumference is divided into eight sectors at 45° angles, forming 40 equal-volume grid cells. Each grid cell in cylindrical coordinates covers the following spatial dimensions: radially, it covers the tank wall thickness; axially, it covers H / H... n The height range covers a 45° fan-shaped area.

[0041] When calculating the spatial distribution state values, all grid cells are included in the statistics. Let the total number of grid cells be G, and the number of sensors in each grid cell be n. i If i = 0, 1, 2, ..., G, then the mean quantity μ n =(Σn i ) / G, quantity standard deviation Spatial distribution state value S=σ n / μ n .

[0042] The physical meaning of the S-value is the uniformity of sensor distribution across the entire storage tank space: when the S-value approaches 0, it indicates that the number of sensors in each grid cell is close to the average, and the sensors are uniformly distributed across the storage tank space. The data in this time period can represent the spatial average state of this indicator across the entire storage tank. When the S-value is large, it indicates that sensors are over-clustered in some grid cells and missing in others, resulting in blind spots in spatial coverage. The data in this time period may lose its representativeness of the overall state of the storage tank due to local oversampling or missing areas. Therefore, the time period with the smallest spatial distribution state value is selected as the representative time period for the entire area to ensure that the initial judgment data uploaded to the cloud comes from the sampling period with the most uniform sensor spatial distribution and the most representative overall state of the storage tank.

[0043] Specifically, the process of filtering the entire representative time segment at the edge includes: The spatial distribution status values ​​of the monitoring sensors within each time period were obtained respectively; The time segment with the smallest spatial distribution state value is selected as the representative time segment of the entire domain.

[0044] Understandably, after dividing the overall space of the storage tank into several grid cells, the ratio of the standard deviation to the mean of the number of sensors in each cell is calculated. In the storage tank monitoring environment, the smaller this ratio, the wider the data coverage and the stronger the ability to represent changes in the overall state. Selecting the time segment with the smallest spatially distributed state value as the representative time segment for the entire area essentially ensures data completeness, guaranteeing that the initial judgment data uploaded to the cloud can cover all locations of the storage tank.

[0045] Specifically, please refer to Figure 2 The diagram illustrates the steps of uploading and storing data from the first and last two time segments to the cloud at the edge of an embodiment of the present invention. The process of uploading and storing data from the first and last two time segments for each type of sensor signal at the edge includes: Step S301: From several time segments of each type of sensor signal, select the time segment with the earliest timestamp as the first time segment and select the time segment with the latest timestamp as the last time segment. Step S302: Generate a beginning data packet and a end data packet for the data in the beginning time segment and the end time segment; Step S303: After associating the first-end data packet and the last-end data packet with the category identifier and timestamp index of the corresponding sensor signal, upload them to cloud storage.

[0046] In this invention, the head-end and tail-end data packets adopt a unified format, containing the following fields: sensor_type: sensor type identifier, enumerated values ​​are "AE" (acoustic emission), "P" (pressure), "T" (temperature), "L" (liquid level); segment_type: segment location identifier, enumerated values ​​are "head" (head end) or "tail" (tail end); start_timestamp: segment start timestamp, millisecond-level Unix timestamp; end_timestamp: segment end timestamp; feature_vector: 1×5-dimensional double-precision floating-point array, in the order of [arithmetic mean, standard deviation, maximum value, minimum value, effective value]; spatial_distribution_value: spatial distribution state value S calculated for this segment; sensor_count: number of valid sensors in this time period. The cloud-based time-series database establishes a composite secondary index based on sensor_type and segment_type, and the cloud-based judgment module accurately retrieves the corresponding segment data through combined queries.

[0047] Specifically, please refer to Figure 3 The diagram illustrates the steps of retrieving data from the beginning and end time segments in the cloud for computation according to an embodiment of the present invention. The process of sequentially retrieving data from the end time segments of various types of signals and the beginning time segments of other types of signals in the cloud for computation includes: Step S401: Based on the category identifier and timestamp index of the sensor signal stored in the cloud; Step S402: In the i-th round of calculation, retrieve the data of the tail time segment of the i-th type of sensor signal and the data of the first time segment of the other N-1 types of sensor signals, and form the i-th heterogeneous data set, i=1, 2, ..., N; Step S403: Perform cross-fusion operation on the i-th heterogeneous data set to obtain the i-th operation result, traverse i=1 to N to obtain N operation results, and form an evaluation set.

