An abnormality monitoring system and method applied to a hydrogen storage tank

Abstract

This invention discloses an anomaly monitoring system and method for hydrogen storage tanks, relating to the field of anomaly monitoring technology. The method includes the following steps: constructing a multi-sensor fusion model and outputting the characteristic parameters of each sensor; matching the characteristic parameters to an anomaly database and fusing them to output structured monitoring conclusions; collecting all historical early warning information that triggers warnings and extracting a set of characteristic early warning information; if the same anomaly event is detected and two not entirely identical target determination information are output consecutively, extracting the corresponding two historical operation records to construct a target operation node; extracting the characteristic information sets of the two records within each node, obtaining two sets of distinguishable characteristic information sets through difference set analysis, calculating their highest similarity with the characteristic early warning information set and the corresponding comprehensive early warning coefficient, screening target operation nodes with early warning coefficients below a threshold, determining the sensor contribution and deviation direction through attribution analysis, and marking them as a basis for equipment maintenance; this invention achieves anomaly monitoring of hydrogen storage tanks.

Description

Technical Field

[0001] This invention relates to the field of anomaly monitoring technology, specifically an anomaly monitoring system and method applied to hydrogen storage tanks. Background Technology

[0002] With the large-scale development of the hydrogen energy industry, hydrogen storage tanks, as core special equipment for hydrogen storage and transportation, have become a key guarantee for the high-quality development of the hydrogen energy industry. As a core link in the safety management and control of storage tanks, hydrogen storage tank anomaly monitoring technology directly affects the operational stability of storage tanks and the safety layout of the entire hydrogen energy industry chain.

[0003] However, traditional methods for monitoring anomalies in hydrogen storage tanks often face the following problems when dealing with complex operating conditions, identifying multiple types of anomalies, and maintaining long-term monitoring systems: First, the accuracy of monitoring and identification is low. Traditional monitoring often uses a single sensor or a simple data fusion method. Anomalies such as hydrogen embrittlement cracks, hydrogen leakage, and tank deformation in hydrogen storage tanks are highly concealed. Single sensor data is easily affected by operating conditions such as high pressure, temperature changes, and electromagnetic interference, which can easily lead to false alarms and missed alarms, making it difficult to determine the type and severity of anomalies. Second, anomaly judgment lacks the ability to trace the source of deviations. Traditional monitoring methods only achieve basic real-time early warning and result output, without conducting in-depth analysis of inconsistent judgment results for the same anomaly event in the same monitoring area. It is impossible to trace the cause of the judgment deviation, nor can it quantify the contribution of each sensor to the judgment deviation. This makes it difficult to locate the root cause of the problem after the early warning, and the operation and maintenance of storage tanks lacks specificity. Summary of the Invention

[0004] The purpose of this invention is to provide an anomaly monitoring system and method for hydrogen storage tanks to solve the problems raised in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an anomaly monitoring method for hydrogen storage tanks, the method comprising the following steps: Collect relevant information about the storage tank and build a multi-sensor fusion model, then output the characteristic parameters of each sensor. The sensor feature parameters are matched with the abnormal pattern library to generate the initial basic probability assignment function of the corresponding abnormal hypothesis category. After being synthesized by DS orthogonal and rule, the Kalman filter state estimate is input into the discriminator to output the structured monitoring conclusion. Collect all historical early warning information that triggers the warning, and extract a set of characteristic early warning information from it; If the multi-sensor fusion model continuously outputs two different target determination information for the same abnormal event in the same monitoring area of ​​the storage tank, then the corresponding two historical operation records are extracted to construct the target operation node. Extract the feature information sets corresponding to the two consecutive historical operation records within the target operation node, obtain the first and second distinguishing feature information sets based on the difference between the two sets, obtain the highest similarity between each feature and each feature in the feature warning information set, and calculate the corresponding first and second comprehensive warning coefficients based on this. Featured operating nodes with a comprehensive early warning coefficient below the threshold are selected, and attribution analysis is used to determine the sensor contribution and deviation direction. The results are then pushed to the operation and maintenance terminal.

[0006] Deploy multimodal sensors to collect relevant information about the storage tank and build a multi-sensor fusion model, outputting the characteristic parameters of each sensor. Specific steps include: Point cloud maps and infrared thermal image feature maps of the storage tank body and surrounding environment are pre-recorded. Spatial coordinate calibration and dynamic tracking of wireless sensor nodes in the tank area are achieved through synchronous positioning and SLAM technology. Multimodal sensing sensors, including acoustic emission sensors, pressure sensors, temperature sensors, electrochemical hydrogen sensors and millimeter-wave radar sensors, are deployed at the storage tank body, valves and pipeline connections to build a multi-sensor fusion model and output the feature parameters of each sensor.

