D-s evidence theory and online learning-based eh oil system health scoring method and system
By combining DS evidence theory with online learning, the problems of insufficient multi-source information fusion and adaptive parameter tuning in EH oil system monitoring and fault diagnosis were solved, achieving efficient and accurate fault diagnosis and real-time monitoring, and reducing operation and maintenance costs.
Patent Information
- Application Number
- CN202611161824.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-03
- Publication Date
- 2026-11-10
AI Technical Summary
Existing EH oil system monitoring and fault diagnosis suffer from poor real-time performance, insufficient fusion of multi-source information, false alarms and missed alarms due to static thresholds, and lack of adaptive capabilities. It cannot effectively integrate mechanistic knowledge with data-driven prediction and lacks adaptive parameter tuning capabilities.
This approach combines Dempster-Shafer evidence theory with online learning. Data is collected by a mechanism engine and a prediction engine and mapped to a basic probability assignment function. Multi-source fusion is performed using Dempster-Shafer evidence theory, and conflict management and adaptive parameter tuning are achieved through an online learning module, including statistical threshold adaptation, data distribution drift detection, and Bayesian parameter optimization.
It achieves efficient fusion of multi-source information, improves diagnostic accuracy, reduces operation and maintenance costs, and has adaptive capabilities, enabling it to automatically adjust diagnostic parameters when equipment ages and operating conditions change, thereby reducing misjudgments and missed reports.
Smart Images

Figure CN123069252A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial equipment condition monitoring and fault diagnosis technology, and more specifically to an EH oil system health scoring method and system based on Dempster-Shafer (DS) evidence theory, multi-source fusion, and online learning adaptive parameter tuning, which is applicable to real-time intelligent diagnosis and health management of steam turbine hydraulic servo systems. Background Technology
[0002] The EH (Electro-Hydraulic) oil system is the core hydraulic servo system of the steam turbine, and its operating status directly determines the safety and stability of the unit. The EH oil system includes key components such as regulating valves, main steam valves, and oil supply devices. It operates under complex conditions and is prone to various types of failures. Common failures include valve jamming, servo valve malfunction, abnormal oil pressure, and oil quality deterioration (excessive particle size, increased water content), etc.
[0003] The existing technologies for EH oil system monitoring and fault diagnosis have the following main shortcomings: 1. Insufficient Integration of Real-Time Oil Quality Monitoring and Diagnosis: Traditional EH oil quality testing methods based on manual sampling and analysis rely on manual collection of oil samples. The sampling frequency is entirely determined by the frequency of manual collection, and the testing time is long (usually several hours to several days), resulting in poor real-time performance and an inability to detect oil quality problems promptly. EH oil consumption is directly proportional to the sampling frequency, leading to high overall costs. Although some systems have deployed online oil quality monitoring equipment (such as particle size sensors and moisture content meters) to collect particle size and moisture content data in real time, this data is only output as independent display indicators to the HMI or DCS system and is not integrated into the overall fault diagnosis framework. Specifically, this manifests as: oil quality deterioration trends cannot be jointly analyzed with other fault signals such as valve jamming and pump malfunctions; preventative maintenance cannot be triggered in the early stages of oil quality deterioration; when both particle size and moisture content exceed limits, the system cannot determine the risk of combined deterioration with higher confidence, making it difficult to provide effective comprehensive diagnostic conclusions in the early stages of oil quality deterioration.
[0004] 2. Narrow Scope: Traditional EH oil system monitoring primarily focuses on the EH oil circuit status, lacking monitoring of supporting equipment. When significant problems arise in the EH oil circuit, the supporting equipment is often already damaged. This prevents proactive equipment maintenance to minimize losses. Furthermore, the single standard for judging the oil circuit status makes it impossible to detect parameters such as bubble rate in real time, easily leading to missed or incorrect diagnoses.
[0005] 3. Limited by Single Diagnostic Methods: Traditional rule-based alarm methods (mechanism rule methods) based on fixed thresholds rely on manual experience to set thresholds, making them unable to adapt to long-term drift due to equipment aging, changes in operating conditions, etc., and prone to false alarms or missed alarms. While anomaly detection methods based on statistics or machine learning can capture deviations in data patterns, they lack the ability to diagnose specific fault types.
[0006] 4. Insufficient fusion of multi-source information: Existing systems often process rule-based alarms and statistical anomalies independently, lacking a mechanism to unify and integrate deterministic rule mechanisms with probabilistic statistical predictions. When the two types of information sources contradict each other (e.g., rule-based alarms are normal but statistical anomalies are high), the system cannot provide a reasonable comprehensive judgment.
[0007] 5. Static thresholds and high parameter tuning costs: Existing systems often have alarm thresholds and health score parameters that are fixed at the factory or manually set once, making it impossible to automatically adjust them based on factors such as equipment aging, seasonal changes, and maintenance history. In large power plants, EH oil system threshold maintenance requires repeated debugging and verification by professional personnel, resulting in high costs.
[0008] 6. Lack of adaptive capability: When the equipment operating conditions undergo structural changes (such as after overhaul or replacement of parts), the existing system cannot automatically identify such distribution drift and recalibrate diagnostic parameters, resulting in a continuous decline in diagnostic accuracy.
[0009] Therefore, there is an urgent need for an EH oil system health scoring method and system that can integrate mechanistic knowledge and data-driven prediction, has conflict handling capabilities, and can adaptively adjust diagnostic parameters online. Summary of the Invention To overcome the shortcomings of the existing technology, this invention discloses a health scoring method and system for EH oil system based on DS evidence theory and online learning, in order to solve the technical problems in the existing technology such as the inability to effectively integrate multi-source diagnostic information, the static threshold leading to false alarms and missed alarms, and the lack of adaptive parameter tuning capabilities.
[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect, this invention provides a health scoring method for the EH oil system based on DS evidence theory and online learning, comprising the following steps: S1. The mechanism engine collects field data from the EH oil system through the first Modbus RTU interface and data from online oil quality monitoring equipment through the second Modbus RTU interface, including particle size and water content indicators. Based on the collected data and configurable logic rules, the mechanism engine performs fault detection in real time and outputs alarm status. Among them, oil quality detection adopts a statistical counting trigger method, and an alarm is output when the number of samplings exceeding the alarm threshold within a set window period reaches a preset threshold. The prediction engine obtains historical monitoring data from the time series database, selects a time series prediction model for multivariate prediction after stationarity and cointegration tests, and performs multiple complementary anomaly detection algorithms on the prediction residuals, and obtains a comprehensive anomaly score through weighted fusion. S2. Map the alarm status output by the mechanism engine to the first basic probability allocation function under the Dempster-Shafer evidence theory framework, and map the comprehensive anomaly score output by the prediction engine to the second basic probability allocation function. S3. The first basic probability allocation function and the second basic probability allocation function are fused using the Dempster orthogonal composition rule, and the fused basic probability allocation function and conflict coefficient are calculated. S4. Implement a graded conflict management strategy based on the conflict coefficient value. Calculate the probability of each fault type based on the fused basic probability allocation function through Pignistic probability transformation, and calculate the health score by combining the severity weight. S5. By periodically executing statistical threshold adaptation, data distribution drift detection, and Bayesian parameter optimization through low-frequency background tasks, diagnostic parameters are automatically updated.