[0048] In this embodiment of the invention, the cloud performs fusion calculations on the uploaded data to obtain an assessment result of the storage tank's status. The fusion calculations can employ any multi-source data fusion algorithm well-known to those skilled in the art, including but not limited to rule-based threshold fusion judgment, statistical learning-based fusion evaluation, or deep neural network-based fusion modeling. When the cloud performs a preliminary assessment of the storage tank's status based on data representing a full-domain time period, a lightweight rule engine can be used for rapid threshold determination: the temporal statistical features of each type of sensor signal are compared item by item with a preset normal baseline range. If all dimensions are within the corresponding baseline range, the storage tank's status is determined to be normal; if any dimension exceeds the baseline range, it is immediately determined to be abnormal and an early warning is triggered.

[0049] During the cross-fusion operation phase, the cloud performs a fusion operation on the i-th heterogeneous data set to generate the i-th operation result. The specific implementation of the fusion operation is performed by a leakage risk assessment model deployed in the cloud. This invention does not limit the specific algorithm architecture of the assessment model, but only requires that the assessment model has the ability to quantify the leakage risk of the input multi-source heterogeneous data.

[0050] In this invention, time-domain statistical features are extracted for each time segment of each type of sensor signal. These time-domain statistical features include at least the arithmetic mean, standard deviation, maximum value, minimum value, and effective value. If a certain type of sensor signal contains data from multiple sensors within the same time segment, the aforementioned time-domain statistical features are first extracted from the data of all sensors within that time segment. Then, the spatial arithmetic mean is taken for the same statistical measure of the same type of sensor to obtain the representative features of that type of signal in that time segment. The representative features of N types of sensor signals are then combined sequentially according to their sorting order to form the input data for the fusion operation.

[0051] This provides an example of cloud-based cross-computation. The cloud-based cross-fusion operation on the i-th heterogeneous data set employs a weighted fusion algorithm based on physical thresholds. Specifically, it includes calculating the anomaly deviation index for representative features of each type of sensor signal, using the formula d. j =|f j -f j0 | / f j0 , where f j f is the effective value of the j-th type of sensor signal in the characteristic represented by that time interval. j0 The historical reference effective value of this type of signal under normal steady state is given; the weight of each type of signal is set as w. j Preferably, the weights of w1=0.4 for the acoustic emission sensing signal, w2=0.3 for the pressure sensing signal, w3=0.2 for the temperature sensing signal, and w4=0.1 for the liquid level sensing signal are summed to 1; the i-th group of calculation results R i =Σ(w j ×d j ) / Σw j , will R i Normalized to the interval [0, 1], this is used as the quantified leakage risk value for that round; the effective value in this invention refers to the root mean square value, which is the value for a time series {x} collected by a sensor within a certain time period. j}, j=1,2,…,M; M is the number of sampling points of the sensor in this time interval. This invention only uses the above weighted fusion as an example. In actual deployment, it can also be replaced by other algorithms with multi-source data fusion and risk quantification output capabilities, such as DS evidence theory, fuzzy comprehensive evaluation, or deep neural networks. The relevant algorithms are well known and will not be described in detail here.

[0052] In the i-th round of computation of this invention, the cloud retrieves data from the tail time segment of the i-th type of sensor signal, i.e., the segment with the latest timestamp in that type of signal, and simultaneously retrieves data from the beginning time segments of the remaining N-1 types of sensor signals, i.e., the segments with the earliest timestamps in each type of signal, together forming the i-th heterogeneous data set. For example, when N=4 and the sorting is [acoustic emission, pressure, temperature, liquid level]: Round 1: Data from the acoustic emission tail time segment + data from the pressure head time segment + data from the temperature head time segment + data from the liquid level head time segment; Round 2: Data from the acoustic emission head time segment + data from the pressure tail time segment + data from the temperature head time segment + data from the liquid level head time segment; Round 3: Data from the first time segment of acoustic emission + data from the first time segment of pressure + data from the last time segment of temperature + data from the first time segment of liquid level; Round 4: Data from the acoustic emission head time segment + data from the pressure head time segment + data from the temperature head time segment + data from the liquid level tail time segment.