[0007] Constructing the multi-sensor fusion model includes the following steps: The raw time-series signals acquired by the multimodal sensing sensors are preprocessed by digital filtering, baseline correction and dimensional normalization. The spatial coordinates of each sensor are calibrated using a SLAM system, and a unified timestamp is added to each frame of data based on the IEEE 1588 time protocol. For continuous numerical time-series data output by pressure and temperature sensors, a state-space model based on Kalman filtering is established. The aligned observation sequence is used as the filter input. Through state prediction and observation update iterative calculation, the real-time optimal estimation sequence of tank wall temperature and pressure is output. For the signals output by the acoustic emission sensor, electrochemical hydrogen sensor, and millimeter-wave radar sensor, corresponding characteristic parameters are extracted respectively, as detailed below: Extract the rise time, ring count, energy, peak frequency, and duration from the acoustic emission sensor; Extract the hydrogen concentration value and its first-order difference from the electrochemical hydrogen sensor; Extract echo amplitude, range profile, and Doppler shift from millimeter-wave radar sensors.

[0008] The sensor feature parameters are matched with an anomaly pattern library to generate an initial basic probability assignment function for the corresponding anomaly hypothesis category; after DS orthogonal and rule synthesis, the Kalman filter state estimate is input into the discriminator to output a structured monitoring conclusion. The specific steps include: The characteristic parameters of each sensor are matched with the preset abnormal mode feature library. Based on the detection probability and false alarm probability obtained by each sensor in the historical calibration test, the basic probability assignment function construction method in the evidence theory is used to generate the initial basic probability assignment function of each sensor for each abnormal hypothesis category in the preset abnormal hypothesis category set in the current decision cycle. The preset abnormal hypothesis category set includes hydrogen embrittlement cracking, hydrogen leakage, and tank deformation. Using the Dempster orthogonal sum rule in the DS evidence synthesis rules, the initial basic probability assignment functions of the three outputs of the acoustic emission sensor, electrochemical hydrogen sensor and millimeter-wave radar sensor within the same decision cycle are synthesized in pairs to generate the joint basic probability assignment function of each anomaly hypothesis category in the preset anomaly hypothesis category set. The pressure and temperature continuous state estimates output by the Kalman filter, together with the joint basic probability assignment function output by the DS evidence synthesis, are input into the top-level logic discriminator to obtain monitoring conclusions in the form of structured data. The monitoring conclusions include operating status category labels, conclusion confidence, the top three sensor identifiers in terms of contribution, and the current timestamp.

[0009] The top-level logic discriminator performs judgments according to the following preset decision logic rules, and the specific steps are as follows: If the pressure estimate exceeds the pressure over-limit threshold or the temperature estimate exceeds the temperature over-limit threshold, the corresponding pressure over-limit or temperature over-limit alarm conclusion will be output directly. If neither pressure nor temperature exceeds the limit, the joint basic probability assignment function value of each abnormal hypothesis category is compared with its respective preset confidence threshold. If the confidence level of only one type of anomalous hypothesis exceeds the threshold, then output the warning conclusion for that type of anomalous hypothesis. If the confidence level of more than one type of anomalous hypothesis exceeds the threshold at the same time, the hypothesis with the highest confidence level will be selected as the output conclusion. If no abnormal assumptions are found and the confidence level exceeds the threshold, then a normal conclusion is output. The final monitoring results are output in the form of structured data, including the operational status category label, the confidence level of the conclusion, the identifiers of the top three contributing sensors, and the current timestamp.

[0010] The process involves collecting all historical early warning information that triggers warnings from the multi-sensor fusion model output, and extracting a set of feature-based early warning information from it. Specific steps include: Full data collection is performed on every historical early warning information output from the multi-sensor fusion model that triggers an early warning or linkage control command. The historical early warning information includes the original sensor data fragments corresponding to each historical early warning time, the joint basic probability assignment function vector output by the fusion model, and the monitoring conclusions. Extract a set of feature-based early warning information from each historical early warning message. The set of feature-based early warning information includes: feature parameters of each sensor, joint basic probability assignment function, and monitoring conclusions.