[0011] Preferably, in step S1, the configuration logic rules include: Rising-edge triggers and falling-edge triggers are used to capture changes in signal state; A turn-on delay timer is used for logical determination of the duration; Combination expressions of Boolean operations, numerical comparisons, and arithmetic operations; For oil quality monitoring, a statistical counting trigger is used: the number of samples exceeding the alarm threshold is accumulated within a preset time window, and an oil quality alarm is output when the number of samples exceeding the threshold within the window period reaches the preset alarm count threshold.
[0012] Preferably, in step S1, the cascading selection strategy of the time series prediction model is: First, an augmented Dickey-Fuller stationarity test was performed on each monitored variable. If the variables are non-stationary, perform an Engle-Granger cointegration test on each pair of variables; If cointegration exists, the Vector Error Correction Model (VECM) is used for modeling. If there is no cointegration relationship, a vector autoregression (VAR) model is used for modeling. Based on the established model, the optimal lag order is selected through information criteria, and the predicted value and the prediction interval calculated based on the residual standard deviation are obtained by forward prediction.
[0013] Preferably, in step S1, the multiple complementary anomaly detection algorithms include: The Grubbs test is used to detect a single most significant outlier. LSCP anomaly detection based on local selection classification is used to capture anomalous patterns in local regions. SRA anomaly detection based on the significance of spectral residuals is used to identify structural anomalies in sequences; The anomaly ratio is calculated for each algorithm separately, and then weighted and fused to obtain a comprehensive anomaly score.
[0014] Preferably, in step S2, the method for converting the alarm state to the basic probability allocation function includes: Count the number of alarms for each fault type and map the alarm signals to the fault type using the configured mapping relationship; When two alarms for excessive particle size and excessive water content in oil quality monitoring are triggered simultaneously, the corresponding fault type mass value is calculated by adding the two alarms together. When there are no alarms, high-confidence quality is assigned to the normal state, and the remainder is assigned to the identification framework set Θ, which represents uncertainty; where the identification framework set Θ is defined as the set of all fault types and normal states. When there is an alarm, the total confidence score is weighted and allocated according to the number of alarms for each fault type, and the remaining quality is allocated to the entire set of the identification framework Θ. Normalization ensures that the sum of the masses of all focal elements is 1.
[0015] Preferably, in step S2, the method for converting the comprehensive anomaly score into the basic probability assignment function includes: Based on the position of the comprehensive anomaly score relative to the preset low and high thresholds, one of three basic probability allocation schemes is selected: Low anomaly range: high probability quality is assigned to normal state, and low probability quality is assigned to the union of faults and the entire set of the identification framework Θ; Medium-abnormal range: medium-probability quality is assigned to the normal state, and higher-probability quality is assigned to the fault union and the entire set of the identification framework Θ; High anomaly interval: low probability quality is assigned to normal state, high probability quality is assigned to fault union, and medium probability quality is assigned to the entire set of recognition frames Θ.
[0016] Preferably, in step S3, the orthogonal synthesis rule is:
[0017]
[0018] in, The basic probability assignment function after fusion; To set the objective; The conflict coefficient; To sum up the elements whose intersection is A; The basic probability assignment function for the mechanism engine; Propositions supported by the mechanism engine; The basic probability assignment function for the prediction engine; Propositions supported by the prediction engine; This is the summation of sets whose intersection is an empty set.
[0019] Preferably, in step S4, the hierarchical conflict management strategy includes a three-level response: Level 1: When the conflict coefficient K is less than the warning threshold, the Dempster fusion result is used directly; Level 2: When the conflict coefficient K is greater than or equal to the warning threshold and less than the critical threshold, a conflict alarm is triggered. The Dempster fusion result is still used, but it is marked as high risk. Level 3: When the conflict coefficient K is greater than or equal to the critical threshold, the Dempster fusion rule fails and a rollback strategy is triggered. One of the following strategies is selected for execution: mechanism priority strategy, weighted average strategy, and separate independent output strategy.
[0020] Preferably, in step S4, the method for calculating the health score includes: Through Pignistic probability transformation: This transforms the basic probability assignment function into reliable probabilities for each basic event; where, For Pignistic probability; Basic events; To identify any subset within the frame; To sum over all subsets A containing x; Assign basic probability values to subset A; Health score = Base score × (1 - Fault deduction - Uncertainty deduction); The fault deduction is calculated as follows: = Σ BetP (fault type) × severity weight. Uncertainty deduction = m(Θ) × uncertainty penalty coefficient; where m(Θ) is the basic probability assignment value of the entire set Θ; The severity of oil deterioration is weighted lower than the severity of valve jamming and oil pump malfunctions, which directly affect unit operation. It also outputs a ranking of fault types, sorted in descending order of Pignatic probability.
[0021] Preferably, in step S5, the statistical threshold adaptation includes: Periodically retrieve historical data for a specified window period from a time-series database; Calculate the mean μ and standard deviation σ of the data; According to the formula new threshold =μ±k×σ Recalculate the upper and lower thresholds, where k is the configured standard deviation multiple; Compare the upper and lower thresholds with the current threshold; if the change exceeds the stabilization percentage, perform an update. The update results are persisted to the cache and configuration files.
[0022] Preferably, in step S5, the data distribution drift detection includes: The data is divided into a reference window and a test window; Perform KS two-sample test and Z-score mean drift detection on each monitoring field; If the p-value of the KS two-sample test is less than the threshold, it is determined that the distribution shape has changed significantly. If the absolute value of the Z-score is greater than the threshold, the mean is considered to have drifted significantly. If any test is significant, a drift is determined to have occurred, triggering a recalculation of the statistical threshold. For oil quality monitoring fields, a more lenient drift judgment threshold is adopted.