[0053] In each round, the above four types of data are combined in a fixed order and input into the cloud, and the corresponding calculation result R is output for that round. i The numbers i = 1, 2, 3, 4 are used to generate N sets of operation results, forming an evaluation set {R1, R2, R3, R4}; the operation result R... i This is a dimensionless quantification value for leakage risk, with its range uniformly normalized to [0, 1].

[0054] In this invention, the cloud-based large-scale model needs to undergo offline calibration before deployment: Multi-source monitoring data of the storage tank under historical normal operating conditions and known pre-leakage states are collected. Input data is generated according to steps S200-S400 and labeled with status tags, with normal operating conditions marked as 0 and abnormal operating conditions marked as 1. The input data and status tags are used to train or calibrate the parameters of the cloud-based large-scale model until the evaluation error on the cloud-based large-scale model validation set converges to a preset threshold range. This invention does not limit the specific training algorithm, network architecture, or optimizer type of the cloud-based large-scale model.

[0055] In this invention, during the actual operation of the storage tank, for the i-th type of sensor signal, its tail time segment represents the data with the latest signal conversion and the most delayed timestamp within the current monitoring period; while the head time segment of the remaining N-1 types of sensor signals represents the data with the earliest conversion and the earliest timestamp among these signals. Combining the two constitutes the heterogeneous data set with the largest time span among all possible pairings within the monitoring period. Since the topological differences in the signal conversion link cannot be eliminated in engineering deployment, this maximum time span objectively exists among multi-source heterogeneous data.

[0056] By performing N rounds of traversal calculations, where each type of sensor signal takes turns as the tail end and the remaining signals are fixed as the head end, the system can comprehensively cover scenarios where each type of signal has served as the most lagging data source. When the storage tank is in a thermodynamic and structural mechanical steady state, regardless of the time span between different types of data, the tank state reflected by different physical quantities such as pressure, temperature, acoustic emission, and liquid level should remain internally consistent, and the leakage risk quantification values ​​output by each round of fusion calculation should tend to be consistent. However, when the storage tank shows signs of leakage, the response speeds of different physical quantities differ significantly: acoustic emission signals respond extremely quickly to crack propagation, pressure signals change gradually with the leakage of the medium, and temperature and liquid level signals have greater thermal and volumetric inertia. This difference in physical response, coupled with time asynchrony, will lead to significant deviations in the calculation results of one or more rounds of cross-combination. Therefore, the traversal cross-fusion calculation is not a simple data permutation and combination, but rather transforms the physical constraint of time misalignment into observable evaluation fluctuations, enabling the system to identify abnormal precursors caused by time asynchrony.

[0057] Specifically, the process of determining whether there is an anomaly in the storage tank's condition based on the comparison of several calculation results includes: Calculate the dispersion index of N sets of operation results in the evaluation set, the dispersion index includes the range value and the standard deviation of the operation result; compare the range value with the preset range reference value, and compare the standard deviation of the operation result with the preset standard deviation reference value; If the range value is greater than the range reference value, or the standard deviation of the calculation result is greater than the standard deviation reference value, the cloud determines that the storage tank status is abnormal and triggers an early warning signal. If the range value is less than or equal to the range reference value, and the standard deviation of the calculation result is less than or equal to the standard deviation reference value, then the cloud determines that the storage tank is in normal condition.

[0058] In the implementation of this invention, data on the steady-state operation of the storage tank without leakage for 72 consecutive hours are collected. Multiple evaluation sets are generated according to the cross-operation process of this invention. The range and standard deviation of all evaluation sets are calculated, and the 95th percentile is taken as the initial range reference value and the initial standard deviation reference value.

