Dam safety state self-diagnosis method and system based on data-mechanism-knowledge combined driving

Abstract

The invention discloses a data-mechanism-knowledge combined driving dam safety state self-diagnosis method and system, and the method comprises the steps: obtaining dam operation state data, extracting a state quantity based on a pre-constructed computable evidence graph, and generating a monitoring evidence; performing inversion correction on mechanism model parameters under the priori constraint of the atlas, and generating mechanism evidence; the monitoring evidence and the mechanism evidence are accessed to a probability graph model for joint inference, and arbitration correction is carried out when a conflict triggering condition is met; and carrying out credibility calibration on the diagnosis result by adopting a conformal prediction method based on the historical calibration sample set, and outputting a security state diagnosis conclusion containing a confidence interval. According to the method, the problems of multi-source evidence conflict, mechanism parameter distortion and low diagnosis result credibility are effectively solved, and credible and automatic diagnosis of dam safety is realized.

Description

Technical Field

[0001] This invention belongs to the field of water conservancy project safety monitoring, and in particular, it is a data-mechanism-knowledge-driven method and system for self-diagnosis of dam safety status. Background Technology

[0002] As a crucial water conservancy infrastructure project, the operational safety status diagnosis of dams is a key aspect of ensuring project safety. Continuous monitoring and analysis of key physical quantities such as deformation and seepage, timely identification of abnormal structural evolution, and implementation of corresponding measures are of great significance for reducing operational risks and ensuring the safe operation of the project.

[0003] Currently, dam safety monitoring primarily utilizes automated monitoring systems to acquire data on deformation, seepage, stress, and strain, which are then combined with regular manual inspections and historical assessment data for comprehensive analysis. Common analytical methods include monitoring trend analysis based on statistical regression (e.g., HST-type models) and structural verification calculation methods based on mechanistic models such as the finite element method. To enhance the comprehensive evaluation capabilities of multi-source information, some research and engineering practices have begun to explore the use of information fusion methods to comprehensively evaluate monitoring data and inspection information.

[0004] However, under complex operating conditions and the coexistence of multi-source heterogeneous information, existing technologies still face several shortcomings: First, inconsistencies in time scales, spatial misalignment, and differences in source credibility may exist among multi-source information, leading to inconsistencies or even conflicts between monitoring data, inspection records, and mechanism analysis results. When there are missing data, outliers, or conflicting evidence, traditional fusion methods often lack explicit mechanisms for handling conflicting evidence, which may cause diagnostic results to tend to average out and affect the sensitivity of abnormal state identification. Second, mechanism model parameters may drift under the influence of factors such as long-term engineering operation and changes in materials and boundary conditions. If parameter inversion is based solely on observation fitting, different parameter combinations may produce similar responses, resulting in insufficient parameter identifiability or parameter values ​​deviating from the physically reasonable range, affecting the ability of mechanism verification results to characterize the actual structural behavior. Third, existing diagnostic results are mostly given in the form of single-level or uncalibrated probabilities, making it difficult to quantify the statistical credibility of the conclusions, which is not conducive to forming verifiable confidence guarantees and a traceable diagnostic interpretation process. Therefore, there is an urgent need for a self-diagnostic technology for dam safety during operation that can comprehensively utilize engineering knowledge, operational data, and mechanistic model information within a unified framework, while also addressing evidence conflict resolution, updating mechanistic parameter constraints, and quantifying the credible expression of diagnostics. Summary of the Invention

[0005] The purpose of this invention is to provide a data-mechanism-knowledge-driven method and system for self-diagnosis of dam safety status, in order to solve the aforementioned problems existing in the prior art.

[0006] The technical solution, a data-mechanism-knowledge jointly driven method for self-diagnosis of dam safety status, includes:

[0007] Acquire operational status data of the dam, extract state variables characterizing the dam's safety status based on a pre-constructed computable evidence map, and map them into monitoring evidence for safety status propositions;

[0008] Under the prior constraints of the computable evidence map, the pre-stored mechanism model parameters are inverted and corrected using monitoring evidence, and the verification calculation is performed accordingly to obtain the mechanism evidence.

[0009] Monitoring evidence and mechanistic evidence are used as observation factors and fed into a pre-configured probabilistic graphical model for joint inference to obtain the initial probability distribution of the safety state. When the preset conflict triggering conditions are met, the weights of monitoring evidence and mechanistic evidence are adjusted using a pre-configured logical rule model and arbitration inference is performed to obtain the corrected posterior probability distribution of the safety state.

[0010] Based on a pre-stored historical calibration sample set, the confidence level of the primary probability distribution or the corrected posterior probability distribution of the safety state is calibrated using a conformal prediction method, resulting in a safety state diagnostic conclusion containing a confidence interval.

[0011] According to one aspect of this application, a data-mechanism-knowledge jointly driven dam safety state self-diagnosis system includes:

[0012] The holographic perception and evidence mapping module is used to acquire the dam's operational status data, extract state variables that characterize the dam's safety status based on a pre-constructed computable evidence map, and map them into monitoring evidence for safety status propositions.

[0013] The mechanism co-inversion module is used to invert and correct the pre-stored mechanism model parameters using monitoring evidence under the prior constraints of the computable evidence map, and to perform verification calculations to obtain mechanism evidence.

[0014] The fusion arbitration inference module is used to input monitoring evidence and mechanistic evidence as observation factors into a pre-configured probabilistic graphical model for joint inference to obtain the initial probability distribution of the safety state; and when the preset conflict triggering conditions are met, it is used to adjust the weights of monitoring evidence and mechanistic evidence using a pre-configured logical rule model and perform arbitration inference to obtain the corrected posterior probability distribution of the safety state.

[0015] The credible decision output module is used to perform credibility calibration on the primary probability distribution of the safety state or the corrected posterior probability distribution of the safety state based on the pre-stored historical calibration sample set and using the conformal prediction method to obtain a safety state diagnostic conclusion containing a confidence interval.

[0016] Beneficial effects: Under the unified engineering evidence framework, this invention realizes a collaborative inference mechanism of monitoring data, mechanism models and engineering knowledge. Through conflict triggering and arbitration processing and statistical confidence calibration, it achieves a structurally consistent expression and credible quantitative output of safety state diagnostic results, thereby improving the consistency, interpretability and engineering applicability of the diagnostic process. Attached Figure Description

[0017] Figure 1 A flowchart illustrating the steps of a data-mechanism-knowledge jointly driven self-diagnosis method for dam safety status provided in this application embodiment.

[0018] Figure 2 This is a flowchart illustrating the steps of mapping state variables to monitoring evidence for security state propositions, as provided in an embodiment of this application.

[0019] Figure 3 This is a flowchart illustrating the steps of inverting and correcting pre-stored mechanistic model parameters using monitoring evidence, as provided in an embodiment of this application.

[0020] Figure 4 A flowchart illustrating the steps for adjusting the weights of monitoring evidence and mechanistic evidence, as provided in this application embodiment.

[0021] Figure 5 The structural diagram of the dam safety state self-diagnosis system driven by data, mechanism and knowledge provided in the embodiments of this application is shown. Detailed Implementation

[0022] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0023] It should be noted that the terms include and have, and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or device.

[0024] like Figure 1 As shown, a data-mechanism-knowledge jointly driven method for self-diagnosis of dam safety status includes the following steps:

[0025] The system acquires operational status data of the dam, extracts state variables characterizing the dam's safety status based on a pre-constructed computable evidence map, and maps these state variables into monitoring evidence for safety status propositions.

[0026] Specifically, operational status data refers to various multi-source heterogeneous data reflecting the physical state and environmental conditions of the dam at the current moment or over a period of time. For example, it may include time-series monitoring data such as deformation, seepage, and stress-strain collected by automated monitoring systems; on-site inspection records such as cracks, leaks, and erosion obtained through manual inspections; and data on past incidents or abnormal events. The computable evidence map not only stores engineering objects and attributes but also predefines inference rules from data characteristics to safety conclusions. In this embodiment, the system cleans and spatiotemporally aligns the acquired data, extracting key state variables based on the rules defined in the map. For example, for piezometer water level monitoring data, state variables can be extracted as the amplitude of the current water level, the recent rate of change of the water level, and the hysteresis correlation coefficient between the water level and the reservoir water level. Using a preset mapping function (such as a Sigmoid scoring function or a piecewise membership function), the numerical state variables are transformed into probabilistic descriptions of safety status propositions. For example, if the rate of change of water level in a piezometer exceeds the warning value, the mapping result may be the proposition: abnormal seepage behavior, probability: 0.85, which constitutes monitoring evidence.

[0027] In some alternative implementations, for non-numerical inspection records, such as the discovery of a new crack, the system will convert it into a qualitative evidence entry based on the semantic matching rules in the graph, and assign it a corresponding confidence level and probability value. For example, the longitudinal crack on the dam crest can be mapped as evidence supporting the proposition of dam slope instability.

[0028] Under the prior constraints of the computable evidence map, the pre-stored mechanistic model parameters are inverted and corrected using monitoring evidence, and the mechanistic evidence is obtained by verification calculation based on the corrected mechanistic model parameters.

[0029] In this embodiment, the mechanistic model typically refers to a mathematical model built upon physical principles, such as a finite element analysis (FEM) model or a mixture statistical model (HST). Traditional parameter inversion often relies solely on fitting observational data, which can easily lead to problems such as different parameters having the same effect or parameter values ​​violating physical common sense. To address this issue, inversion can be performed under the prior constraints of a computable evidence spectrum.

[0030] Specifically, the system extracts reasonable physical ranges or prior distributions of mechanistic parameters (such as permeability coefficient and elastic modulus) from the graphs and incorporates them as regularization terms into the inversion objective function. By minimizing the residuals between observed and calculated values, as well as the deviations between parameters and prior values, corrected parameters that conform to engineering realities are obtained. Based on the corrected parameters, the system runs the mechanistic model again for verification calculations, obtaining key verification indicators such as the anti-sliding stability safety factor and seepage gradient. According to the safety criteria required by the specifications, the verification indicators are transformed into probabilistic mechanistic evidence. For example, if the calculated anti-sliding stability safety factor is greater than the allowable value in the specifications, then the proposition "anti-sliding stability" with a probability of 0.99 is generated, ensuring the consistency between the mechanistic analysis and the current actual monitoring status.

[0031] Monitoring evidence and mechanistic evidence are used as observation factors and fed into a pre-configured probabilistic graphical model for joint inference to obtain the primary probability distribution of the safety state. When the preset conflict triggering conditions are met, the weights of monitoring evidence and mechanistic evidence are adjusted using a pre-configured logical rule model and arbitration inference is performed to obtain the corrected posterior probability distribution of the safety state.

[0032] Alternatively, it can be described as follows: a probabilistic graphical model is constructed with the safety state proposition as a latent variable. Monitoring evidence and mechanistic evidence are used as observation factors and fed into the probabilistic graphical model for joint inference to obtain the primary probability distribution of the safety state; it is then determined whether the preset conflict triggering conditions are met; when the preset conflict triggering conditions are met, the weights of the monitoring evidence and mechanistic evidence are adjusted using a pre-configured logical rule model and arbitration inference is performed to obtain the corrected posterior probability distribution of the safety state; if the preset conflict triggering conditions are not met, the primary probability distribution of the safety state is used as the probability distribution to be calibrated.