[0011] If the multi-sensor fusion model is detected to output two different target determination information for the same abnormal event in the same monitoring area of ​​the storage tank, the two historical operation records generated by the multi-sensor fusion model when outputting two different target determination information are extracted, and a target operation node is constructed based on the abnormal event. The specific steps include: The not entirely identical target determination information includes one of the following situations: different abnormal hypothesis categories are output for the same abnormal event; the joint confidence difference of the same abnormal hypothesis category exceeds a preset deviation threshold; the output positioning coordinate deviation exceeds a preset distance threshold. The historical operation record includes the first timestamp, the first joint basic probability assignment function vector, and the first final monitoring conclusion corresponding to the first output, and the second timestamp, the second joint basic probability assignment function vector, and the second final monitoring conclusion corresponding to the second output. The time window of the abnormal event, the monitoring area identifier, and the set of participating sensors are used as the identification information of the target operating node.

[0012] Extract the feature information sets corresponding to the two consecutive historical operation records within the target running node. Obtain the first and second distinguishing feature information sets based on the difference between the two sets, and obtain the highest similarity between each feature and each feature in the feature warning information set. The specific steps include: If target determination information A and B are extracted from the first historical operation record P(A) and the second historical operation record P(B) within a certain target operation node, respectively, and the first historical operation record P(A) was generated earlier than the second historical operation record P(B); feature extraction is performed on the target determination information A and B respectively to obtain the feature information sets D(A) and D(B) of the corresponding target determination information A and B; the feature information sets include the feature parameters of each sensor and the joint basic probability assignment function value of each anomaly hypothesis category in the preset anomaly hypothesis category set obtained after DS evidence synthesis; Extract a first set of distinguishing feature information F(A) = D(A) - D(A) ∩ D(B) and a second set of distinguishing feature information F(B) = D(B) - D(A) ∩ D(B) for the target running node; and obtain the highest similarity value presented after calculating the similarity with each feature information in the first set of distinguishing feature information F(A) and the second set of distinguishing feature information F(B) in turn.

[0013] The first and second comprehensive early warning coefficients are calculated based on the first and second sets of distinguishing feature information, respectively. The specific steps include: Calculate the first comprehensive early warning coefficient based on the first distinguishing feature information set F(A) for the target running node: β1=θ1×θ2×...×θ n ; Among them, θ1, θ2,..., θ n These represent the 1st, 2nd, ..., nth feature information in the first distinguishing feature information set F(A), respectively, and the highest similarity value presented after calculating the similarity with each feature information in the feature warning information set in turn; Calculate the second comprehensive early warning coefficient based on the second distinguishing feature information set F(B) for the target running node: β2=θ1 ’ ×θ2 ’ ×...×θ m ’ ; where θ1 ’ θ2 ’ ..., θ m ’ These represent the highest similarity values ​​obtained by calculating the similarity between the 1st, 2nd, ..., mth features in the second distinguishing feature information set F(B) and each feature in the feature warning information set.

[0014] Based on the comprehensive early warning coefficient of the target operating nodes, characteristic operating nodes below a preset threshold are selected. For each characteristic operating node, attribution analysis is used to quantify the contribution of each sensor to the judgment deviation, and the direction and weight of the deviation are determined. Sensors whose contribution exceeds the preset threshold are marked and recorded, and pushed to the operation and maintenance terminal as a maintenance basis. The specific steps include: Based on the changes in the comprehensive early warning coefficient presented in each target operating node, characteristic operating nodes with a comprehensive early warning coefficient lower than a preset reliability threshold are selected. The preset reliability threshold is determined based on the statistical distribution of the comprehensive early warning coefficients of historical target operating nodes, specifically by taking the 10th percentile of the comprehensive early warning coefficients of all historical target operating nodes as the threshold. Based on each characteristic operating node, the contribution of each sensing sensor to the judgment deviation of that node is quantified using attribution analysis methods, including Shapley value attribution analysis or gradient contribution calculation. According to the magnitude and sign of each sensor's contribution, combined with the deviation direction of the sensor's feature value from the mean of the corresponding feature in the historical feature warning information set, the deviation direction and contribution weight of the sensor are determined. Sensors whose contribution exceeds a preset contribution threshold are marked as features, and their deviation direction and contribution weight are recorded to form feature marking records. The feature marker records are pushed to the storage tank operation and maintenance management terminal as a basis for sensor performance maintenance.