[0023] Preferably, in step S5, the Bayesian parameter optimization includes: The optimal combination of predictive threshold parameters is searched using a Gaussian process optimization framework and an EI acquisition function. Two objective functions are provided: the tagged mode uses historical maintenance records and a weighted F1 score as the objective function; the untagged agent mode comprehensively evaluates three components: parameter consistency, parameter stability, and the rationality of alarm frequency. After optimization, parameters that can be automatically modified are applied directly, while suggestions are only provided for parameters that require manual confirmation.
[0024] Preferably, the method further includes four types of threshold hierarchical management: Statistical threshold A: Automatically calculated based on sliding window statistical data, and can be automatically modified; Predictive threshold B: Automatically searched by Bayesian optimization parameter tuning, allowing for automatic modification; Mechanism-based threshold C: A fixed threshold used in the mechanism engine logic, maintained only manually; Security Boundary Threshold D: Security-related hard limits, marked as read-only, and prohibited from any automatic modification; Each threshold entry includes metadata such as name, type, current value, associated field, direction, standard deviation multiple, window period, minimum number of samples, stabilization percentage, search space, whether automatic modification is allowed, and whether manual confirmation is required.
[0025] Secondly, this invention also provides a health scoring system for the EH oil system based on DS evidence theory and online learning, including: The mechanism engine module, deployed on the ARM64 edge computing device, collects EH oil system data through the first and second Modbus RTU interfaces, including control valve position, oil pressure, oil temperature, vibration, liquid level, particle size and water content signals. It performs fault detection in real time through configurable logic rules and outputs deterministic alarm status. The prediction engine module, deployed on the data analysis server, retrieves historical monitoring data from the time series database, performs time series prediction and multiple complementary anomaly detections, and outputs a comprehensive anomaly score. The DS fusion engine module converts mechanism alarms and predicted anomalies into basic probability assignment functions, performs multi-source fusion through Dempster orthogonal synthesis rules, and executes a hierarchical conflict management strategy. The health scoring module calculates health scores and fault rankings based on the fused basic probability allocation function through Pignistic probability transformation. The online learning module includes a statistical threshold adaptive submodule, a data distribution drift detection submodule, and a Bayesian parameter optimization submodule, which periodically and automatically updates diagnostic parameters. The communication middleware enables decoupled communication between modules through message queues and cache databases.
[0026] The beneficial effects of this invention are: 1. Multi-source information fusion improves diagnostic accuracy: The deterministic rule diagnosis of the mechanism engine and the probabilistic statistical prediction of the prediction engine are unified and integrated through DS evidence theory to overcome the limitations of a single method and achieve complementary verification on multiple fault types.
[0027] 2. Conflict detection and intelligent backoff mechanism: The conflict coefficient K built into DS fusion can quantify the degree of contradiction between two information sources. The three-level conflict management strategy ensures that reasonable diagnostic results can still be output when there is contradiction between information sources, avoiding misjudgment caused by the inaccuracy of a single information source.
[0028] 3. Online Adaptation Reduces Operation and Maintenance Costs: The three-layer online learning architecture of statistical threshold adaptation, data drift detection, and Bayesian optimization parameter tuning enables the system to automatically adapt to equipment aging and changes in operating conditions, significantly reducing manual intervention and parameter tuning workload.
[0029] 4. Edge-cloud collaborative architecture: The mechanism engine is deployed on ARM64 edge devices to achieve millisecond-level real-time detection, while the prediction engine uses statistical learning models to provide deep analysis. The two communicate decoupled through a cache database and MQTT message bus, balancing real-time performance and computing power.
[0030] 5. Configurable and Scalable: The logic rules of the mechanism engine are managed through configuration files, decoupled from the code. When adding fault types or adjusting logic, there is no need to modify the code and recompile, reducing the cost of expansion projects.
[0031] 6. Deep Integration of Oil Quality Monitoring and Overall Diagnostic Framework: Particle size and water content data collected by online oil quality monitoring equipment are not only output to the HMI or DCS system as independent display indicators, but also participate in DS multi-source fusion as key inputs to the mechanism engine. When both particle size and water content exceed the standard, the system uses a mass accumulation mechanism to determine oil deterioration faults with higher confidence. It combines the oil deterioration trend with other fault signals such as valve jamming and oil pump abnormalities for joint analysis, triggering preventive maintenance recommendations in the early stages of oil deterioration. Attached Figure Description
[0032] Figure 1 This is a diagram of the overall system architecture of the present invention, showing the dual-engine architecture of the mechanism engine and the prediction engine, as well as the data flow of Redis, TDengine, and MQTT; Figure 2 This is a flowchart of the dual-engine parallel diagnostic process of the present invention; Figure 3 This is a flowchart of the DS evidence theory fusion and conflict management process of the present invention; Figure 4 This is a flowchart of the online adaptive learning parameter tuning process of the present invention; Figure 5 This is a flowchart of the health score calculation process of the present invention; Figure 6 This is a schematic diagram of the four types of threshold hierarchical management of the present invention. Detailed Implementation
[0033] The following will provide a clear and complete description of the concept, specific structure, and technical effects of the present invention in conjunction with the embodiments and accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention.
[0034] Example 1 A health scoring method for the EH oil system based on DS evidence theory and online learning, such as Figures 1-6 As shown, it includes the following steps: Step S1: Dual-engine data acquisition and parallel diagnostics 1. The mechanism engine collects field data from the EH oil system via the first Modbus RTU protocol, including signals such as control valve position, oil pressure, oil temperature, vibration, and liquid level; and independently collects data from online oil quality monitoring equipment via the second Modbus RTU protocol, including particle size and water content indicators. Based on configurable logic rules (consistent with PLC ladder logic), it performs real-time logical judgments such as valve jamming detection, servo valve jamming detection, unloading valve anomaly detection, and oil pressure anomaly detection. For oil quality monitoring, a statistical counting trigger method is used—the number of samples exceeding the alarm threshold is counted within a preset time window. An alarm is only output when the number of samples exceeding the threshold within the window reaches the preset alarm count threshold, avoiding false alarms caused by occasional anomalies; a deterministic alarm status is output (0=normal, 1=alarm). 2. The prediction engine obtains historical monitoring data from the time series database, uses the augmented Dickey-Fuller (ADF) test and Engle-Granger cointegration test to determine whether there is a long-term equilibrium relationship between variables, and performs multivariate time series prediction based on the Vector Error Correction Model (VECM) or VAR model to obtain the predicted value and prediction interval for each monitoring point. 3. Three complementary anomaly detection algorithms are applied to the predicted residuals: Grubbs test to detect single outliers, Locally Selective Classification by Probability (LSCP) to detect local anomalies, and Sparse Representation-based Anomaly Detection (SRA) to detect sparse reconstruction anomalies; the anomaly ratios of the three algorithms are combined to obtain a comprehensive anomaly score.