[0033] Specifically, the probabilistic graphical model preferably adopts the factor graph model. In the factor graph model, safety-state propositions such as the overall safety of the dam and local seepage anomalies are represented as latent variable nodes, while monitoring evidence and mechanistic evidence are represented as factor nodes, connected to the corresponding variable nodes through potential functions. The system performs joint inference through the sum-product algorithm or the belief propagation algorithm, fusing multi-source information to obtain a preliminary probability distribution. However, in actual engineering, there may be conflicts between data sources, such as normal monitoring data but abnormal mechanistic calculations, or extremely poor data quality. In this case, forced fusion may lead to distorted conclusions. Therefore, a conflict triggering mechanism can be introduced. When the data missing rate is detected to be too high or the degree of conflict between evidence exceeds a threshold, the system will automatically switch to arbitration mode. In arbitration mode, a logical rule model, such as probabilistic soft logic (PSL), is used to introduce higher-order logical constraints. For example, if piping occurs, it must be a dangerous situation, and the weight of evidence is dynamically adjusted according to the type of conflict, such as reducing the weight of low-quality evidence. By solving for the maximum posterior probability or minimizing the hinge loss in the logical model, the system outputs a logically corrected posterior probability distribution of the safe state, providing diagnostic results that conform to engineering logic even under extreme conditions.

[0034] Based on a pre-stored historical calibration sample set, the confidence level of the primary probability distribution or the corrected posterior probability distribution of the safety state is calibrated using a conformal prediction method, resulting in a safety state diagnostic conclusion containing a confidence interval.

[0035] In this embodiment, although the probability distribution has incorporated multi-source information, probability values ​​such as 0.9 are often statistically uncalibrated, meaning they do not strictly represent a 90% accuracy rate. To provide dam managers with a legally valid or risk-based decision-making basis, a conformal prediction method is introduced.

[0036] Specifically, the system utilizes a pre-collected historical calibration sample set—that is, historical data on known true safety states—to calculate a non-conformity measure for the current diagnostic result, measuring the deviation of the current sample from historical normal samples. Based on a user-defined confidence level, such as 95%, a quantile threshold is calculated, and a prediction set containing the true label, i.e., a confidence interval, is constructed accordingly. The final output safety state diagnostic conclusion is no longer a single label, but a set with a confidence guarantee; for example, the probability that the current state belongs to the set {normal, slightly abnormal} is 95%. This effectively quantifies the uncertainty risk of the diagnosis, achieving reliable diagnosis.

[0037] This embodiment addresses the problems of dam safety assessment relying on human experience, difficulties in integrating multi-source information, and lack of reliable quantification of diagnostic results. It proposes an automated diagnostic framework that integrates engineering knowledge, monitoring data, and mechanistic models. This framework can run on the server side of the dam safety monitoring and management system and is applicable to the operational safety diagnosis of various dam types, such as earth-rock dams and concrete dams.

[0038] In one possible embodiment, the data-mechanism-knowledge jointly driven dam safety state self-diagnosis method can also be:

[0039] Obtain operational status data of the dam;

[0040] State variables characterizing the dam's safety state are extracted based on a pre-constructed computable evidence map, and these state variables are mapped to monitoring evidence for the safety state proposition.

[0041] Under the prior interval of physical parameters defined by the computable evidence spectrum and the engineering physical constraints, the monitoring evidence is used to invert and correct the pre-stored mechanism model parameters, and the corrected mechanism model is used to perform verification calculations to generate mechanism evidence.

[0042] A probabilistic graphical model containing safety-state propositional variable nodes and corresponding evidence factor nodes is constructed. The monitoring evidence and mechanism evidence are used as observation factors and connected to the probabilistic graphical model for joint inference to obtain the primary probability distribution of safety-state.

[0043] When at least one conflict triggering function outputs a preset triggering condition, the weights of the monitoring evidence and the mechanism evidence are adjusted using a logical rule model, and arbitration inference is performed based on the adjusted weights to obtain the corrected security state posterior probability distribution.

[0044] Based on a pre-stored historical calibration sample set, the conformal prediction method is used to statistically calibrate the primary probability distribution of the safety state or the corrected posterior probability distribution of the safety state, construct a prediction set that meets a preset confidence level, and output a safety state diagnostic conclusion containing a confidence interval.

[0045] To further illustrate the implementation mechanism of self-diagnosis in this embodiment, the following explanation is provided: In this embodiment, self-diagnosis does not refer to the automatic classification of a single model, but rather to the system's ability to automatically complete a closed-loop inference process—including security state inference, conflict resolution, and credibility calibration—under multi-source information constraints without human intervention. The self-diagnosis process includes at least the following technical mechanisms:

[0046] The system automatically extracts state variables based on operation monitoring data, inspection records, and historical event data, and maps the state variables into probabilistic evidence representations of safety state propositions, thereby realizing the autonomous generation and structured expression of evidence. This process does not rely on manual scoring or expert subjective judgment.

[0047] The constructed computable evidence graph is used not only to store engineering knowledge but also to provide prior constraints on mechanistic model parameters and a consistency structure among evidence. During the mechanistic parameter inversion process, the system updates the model parameters under the constraints of the graph and outputs parameter uncertainty information, enabling the mechanistic model to adaptively match the current operating state.

[0048] During the multi-source evidence fusion phase, the system constructs a unified probabilistic inference model based on the graph structure and performs joint inference on evidence from monitoring data, inspection information, historical events, and mechanism verification. When evidence conflicts or data quality anomalies are detected, the system automatically adjusts the evidence weights and performs arbitration processing according to preset engineering trigger conditions, without manual intervention.

[0049] After obtaining the fused security state probability distribution, the system introduces a statistical calibration mechanism to ensure the confidence level of the diagnostic results, outputs a prediction set that meets the preset confidence level, and simultaneously generates evidence path information that supports the diagnostic conclusion.

[0050] Through the synergistic effect of the above multi-layered mechanisms, a closed-loop collaborative inference of the data layer, mechanism layer and knowledge layer is realized, enabling the safety-state diagnosis process to have autonomy, structural consistency and statistical reliability, thus constituting a self-diagnostic capability in an engineering sense.

[0051] In one possible implementation, the computable evidence graph includes an ontology layer and an evidence layer;

[0052] The ontology layer defines the topological relationships between engineering objects, structural parts, safety indicators, disease types, and sets of safety-state propositions;

[0053] The evidence layer defines the association path between evidence items and the set of security-state propositions, and the edge weights of the association path are composed of probability components and credibility components.

[0054] Specifically, the ontology layer forms the skeleton of the graph, used for semantic modeling of knowledge in the field of dam safety. In this embodiment, the engineering objects defined by the ontology layer can be concretized into entity classes such as earth-rock dams and concrete gravity dams; structural parts can include entities such as dam foundations, dam body cutoff walls, and spillways; safety indicators cover monitored physical quantities such as horizontal displacement, seepage flow, and piezometer water levels; and disease types include common engineering diseases such as piping, soil erosion, and uneven settlement. The set of safety state propositions is the final output target of the diagnosis, defined for example as the set {normal state, slight anomaly, severe anomaly, dangerous situation}. The ontology layer also defines the topological relationships between these entities, such as the relation Monitor_at (monitoring at) connecting safety indicators and structural parts, and the relation Indicate (representation) connecting disease types and safety state propositions. Topological relationships ensure that the reasoning process conforms to engineering logic; for example, an anomaly in seepage flow at the dam foundation can only support the proposition of anomaly in the dam foundation seepage, and will not incorrectly point to the proposition of dam crest deformation.

[0055] Building upon this, the evidence layer is used to instantiate specific evidence items and establish their association with propositions. Each piece of evidence, whether from sensor monitoring or mechanistic models, is represented as an evidence node in the graph. This node is connected to the security-state proposition node in the ontology layer via weighted edges. The edge weights of the association paths are not single numerical values, but rather two-dimensional vectors or composite scalars composed of probability and confidence components.

[0056] In some implementations, the edge weight W edge It can be represented in linear combination form:

[0057] W edge =α*P evidence +β*C credibility ;

[0058] Among them W edge P represents the edge weights used for inference. evidence C represents a probability component, indicating the probability that the evidence supports the corresponding proposition, for example, 0.8; credibility The confidence component represents the reliability of the evidence itself, for example, 0.9; α and β are preset adjustment coefficients, usually satisfying α+β=1. This ensures that the system considers not only the probability of the evidence but also its confidence when fusing evidence, effectively avoiding the misleading influence of low-quality, high-probability evidence on diagnostic results.

[0059] In an optional embodiment, the credibility component is weighted and quantified based on the authority level of the data source and the data quality index; the degree of evidence conflict is calculated based on the conflict coefficient or Jousselme distance in the Dempster-Shafer evidence theory to characterize the distributional difference between the monitoring evidence probability vector and the mechanistic evidence probability vector.

[0060] In this embodiment, to achieve the computability of the graph, a clear quantitative definition is given for the confidence component. Specifically, the confidence component C... credibility The calculation formula is as follows:

[0061] C credibility =w auth *L auth +w qual *I qual ;

[0062] Where L auth The authority level of the data source is a normalized scalar, for example, defined as: L of a calibrated precision instrument. auth =1.0, Regular Inspection Record L auth =0.8, unverified tip L auth =0.5; I qual Data quality metrics, such as data integrity or signal-to-noise ratio, are calculated based on the data cleaning process; auth and w qual These are the corresponding weighting coefficients. The system can automatically assign objective credibility labels to evidence from different sources and of different qualities.

[0063] Furthermore, to support the subsequent conflict triggering mechanism, this embodiment also defines a method for calculating the degree of evidence conflict. When monitoring evidence and mechanistic evidence respectively provide probability distributions for the same set of propositions, it is necessary to quantify the degree of conflict between them. In a preferred embodiment, the system uses the conflict coefficient K in Dempster-Shafer (DS) evidence theory to calculate the degree of conflict, and the formula is as follows:

[0064] K=Σ(m1(A i )*m2(B j For all satisfying A i ∩B j =Sum of (i, j) in the empty set;

[0065] Where K is the conflict coefficient, m1(A) i To monitor evidence for proposition A i The basic probability distribution, m2(B j ) as mechanistic evidence for proposition B jThe basic probability assignment. When K is close to 1, it indicates that the two are highly conflicting, for example, one points to normal and the other points to danger.

[0066] In another alternative implementation, to overcome the paradox of DS theory under highly conflicting conditions, the system can employ the Jousselme distance d J To measure conflict, the calculation formula is as follows:

[0067] d J =sqrt(0.5*(V m -V p ) T *D mat *(V m -V p ));

[0068] Where d J Let V be the Jousselme distance. m V is the probability vector of the evidence being monitored. p Let be the probability vector of the mechanistic evidence. T D represents the vector transpose. mat This is a distance matrix constructed based on proposition similarity. If the propositions are independent, D... mat It can be simplified to an identity matrix. The Jousselme distance can more precisely characterize the differences between two probability distributions in geometric space, providing a precise numerical basis for conflict triggering.