[0015] An anomaly monitoring system for hydrogen storage tanks is provided, which is applicable to the aforementioned anomaly monitoring method for hydrogen storage tanks.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. By deploying multimodal sensing sensors and building a fusion model, and combining DS orthogonal and regular and Kalman filtering, the spatiotemporal alignment and decision-level fusion of heterogeneous sensing data are achieved. Unlike the single sensor monitoring or simple data fusion methods in the existing technology, this invention can match the abnormal pattern library and output structured monitoring conclusions, capture the characteristic differences of various anomalies in the storage tank, and realize the identification of hydrogen embrittlement cracks, hydrogen leakage and tank deformation. 2. This invention collects all historical early warning information and extracts a set of characteristic early warning information. It constructs a target operating node for inconsistent judgment results of the same abnormal event, obtains a set of distinguishing characteristic information through difference analysis, and calculates a comprehensive early warning coefficient. Unlike existing technologies that only output early warning results and have no ability to trace the source of judgment deviations, this invention can trace the causes of abnormal judgment deviations and realize the source of judgment deviations in the monitoring of abnormalities in storage tanks. Attached Figure Description

[0017] Figure 1 This is a schematic flowchart of an anomaly monitoring method for hydrogen storage tanks according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] like Figure 1 As shown, the present invention provides a technical solution, a method for anomaly monitoring applied to hydrogen storage tanks, the method comprising the following steps: Collect relevant information about the storage tank and build a multi-sensor fusion model, then output the characteristic parameters of each sensor. The sensor feature parameters are matched with the abnormal pattern library to generate the initial basic probability assignment function of the corresponding abnormal hypothesis category. After being synthesized by DS orthogonal and rule, the Kalman filter state estimate is input into the discriminator to output the structured monitoring conclusion. Collect all historical early warning information that triggers the warning, and extract a set of characteristic early warning information from it; If the multi-sensor fusion model continuously outputs two different target determination information for the same abnormal event in the same monitoring area of ​​the storage tank, then the corresponding two historical operation records are extracted to construct the target operation node. Extract the feature information sets corresponding to the two consecutive historical operation records within the target operation node, obtain the first and second distinguishing feature information sets based on the difference between the two sets, obtain the highest similarity between each feature and each feature in the feature warning information set, and calculate the corresponding first and second comprehensive warning coefficients based on this. Featured operating nodes with a comprehensive early warning coefficient below the threshold are selected, and attribution analysis is used to determine the sensor contribution and deviation direction. The results are then pushed to the operation and maintenance terminal.

[0020] Deploy multimodal sensors to collect relevant information about the storage tank and build a multi-sensor fusion model, outputting the characteristic parameters of each sensor. Specific steps include: Point cloud maps and infrared thermal image feature maps of the storage tank body and surrounding environment are pre-recorded. Spatial coordinate calibration and dynamic tracking of wireless sensor nodes in the tank area are achieved through synchronous positioning and SLAM technology. Multimodal sensing sensors, including acoustic emission sensors, pressure sensors, temperature sensors, electrochemical hydrogen sensors and millimeter-wave radar sensors, are deployed at the storage tank body, valves and pipeline connections to build a multi-sensor fusion model and output the feature parameters of each sensor.

[0021] Constructing the multi-sensor fusion model includes the following steps: The raw time-series signals acquired by the multimodal sensing sensors are preprocessed by digital filtering, baseline correction and dimensional normalization. The spatial coordinates of each sensor are calibrated using a SLAM system, and a unified timestamp is added to each frame of data based on the IEEE 1588 time protocol. For continuous numerical time-series data output by pressure and temperature sensors, a state-space model based on Kalman filtering is established. The aligned observation sequence is used as the filter input. Through state prediction and observation update iterative calculation, the real-time optimal estimation sequence of tank wall temperature and pressure is output. For the signals output by the acoustic emission sensor, electrochemical hydrogen sensor, and millimeter-wave radar sensor, corresponding characteristic parameters are extracted respectively, as detailed below: Extract the rise time, ring count, energy, peak frequency, and duration from the acoustic emission sensor; Extract the hydrogen concentration value and its first-order difference from the electrochemical hydrogen sensor; Extract echo amplitude, range profile, and Doppler shift from millimeter-wave radar sensors.

[0022] The sensor feature parameters are matched with an anomaly pattern library to generate an initial basic probability assignment function for the corresponding anomaly hypothesis category; after DS orthogonal and rule synthesis, the Kalman filter state estimate is input into the discriminator to output a structured monitoring conclusion. The specific steps include: The characteristic parameters of each sensor are matched with the preset abnormal mode feature library. Based on the detection probability and false alarm probability obtained by each sensor in the historical calibration test, the basic probability assignment function construction method in the evidence theory is used to generate the initial basic probability assignment function of each sensor for each abnormal hypothesis category in the preset abnormal hypothesis category set in the current decision cycle. The preset abnormal hypothesis category set includes hydrogen embrittlement cracking, hydrogen leakage, and tank deformation. Using the Dempster orthogonal sum rule in the DS evidence synthesis rules, the initial basic probability assignment functions of the three outputs of the acoustic emission sensor, electrochemical hydrogen sensor and millimeter-wave radar sensor within the same decision cycle are synthesized in pairs to generate the joint basic probability assignment function of each anomaly hypothesis category in the preset anomaly hypothesis category set. The pressure and temperature continuous state estimates output by the Kalman filter, together with the joint basic probability assignment function output by the DS evidence synthesis, are input into the top-level logic discriminator to obtain monitoring conclusions in the form of structured data. The monitoring conclusions include operating status category labels, conclusion confidence, the top three sensor identifiers in terms of contribution, and the current timestamp.