[0035] Step S2: Conversion from Mechanism Alarm to Basic Probability Assignment (BPA) Map the alarm state output by the mechanism engine in step S1 to the basic probability allocation Mass function under the Dempster-Shafer framework: - The Frame of Discernment is defined as the set of all fault types and normal states: Θ = {Normal, Valve jamming, Servo valve jamming, Unloading valve jamming, Abnormal oil pressure, Abnormal oil temperature, Abnormal oil level, Abnormal oil pump, Oil deterioration}; where oil deterioration is a composite fault type, including two monitoring dimensions: excessive particle size and excessive water content. - Count alarms for each fault type and assign mass values to the corresponding fault type based on the number of alarms; - When two alarms from oil quality monitoring (excessive particle size and excessive water content) are triggered simultaneously, the mass value of the corresponding oil deterioration fault type is calculated by summing the two alarms, reflecting a higher confidence level of multi-parameter joint deterioration; - The remaining mass is assigned to the entire set Θ of the recognition framework, representing uncertainty; - Assign the high confidence mass to the "normal" state when there are no alarms.
[0036] Step S3: Conversion of predicted anomalies to Basic Probability Assignment (BPA) The three anomaly ratios output by the prediction engine in step S1 are weighted and fused together: - Overall anomaly score = w_grubbs × r_grubbs + w_lscp × r_lscp + w_sra × r_sra; - Based on the position of the comprehensive anomaly score relative to the preset low and high thresholds, select one of three BPA allocation schemes: - Low anomaly scheme: high mass is assigned to {normal}, and low mass is assigned to the union of faults and Θ; - Medium anomaly scheme: medium mass is assigned to {normal}, and higher mass is assigned to the union of faults and Θ; - High anomaly scheme: low mass is assigned to {normal}, high mass is assigned to the union of faults, and medium mass is assigned to Θ.
[0037] Step S4: Dempster-Shafer Evidence Theory Multi-Source Fusion The mechanistic BPA and predicted BPA obtained in steps S2 and S3 are fused using Dempster orthogonality and synthesis rules: - m_fused(A) = (1 / (1-K)) × Σ_{B∩C=A} m_mech(B) × m_pred(C), where K= Σ_{B∩C= } m_mech(B) × m_pred(C) is the conflict coefficient; - Output the merged BPA, conflict coefficient K, and whether the fusion was rejected.
[0038] Step S5: Three-level conflict management strategy Based on the value of the conflict coefficient K, a three-level response strategy is implemented: - Level 1 (K < warning threshold, default 0.3): Directly use the Dempster fusion result; - Level 2 (Warning threshold ≤ K < Critical threshold, default 0.6): Triggers a conflict alarm, still uses the Dempster fusion result but marks it as high risk; - Level 3 (K ≥ Critical Threshold): Dempster fusion rule fails (denominator 1-K approaches 0), triggering a fallback strategy, choosing one of the following three strategies: (a) Mechanism Priority: Directly use the mechanism engine BPA; (b) Weighted Average: Weight the two BPAs according to the source weights; (c) Separate Processing: Output the mechanism and prediction results independently.
[0039] Step S6: Health Score Calculation Based on the fused BPA, the belief function is transformed into a reliable probability through Pignistic probability transformation: - BetP(x) = Σ_{A x} m(A) / |A|, where A iterates over all focal elements; - Health score = Base score (100) × (1 - Fault deduction - Uncertainty deduction); - Fault deduction = Σ_{fault type i} BetP(fault type i) × severity weight i; where the severity weight of oil deterioration is set to 0.8, which is lower than fault types that directly affect unit operation, such as valve jamming (1.0) and oil pump abnormality (1.1), because oil quality problems are progressive deterioration and have a longer treatment window. - Uncertainty deduction = m(Θ) × Uncertainty penalty coefficient (0.3).
[0040] It also outputs a ranking of fault types (arranged in descending order of belief) for maintenance personnel to refer to.
[0041] Step S7: Online learning for adaptive parameter tuning The diagnostic parameters are automatically optimized through low-frequency background tasks, including three parallel sub-modules: (a) Statistical threshold adaptation (executed once every 24 hours) - Retrieve historical data for a specified window period (e.g., the last 90 days) from a time series database; - Calculate the mean and standard deviation of the data; - Recalculate the threshold using the formula new_threshold = mean + σ × std (upper threshold) or mean - σ × std (lower threshold); - Update only when the change exceeds the stabilization percentage (default 5%) compared to the current threshold, to avoid parameter jitter; - Update results and persist them to the cache and configuration files.
[0042] (b) Data distribution drift detection (executed every 6 hours) - Divide the data into a reference window (e.g., 90 days ago to 7 days ago) and a test window (the last 7 days); - Perform two statistical tests on each monitored field: - KS two-sample test: Tests whether two sets of data come from the same distribution. If the p-value is less than the threshold (default 0.01), it is determined that the distribution shape has changed significantly. - Z-score mean drift: |(test_mean - ref_mean) / ref_std| is considered to have a significant mean drift if it is greater than the threshold (default 3.0); - If any test is significant, a drift is determined to have occurred, triggering a recalculation of the statistical threshold. - The drift results are persisted to the cache.
[0043] - For oil quality monitoring fields (particle size, water content), a more lenient drift judgment threshold is adopted because the trend of increasing oil quality parameters represents the normal trend of equipment aging or oil deterioration, rather than a structural change in operating conditions, thus avoiding frequent false triggering of threshold recalculation.
[0044] (c) Bayesian optimization and parameter tuning (executed every 168 hours / week) - Using the Gaussian process optimization framework, the optimal combination of predictive threshold parameters is searched using the EI (Expected Improvement) acquisition function; - Two objective functions are provided: - Labeled mode: Utilizes historical maintenance records as true labels, with the F1 score (0.5×F1 + 0.3×(1-FPR) + 0.2×Recall) as the objective function; - Unlabeled proxy mode: Uses a proxy objective function to comprehensively evaluate three components: parameter consistency (conflict coefficient proxy), parameter stability (degree of deviation from default value), and the rationality of alarm frequency; - After optimization, the optimal parameters will be automatically applied to the registry, and suggestions will only be made for thresholds marked as requiring confirmation.
[0045] Example 2 This embodiment further elaborates on the overall system architecture based on the above embodiments.