[0059] In engineering practice, when a storage tank is operating normally in a steady state, its physical parameters change slowly. Even if there are time misalignments in the various sensors due to differences in the data links, the fusion evaluation results should be concentrated within a small range. At this time, both the range and the standard deviation of the calculation results are at a low level. When the storage tank shows signs of leakage, at least one type of sensor signal will first capture the abnormal change on the time axis, causing the cross-fusion calculation results of the corresponding round to deviate significantly from those of other rounds, which is manifested as an increase in the range. If the physical responses among the multi-source data show overall inconsistency, or if the storage tank is in a non-steady-state transition phase, the overall fluctuation of the N sets of calculation results will intensify, which is manifested as an increase in the standard deviation of the calculation results.

[0060] By comparing the range value with a preset range reference value and the standard deviation of the calculation result with a preset standard deviation reference value, the system statistically quantifies the physical process of fusion evaluation of the impact of time misalignment. If the range value is greater than the range reference value, or the standard deviation of the calculation result is greater than the standard deviation reference value, it is determined that the fusion evaluation results of multi-source heterogeneous data under the condition of maximum time asynchrony are significantly inconsistent, indicating an anomaly in the tank status and triggering an early warning signal; otherwise, it is determined that the results of each round of calculation are temporally consistent, and the tank status is not abnormal. This judgment mechanism does not rely on the internal parameters of a single fusion model, but can achieve anomaly identification only through the statistical dispersion of the results of multiple rounds of cross-calculation, thereby significantly enhancing the accuracy and robustness of multi-source data fusion while improving the timeliness of prediction.

[0061] This embodiment provides a computer-readable storage medium storing computer program code. When the computer program code is run on a computer, the computer executes the tank leakage prediction method based on large model multi-source data fusion provided in the above embodiment.

[0062] This document provides an embodiment of the cloud-based large model data input and processing process, where the input features include 5-dimensional time-domain statistical features of each of the four types of sensors; The hidden layer is set with 3 non-linear activation layers, and the output layer outputs a normalized leakage risk quantification value in the range of 0 to 1; the offline calibration dataset contains two types of samples: samples of normal steady-state operation of the storage tank, labeled 0; and samples of artificially simulated micro-leakage and early signs of tank wall cracks, labeled 1; after training, the model weights are fixed and deployed in the cloud inference module. Cloud-based inference call logic: Each set of heterogeneous data is uniformly organized into fixed-dimensional feature vectors and then input into a large model in batches. A single inference outputs a single set of risk values.

[0063] Specifically, computer-readable storage media include, but are not limited to: read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), electrically erasable programmable read-only memory (EEPROM), flash memory, solid-state drive (SSD), optical disc (CD-ROM), digital versatile optical disc (DVD) or other optical storage media, magnetic tape, disk storage or other magnetic storage media, and any medium that can be used to store the computer program code.

[0064] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0065] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting tank leakage based on large-scale model multi-source data fusion, characterized in that, include: The edge end pre-collects the data transmission link parameters of various monitoring sensors in the storage tank, counts the number of signal conversions for each type of sensor signal, and records the completion timestamp of each signal conversion; The edge end sorts the various sensor signals according to the number of signal conversions, and performs edge-side decoupling and diversion on the various sensor signals according to the sorting order. This includes dividing a single type of sensor signal into several different time segments according to the completion timestamp, and selecting representative time segments of the whole domain based on the spatial distribution status of the monitoring sensors corresponding to the sensor signals in each time segment. The edge device prioritizes uploading data representing the entire time segment to the cloud for computation. In response to the cloud's initial assessment that the tank status is normal, the edge device filters the data from the first and last two time segments of each type of sensor signal and uploads and stores the data to the cloud. The cloud sequentially retrieves data from the tail time segment of various types of sensor signals and performs cross-fusion calculations with the data from the beginning segment of other types of sensor signals to obtain multiple sets of calculation results. The multiple sets of calculation results are then compared for consistency, and the tank status is determined based on the comparison results.