[0023] The top-level logic discriminator performs judgments according to the following preset decision logic rules, and the specific steps are as follows: If the pressure estimate exceeds the pressure over-limit threshold or the temperature estimate exceeds the temperature over-limit threshold, the corresponding pressure over-limit or temperature over-limit alarm conclusion will be output directly. If neither pressure nor temperature exceeds the limit, the joint basic probability assignment function value of each abnormal hypothesis category is compared with its respective preset confidence threshold. If the confidence level of only one type of anomalous hypothesis exceeds the threshold, then output the warning conclusion for that type of anomalous hypothesis. If the confidence level of more than one type of anomalous hypothesis exceeds the threshold at the same time, the hypothesis with the highest confidence level will be selected as the output conclusion. If no abnormal assumptions are found and the confidence level exceeds the threshold, then a normal conclusion is output. The final monitoring results are output in the form of structured data, including the operational status category label, the confidence level of the conclusion, the identifiers of the top three contributing sensors, and the current timestamp.

[0024] The process involves collecting all historical early warning information that triggers warnings from the multi-sensor fusion model output, and extracting a set of feature-based early warning information from it. Specific steps include: Full data collection is performed on every historical early warning information output from the multi-sensor fusion model that triggers an early warning or linkage control command. The historical early warning information includes the original sensor data fragments corresponding to each historical early warning time, the joint basic probability assignment function vector output by the fusion model, and the monitoring conclusions. Extract a set of feature-based early warning information from each historical early warning message. The set of feature-based early warning information includes: feature parameters of each sensor, joint basic probability assignment function, and monitoring conclusions.

[0025] If the multi-sensor fusion model is detected to output two different target determination information for the same abnormal event in the same monitoring area of ​​the storage tank, the two historical operation records generated by the multi-sensor fusion model when outputting two different target determination information are extracted, and a target operation node is constructed based on the abnormal event. The specific steps include: The not entirely identical target determination information includes one of the following situations: different abnormal hypothesis categories are output for the same abnormal event; the joint confidence difference of the same abnormal hypothesis category exceeds a preset deviation threshold; the output positioning coordinate deviation exceeds a preset distance threshold. The historical operation record includes the first timestamp, the first joint basic probability assignment function vector, and the first final monitoring conclusion corresponding to the first output, and the second timestamp, the second joint basic probability assignment function vector, and the second final monitoring conclusion corresponding to the second output. The time window of the abnormal event, the monitoring area identifier, and the set of participating sensors are used as the identification information of the target operating node.

[0026] Extract the feature information sets corresponding to the two consecutive historical operation records within the target running node. Obtain the first and second distinguishing feature information sets based on the difference between the two sets, and obtain the highest similarity between each feature and each feature in the feature warning information set. The specific steps include: If target determination information A and B are extracted from the first historical operation record P(A) and the second historical operation record P(B) within a certain target operation node, respectively, and the first historical operation record P(A) was generated earlier than the second historical operation record P(B); feature extraction is performed on the target determination information A and B respectively to obtain the feature information sets D(A) and D(B) of the corresponding target determination information A and B; the feature information sets include the feature parameters of each sensor and the joint basic probability assignment function value of each anomaly hypothesis category in the preset anomaly hypothesis category set obtained after DS evidence synthesis; Extract a first set of distinguishing feature information F(A) = D(A) - D(A) ∩ D(B) and a second set of distinguishing feature information F(B) = D(B) - D(A) ∩ D(B) for the target running node; and obtain the highest similarity value presented after calculating the similarity with each feature information in the first set of distinguishing feature information F(A) and the second set of distinguishing feature information F(B) in turn.

[0027] The first and second comprehensive early warning coefficients are calculated based on the first and second sets of distinguishing feature information, respectively. The specific steps include: Calculate the first comprehensive early warning coefficient based on the first distinguishing feature information set F(A) for the target running node: β1=θ1×θ2×...×θ n ; Among them, θ1, θ2,..., θ n These represent the 1st, 2nd, ..., nth feature information in the first distinguishing feature information set F(A), respectively, and the highest similarity value presented after calculating the similarity with each feature information in the feature warning information set in turn; Calculate the second comprehensive early warning coefficient based on the second distinguishing feature information set F(B) for the target running node: β2=θ1 ’ ×θ2 ’ ×...×θ m ’ ; where θ1 ’ θ2 ’ ..., θ m ’ These represent the highest similarity values ​​obtained by calculating the similarity between the 1st, 2nd, ..., mth features in the second distinguishing feature information set F(B) and each feature in the feature warning information set.