[0046] The intelligent monitoring system for the EH oil system in this embodiment is deployed at the turbine site, such as... Figure 1 As shown, a two-tier architecture of "edge computing + data analysis" is adopted: Edge computing layer (ARM64 devices, such as Phytium / Kunpeng platforms) - The mechanism engine module runs as compiled native code, with an execution cycle of 10 seconds. - Data from the PLC is acquired via the first Modbus RTU protocol, including signals from key components such as the control valve group, main steam valve group, and fuel supply device. - Data from the online oil quality monitoring device (particle size sensor) is independently acquired via a second Modbus RTU protocol. The register address range is distinguished from the data from the oil supply device. The raw values are converted by a 10-fold ratio before being used in logical judgments. - Communicating with the prediction engine via the MQTT protocol - Alert results via Redis storage mechanism Data analysis layer (Linux server) - Prediction Engine Module: Performs data acquisition, VECM / VAR time series prediction, and triple anomaly detection. - Configuration-based logic engine module: Parses YAML rules and executes supplementary logic checks. - Health scoring module: Integrates both DS fusion and raw weighted scoring modes. - DS Fusion Engine Module: Performs Dempster orthogonal composition and conflict management - Online learning module: includes three sub-modules: statistical adaptive learning, drift detection, and Bayesian optimization. - TDengine Time Series Database: Stores historical monitoring data - Redis cache database: stores real-time alert status and online learning metadata. - MQTT Broker: A message bus serving as both the edge layer and the data analytics layer. Configuration Management - Mechanism and logic configuration: One set of YAML configurations each for the regulating valve, main steam valve, and fuel supply device. - DS fusion parameter configuration: recognition framework, source weights, BPA transformation parameters, conflict management threshold - Online learning for parameter configuration: threshold registry, scheduling cycle, drift detection parameters, Bayesian optimization parameters The fault types and severity weights included in the identification framework Θ are shown in the table below: Table 2.1 Fault types and severity weights included in the identification framework Θ
[0047] Example 3 This embodiment further elaborates on the mechanism engine diagnostic process based on the above embodiments.
[0048] The mechanism engine runs in 10-second cycles on ARM64 edge devices, such as... Figure 2 As shown, execute the following process: 1. Data Reading: Collect field data of the EH oil system through the first Modbus RTU module, including: - Control valve group (e.g., 6 groups): valve position feedback, oil pressure, safety oil pressure, return oil temperature, command signal, etc.; - Main steam valve assembly (e.g., 2 assemblies): oil pressure, safety oil pressure, return oil temperature, etc.; - Fixed points of the oil supply device: main pump vibration X / Y, oil tank temperature, oil tank level, system oil pressure, pump current, high pressure cut-off test pressure, etc.; Data from the online oil quality monitoring equipment is collected via the second Modbus RTU module, including: - Particle size classification: ISO / NAS / GJB standard classification, converted by 10 from the original value; - Moisture content: An indicator of the water content in oil; 2. Logical Judgment: Executes logical operations equivalent to PLC ladder diagrams through a configurable logic engine, supporting: - Triggers: Rising edge trigger (R_trig), Falling edge trigger (F_trig); - Timer: On-Delay Timer (TON); - Boolean operations: AND, OR, NOT; - Numerical comparisons: <, >, ≤, ≥, ==, !=; - Arithmetic operations: +, -, ×, ÷.
[0049] Statistical counting trigger: For oil quality monitoring, the number of samples exceeding the alarm threshold is counted within a preset time window. When the number of samples exceeding the threshold within the window reaches the preset alarm count threshold, an alarm is output. Specific parameters are: window period of 2 days, alarm count threshold of 10 times per day for two consecutive days.
[0050] 3. Alarm Output: When the logical result is "Alarm", the alarm status is written to Redis. The key format is "{Project Name}{Unit Number}:Mechanism:{Logic Template}", which includes field values, logical status, timestamp and other information.
[0051] 4. MQTT Publishing: Alarm information is also published to a specified topic via MQTT for the prediction engine to subscribe to.
[0052] Example 4 This embodiment further elaborates on the anomaly detection of the prediction engine based on the above embodiments.
[0053] like Figure 2 As shown, the prediction engine executes the following process: 1. Data Acquisition: Query the prediction data for the most recent steps and the anomaly detection data for the most recent window period from the historical table through the TDengine connector.
[0054] 2. Cointegration test: Perform Engle-Granger cointegration tests on each pair of monitored variables to determine whether a long-term equilibrium relationship exists.
[0055] 3. VECM / VAR modeling: - If the variables have a cointegration relationship, use VECM (Vector Error Correction Model) to capture long-term equilibrium and short-term fluctuations; - If there is no cointegration relationship, use the VAR (Vector Autoregression) model; - Select the optimal lag order using the AIC / BIC criterion.
[0056] 4. Time series prediction: Use the trained VECM / VAR model to predict several steps forward to obtain the predicted value and the prediction interval calculated based on the residual standard deviation.
[0057] 5. Triple anomaly detection: - Grubbs test: Detects a single most significant outlier and is suitable for normally distributed data; - LSCP Anomaly Detection: Anomaly detection based on local selection classification, integrating four base detectors: LOF, COF, INNE, and ECOD, to capture anomaly patterns in local regions; - SRA Anomaly Detection: Anomaly detection based on the significance of spectral residuals, identifying structural anomalies in sequences through frequency domain reconstruction errors.
[0058] 6. Anomaly Ratio Calculation: For each algorithm, the proportion of samples detected as anomalies is statistically analyzed to obtain three anomaly ratios: r_grubbs, r_lscp, and r_sra (all within the range of [0,1]).
[0059] 7. MQTT Publishing: Publish prediction results, anomaly rates, and health scores via MQTT topics.
[0060] Example 5 This embodiment further elaborates on the DS evidence theory fusion engine based on the above embodiments.
[0061] like Figure 3 As shown, the DS fusion engine is the core component for realizing multi-source fusion, consisting of three parts: BPA converter, Dempster synthesizer, and conflict manager.
[0062] 5.1 BPA Converter The BPA converter reads the identification framework and mapping relationship from the configuration file and performs two types of conversions: Mechanism alarm → BPA conversion Input: Alarm status dictionary, where the key is the logic template identifier (e.g., "control valve: logic 1") and the value is 0 or 1.