[0069] According to one aspect of this application, a computable formula for the credibility of evidence is constructed, which integrates data quality, parameter uncertainty, and degree of evidence conflict into a unified credibility index.

[0070] Specifically, the credibility components need to be further specified into computable mathematical expressions. This embodiment provides a method for calculating the credibility of evidence that integrates multiple factors. Credibility of evidence c e The calculation formula is as follows:

[0071] c e =σ(α0+α1×Q e -α2×U θ -α3×D e_avg );

[0072] Where c e For the credibility of evidence e, σ represents the Sigmoid function used to normalize the result to the zero-one interval, α0 is the bias term, and α1, α2, and α3 are the weight coefficients of each factor, respectively. Q e As the quality metric of the data source corresponding to evidence e, U θ D is the uncertainty index for the inversion parameters. e_avgLet Q be the average degree of conflict between evidence e and other evidence. Higher data quality equates to higher credibility, reflected in Q. e The positive coefficient of the term; the greater the parameter uncertainty, the lower the credibility of the mechanistic evidence, which is reflected in U. θ The negative coefficient of the item; the greater the conflict between the evidence and other evidence, the lower the credibility, which is reflected in D. e_avg The negative coefficients of the terms. The linear combination result is mapped to the zero-one interval using the Sigmoid function to ensure the reasonableness of the confidence value. The default values ​​of the weight coefficients can be set based on engineering experience or calibrated using historical data. In a preferred embodiment, α0 is 1.0, α1 is 2.0, α2 is 1.5, and α3 is 1.5.

[0073] The computable evidence graph in this embodiment is not merely a static repository of engineering information, but also a computable bridge connecting multi-source heterogeneous data with security-state propositions. By defining the topological structure of the ontology layer and the quantified weights of the evidence layer, it achieves the transformation from qualitative knowledge to quantitative reasoning. The computable evidence graph can be stored in a graph database or a Resource Description Framework (RDF) for real-time access by the diagnostic system.

[0074] In one exemplary embodiment, the operational status data includes monitoring time-series data, inspection record data, and historical event data.

[0075] Specifically, monitoring time-series data refers to a sequence of physical quantities collected at a fixed frequency by the dam's automated monitoring system. For example, this may include: environmental quantities (reservoir water level, air temperature, rainfall), deformation quantities (surface displacement, internal settlement), seepage flow (piezometer water level, seepage around the dam), and stress strain quantities (reinforcement stress, earth pressure). In this embodiment, to assess data availability, the system can calculate data integrity rate as a key quality indicator. Data integrity rate R c The calculation formula is as follows:

[0076] R c =N valid / N total ;

[0077] Where N valid N represents the number of valid data points actually acquired. total This represents the theoretical total number of data points that should be collected. If R c If the value is below a preset threshold, such as 0.8, the subsequent conflict handling mechanism will be triggered or the interpolation completion procedure will be started.

[0078] Inspection record data refers to written descriptions, photographs, or video materials generated by operators during regular or irregular visual inspections of the dam. For example, a wetland of approximately 2 square meters was found at chainage 0+150 on the downstream face of the dam. Historical event data refers to records of special events that occur throughout the dam's entire life cycle, including records of all highest water levels, earthquake records, and records of past emergency responses, such as the piping failure on the back slope in 2010. Unstructured data contains empirical information that mechanistic models cannot cover, and is an important supplement to achieving holistic diagnosis.

[0079] like Figure 2 As shown, mapping state variables to monitoring evidence for security-state propositions includes:

[0080] Amplitude features, trend features, and environmental response features are extracted from the monitoring time series data as state variables. The state variables are then converted into probability vectors for safety state propositions using a preset scoring function.

[0081] The inspection record data and historical event data are converted into semantic evidence items according to structural parts and preset hierarchical rules, and then written into monitoring evidence in combination with probability vectors.

[0082] In this embodiment, for numerical monitoring time-series data, simple observations are insufficient to directly reflect the safety status; it is necessary to extract state quantities with physical meaning. Specifically, the system extracts the following three types of key features:

[0083] Amplitude characteristics: including the current observed value, recent extreme values, and margin relative to the design warning value.

[0084] Trend characteristics: To eliminate the interference of random noise on trend judgment, the Theil-Sen estimation algorithm is preferred to calculate the rate of change of the monitored data, i.e., the slope. The calculation formula is:

[0085] k TS =median((x j -x i ) / (ji));

[0086] Where k TS For a robust regression slope, median indicates the median operation, x j and x i These are the monitoring data values ​​at time points j and i, respectively (j>i). Compared to the traditional least squares method, the Theil-Sen slope is insensitive to outliers and can more realistically reflect the long-term evolution trend of dam behavior.

[0087] Environmental response characteristics: The hysteresis correlation coefficient between monitored effects (such as displacement) and environmental quantities (such as water level) is calculated to characterize the response sensitivity of the dam structure.

[0088] After extracting the state variables, the system maps them into probability vectors using a pre-defined scoring function. For example, for the state variable of seepage pressure, a sigmoid function can be used to map the pressure value into the probability P of the seepage anomaly proposition. abnormal :

[0089] P abnormal =1 / (1+exp(-k shape *(x val -x threshold )));

[0090] Where x val x is the value of the currently extracted state variable. threshold k is the preset anomaly detection threshold. shape The shape parameter controls the steepness of the curve. When x val Much greater than x threshold At that time, P abnormal Approaching 1; when x val Much smaller than x threshold At that time, P abnormal Approaching 0. Continuous probability mappings reflect the gradual change of state better than simple 0 / 1 hard threshold alarms.

[0091] For non-numerical inspection and historical data, the system employs a semantic transformation strategy. Specifically, based on the ontology layer definition in the evidence graph, text records are segmented and entity recognized to extract structural parts and descriptions of defects. According to the grading rules in industry standards, defect descriptions are mapped to severity levels (Level I, Level II, Level III). These levels are then converted into evidence items for specific propositions. For example, an inspection record describing severe piping is transformed into the evidence item: {Proposition: seepage failure, probability: 0.95, confidence level: 0.9}. This achieves seamless integration of qualitative description and quantitative reasoning.

[0092] This embodiment introduces robust statistical methods and semantic transformation rules, which makes the evidence input into the diagnostic system highly accurate and consistent.

[0093] like Figure 3 As shown, according to one aspect of this application, the inversion correction of pre-stored mechanistic model parameters using monitoring evidence includes:

[0094] The prior intervals of physical parameters and engineering physical constraints of the mechanism model parameters are extracted from the ontology layer of the computable evidence map.

[0095] Construct an inversion objective function, which includes an observation consistency term that characterizes the deviation between the mechanism model response and the monitoring evidence, and a priori constraint term that characterizes the deviation of the mechanism model parameters from the prior interval of the physical parameters.

[0096] The objective function of the inversion is optimized under the condition of satisfying the engineering physical constraints to obtain the parameter estimates of the mechanism model.

[0097] In other words, the prior intervals of physical parameters and engineering physical constraints of the mechanism model parameters are extracted from the ontology layer of the computable evidence map; an inversion objective function is constructed, which includes an observation consistency term representing the deviation between the mechanism model response obtained by running the mechanism model based on the current parameter values ​​and the monitoring evidence, and a prior constraint term representing the deviation of the mechanism model parameters from the prior intervals of physical parameters; the inversion objective function is optimized under the condition of satisfying the engineering physical constraints to obtain the parameter estimates of the mechanism model parameters, and the parameter estimates are used as the corrected mechanism model parameters.

[0098] Alternatively, the mechanism model parameter inversion can be described as follows: constructing an inversion objective function that includes observation consistency terms and parameter prior constraint terms; and optimizing the inversion objective function under the condition of satisfying engineering physical constraints to obtain the parameter estimates of the mechanism model parameters.

[0099] Specifically, in this embodiment, the mechanistic model can be a finite element model for seepage analysis or a hybrid statistical model for deformation analysis. The system accesses a computable evidence map and, based on the currently inverted object, such as the permeability coefficient of an earth-rock dam, retrieves the physical prior information of that parameter from the ontological layer. For example, for the permeability coefficient k of a clay core wall, the prior mean μ recorded in the map... k =1.0e -7 m / s, prior variance σ k 2 =1.0e -15 , and the physical constraint interval [1.0e -9 1.0e -5 The regularized inversion objective function is constructed. Unlike the traditional method that only minimizes the sum of squared residuals, the objective function J(θ) in this embodiment explicitly includes prior constraints from the graph. Its mathematical expression is as follows:

[0100] J(θ)=(yh(θ)) T *R -1 *(yh(θ))+λ*(θ-μ) T *P -1 *(θ-μ);

[0101] Where J(θ) is the inversion objective function value; θ is the parameter vector to be inverted, such as the permeability coefficient of each zone; y is the measured vector corresponding to the monitoring evidence, such as the water level observation value of the piezometer; h(θ) is the calculated response vector of the mechanism model under parameter θ; Tdenoted as transpose; R is the observation noise covariance matrix, reflecting the reliability of the monitoring data; μ is the parameter prior mean vector; P is the parameter prior covariance matrix, reflecting the uncertainty of the spectral knowledge; λ is the regularization coefficient that balances the weights of the observation and constraint terms.

[0102] In the objective function, the first term (yh(θ)) T *R -1 *(yh(θ)) is the observation consistency term, which forces the model output to approximate the measured data as closely as possible; the second term λ*(θ-μ) T *P -1 *(θ-μ) represents the prior constraint, ensuring that the inversion parameters do not deviate too far from the physical principles given by the graph. Using nonlinear optimization algorithms, such as Sequential Quadratic Programming (SQP) or the trust region algorithm, the optimal parameter estimate θ that minimizes J(θ) is found, while satisfying the physical constraints (e.g., k>0). hat .

[0103] To illustrate more intuitively, suppose the measured water level in the piezometer at a certain moment is y = 100.0 m, the calculated value from the model is h(θ), and the prior mean is μ = 100.5 m. If only the observed value is considered, the inversion algorithm might adjust the parameters to unreasonable values ​​in order to fit 100.0 m; however, by introducing prior constraints, the algorithm will find the optimal balance between fitting the data and conforming to common sense, resulting in more robust parameter estimates.

[0104] In a further embodiment, the method further includes outputting an uncertainty interface characterizing the inversion confidence level, the uncertainty interface comprising the following metrics:

[0105] The posterior covariance matrix of the parameters, calculated based on the second derivative information of the inverted objective function, is used to characterize the distribution range of the parameter estimates.

[0106] The residual statistics calculated based on the residual sequence after inversion convergence are used to characterize the explanatory power of the mechanistic model for monitoring evidence.

[0107] The identifiability index, calculated based on the parameter sensitivity matrix, is used to characterize the invertibility of mechanistic model parameters under current monitoring evidence.

[0108] In other words, the method also includes an uncertainty interface that outputs an uncertainty characterization of the inversion confidence level, the uncertainty interface including at least: the parameter posterior covariance matrix; residual statistics; and parameter identifiability index.

[0109] Furthermore, the uncertainty interface is used as a basis for the credibility weight of mechanistic evidence in the joint inference of probabilistic graphical models.