[0028] Based on the comprehensive early warning coefficient of the target operating nodes, characteristic operating nodes below a preset threshold are selected. For each characteristic operating node, attribution analysis is used to quantify the contribution of each sensor to the judgment deviation, and the direction and weight of the deviation are determined. Sensors whose contribution exceeds the preset threshold are marked and recorded, and pushed to the operation and maintenance terminal as a maintenance basis. The specific steps include: Based on the changes in the comprehensive early warning coefficient presented in each target operating node, characteristic operating nodes with a comprehensive early warning coefficient lower than a preset reliability threshold are selected. The preset reliability threshold is determined based on the statistical distribution of the comprehensive early warning coefficients of historical target operating nodes, specifically by taking the 10th percentile of the comprehensive early warning coefficients of all historical target operating nodes as the threshold. Based on each characteristic operating node, the contribution of each sensing sensor to the judgment deviation of that node is quantified using attribution analysis methods, including Shapley value attribution analysis or gradient contribution calculation. According to the magnitude and sign of each sensor's contribution, combined with the deviation direction of the sensor's feature value from the mean of the corresponding feature in the historical feature warning information set, the deviation direction and contribution weight of the sensor are determined. Sensors whose contribution exceeds a preset contribution threshold are marked as features, and their deviation direction and contribution weight are recorded to form feature marking records. The feature marker records are pushed to the storage tank operation and maintenance management terminal as a basis for sensor performance maintenance.

[0029] In Example 1: Spatial calibration of sensor nodes in the tank area and deployment of multimodal sensors are carried out. Spatial point cloud features and infrared thermal imaging features of the tank body and surrounding environment are collected in advance. Based on synchronous positioning and scene modeling technology, spatial coordinate calibration and dynamic tracking of wireless sensor nodes in the tank area are completed. Multiple types of sensing sensors, including acoustic emission sensors, pressure sensors, temperature sensors, electrochemical hydrogen sensors and millimeter-wave radar sensors, are deployed in key parts with high anomaly occurrence, such as the tank body, valves and pipeline connections. According to the acquisition characteristics and monitoring dimensions of various sensors, a multi-sensor fusion model is constructed. Feature extraction is performed on the raw data collected by each sensor, and the feature parameters corresponding to each sensor are output. The extracted sensor feature parameters are matched with a preset abnormal pattern feature library. Combined with the detection performance and false alarm situation of various sensors in historical calibration tests, the initial judgment criteria corresponding to various abnormal hypothesis categories are constructed. The initial judgment criteria of three core sensors, namely acoustic emission, electrochemical hydrogen, and millimeter-wave radar, are fused and processed. Then, combined with the continuous state evaluation results of tank pressure and temperature, they are input into the top-level logic discrimination module. Finally, a structured monitoring conclusion is output. The conclusion includes the tank operating status category, the confidence level of the judgment result, the sensor identification that has a significant impact on the judgment result, and the monitoring time information. All historical early warning information output by the multi-sensor fusion model that triggers early warning or linkage control commands is comprehensively collected. Sensor feature parameters, fusion judgment criteria, and final monitoring conclusions are extracted from each historical early warning information and integrated to form a feature early warning information set. If the fusion model continuously outputs different judgment results for the same abnormal event in the same monitoring area of ​​the storage tank, the corresponding two historical operation records are extracted. The target operation node is constructed with the abnormal event as the core. Feature information sets of the two records in the node are extracted respectively. Two sets of differentiated feature information sets are obtained through comparative analysis. The two sets of differentiated feature information are matched with the feature early warning information set respectively to obtain the highest matching degree of each differentiated feature and calculate the comprehensive early warning coefficient. Featured operating nodes whose comprehensive early warning coefficients do not meet the preset standards are selected. Attribution analysis is used to quantify the influence of each sensor on the judgment deviation. By combining the deviation of sensor feature values ​​from historical feature averages, the deviation direction of each sensor is clarified. Sensors with excessive influence are marked, and their deviation direction and influence weight are recorded. After forming feature marking records, they are pushed to the storage tank operation and maintenance management terminal as the basis for sensor performance maintenance and calibration.