[0063] Conversion logic: - Count the number of alarms for each fault type and map the logic template identifier to the fault type through the configured fault mapping relationship (e.g., "Control valve: Logic 1" → valve jamming). - When two alarms from the oil quality monitoring system (excessive particle size and excessive water content) are triggered simultaneously, the mass value of the corresponding oil deterioration fault type is calculated by summing the values of the two alarms. - When there are no alarms: Normal state quality = configured normal state quality without alarms (default 0.85), the rest are assigned to Θ (uncertainty); - When there are alarms: Total confidence = min(confidence × number of alarms / total number of logic, confidence), weighted according to the number of alarms for each fault type; the remaining quality is assigned to Θ; - Normalization ensures that sum(mass) = 1.
[0064] Predicting anomalies → BPA conversion Input: r_grubbs, r_lscp, r_sra (all in [0,1]).
[0065] Conversion logic: -Comprehensive anomaly score = w_grubbs×r_grubbs + w_lscp×r_lscp + w_sra×r_sra (default weights 0.3 / 0.3 / 0.4); - Based on whether the score falls into one of the three ranges: low (< low threshold), medium (low threshold ~ high threshold), or high (> high threshold), select the corresponding BPA allocation scheme: - Low anomaly scheme: normal=0.80, fault_union=0.00, Θ=0.20 - Exception handling scheme: normal=0.30, fault_union=0.40, Θ=0.30 - High anomaly scheme: normal=0.05, fault_union=0.80, Θ=0.15 5.2 Dempster Orthogonal Assortment Rules The Dempster synthesizer performs dual-source fusion, and the algorithm flow is as follows: - Traverse all focal element pairs of the first source m1 and the second source m2: - Calculate the intersection of two focal elements; - If the intersection is empty, the cumulative conflict amount = m1(focal element 1) × m2(focal element 2); - If the intersection is not empty, the cumulative composite quantity [intersection] += m1(joule 1) × m2(joule 2); - If the conflict is ≥ 1.0 (complete conflict), return a uniform distribution; - Otherwise, perform normalization for each focal element: combined[focal element] / = (1 - conflict); - Returns the synthesis result and conflict coefficient.
[0066] 5.3 Conflict Management The fusion engine executes the complete fusion process: 1. Convert mechanism-based alarms and predicted anomalies into BPAs respectively; 2. Perform Dempster orthogonal synthesis to fuse the two BPAs and obtain the conflict coefficient K; 3. Compare K with the two-level thresholds: - K<0.3: Normal fusion, output fused BPA; - 0.3 ≤ K<0.6: Conflict warning, still outputs merged BPA but marked conflict_warning; - K ≥ 0.6: Fusion rejected, rollback strategy executed.
[0067] 5.4 Pignistic Probability Transformation and Health Score like Figure 5 As shown, the Pignistic probability transformation converts the BPA containing uncertainty mass into a reliable probability for each fundamental event: BetP(x) = Σ_{A: x∈A} m(A) / |A| Health score calculation: - Normal probability P_normal = BetP(normal) - Fault penalty = Σ BetP(fault_i) × severity_weight(fault_i) - Uncertainty penalty = m(Θ) × uncertainty_penalty - Health score = 100 × (1 - Fault deduction - Uncertainty deduction), range [0, 100].
[0068] Example 6 This embodiment further elaborates on the online learning adaptive parameter tuning system based on the above embodiments.
[0069] The online learning module is a complete adaptive parameter management framework.
[0070] 6.1 Threshold Registry The threshold registry manages four types of thresholds, such as Figure 6 As shown, the attributes of various thresholds are as follows: Table 6.1 Thresholds: Four Categories of Thresholds in Registry Management
[0071] Each threshold entry contains the following metadata: - Name, type (statistical / predictive / mechanistic / safety), current value; - Associated TDengine fields (single or multiple fields), direction (upper threshold / lower threshold); - Sigma (standard deviation multiple), data backtracking window days, minimum effective sample size; - Anti-shake percentage (not updated if the change is less than this value), search space [low, high] (used for Bayesian optimization); - Whether automatic modification is allowed, and whether manual confirmation is required.
[0072] The update_value() method performs security checks before updating: it rejects automatic modification based on security boundaries, performs debouncing checks to avoid parameter jitter, and writes the updated data to Redis and the configuration file.
[0073] 6.2 Adaptive Statistical Threshold like Figure 4 As shown, the statistical adapter executes once every 24 hours: 1. Retrieve all threshold entries with type=statistical from the threshold registry; 2. For each entry, query the historical data for the most recent window_days days from TDengine (satisfying the minimum number of samples min_samples). 3. Calculate the mean and standard deviation (std); 4. Calculate the new threshold by direction: - Upper threshold: new = mean + sigma × std - Lower threshold: new = mean - sigma × std 5. Round to one decimal place; 6. Update only if the change exceeds change_threshold_pct% compared to the current value; 7. Update the results and write them to Redis and YAML.
[0074] The compute interface provides a pure compute version (without database access) for unit testing and offline verification.
[0075] 6.3 Data Distribution Drift Detection like Figure 4 As shown, the drift detector executes once every 6 hours: 1. Obtain the list of fields to be monitored from the configuration; 2. For each field: - Get reference window data (ref_days days ago to test_days days ago, default 90 days to 7 days); - Retrieve test window data (last test_days days, default 7 days); - Perform a two-sample KS test to obtain the statistic and p-value; - Calculate Z-score: z = (test_mean - ref_mean) / ref_std; - Detect drift: ks_p<ks_p_threshold(0.01) OR |z|> zscore_threshold(3.0); - For oil quality monitoring fields (particle size, moisture content), a more lenient drift judgment threshold is adopted; 3. If drift is detected, the statistical threshold will be automatically recalculated; 4. Write the drift result to Redis, with the key format: {PROJECT}{UNIT}:OL:Drift:{field}.
[0076] The testing interface provides a pure computational version for unit testing.
[0077] 6.4 Bayesian Optimization and Parameter Tuning like Figure 4 As shown, the Bayesian optimizer executes once every 168 hours (week): 1. Extract all thresholds with type=predictive and search space from the threshold registry as optimization parameters; 2. Construct a real-dimensional search space for Gaussian process optimization; 3. Choose the objective function: - Labeled objective function: Labeled mode, replaying records maintained in history, with F1 score plus false positive penalty as the objective (0.5×F1 + 0.3×(1-FPR) + 0.2×Recall); - Agent objective function: Tagless agent mode, comprehensively evaluate parameter consistency (0.4), parameter stability (0.3), and alarm frequency rationality (0.3); 4. Call the Gaussian process optimizer to perform optimization (EI acquisition function, default 30 iterations, 8 random initial explorations); 5. Application optimization results: Parameters that do not require manual confirmation are automatically written; otherwise, only log alerts are issued.
[0078] The results application supports both automatic write and suggestion-only modes.