[0110] Specifically, traditional inversion typically only outputs definite parameter values, while this embodiment further outputs the statistical properties of the parameters, i.e., the uncertainty interface, which is crucial for subsequent probability fusion. In the optimal solution θ... hat Near the target function J(θ), the posterior covariance matrix C of the parameters can be approximately obtained by performing a second-order Taylor expansion. p :

[0111] C p ≈(J''(θ hat )+P -1 ) -1 ;

[0112] Where J''(θ) hat Let be the Hessian matrix of the objective function at the optimal solution, i.e., the second derivative matrix. p The diagonal elements represent the posterior variance of each inversion parameter. If the posterior variance of a parameter is large, it indicates that the parameter has high uncertainty under the current data, and the credibility of the subsequent mechanistic evidence should be reduced accordingly.

[0113] The root mean square error (RMSE) is typically used to calculate residual statistics.

[0114] RMSE=sqrt(mean((yh(θ hat )) 2 ));

[0115] RMSE, or Residual Statistic, reflects whether the corrected mechanistic model truly explains the monitoring data. If the RMSE remains large after inversion, it indicates that there may be a structural bias in the mechanistic model or that there are abnormal loads that were not considered by the model.

[0116] The identifiability index is typically calculated using singular value decomposition (SVD) based on the parameter sensitivity matrix S = Ψh / Ψθ, where Ψ is the partial derivative. If the smallest singular value of S approaches 0, it indicates that the data is insensitive to changes in certain parameters, which are unidentifiable. In this case, the uncertainty interface will issue an alert, prompting the diagnostic system not to over-rely on mechanistic analysis conclusions involving these parameters. By outputting interface information, a transition from point estimation to distribution estimation is achieved.

[0117] According to one aspect of this application, the inversion objective function also includes an evidence consistency constraint term and a fusion posterior guiding term;

[0118] The evidence consistency constraint term is used to ensure that the estimated values ​​of the constraint parameters fall within the constraint set automatically inferred from the computable evidence graph. It is calculated by summing the product of the confidence weight of each constraint path and the square of the distance from the parameter to the constraint set.

[0119] The fusion posterior guiding term is used to feed back the safety state fusion posterior probability distribution of the current iteration into the parameter inversion process as a guiding signal. It is calculated as the negative log-likelihood of the parameter-proposition mapping result under the fusion posterior distribution.

[0120] In other words, based on the inversion objective function, an evidence consistency constraint term and a fusion posterior guiding term are further introduced to construct a collaborative inversion objective function, so as to achieve bidirectional coupling between mechanism parameter inversion and evidence fusion.

[0121] Specifically, traditional inversion objective functions only include observation consistency terms and prior constraint terms, which suffers from the problem of one-way error propagation: when there is systematic bias in the monitoring data, the inversion parameters will deviate from the true values, and the error will inevitably propagate to the mechanistic evidence, ultimately contaminating the fusion results. To solve this problem, the original objective function can be extended to a collaborative inversion objective function. The mathematical expression of the collaborative inversion objective function is as follows:

[0122] J total (θ)=L fit (θ)+λ1×R prior (θ)+λ2×C evid (θ)+λ3×G post (θ);

[0123] J total (θ) represents the objective function value of the collaborative inversion, θ is the parameter vector of the mechanism model to be inverted, and L fit (θ) is the observation consistency term, R prior (θ) represents the prior constraint term, C evid (θ) is the evidence consistency constraint term, G post (θ) represents the fusion posterior guiding term, and λ1, λ2, and λ3 are the weight coefficients of each term. The observation consistency term L... fit (θ) is used to approximate the measured data with the response of the constraint mechanism model. The prior constraint term R... prior (θ) is used to constrain the parameters from deviating too far from the prior mean given by the spectrum.

[0124] Consistency of Evidence Constraint C evid (θ) is a newly added constraint term, representing the application of soft constraints to the inversion parameters using historical diagnostic conclusions, expert experience rules, and engineering constraint information stored in the computable evidence map, ensuring that the parameter estimation results are consistent with engineering knowledge. For example, the formula for calculating the evidence consistency constraint term is as follows:

[0125] C evid (θ)=Σ k (β k ×d(θ,A k ) 2);

[0126] Where C evid (θ) represents the value of the evidence consistency constraint term, Σ k This represents the summation over all constrained paths k, β k Let A be the credibility weight of the k-th evidence path. k Let d(θ, A) be the set of constraints imposed on parameter θ by the k-th evidence path. k Let θ be the parameter vector to the constraint set A. k The distance function.

[0127] Constraint set A k The specific form varies depending on the constraint type. When the constraint type is an interval constraint, the constraint set is represented as the allowed range of parameter values, and the distance function is defined as:

[0128] d(θ i A k )=max(0,θ i lo -θ i θ i -θ i hi );

[0129] Where d(θ) i A k ) is the parameter θ i To interval constraint A k The distance, θ i lo θ is the lower bound of the constraint interval. i hi This is the upper bound of the constraint interval. The distance is zero when the parameter value falls within the interval, and the distance is the excess amount when the parameter value exceeds the interval.

[0130] Fusion of the posterior guiding term G post (θ) is a newly added guiding term. It represents the use of the safety-state fusion posterior probability distribution of the current iteration as a guiding signal fed back to the parameter inversion process. This allows the inversion process to be aware of the downstream fusion results, and when monitoring data causes inversion deviations, the constraints of the fusion posterior can partially correct these deviations. For example, the calculation formula for the fusion posterior guiding term is as follows:

[0131] G post (θ)=-log(q n (P=φ(θ)));

[0132] Among them G post (θ) represents the value of the fusion posterior guiding term, log represents the natural logarithm, and q n(P) represents the posterior probability distribution of the safety state fusion obtained in the nth iteration, and φ(θ) is the parameter-proposition mapping function, used to map the mechanistic model parameter θ to the corresponding safety state proposition P. The role of the parameter-proposition mapping function φ(θ) is to establish the correspondence between mechanistic parameters and safety states. For example, when the inverted permeability coefficient θ makes the seepage stability safety factor greater than the allowable value in the specification, the mapping result is a normal seepage behavior proposition. This mapping is determined jointly by mechanistic verification calculation and safety criteria.

[0133] By introducing the two new terms mentioned above, the objective function of the collaborative inversion achieves the following: the evidence consistency constraint term enables the inversion process to utilize engineering knowledge in the graph, avoiding parameter estimation from violating physical common sense; the fusion posterior guidance term establishes a feedback channel from the fusion result to the inversion process, enabling the overall system to form a closed loop of data, inversion, mechanistic evidence, fusion, feedback, and inversion, thereby improving the robustness of the diagnostic results.

[0134] The constraint set A is automatically derived from the computable evidence graph. k and its credibility weight β k .

[0135] In this embodiment, constraint set A k Instead of being manually specified, the constraints are automatically obtained through graph querying and reasoning. For example, a method for automatically extracting parameter constraints from the evidence graph is provided, including four sub-steps: parameter-entity mapping, constraint path querying, constraint formal construction, and credibility weight calculation.

[0136] Sub-step one: Establish the mechanism model parameters θ i The semantic correspondence between the parameter and the entity nodes in the evidence map is determined. The semantic label is determined based on the physical meaning of the parameter, and the associated map entity type is determined based on the semantic label. For example, for the permeability coefficient parameter k, its semantic label is seepage characteristic parameter, and the associated map entity types include material nodes, seepage index nodes, and seepage disease nodes.

[0137] Sub-step two involves performing a breadth-first search in the evidence graph, starting with the parameter-associated entity nodes, to identify terminating nodes that satisfy the constraint semantics. Terminating nodes fall into three categories: numerical constraint nodes containing numerical ranges, historical conclusion nodes containing historical diagnostic conclusions, and parameter-associated nodes containing associated parameters. The search depth is limited to a preset value to control computational complexity. After the search is complete, the path and edge sequence from the starting node to the terminating node are recorded.

[0138] Sub-step three involves transforming the queried path into a computable constraint set based on the type of the path's termination node. Four constraint types are specifically defined:

[0139] Interval constraint: When the path terminates at a node within a numerical range, the allowed range of values ​​for the construction parameter is defined as A. k ={θ i ∈[θ i lo θ i hi The sources of interval boundaries include nominal values ​​in design documents, recommended ranges in specifications, or historical inversion calibration values.

[0140] Ratio constraint: When the path contains physical relationships between parameters, the allowable range of parameter ratios is denoted as A. k ={(θ i / θ j )∈[r lo r hi ]}, where r lo and r hi These are the lower and upper bounds of the ratio, respectively.

[0141] Conditional constraints: When the path contains load condition nodes, construct parameter constraints that depend on the load conditions, denoted as A. k ={θ i ∈[θ i lo (c), θ i hi (c)], when the operating condition is c}, where c is the operating condition variable.

[0142] Historical consistency constraint: When the path terminates at a historical diagnostic conclusion node, an implicit constraint is constructed that requires the inversion parameters to make the mechanism verification result consistent with the historical conclusion. Its distance function is calculated using KL divergence and is expressed as d(θ, A). k ) 2 =D KL (P mech (θ)||P hist ), where P mech (θ) is the mechanism kernel probability distribution calculated based on parameter θ, P hist Let D be the probability distribution corresponding to historical diagnostic conclusions. KL This represents the KL divergence.

[0143] Sub-step four: Calculate the confidence weight β based on the attributes of the constraint path. k The calculation formula is as follows:

[0144] β k =σ(w1×Q source +w2×R time +w3×S path -w0);

[0145] Where β kLet Q be the confidence weight of the k-th constraint path, σ represent the Sigmoid function used to normalize the result to the zero-one interval, and Q be the... source R is a score indicating the authority level of the data source. time S is the score for timeliness. path The path strength score is represented by w0, which is the bias term, and w1, w2, and w3 are the corresponding weight coefficients.

[0146] Timeliness score R time The calculation method is as follows:

[0147] R time =exp(-λ t ×Δt);

[0148] Where R time For timeliness score, exp represents the exponential function, λ t Δt represents the time-dependent decay coefficient, where Δt is the time interval between the constraint source and the current moment. This reflects the engineering experience that more recent constraint information has higher reliability.

[0149] An alternating iterative algorithm is used to solve the collaborative inversion objective function, achieving bidirectional coupling between mechanism inversion and evidence fusion.

[0150] Specifically, due to the fusion posterior guiding term G in the collaborative inversion objective function post (θ) depends on the fused posterior distribution q(P), which in turn depends on the mechanistic evidence, which in turn depends on the inversion parameter θ. Therefore, it cannot be solved directly. An alternating iterative approach is preferred, alternating between parameter updates and posterior updates until convergence. The initialization settings for the alternating iterative algorithm are as follows: Initialize the parameter θ... 0 Set as prior mean μ θ , the initial fusion post-test q 0 (P) is set to a uniform distribution, and the convergence threshold ε' and the maximum number of iterations N are set. max .

[0151] For example, the iterative steps of the alternating iterative algorithm are as follows, assuming the current iteration round is n: fix the fused posterior distribution q from the previous round. n-1 (P), find the parameter values ​​that minimize the joint inversion objective function:

[0152] θ n =argmin θ (J total (θ;q n-1 ));

[0153] Where θ n argmin represents the parameter estimate obtained in the nth iteration. θJ represents the value that minimizes the objective function with respect to parameter θ. total (θ;q n-1 ) represents the result after the previous round of fusion, q n-1 The objective function value for the joint inversion under given conditions can be obtained using the Levenberg-Marquardt algorithm or sequential quadratic programming.