[0030] An anomaly monitoring system for hydrogen storage tanks is provided, which is applicable to the aforementioned anomaly monitoring method for hydrogen storage tanks.

[0031] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. An anomaly monitoring method applied to hydrogen storage tanks, characterized in that: The method includes the following steps: Collect relevant information about the storage tank and build a multi-sensor fusion model, then output the characteristic parameters of each sensor. The sensor feature parameters are matched with the abnormal pattern library to generate the initial basic probability assignment function of the corresponding abnormal hypothesis category. After being synthesized by DS orthogonal and rule, the Kalman filter state estimate is input into the discriminator to output the structured monitoring conclusion. Collect all historical early warning information that triggers the warning, and extract a set of characteristic early warning information from it; If the multi-sensor fusion model continuously outputs two different target determination information for the same abnormal event in the same monitoring area of ​​the storage tank, then the corresponding two historical operation records are extracted to construct the target operation node. Extract the feature information sets corresponding to the two consecutive historical operation records within the target operation node, obtain the first and second distinguishing feature information sets based on the difference between the two sets, obtain the highest similarity between each feature and each feature in the feature warning information set, and calculate the corresponding first and second comprehensive warning coefficients based on this. Featured operating nodes with a comprehensive early warning coefficient below the threshold are selected, and attribution analysis is used to determine the sensor contribution and deviation direction. The results are then pushed to the operation and maintenance terminal.

2. The anomaly monitoring method for hydrogen storage tanks according to claim 1, characterized in that: Deploy multimodal sensors to collect relevant information about the storage tank and build a multi-sensor fusion model, outputting the characteristic parameters of each sensor. Specific steps include: Point cloud maps and infrared thermal image feature maps of the storage tank body and surrounding environment are pre-recorded. Spatial coordinate calibration and dynamic tracking of wireless sensor nodes in the tank area are achieved through synchronous positioning and SLAM technology. Multimodal sensing sensors, including acoustic emission sensors, pressure sensors, temperature sensors, electrochemical hydrogen sensors and millimeter-wave radar sensors, are deployed at the storage tank body, valves and pipeline connections to build a multi-sensor fusion model and output the feature parameters of each sensor.

3. The anomaly monitoring method for hydrogen storage tanks according to claim 2, characterized in that: The sensor feature parameters are matched with the anomaly pattern library to generate the initial basic probability assignment function for the corresponding anomaly hypothesis category; After orthogonal and rule-based synthesis using DS, the Kalman filter state estimate is input into the discriminator to output structured monitoring conclusions. Specific steps include: The characteristic parameters of each sensor are matched with the preset abnormal mode feature library. Based on the detection probability and false alarm probability obtained by each sensor in the historical calibration test, the basic probability assignment function construction method in the evidence theory is used to generate the initial basic probability assignment function of each sensor for each abnormal hypothesis category in the preset abnormal hypothesis category set in the current decision cycle. The preset abnormal hypothesis category set includes hydrogen embrittlement cracking, hydrogen leakage, and tank deformation. Using the Dempster orthogonal sum rule in the DS evidence synthesis rules, the initial basic probability assignment functions of the three outputs of the acoustic emission sensor, electrochemical hydrogen sensor and millimeter-wave radar sensor within the same decision cycle are synthesized in pairs to generate the joint basic probability assignment function of each anomaly hypothesis category in the preset anomaly hypothesis category set. The pressure and temperature continuous state estimates output by the Kalman filter, together with the joint basic probability assignment function output by the DS evidence synthesis, are input into the top-level logic discriminator to obtain monitoring conclusions in the form of structured data. The monitoring conclusions include operating status category labels, conclusion confidence, the top three sensor identifiers in terms of contribution, and the current timestamp.

4. The anomaly monitoring method for hydrogen storage tanks according to claim 3, characterized in that: The process involves collecting all historical early warning information that triggers warnings from the multi-sensor fusion model output, and extracting a set of feature-based early warning information from it. Specific steps include: Full data collection is performed on every historical early warning information output from the multi-sensor fusion model that triggers an early warning or linkage control command. The historical early warning information includes the original sensor data fragments corresponding to each historical early warning time, the joint basic probability assignment function vector output by the fusion model, and the monitoring conclusions. Extract a set of feature-based early warning information from each historical early warning message. The set of feature-based early warning information includes: feature parameters of each sensor, joint basic probability assignment function, and monitoring conclusions.