[0079] 6.5 Online Learning Scheduler The online scheduler is invoked every tick in the main loop, internally triggering low-frequency tasks according to a counter / time interval: - Statistical update: every 8640 ticks (24 hours @ 10 seconds / tick) - Drift detection: every 2160 ticks (6 hours) - Bayesian optimization: per 60480 ticks (168 hours / week) All tasks are executed asynchronously in a thread pool (maximum number of worker threads = 2), without blocking the main loop. A manual trigger interface is provided for debugging.
[0080] Example 7 This embodiment further elaborates on data flow and communication based on the above embodiments.
[0081] The system decouples its modules through a three-layer communication protocol: 1. Modbus RTU (Edge Side → Field Device): The mechanism engine collects PLC register data through the first Modbus RTU (sampling frequency synchronized with PLC scan cycle, typical value 50-100ms), and provides it for logic judgment after debouncing and filtering; it independently collects granular sensor data (holding register FC03) through the second Modbus RTU, and the original value is converted by ÷10 and then used in statistical counting trigger logic.
[0082] 2. Caching database (Redis, edge computing) Data Analysis Layer): - Mechanism engine write: Store the judgment status, field values, and timestamps of each logic in the key format of "project name + unit number + logic template"; - The prediction engine reads the mechanism alert status in Redis and combines it with historical data in the time series database to perform anomaly detection; - Online learning data: Stores dynamic thresholds, drift detection results, and Bayesian optimization reports.
[0083] 3. MQTT message bus (data analysis layer → external system): - Health Rating Topic: Publish health rating data, including ratings and timestamps; - DS Fusion Topic: Publish fusion results, including health score, conflict coefficient, conflict warning, fusion rejection status, fault ranking, pignistic probability distribution, and source BPA.
[0084] Example 8 This embodiment further elaborates on the key configuration parameters based on the above embodiments.
[0085] The key parameters for DS fusion configuration are shown in the table below: Table 8.1 Identification Frame Configuration
[0086] Table 8.2 BPA Conversion Parameters
[0087] Table 8.3 Source Weights and Conflict Management
[0088] Table 8.4 Health Score Parameters
[0089] Table 8.5 Severity Weights for Each Fault Type
[0090] The key parameters for configuring online learning are shown in the table below: Table 8.6 Scheduling Cycle
[0091] Table 8.7 Drift Detection Parameters
[0092] Table 8.8 Bayesian Optimization Parameters
[0093] The embodiments of the present invention have been described in detail above, but the present invention is not limited to the described embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention, and these equivalents or substitutions are all included within the scope defined by the claims of the present invention.
Claims
1. A method for health scoring of EH oil systems based on DS evidence theory and online learning, characterized in that, Includes the following steps: S1. The mechanism engine collects field data from the EH oil system through the first Modbus RTU interface and data from online oil quality monitoring equipment through the second Modbus RTU interface, including particle size and water content indicators. Based on the collected data and configurable logic rules, the mechanism engine performs fault detection in real time and outputs alarm status. Among them, oil quality detection adopts a statistical counting trigger method, and an alarm is output when the number of samplings exceeding the alarm threshold within a set window period reaches a preset threshold. The prediction engine obtains historical monitoring data from the time series database, selects a time series prediction model for multivariate prediction after stationarity and cointegration tests, and performs multiple complementary anomaly detection algorithms on the prediction residuals, and obtains a comprehensive anomaly score through weighted fusion. S2. Map the alarm status output by the mechanism engine to the first basic probability allocation function under the Dempster-Shafer evidence theory framework, and map the comprehensive anomaly score output by the prediction engine to the second basic probability allocation function. S3. The first basic probability allocation function and the second basic probability allocation function are fused using the Dempster orthogonal composition rule, and the fused basic probability allocation function and conflict coefficient are calculated. S4. Implement a graded conflict management strategy based on the conflict coefficient value. Calculate the probability of each fault type based on the fused basic probability allocation function through Pignistic probability transformation, and calculate the health score by combining the severity weight. S5. By periodically executing statistical threshold adaptation, data distribution drift detection, and Bayesian parameter optimization through low-frequency background tasks, diagnostic parameters are automatically updated.
2. The DS evidence theory and online learning-based EH oil system health scoring method as described in claim 1, characterized in that, In step S1, the configuration logic rules include: Rising-edge triggers and falling-edge triggers are used to capture changes in signal state; A turn-on delay timer is used for logical determination of the duration; Combination expressions of Boolean operations, numerical comparisons, and arithmetic operations; For oil quality monitoring, a statistical counting trigger is used: the number of samples exceeding the alarm threshold is accumulated within a preset time window, and an oil quality alarm is output when the number of samples exceeding the threshold within the window period reaches the preset alarm count threshold.
3. The EH oil system health scoring method based on DS evidence theory and online learning as described in claim 1, characterized in that, In step S1, the cascade selection strategy of the time series prediction model is as follows: First, an augmented Dickey-Fuller stationarity test was performed on each monitored variable. If the variables are non-stationary, perform an Engle-Granger cointegration test on each pair of variables; If cointegration exists, the Vector Error Correction Model (VECM) is used for modeling. If there is no cointegration relationship, a vector autoregression (VAR) model is used for modeling. Based on the established model, the optimal lag order is selected through information criteria, and the predicted value and the prediction interval calculated based on the residual standard deviation are obtained by forward prediction.
4. The EH oil system health scoring method based on DS evidence theory and online learning as described in claim 1, characterized in that, In step S1, the multiple complementary anomaly detection algorithms include: The Grubbs test is used to detect a single most significant outlier. LSCP anomaly detection based on local selection classification is used to capture anomalous patterns in local regions. SRA anomaly detection based on the significance of spectral residuals is used to identify structural anomalies in sequences; The anomaly ratio is calculated for each algorithm separately, and then weighted and fused to obtain a comprehensive anomaly score.
5. The EH oil system health scoring method based on DS evidence theory and online learning as described in claim 1, characterized in that, In step S2, the methods for converting alarm states to the basic probability assignment function include: Count the number of alarms for each fault type and map the alarm signals to the fault type using the configured mapping relationship; When two alarms for excessive particle size and excessive water content in oil quality monitoring are triggered simultaneously, the corresponding fault type mass value is calculated by adding the two alarms together. When there are no alarms, high-confidence quality is assigned to the normal state, and the remainder is assigned to the identification framework set Θ, which represents uncertainty; where the identification framework set Θ is defined as the set of all fault types and normal states. When there is an alarm, the total confidence is weighted and allocated according to the number of alarms for each fault type, and the remaining quality is allocated to the entire set Θ of the identification framework; Normalization ensures that the sum of the masses of all focal elements is 1.