[0154] Based on the updated parameter θ n The mechanism verification calculation was performed to obtain the verification index and its uncertainty:

[0155] I mech n =f check (θ n );

[0156] Among them I mech n f is the mechanism verification index obtained in the nth iteration. check This represents the mechanism verification calculation function. It also calculates the uncertainty of the verification index.

[0157] σ mech n =sqrt((J θ ) T ×Σ θ n ×J θ );

[0158] Where σ mech n Let J be the uncertainty of the mechanism verification index for the nth iteration, sqrt denotes the square root operation, and J is the value of J. θ Let Σ be the Jacobian matrix of the complex kernel function with respect to the parameters. θ n Let be the posterior covariance matrix of the parameters in the nth iteration. T This represents the matrix transpose. The propositional probability vector p is generated based on the verification index and security criteria to establish the mechanism evidence. mech n .

[0159] By jointly inferring the updated mechanistic evidence with other evidence, a new fused posterior distribution is obtained:

[0160] q n (P)=(1 / Z n )×Π e (ψ e (P;E) e ))×ψ mech (P;E) mech n )×ψ KG (P);

[0161] Where q n (P) represents the posterior probability distribution of the security-state fusion obtained in the nth iteration, Z n Π is the normalization constant for the nth iteration. e Let ψ represent the product of the potential functions of all non-mechanistic evidence e. e (P;E) e Let ψ be the potential function of evidence e with respect to proposition P. mech (P;E) mech n Let ψ be the mechanistic evidence potential function for the nth iteration. KG (P) is the consistency constraint potential function provided by the evidence graph.

[0162] Calculate the change between the current iteration and the previous iteration as the convergence metric:

[0163] Δ n =||θ n -θ n-1 ||2+D KL (q n (P)||q n-1 (P));

[0164] Where Δ n Let D be the convergence index for the nth iteration, |||2 denotes the L2 norm of the vector, and D KL The KL divergence is used to measure the difference between two probability distributions. When Δ n When the value is less than the preset threshold ε', it is considered convergent, and the final result θ*=θ is output. n and q*(P)=q n (P); otherwise, continue to the next iteration until the maximum number of iterations N is reached. max .

[0165] During the alternating iteration process, the consistency between mechanistic evidence and other evidence is monitored in real time, and the weight coefficients of the fused posterior guiding term are adaptively adjusted accordingly.

[0166] Specifically, the weight coefficient λ3 of the fusion posterior guiding term in the collaborative inversion objective function determines the strength of the influence of the fusion result on the inversion process. When the mechanistic evidence is highly consistent with other evidence, this weight should be increased to enhance the synergistic effect; when the mechanistic evidence conflicts with other evidence, this weight should be decreased to avoid misguidance. This embodiment provides an adaptive adjustment mechanism based on a consistency monitoring index ρ. n The formula used to quantify the degree of agreement between mechanistic evidence and other evidence in the nth iteration is as follows:

[0167] ρ n =1-(1 / |Eother |)×Σ e (||p mech n -p e ||2×min(β mech n ,β e ));

[0168] Where ρ n Let |E be the consistency monitoring metric for the nth iteration. other | represents the number of pieces of evidence other than mechanistic evidence, Σ e p represents the summation over all other evidence e. mech n Let p be the probability vector of the mechanistic evidence proposition in the nth iteration. e Let be the propositional probability vector of evidence e, |||2 denotes the vector's L2 norm, min denotes the minimum value operation, and β mech n For the credibility of mechanistic evidence, β e The credibility of evidence e is considered. The consistency monitoring index ranges from zero to one, with a higher value indicating higher consistency.

[0169] Based on the consistency monitoring index, the adaptive adjustment formula for the weighting coefficient λ3 is as follows:

[0170] λ3 n+1 =λ3 0 ×σ(α×(ρ n -ρ th ));

[0171] Where λ3 n+1 λ3 is the weighting coefficient used in the next iteration. 0 Here, σ represents the initial value of the weighting coefficients, α represents the Sigmoid function, α is the sensitivity adjustment parameter, and ρ is the initial value of the weighting coefficients. n ρ is the consistency monitoring metric for the current iteration. th This is the consistency threshold.

[0172] When the consistency index ρ n Greater than the threshold ρ th When the Sigmoid function output is greater than 0.5, the weight coefficient λ3 increases, and the guiding effect of the fused posterior on the inversion is enhanced; when the consistency index ρ n Less than the threshold ρ th When the output of the Sigmoid function is less than 0.5, the weight coefficient λ3 decreases, thus avoiding erroneous guidance from the fused posterior pair during inversion in case of conflict.

[0173] According to another aspect of this application, the method further includes: when an inversion closed-loop update trigger signal is received, re-executing parameter inversion using a reliability-weighted inversion objective function.

[0174] Specifically, when flag I is triggered fb When the value equals 1, the system needs to re-execute the parameter inversion. Unlike conventional inversion, this embodiment introduces an evidence reliability weight in the observation consistency term to suppress the impact of low-quality data on the inversion results.

[0175] The mathematical expression for the reliability-weighted inversion objective function is as follows:

[0176] J weighted (θ)=Σ t (ω t ×||y t -h t (θ)||2 2 )+λ×(θ-μ0) T ×Σ0 -1 ×(θ-μ0);

[0177] J weighted (θ) represents the reliability-weighted inversion objective function value, Σ t This represents the summation over all times t, ω t Let y be the reliability weight corresponding to the monitoring data at time t. t Let h be the monitoring response observation value at time t. t (θ) represents the predicted response of the mechanistic model at time t under parameter θ, |||2 represents the vector L2 norm, λ is the regularization coefficient, θ is the parameter vector to be inverted, μ0 is the prior mean vector of the parameters, and Σ0 is the prior covariance matrix of the parameters. T Indicates matrix transpose. -1 This represents finding the inverse of a matrix.

[0178] The squared residual term at each time step in the observation consistency term of the reliability-weighted inversion objective function is multiplied by the corresponding reliability weight ω. t When data at a certain moment is deemed to have a low reliability weight due to sensor malfunction, its contribution to the objective function is suppressed; conversely, when data at a certain moment has a high reliability weight, its contribution to the objective function is normally included. Through this weighting mechanism, the inversion process can automatically suppress the influence of bad data and improve the robustness of parameter estimation.

[0179] This embodiment introduces an evidence graph as a prior knowledge source, constructs an inversion framework with regularization constraints, and outputs an interface for parameter uncertainty, providing the necessary statistical basis for subsequent probability fusion.

[0180] In one possible embodiment, the probabilistic graphical model is specifically a factor graphical model;

[0181] The construction of probabilistic graphical models includes:

[0182] Establish variable nodes that correspond one-to-one with the safety state proposition, as well as factor nodes that correspond to monitoring evidence and mechanistic evidence, respectively;

[0183] Based on the edge weights defined in the computable evidence graph, a potential function is constructed to connect variable nodes and factor nodes, which is used to quantify the degree to which the evidence supports the proposition.

[0184] In other words, the probabilistic graphical model is specifically a factor graphical model, which includes: variable nodes that correspond one-to-one with the security state proposition; factor nodes that correspond to monitoring evidence and mechanistic evidence, respectively; and a potential function that connects the variable nodes and factor nodes based on the edge weights defined in the computable evidence graph.

[0185] In this embodiment, the factor graph model is a bipartite graph containing two types of nodes: variable nodes and factor nodes. Specifically, each safety-state proposition, such as normal seepage or crack propagation, is instantiated as a variable node X. i Its value space is {0, 1} or a continuous probability interval [0, 1]. Each input piece of evidence, such as monitoring evidence E... m Or mechanistic evidence E p All are instantiated as a factor node F j The factor nodes are connected to their supporting variable nodes via edges. The potential function defines the influence of the factor nodes on the variable nodes. This is based on the edge weights W. edge (Including probability and confidence), a log-linear potential function of the following form can be constructed:

[0186] psi(X i )=exp(W edge *phi(X i E j ));

[0187] Among them, psi(X) i ) represents the potential function value, phi(X) i E j ) is the characteristic function, when variable X i State and Evidence E j The value is 1 if there is agreement, and 0 otherwise. By running the sum-product algorithm on the entire graph, the system can calculate the marginal probability distribution of each variable node, i.e., the primary probability distribution of the safe state.

[0188] Furthermore, to incorporate both the credibility and probability components into the potential function, a credibility-probability joint weighted potential function is constructed. The form of the joint weighted potential function is as follows:

[0189] ψ e (y)=exp(c e ×log(p e (y)+ε*));

[0190] Where ψ e (y) represents the potential function value of evidence e with respect to proposition y, exp represents the exponential function, and c e For the credibility of evidence e, log represents the natural logarithm function, p e (y) represents the probability component of evidence e with respect to proposition y, and ε* is a small positive number to prevent overflow in logarithmic operations, typically taking the value 1 × 10^2. -10 .

[0191] When the credibility of the evidence is c e When the value is high, the joint weighting potential function affects the probability component p. e (y) has a strong amplification effect, and this evidence has a significant influence in fusion inference; when the credibility of the evidence is c e At lower values, the joint weighting potential function has a weaker amplification effect on the probability components, thus suppressing the influence of this evidence in the fusion inference. This allows the factor graph fusion process to simultaneously consider both the probability and credibility of the evidence.

[0192] In one possible implementation, the inversion correction adopts an alternating iterative solution method, including: parameter update step, mechanism evidence update step, fusion posterior update step, and convergence determination step; during the iteration process, the weight coefficient of the fusion posterior guide term is adaptively adjusted based on the consistency monitoring index of mechanism evidence and other evidence. When the consistency index is higher than a preset threshold, the weight coefficient is increased to enhance the synergistic effect, and when the consistency index is lower than the preset threshold, the weight coefficient is decreased to avoid misguided guidance.

[0193] In one embodiment of this application, the preset conflict triggering condition includes at least one of the following conditions:

[0194] The data integrity rate calculated based on the running status data is lower than the preset integrity threshold, or the data anomaly rate is higher than the preset anomaly rate threshold.

[0195] The degree of evidence conflict between monitoring evidence and mechanistic evidence calculated using evidence theory or distance metric functions is higher than the preset conflict threshold.

[0196] Alternatively, the data integrity rate can be calculated based on the ratio of the number of valid data points in the operational status data to the total number of data points to be collected. If the data integrity rate is lower than the preset integrity threshold, the data anomaly rate can be calculated based on the ratio of the number of abnormal data points in the operational status data to the total number of data points. If the data anomaly rate is higher than the preset anomaly rate threshold, the data anomaly rate can be calculated.

[0197] Specifically, the preset conflict trigger condition is the decision point at which the system switches from automatic fusion to logical arbitration. The system monitors the following two types of indicators in real time:

[0198] Data quality metric: Data integrity rate R obtained from reading and calculation. c If R c If the value is less than 0.8, it is determined to be a test failure trigger.