2. The method for predicting tank leakage based on large-scale model multi-source data fusion according to claim 1, characterized in that, The process of pre-collecting data transmission link parameters of various monitoring sensors in the storage tank at the edge and counting the number of signal conversions for each type of sensor signal includes: Obtain the signal acquisition link topology of various monitoring sensors, identify the number of nodes in the signal acquisition link topology that convert analog signals to digital signals, and record the number of nodes as the number of signal conversions.

3. The method for predicting tank leakage based on large-scale model multi-source data fusion according to claim 2, characterized in that, The monitoring sensors include pressure sensors, acoustic emission sensors, temperature sensors, and liquid level sensors; the number of signal conversions covers all signal conversion stages under the corresponding monitoring indicators.

4. The method for predicting tank leakage based on large-scale model multi-source data fusion according to claim 2, characterized in that, The process of sorting edge-end sensors based on the number of signal conversions for various sensor signals includes: All categories of sensor signals are initially sorted according to the number of signal conversions from high to low. If there are multiple types of sensor signals with the same number of signal conversions, then the signals are sorted in descending order of their historical average transmission delay time to generate a sorting order.

5. The method for predicting tank leakage based on large-scale model multi-source data fusion according to claim 4, characterized in that, The process of dividing a single type of signal into several different time segments based on the completion timestamp at the edge includes: Using the completion timestamp as the dividing boundary, the single-type sensor signal is divided into different time segments at equal time intervals according to the order of time; The earlier the sorting order, the more time segments are divided.

6. The method for predicting tank leakage based on large-scale model multi-source data fusion according to claim 1, characterized in that, The process of determining the spatial distribution of monitoring sensors corresponding to sensing signals within each time period at the edge includes: The entire space of the storage tank is divided into several grid units, and the number of sensors falling into each grid unit is counted. Calculate the mean and standard deviation of the number of sensors in all grid cells, and determine the spatial distribution state value by the ratio of the standard deviation to the mean.

7. The method for predicting tank leakage based on large-scale model multi-source data fusion according to claim 6, characterized in that, The process of filtering the entire representative time segment at the edge includes: The spatial distribution status values ​​of the monitoring sensors within each time period were obtained respectively; The time segment with the smallest spatial distribution state value is selected as the representative time segment of the entire domain.

8. The method for predicting tank leakage based on large-scale model multi-source data fusion according to claim 1, characterized in that, The process of uploading and storing data from the first and last two time segments of each type of sensor signal at the edge to the cloud includes: From several time segments of each type of sensor signal, the time segment with the earliest timestamp is selected as the first time segment, and the time segment with the latest timestamp is selected as the last time segment. Generate a beginning data packet and a end data packet for the data in the beginning time segment and the end time segment; After associating the first-end data packet and the last-end data packet with the corresponding sensor signal category identifier and timestamp index, they are uploaded to cloud storage.

9. The method for predicting tank leakage based on large-scale model multi-source data fusion according to claim 8, characterized in that, The process of sequentially retrieving data from the tail time segments of various types of signals and the beginning time segments of other types of signals in the cloud for calculation includes: Based on the category identifier and timestamp index of the sensor signals stored in the cloud; In the i-th round of operation, the data of the tail time segment of the i-th type of sensor signal and the data of the head time segment of the other N-1 types of sensor signals are retrieved to form the i-th heterogeneous data set, i=1,2,...,N; Perform cross-fusion operation on the i-th heterogeneous data set to obtain the i-th operation result, traverse i=1 to N to obtain N operation results, and form an evaluation set.

10. The method for predicting tank leakage based on large-scale model multi-source data fusion according to claim 9, characterized in that, The process of determining whether there is an anomaly in the storage tank's condition based on the comparison of several calculation results includes: Calculate the dispersion index of N sets of operation results in the evaluation set, the dispersion index includes the range value and the standard deviation of the operation result; compare the range value with the preset range reference value, and compare the standard deviation of the operation result with the preset standard deviation reference value; If the range value is greater than the range reference value, or the standard deviation of the calculation result is greater than the standard deviation reference value, the cloud determines that the storage tank status is abnormal and triggers an early warning signal. If the range value is less than or equal to the range reference value, and the standard deviation of the calculation result is less than or equal to the standard deviation reference value, then the cloud determines that the storage tank is in normal condition.

Images