5. The anomaly monitoring method for hydrogen storage tanks according to claim 4, characterized in that: If the multi-sensor fusion model is detected to output two different target determination information for the same abnormal event in the same monitoring area of ​​the storage tank, the two historical operation records generated by the multi-sensor fusion model when outputting two different target determination information are extracted, and a target operation node is constructed based on the abnormal event. The specific steps include: The not entirely identical target determination information includes one of the following situations: different abnormal hypothesis categories are output for the same abnormal event; the joint confidence difference of the same abnormal hypothesis category exceeds a preset deviation threshold; the output positioning coordinate deviation exceeds a preset distance threshold. The historical operation record includes the first timestamp, the first joint basic probability assignment function vector, and the first final monitoring conclusion corresponding to the first output, and the second timestamp, the second joint basic probability assignment function vector, and the second final monitoring conclusion corresponding to the second output. The time window of the abnormal event, the monitoring area identifier, and the set of participating sensors are used as the identification information of the target operating node.

6. The anomaly monitoring method for hydrogen storage tanks according to claim 5, characterized in that: Extract the feature information sets corresponding to the two consecutive historical operation records within the target running node. Obtain the first and second distinguishing feature information sets based on the difference between the two sets, and obtain the highest similarity between each feature and each feature in the feature warning information set. The specific steps include: If target determination information A and B are extracted from the first historical operation record P(A) and the second historical operation record P(B) within a certain target operation node, respectively, and the first historical operation record P(A) was generated earlier than the second historical operation record P(B); feature extraction is performed on the target determination information A and B respectively to obtain the feature information sets D(A) and D(B) of the corresponding target determination information A and B; the feature information sets include the feature parameters of each sensor and the joint basic probability assignment function value of each anomaly hypothesis category in the preset anomaly hypothesis category set obtained after DS evidence synthesis; Extract a first set of distinguishing feature information F(A) = D(A) - D(A) ∩ D(B) and a second set of distinguishing feature information F(B) = D(B) - D(A) ∩ D(B) for the target running node; and obtain the highest similarity value presented after calculating the similarity with each feature information in the first set of distinguishing feature information F(A) and the second set of distinguishing feature information F(B) in turn.

7. The anomaly monitoring method for hydrogen storage tanks according to claim 6, characterized in that: The first and second comprehensive early warning coefficients are calculated based on the first and second sets of distinguishing feature information, respectively. The specific steps include: Calculate the first comprehensive early warning coefficient based on the first distinguishing feature information set F(A) for the target running node: β1=θ1×θ2×...×θ n ; Among them, θ1, θ2,..., θ n These represent the 1st, 2nd, ..., nth feature information in the first distinguishing feature information set F(A), respectively, and the highest similarity value presented after calculating the similarity with each feature information in the feature warning information set in turn; Calculate the second comprehensive early warning coefficient based on the second distinguishing feature information set F(B) for the target running node: β2=θ1 ’ ×θ2 ’ ×...×θ m ’ ; where θ1 ’ θ2 ’ ..., θ m ’ These represent the highest similarity values ​​obtained by calculating the similarity between the 1st, 2nd, ..., mth features in the second distinguishing feature information set F(B) and each feature in the feature warning information set.

8. The anomaly monitoring method for hydrogen storage tanks according to claim 7, characterized in that: Based on the comprehensive early warning coefficient of the target operating node, characteristic operating nodes below the preset threshold are selected. For each characteristic operating node, attribution analysis is used to quantify the contribution of each sensor to the judgment deviation and determine the deviation direction and weight. Sensors whose contribution exceeds a preset threshold are marked and recorded, and then pushed to the maintenance terminal as a maintenance basis. The specific steps include: Based on the changes in the comprehensive early warning coefficient presented in each target operating node, characteristic operating nodes with a comprehensive early warning coefficient lower than a preset reliability threshold are selected. The preset reliability threshold is determined based on the statistical distribution of the comprehensive early warning coefficients of historical target operating nodes, specifically by taking the quantile corresponding to the preset quantile parameter of the comprehensive early warning coefficients of all historical target operating nodes as the threshold. Based on each characteristic operating node, the contribution of each sensing sensor to the judgment deviation of that node is quantified using attribution analysis methods, including Shapley value attribution analysis or gradient contribution calculation. According to the magnitude and sign of each sensor's contribution, combined with the deviation direction of the sensor's feature value from the corresponding feature mean in the historical feature warning information set, the deviation direction and contribution weight of the sensor are determined. Sensors whose contribution exceeds a preset contribution threshold are marked as features, and their deviation direction and contribution weight are recorded to form feature marking records. The feature marker records are pushed to the storage tank operation and maintenance management terminal as a basis for sensor performance maintenance.

9. An anomaly monitoring system for hydrogen storage tanks, characterized in that: The method for anomaly monitoring of a hydrogen storage tank as described in any one of claims 1-8.

Images