[0199] Evidence conflict index: Read the calculated conflict coefficient K or Jousselme distance d J If d J A value >0.4 indicates a significant discrepancy between monitoring and mechanistic conclusions. For example, monitoring might show a stable water level, but the mechanistic model suggests a large fluctuation, indicating a conflict trigger. Figure 4 As shown, in a preferred implementation, the weights of monitoring evidence and mechanistic evidence are adjusted according to the conflict fractal-driven principle, specifically as follows:

[0200] When the conflict is triggered by the data integrity rate or anomaly rate not meeting the requirements, a linear decay function is constructed based on the proportion of missing or anomaly data to reduce the weight of the monitoring evidence.

[0201] When the conflict trigger condition is that the degree of evidence conflict is higher than the conflict threshold, the consistency score of each piece of evidence is calculated based on the logical rule model, and the weights of monitoring evidence and mechanism evidence are redistributed according to the consistency score.

[0202] In this embodiment, categorized governance is implemented. For the first scenario, where the data source itself is of poor quality, the system performs a weight reduction operation. Preferably, a linear decay function is used to assign weights W to the monitored evidence. m Reduce the amount:

[0203] W m_new =W m_old *(1-λ*R missing );

[0204] Among them W m_new The weight after reduction; W m_old λ represents the original weights; λ is the adjustment coefficient, for example, 1.5; R missing The missing data rate, i.e., 1-R c This means that the worse the data quality, the less say it has in the integration process.

[0205] In the second scenario, where the source data quality is acceptable but the conclusions conflict, the system performs a weight reallocation. Instead of blindly reducing the weight of one side, the decision-making power is delegated to the subsequent Probabilistic Soft Logic (PSL) model. After the PSL model runs, it calculates the degree to which each piece of evidence satisfies the logical rules, i.e., the consistency score. Evidence with higher scores—that is, evidence more consistent with physical logic and expert experience—will have its weight dynamically increased, and vice versa.

[0206] In another embodiment of this application, the conflict triggering function includes at least: a data integrity rate triggering function; a data anomaly rate triggering function; an evidence conflict degree triggering function; and a parameter identifiability triggering function.

[0207] The logical rule model adaptively adjusts the weights of monitoring evidence and mechanistic evidence based on the output of the conflict triggering function. The adaptive adjustment includes weight reduction or weight reallocation.

[0208] The weight adjustment follows the conflict classification-driven principle, including: when the conflict triggering condition is that the data integrity rate or anomaly rate does not meet the requirements, a linear decay function is constructed based on the proportion of missing or anomaly data to reduce the weight of the monitoring evidence; when the conflict triggering condition is that the degree of evidence conflict is higher than the conflict threshold, the consistency score of each piece of evidence is calculated based on the logical rule model, and the weights of the monitoring evidence and the mechanism evidence are redistributed accordingly; when the conflict triggering function corresponds to the identifiability index of the mechanism model parameters, the weight of the mechanism evidence is decayed or adjusted according to the degree of parameter identifiability.

[0209] Specifically, in the conflict triggering mechanism, the system considers not only conflicts caused by differences in evidence distribution but also conflicts caused by the stability of the mechanism model itself. In this embodiment, after the mechanism model parameters are inverted, the system calculates a parameter identifiability index to measure the degree of invertibility of the mechanism model parameters under the current monitored evidence conditions. When the identifiability index exceeds a preset threshold, it is determined that the current parameters have an unidentifiable risk, meaning that different parameter combinations may produce similar model responses. At this time, the system triggers a parameter identifiability conflict handling process, reduces the weight of the mechanism evidence, and writes the parameter identifiability state as a conflict type into the evidence map for subsequent fusion inference.

[0210] In one possible embodiment, the logic rule model adopts a probabilistic soft logic model;

[0211] By adjusting the weights of monitoring evidence and mechanistic evidence using a logical rule model and performing arbitration inference, a modified posterior probability distribution of the safety state is obtained, including:

[0212] Map the truth values ​​of security-state propositions to continuous variables on the closed interval [0, 1].

[0213] Based on the Lukasiewicz fuzzy logic operator combined with continuous variables, a weighted set of logic rules containing monitoring evidence and mechanistic evidence is constructed.

[0214] By minimizing the hinge loss distance of the weighted logic rule set on all security-state propositions using a convex optimization algorithm, the corrected security-state posterior probability distribution is inferred.

[0215] Specifically, unlike traditional Boolean logic (true / false), probabilistic soft logic (PSL) allows the truth value I(x) of a proposition to vary continuously within the interval [0, 1], making it suitable for handling fuzzy issues such as dam safety. In this embodiment, the system defines a set of weighted logic rules. For example, the rules regarding piping could be:

[0216] Rule R1: Monitor(Permeability_High) AND Mechanism(Gradient_High) -> Risk(Piping). Weight: 10.0.

[0217] Among them, Monitor(Permeability_High) indicates that the permeability coefficient (or permeability gradient) is high; Mechanism(Gradient_High) indicates that the permeability gradient is high based on mechanism analysis; Risk(Piping) indicates the risk of piping.

[0218] The system uses the Lukasiewicz t-norm to compute the truth values ​​of logical operators. For example, the truth value I(A AND B) of the logical AND operation is calculated as follows:

[0219] I(A AND B)=max(0,I(A)+I(B)-1);

[0220] Where I(A) is the truth value of the rule's antecedent and I(B) is the truth value of the rule's consequent.

[0221] The formula for calculating the truth value I(A -> B) of logical implication (->) is:

[0222] I(A -> B)=min(1,1-I(A)+I(B));

[0223] To satisfy rule R1, we want I(A->B) to be as close to 1 as possible, i.e., we want I(A) - I(B) to be as small as possible. PSL achieves this by minimizing the hinge loss of distance satisfaction. For rule R... j Its loss function L j Defined as:

[0224] L j=w j *max(0,I(r body )-I(r head )) p ;

[0225] Where w j For the rule weights, I(r) body ) represents the truth value of the rule's antecedent (IF part), I(r) head ) represents the truth value of the rule consequent (THEN part), and p is usually 1 or 2.

[0226] The system sums the loss functions of all rules to construct a global objective function, and then uses a convex optimization algorithm (such as the Alternating Direction Multiplier Method, ADMM) to find the truth distribution of propositions that minimizes the total loss. The optimal solution is the corrected safe-state posterior probability distribution.

[0227] Furthermore, this embodiment specifically provides an explanation chain output function. After PSL inference is completed, the system will backtrack to the logic rules that are in an active state (i.e., have a truth value significantly greater than 0 and a high weight), and string them together to form an explanation text. For example: due to the detection of an abnormal seepage gradient (truth value 0.9) and the mechanism review confirming soil parameter deterioration (truth value 0.8), according to rule R1, a piping risk is determined to exist (truth value 0.85). This improves the understandability of the diagnostic results.

[0228] According to one aspect of this application, the method further includes: calculating a comprehensive conflict trigger score, and determining whether to trigger a conflict handling process based on the comprehensive conflict trigger score.

[0229] In this embodiment, the two types of triggering conditions (data quality triggering and evidence conflict triggering) described in the above embodiments can be unified into a comprehensive scoring mechanism. For example, the formula for calculating the comprehensive conflict triggering score T is as follows:

[0230] T=β1×(1-Q mon )+β2×D mon_mech +β3×U θ ;

[0231] Where T is the overall conflict trigger score, β1, β2, and β3 are the weight coefficients of each sub-item, and Q is the overall conflict trigger score. mon D is a comprehensive quality indicator for monitoring data. mon_mech To monitor the average degree of conflict between evidence and mechanistic evidence, U θ This is the uncertainty index for the inversion parameters.

[0232] The average degree of conflict D between monitoring evidence and mechanistic evidence mon_mech The calculation formula is as follows:

[0233] D mon_mech=(1 / (|E mon |×|E mech |))×Σ i Σ j D(e i e j );

[0234] Where D mon_mech To monitor the mechanism-average conflict degree, |E mon |E represents the amount of evidence monitored. mech | represents the amount of mechanistic evidence, Σ i This indicates that all monitoring evidence e i Summation, Σ j This indicates that all mechanistic evidence e j Summation, D(e i e j ) is evidence e i With e j The degree of conflict between them.

[0235] Degree of conflict of evidence D(e) i e j The calculation formula for ) is as follows:

[0236] D(e i e j )=(1 / 2)×Σ y |p i (y)-p j (y)|;

[0237] Where D(e) i e j ) is evidence e i With evidence e j The degree of conflict between them, Σ y p represents the summation over all security-state propositions y. i (y) is evidence e i For the probability components of proposition y, p j (y) is evidence e j For the probability components of proposition y, || represents taking the absolute value. The degree of conflict of evidence ranges from zero to one; the larger the value, the greater the difference in distribution between the two pieces of evidence.

[0238] When the overall conflict trigger score T is greater than or equal to the preset trigger threshold τ T If the condition is met, the system determines that a conflict handling process has been triggered and proceeds to conflict classification and arbitration; otherwise, it directly outputs the primary probability distribution of the safe state as the final result.

[0239] According to another aspect of this application, the conflict classification-driven principle also includes: performing conflict classification on the evidence that triggers conflict resolution to determine the root cause category of the conflict.

[0240] In a preferred implementation, conflict classification includes four types: sensor failure, data missing, model mismatch, and structural anomaly.

[0241] The method constructs a feature vector based on data quality indicators, inversion residuals, parameter uncertainty, and evidence conflict degree, and uses a logical rule model to infer the probability distribution of each conflict type.

[0242] When the cumulative probability of model mismatch conflicts exceeds a preset threshold, the updated evidence reliability weights are fed back to the inversion correction step, and the reliability-weighted parameter inversion is re-executed to form a closed loop.

[0243] Specifically, when a conflict of evidence is detected, there may be multiple reasons for the conflict, and different reasons require different handling strategies. For example, four conflict types are defined, and a classification method based on feature vectors is provided. The four conflict types include:

[0244] The first type is sensor fault type, denoted as sensor_fault, which indicates that the conflict is caused by sensor equipment failure leading to distortion of monitoring data. The handling strategy is to reduce the weight of the monitoring evidence.

[0245] The second type is the data missing type, denoted as data_missing. This indicates that the conflict is caused by incomplete evidence information due to a large area of ​​missing monitoring data. The handling strategy is to reduce the weight of the monitoring evidence according to the proportion of missing data.

[0246] The third type is model mismatch, denoted as model_mismatch, which indicates that the conflict is caused by the mismatch between the mechanism model structure or parameters and the actual engineering state. The handling strategy is to trigger the parameter recalibration process and feed the updated evidence reliability weights back to the inversion step.

[0247] The fourth type is structural anomalous type, denoted as struct_abnormal, which indicates that the conflict is caused by an abnormal state in the dam structure. The handling strategy is to increase the weight of monitoring evidence and trigger an early warning.

[0248] For each piece of evidence e that triggers conflict resolution, construct a conflict type feature vector x. e Its composition is as follows:

[0249] x e =[Q e r e U θ D e_avg ];

[0250] Where x e Let Q be the fractal eigenvector of evidence e.e For the quality metric of the data source corresponding to evidence e, r e U is the residual statistic between evidence e and the response of the mechanistic model. θ D represents the uncertainty of the inversion parameters. e _ avg The average degree of conflict between evidence e and other evidence.