6. The EH oil system health scoring method based on DS evidence theory and online learning as described in claim 1, characterized in that, In step S2, the method for converting the comprehensive anomaly score into the basic probability assignment function includes: Based on the position of the comprehensive anomaly score relative to the preset low and high thresholds, one of three basic probability allocation schemes is selected: Low anomaly range: high probability quality is assigned to normal state, and low probability quality is assigned to the union of faults and the entire set of the identification framework Θ; Medium-abnormal range: medium-probability quality is assigned to the normal state, and higher-probability quality is assigned to the fault union and the entire set of the identification framework Θ; High anomaly interval: low probability quality is assigned to normal state, high probability quality is assigned to fault union, and medium probability quality is assigned to the entire set of recognition frames Θ.
7. The EH oil system health scoring method based on DS evidence theory and online learning as described in claim 1, characterized in that, In step S3, the orthogonal synthesis rule is as follows: in, The basic probability assignment function after fusion; To set the objective; The conflict coefficient; To sum up the elements whose intersection is A; The basic probability assignment function for the mechanism engine; Propositions supported by the mechanism engine; The basic probability assignment function for the prediction engine; Propositions supported by the prediction engine; This is the summation of sets whose intersection is an empty set.
8. The EH oil system health scoring method based on DS evidence theory and online learning as described in claim 1, characterized in that, In step S4, the hierarchical conflict management strategy includes a three-level response: Level 1: When the conflict coefficient K is less than the warning threshold, the Dempster fusion result is used directly; Level 2: When the conflict coefficient K is greater than or equal to the warning threshold and less than the critical threshold, a conflict alarm is triggered. The Dempster fusion result is still used, but it is marked as high risk. Level 3: When the conflict coefficient K is greater than or equal to the critical threshold, the Dempster fusion rule fails and a rollback strategy is triggered. One of the following strategies is selected for execution: mechanism priority strategy, weighted average strategy, and separate independent output strategy.
9. The EH oil system health scoring method based on DS evidence theory and online learning as described in claim 1, characterized in that, In step S4, the method for calculating the health score includes: Through Pignistic probability transformation: This transforms the basic probability assignment function into reliable probabilities for each basic event; where, For Pignistic probability; Basic events; To identify any subset within the frame; To sum over all subsets A containing x; Assign basic probability values to subset A; Health score = Base score × (1 - Fault deduction - Uncertainty deduction); The fault deduction is calculated as follows: = Σ BetP (fault type) × severity weight. Uncertainty deduction = m(Θ) × uncertainty penalty coefficient; where m(Θ) is the basic probability assignment value of the entire set Θ; The severity of oil deterioration is weighted lower than the severity of valve jamming and oil pump malfunctions, which directly affect unit operation. It also outputs a ranking of fault types, sorted in descending order of Pignatic probability.
10. The EH oil system health scoring method based on DS evidence theory and online learning as described in claim 1, characterized in that, In step S5, the statistical threshold adaptation includes: Periodically retrieve historical data for a specified window period from a time-series database; Calculate the mean μ and standard deviation σ of the data; According to the formula new threshold =μ±k×σ Recalculate the upper and lower thresholds, where k is the configured standard deviation multiple; Compare the upper and lower thresholds with the current threshold; if the change exceeds the stabilization percentage, perform an update. The update results are persisted to the cache and configuration files.
11. The EH oil system health scoring method based on DS evidence theory and online learning as described in claim 1, characterized in that, In step S5, the data distribution drift detection includes: The data is divided into a reference window and a test window; Perform KS two-sample test and Z-score mean drift detection on each monitoring field; If the p-value of the KS two-sample test is less than the threshold, it is determined that the distribution shape has changed significantly. If the absolute value of the Z-score is greater than the threshold, the mean is considered to have drifted significantly. If any test is significant, a drift is determined to have occurred, triggering a recalculation of the statistical threshold. For oil quality monitoring fields, a more lenient drift judgment threshold is adopted.
12. The EH oil system health scoring method based on DS evidence theory and online learning as described in claim 1, characterized in that, In step S5, the Bayesian parameter optimization includes: The optimal combination of predictive threshold parameters is searched using a Gaussian process optimization framework and an EI acquisition function. Two objective functions are provided: the tagged mode uses historical maintenance records and a weighted F1 score as the objective function; the untagged agent mode comprehensively evaluates three components: parameter consistency, parameter stability, and the rationality of alarm frequency. After optimization, parameters that can be automatically modified are applied directly, while suggestions are only provided for parameters that require manual confirmation.
13. The EH oil system health scoring method based on DS evidence theory and online learning as described in claim 1, characterized in that, The method also includes four types of threshold hierarchical management: Statistical threshold A: Automatically calculated based on sliding window statistical data, and can be automatically modified; Predictive threshold B: Automatically searched by Bayesian optimization parameter tuning, allowing for automatic modification; Mechanism-based threshold C: A fixed threshold used in the mechanism engine logic, maintained only manually; Security Boundary Threshold D: Security-related hard limits, marked as read-only, and prohibited from any automatic modification; Each threshold entry includes metadata such as name, type, current value, associated field, direction, standard deviation multiple, window period, minimum number of samples, stabilization percentage, search space, whether automatic modification is allowed, and whether manual confirmation is required.
14. A health scoring system for the EH oil system based on DS evidence theory and online learning, characterized in that... The EH oil system health scoring method applied to any one of claims 1-13 includes: The mechanism engine module, deployed on the ARM64 edge computing device, collects EH oil system data through the first and second Modbus RTU interfaces, including control valve position, oil pressure, oil temperature, vibration, liquid level, particle size and water content signals. It performs fault detection in real time through configurable logic rules and outputs deterministic alarm status. The prediction engine module, deployed on the data analysis server, retrieves historical monitoring data from the time series database, performs time series prediction and multiple complementary anomaly detections, and outputs a comprehensive anomaly score. The DS fusion engine module converts mechanism alarms and predicted anomalies into basic probability assignment functions, performs multi-source fusion through Dempster orthogonal synthesis rules, and executes a hierarchical conflict management strategy. The health scoring module calculates health scores and fault rankings based on the fused basic probability allocation function through Pignistic probability transformation. The online learning module includes a statistical threshold adaptive submodule, a data distribution drift detection submodule, and a Bayesian parameter optimization submodule, which periodically and automatically updates diagnostic parameters. The communication middleware enables decoupled communication between modules through message queues and cache databases.