[0251] Based on the fractal feature vector x e The probability distribution of each conflict type is calculated using a probabilistic soft logic model, and the output is:

[0252] π e k =P(z e =k|x e );

[0253] Where π e k Let z be the probability that evidence e belongs to the k-th conflict type. e Let P(z) be the conflict-figurative random variable of evidence e. e =k|x e ) represents the situation where, given the eigenvector x e The conditional probability of conflict type k under given conditions.

[0254] The evidence reliability weight is calculated based on the conflict classification results, and it is determined whether the inversion closed-loop update is triggered.

[0255] Specifically, traditional methods often adjust the weights of conflicting evidence by a fixed percentage reduction, lacking objective basis. This embodiment calculates the reliability weight of each piece of evidence based on the conflict classification probability, achieving refined weight adjustment driven by classification. Evidence reliability weight ω e The calculation formula is as follows:

[0256] ω e =1-π e sensor_fault -γ×π e data_missing ;

[0257] Where ω e π is the reliability weight of evidence e. e sensor_fault As evidence of the probability that e is classified as a sensor failure, π e data_missing Let γ be the probability that evidence e is classified as missing data, and let γ be the reduction factor for missing data, ranging from zero to one.

[0258] The higher the probability that the evidence is determined to be a sensor malfunction, the lower its reliability weight; the higher the probability that the evidence is determined to be missing data, the lower its reliability weight is proportionally reduced; when the evidence is determined to be structurally abnormal, its reliability weight remains unchanged or increases to ensure that genuine abnormal signals are not erroneously suppressed.

[0259] Furthermore, the triggering condition for the inversion closed-loop update is defined. When the cumulative probability of model mismatch exceeds a preset threshold, it indicates that the mechanistic model needs to be recalibrated, and the updated evidence reliability weights should be fed back into the inversion step. Inversion closed-loop update trigger flag I fb The calculation formula is as follows:

[0260] I fb =I[Σ e π e model_mismatch ≥τ M ];

[0261] Among them I fb The inversion closed-loop update trigger flag; I[ ] is an indicator function, which takes the value 1 when the condition in parentheses is true and 0 otherwise; Σ e This represents the summation over all triggering evidence; π e model_mismatch τ represents the probability that evidence e is classified as model mismatched. M This is the threshold for triggering model mismatch.

[0262] When I fb When the value equals 1, the system executes the following closed-loop update process: The updated evidence reliability weight vector Ω = {ω} e The data serves as input for the reliability-weighted inversion; based on the weighted monitoring data, parameter inversion is re-executed; mechanistic evidence is regenerated based on the updated parameters; and the evidence fusion step is repeated. This closed-loop mechanism enables the system to automatically recover from a contaminated state to a self-cleaning state.

[0263] According to another aspect of this application, the method further includes: defining the structure and key rule content of the probabilistic soft logic rule base.

[0264] In this embodiment, a weighted logic rule set is predefined. A probabilistic soft logic rule base for dam safety state diagnosis is provided, comprising a three-layer structure: a conflict detection rule layer, a conflict classification rule layer, and an arbitration decision rule layer.

[0265] Specifically, the first layer is the conflict detection rule layer, used to determine whether there are conflicts of evidence requiring arbitration. Representative rules include:

[0266] Rule R1.1: When the degree of conflict between surveillance evidence and mechanistic evidence exceeds a threshold, it is considered highly conflicting. The PSL expression for this rule is: D(emon e mech ) greater than τ D Contains High Conflict (e mon e mech ), with a rule weight of 1.0. Where τ D This represents the conflict threshold. mon To monitor evidence, e mech For mechanistic evidence, D() is the evidence conflict degree function, and High Conflict() represents a highly conflicted state.

[0267] Rule R1.2: When the missing data rate exceeds a threshold, it is considered a data quality issue. The PSL expression for this rule is: Missing Rate(sensor) greater than τ. miss It contains a Data Quality Issue (sensor), and the rule weight is 1.0. Where τ miss Here, is the missing rate threshold, Missing Rate() is the missing data rate function, sensor is the sensor, and DataQualityIssue() is the data quality issue status.

[0268] The second layer is the conflict classification rule layer, used to locate the root cause of conflict. Representative rules include:

[0269] Rule R2.1: When a sensor output undergoes a sudden change but there is no load change, it is considered a sensor fault. The PSL expression for this rule is: Sudden Jump(sensor) and No Load Change imply Fault Type(sensor, sensor_fault), with a rule weight of 0.9. Where: Sudden Jump() represents the sensor output sudden change state, No Load Change represents the state without load change, and Fault Type() represents the fault type determination.

[0270] Rule R2.2: When the residual is large but all sensors are in normal condition, it is considered a model mismatch. The PSL expression for this rule is: High Residual(θ) and All Sensor Normal imply Fault Type(model, model_mismatch), with a rule weight of 0.85. High Residual() represents the high residual state, and All Sensor Normal represents the normal state of all sensors.

[0271] Rule R2.3: When multiple sensors in the same area malfunction simultaneously, it is considered a structural anomaly. The PSL expression for this rule is: Spatial Correlation(abnormal_sensors) and InSame Zone implies Fault Type(structure, struct_abnormal), with a rule weight of 0.95. Here, Spatial Correlation() represents spatial correlation, abnormal_sensors is the set of anomalous sensors, InSame Zone represents the state of the same area, and structure represents the dam structure.

[0272] Rule R2.4: When a defect discovered during inspection matches the location of a monitored anomaly, it is considered a structural anomaly. The PSL expression for this rule is: Inspection Finding(location, defect) and Monitor Anomaly(location) imply Fault Type(structure, struct_abnormal), with a rule weight of 0.98. Here, Inspection Finding() represents the defect event discovered during inspection, location is the dam project location, defect is the structural defect, and Monitor Anomaly() represents the monitored anomaly state.

[0273] The third layer is the arbitration decision-making rule layer, used to determine the final diagnostic conclusion and evidence weighting adjustment strategy. Representative rules include:

[0274] Rule R3.1: When evidence is determined to be a sensor fault, reduce the weight of that evidence. The PSL expression for this rule is: Fault Type(e, sensor_fault) implies Reduce Weight(e), and the rule weight is 1.0. ReduceWeight() is the action to reduce the weight.

[0275] Rule R3.2: When the cumulative model mismatch probability exceeds a threshold, an inversion update is triggered. The PSL expression for this rule is: Σπ model_mismatch Greater than τ M It contains a Trigger Inversion Update rule with a weight of 1.0. The TriggerInversion Update rule triggers the inversion update action.

[0276] Rule R3.3: Output a high-confidence conclusion when all evidence is consistent. The PSL expression for this rule is: AllEvidence Consistent implies High Confidence Conclusion, with a rule weight of 0.95. Here, AllEvidence Consistent represents the consistency of all evidence, and High Confidence Conclusion represents the high-confidence conclusion.

[0277] Rule R3.4: When a conflict cannot be resolved and involves a safety-critical proposition, a conservative conclusion shall be adopted. The PSL expression for this rule is: Conflict Unresolved and Safety Critical imply a Conservative Conclusion, with a rule weight of 1.0. Here, Conflict Unresolved represents the unresolved conflict state, Safety Critical represents the safety-critical proposition, and Conservative Conclusion represents the conservative conclusion.

[0278] The weights of the above rules can be set based on engineering experience, or learned through the maximum likelihood estimation method using historical diagnostic cases. During PSL inference, the system transforms the above rules into a hinge loss function and uses a convex optimization algorithm to solve for the optimal proposition truth distribution that satisfies all rules.

[0279] In this embodiment, under normal circumstances, factor graphs are used for efficient probability fusion; once a conflict or anomaly is triggered, the system immediately activates the PSL arbitration mechanism, introduces higher-order logic rules to correct the evidence, and ensures the robustness of the diagnostic conclusion.

[0280] In another embodiment of this application, while PSL has advantages in handling continuous truth values, Markov logic networks provide an alternative implementation based on a probabilistic graphical model for scenarios with extremely strict logical constraints, such as red-line alarms. Specifically, the logical rule model employs a Markov logic network (MLN). An MLN can be defined by first-order logic formulas (F... i ) and the corresponding real weights (w) i )composition.

[0281] Specifically, the system constructs a set of logical formulas based on expert knowledge in the dam field, describing the constraint relationships between evidence and propositions, for example:

[0282] Formula F1:!Conflict(Evidence_A, Evidence_B) v Consistent(Conclusion).

[0283] Formula F2: High_Quality(Source_A)≥Trust(Source_A).

[0284] Where ! represents the logical NOT operator; v represents the logical OR operator; Conflict() indicates a conflict of evidence, Evidence_A and Evidence_B represent evidence A and evidence B respectively; Consistent() indicates a consistent conclusion; Conclusion represents the diagnostic conclusion; High_Quality() indicates a high-quality source of evidence; Source_A represents source A; Trust() indicates the credibility of the source of evidence.

[0285] In MLNs, the logistic formulas are transformed into characteristic functions in the Markov network. For a given set of safe-state propositions X (i.e., possible worlds), its joint probability distribution P(X) is given by:

[0286] P(X)=(1 / Z)*exp(∑(w i *n i (X)));

[0287] Where P(X) is the probability of the proposition combination occurring under a specific truth assignment; Z is the normalized partition function; ∑ represents the summation over all rules i; w i The weight of the i-th rule reflects its strength or reliability; n i (X) represents the number of times the i-th rule is satisfied or the number of instantiations under the current truth assignment X.

[0288] When executing arbitration inferences, the system fixes monitoring evidence and mechanistic evidence as observed variables, i.e., evidence variables. The task then becomes finding the truth assignment of the hidden variable that maximizes P(X), i.e., the security-state proposition. This is equivalent to solving the weighted maximum satisfiability problem. The system uses Max Walk SAT (Max Walk Satisfiability) or MCMC (Markov Chain Monte Carlo) sampling algorithms for this solution.

[0289] If the Max Walk SAT algorithm is used, the system randomly flips the truth value (true / false) of the proposition. If the flip increases the total number of high-weight rules that are satisfied, the flip is retained. After multiple iterations, the system converges to the most likely truth state, which is the corrected posterior probability distribution of the safety state.

[0290] Compared to PSL, MLN places greater emphasis on the probabilistic nature of discrete logic. In some implementations, if it is necessary to handle continuous evidence probabilities, this embodiment can also employ a hybrid Markov logic network, embedding continuous evidence as numerical nodes into the network, and constraining the relationship between numerical variables and logical variables by defining a Gaussian potential function to achieve a more refined arbitration effect.

[0291] According to one aspect of this application, in existing deep learning or Bayesian diagnostics, the probability values ​​output by the model often exhibit overconfidence, meaning the model believes it is very certain, but in reality, the probability of prediction error is much greater than 0.01. Preferably, by introducing a statistical conformal prediction framework and calibrating the model output using historical data, the confidence interval of the final output can cover the true safe state with a mathematically provable probability.

[0292] Specifically, based on a historical calibration sample set, the confidence level of the primary probability distribution or the corrected posterior probability distribution of the safety status is calibrated using a conformal prediction method to obtain a safety status diagnostic conclusion containing confidence intervals, including:

[0293] Construct an inconsistency metric function to measure the difference between the current safety state primary probability distribution or the corrected safety state posterior probability distribution and the true labels in the historical calibration sample set; calculate the inconsistency score distribution of the historical calibration sample set based on the inconsistency metric function; calculate quantile thresholds according to the preset confidence level and inconsistency score distribution, construct a prediction set containing the true safety state proposition, and output the prediction set as a confidence interval.

[0294] Alternatively, a non-consistency metric function can be constructed to measure the degree of difference between the current security state probability distribution and the true labels in the historical calibration sample set.

[0295] In this embodiment, the system divides the historical dataset into a training set and a calibration set. After the fusion arbitration model completes parameter optimization on the training set, it uses the calibration set to calculate the inconsistency score. An inconsistency metric function is used to quantify the oddity of the prediction results. For the i-th sample (xi) in the calibration set... i y i ), where x i To input evidence, y i Label the actual safety state, such as normal. Inconsistency score ξ i The calculation formula is as follows:

[0296] ξ i =1-P hat (y i |x i );

[0297] Where Phat (y i |x i ) is the model (i.e., the posterior probability distribution) relative to the true label y. i The predicted probability. Clearly, if the model predicts accurately and with confidence, i.e., P... hat If ξ is close to 1, then i ξ is close to 0; conversely, if the model predicts incorrectly or lacks confidence, ξ i It will be very big.

[0298] The system calculates the inconsistency scores of all n samples in the calibration set, obtaining the distribution sequence {ξ1, ..., ξ}. n Based on the user-defined confidence level, for example, confidence level 1-ε=95%, the system calculates the quantile threshold Q of the score distribution. ε The calculation formula is as follows:

[0299] Q ε =Quantile(1-ε,{ξ1,...,ξ n})*(1+1 / n);

[0300] Where Quantile represents the quantile operation, and (1+1 / n) is a statistical correction factor for a finite sample size. Threshold Q ε This means that in 95% of cases, the inconsistency score will not exceed this value.

[0301] For the new sample x to be diagnosed at the current moment new The system constructs a prediction set C(x) new This includes all candidate propositions y that satisfy the inconsistency test:

[0302] C(x new )={y|1-P hat (y|x new )≤Q ε};

[0303] In other words, the system will include all propositions with inconsistency scores below a threshold in the prediction set. For example, if the model outputs probabilities as: normal (0.6), slightly abnormal (0.35), and severely abnormal (0.05), and the calculated threshold Q... ε =0.4. Therefore:

[0304] For normal values: 1 - 0.6 = 0.4 ≤ 0.4, include them in the set.

[0305] For minor abnormalities: 1 - 0.35 = 0.65 > 0.4, so it is excluded.

[0306] For severe abnormalities: 1 - 0.05 = 0.95 > 0.4, exclude.

[0307] At this point, the system outputs the conclusion that, at a 95% confidence level, the dam's current safety state is {normal}.

[0308] However, if the model output is ambiguous, for example: normal (0.45), slightly abnormal (0.40), severely abnormal (0.15).

[0309] The inconsistency scores of the first two propositions (0.55 and 0.6) may both be less than a certain lenient threshold, in which case the output set is: {normal, slightly abnormal}.

[0310] By using aggregated outputs, the scope of uncertainty in the current diagnosis is explicitly communicated to dam managers, thus avoiding the risk of misleading conclusions.

[0311] According to one aspect of this application, the conformal prediction method includes:

[0312] Calculate the distribution of inconsistent scores based on the historical calibration sample set;

[0313] A quantile threshold is determined from the inconsistent score distribution based on a preset confidence level;

[0314] For the current sample, calculate the corresponding inconsistency score for each candidate safety status label and compare it with the quantile threshold. Include safety status labels with inconsistency scores less than or equal to the quantile threshold into the prediction set to obtain a prediction set containing the true safety status labels, and output it as the safety status diagnosis conclusion.

[0315] In one embodiment of this application, a data-mechanism-knowledge jointly driven method for self-diagnosis of dam safety status may further include:

[0316] S1. Acquire multi-source engineering data and operational status data, including monitoring time series data, inspection record data, and historical event data; perform time alignment, spatial part coding unification, and missing measurement and anomaly handling on the operational status data to obtain standardized operational data; and calculate data quality indicators.

[0317] S2. Construct a computable evidence map based on multi-source engineering data. The computable evidence map includes an ontology layer and an evidence layer. The ontology layer is used to describe engineering objects, structural parts, indicators, defects, working conditions, treatment measures, and the set of safety-state propositions and their relationships. The evidence layer is used to describe the relationship between evidence items and the set of safety-state propositions, and provides an evidence conflict degree interface and an interpretation chain output interface.

[0318] S3. Extract state variables representing the dam's safety status from standardized operational data, and map the state variables into proposition probability vectors for the set of safety status propositions; generate evidence entries based on monitoring evidence, inspection evidence, and historical evidence, and write the evidence entries into the evidence layer.

[0319] S4. Extract the prior constraints and physical constraints of the mechanism model parameters from the ontology layer, construct the inversion objective function based on the monitoring response and the mechanism model response, perform inversion correction on the mechanism model parameters, and obtain parameter estimates; and output the uncertainty interface that characterizes the inversion credibility, wherein the uncertainty interface includes at least the parameter posterior uncertainty and the inversion residual / identifiability index.

[0320] S5. Based on parameter estimation, perform mechanism verification calculations to obtain key verification indicators and their uncertainties; map the key verification indicators to mechanism proposition probability vectors for the set of safety-state propositions according to preset criteria, generate mechanism evidence entries, and write the mechanism evidence entries into the evidence layer.

[0321] S6. Construct a factor graph model with the security state proposition as the latent variable and monitoring evidence, inspection evidence, historical evidence and mechanism evidence as the observed variables, and use the consistency constraint of the computable evidence graph as the prior constraint factor; based on the factor graph model, infer the fused security state posterior probability distribution and output the corresponding explanatory chain.

[0322] S7. Based on the data quality indicators, uncertainty interface, and evidence conflict degree interface output, determine whether the set of enumerable engineering conflict triggering conditions is met. When the triggering conditions are met, reduce or reweight the corresponding evidence weights, and use probabilistic soft logic or Markov logic models to arbitrate the conflicting evidence, and obtain the security state probability distribution and trigger / arbitration record after arbitration.

[0323] S8. Based on the historical calibration sample set, perform conformal prediction credibility calibration on the fused safety state probability distribution or the arbitration probability distribution to obtain the final diagnostic conclusion and its credibility; and output the explanatory chain supporting the diagnostic conclusion and the credibility information of the evidence to realize the self-diagnosis of the dam safety state.

[0324] This embodiment summarizes the process of the data-mechanism-knowledge jointly driven dam safety state self-diagnosis method into eight steps, namely: S1 is used to complete the access, preprocessing and quality quantification of multi-source operation data, and output standardized operation data and data quality indicators; S2 is used to construct a computable evidence graph, and output the ontology layer / evidence layer structure and computable interface; S3 is used for state quantity extraction and evidence mapping, and output monitoring / inspection / historical evidence items and their proposition probability vectors and credibility; S4 is used for mechanism parameter inversion correction under the prior constraints of the graph, and output parameter estimates and uncertainty interfaces; S5 is used for mechanism verification calculation and evidence, and output mechanism evidence items; S6 is used for factor graph joint inference, and output the primary posterior probability distribution of safety state and interpretation chain; S7 is used for conflict triggering, classification and logical arbitration inference, and output the corrected posterior probability distribution and trigger / arbitration records; S8 is used for conformal prediction credibility calibration, and output the final diagnostic conclusion containing confidence intervals. Among them, monitoring evidence, inspection evidence, historical evidence and mechanism evidence are all written into the evidence layer of the evidence graph in the form of evidence items, and participate in subsequent inferences through edge weights.

[0325] According to one aspect of this application, a data-mechanism-knowledge jointly driven dam safety state self-diagnosis system is also provided, comprising:

[0326] The holographic perception and evidence mapping module is used to acquire the dam's operational status data, extract state variables that characterize the dam's safety status based on a pre-constructed computable evidence map, and map the state variables into monitoring evidence for safety status propositions.

[0327] The mechanism co-inversion module is used to invert and correct the mechanism model parameters using monitoring evidence under the prior constraints of the computable evidence map, and to perform verification calculations based on the corrected mechanism model parameters to obtain mechanism evidence.

[0328] The integrated arbitration inference module is used to construct a probabilistic graphical model with the security status proposition as the latent variable. Monitoring evidence and mechanistic evidence are used as observation factors to enter the probabilistic graphical model for joint inference to obtain the initial probability distribution of the security status. When the preset conflict triggering conditions are met, the module uses a logical rule model to adjust the weights of monitoring evidence and mechanistic evidence and execute arbitration inference to obtain the corrected posterior probability distribution of the security status.

[0329] The credible decision output module is used to perform credibility calibration on the primary probability distribution of the safety state or the corrected posterior probability distribution of the safety state based on the historical calibration sample set and using the conformal prediction method, so as to obtain a safety state diagnostic conclusion containing a confidence interval.

[0330] In this embodiment, the system is typically deployed at the physical level on a server cluster or cloud platform at the dam management center. The holographic perception and evidence mapping module connects to the data acquisition unit (DAU) at the dam site via an industrial Ethernet interface, receiving sensor data streams in real time; it also provides a web interface for inspection personnel to enter inspection forms. The holographic perception and evidence mapping module integrates a preprocessing algorithm library and an evidence graph analysis engine. The mechanism co-inversion module integrates a finite element calculation kernel (such as ANSYS or a self-developed solver) and a nonlinear optimization algorithm library. The mechanism co-inversion module utilizes high-performance computing (HPC) resources to execute parameter inversion tasks in parallel. The fusion arbitration inference module is the intelligent core of the system, running a factor graph inference engine and a PSL solver. The fusion arbitration inference module supports dynamically loading new logical rules to adapt to changes in the engineering environment. The credible decision output module connects to a human-computer interaction interface (GUI), displaying not only the final diagnostic conclusion (confidence interval) but also the generated explanatory chain through visual charts. For example, the screen highlights: "Due to a conflict between monitoring data A and mechanism calculation B, PSL rule C is triggered for arbitration, and the judgment is biased towards the mechanism side." This provides an intuitive basis for engineering technicians to review and diagnose results.

[0331] like Figure 5 As shown, according to another aspect of this application, the data-mechanism-knowledge jointly driven dam safety state self-diagnosis system includes at least a holographic perception and evidence mapping module, a mechanism collaborative inversion module, a fusion arbitration inference module, and a credible decision output module. Each module interacts with the evidence map service through a data bus to form a traceable self-diagnosis closed-loop link.

[0332] The holographic perception and evidence mapping module is used to access dam operation status data and generate monitoring evidence. Its inputs include: monitoring time-series data collected by the automated monitoring system, manual inspection records, and historical anomaly data; its processing includes: spatiotemporal alignment, handling of missing data and anomalies, calculation of data quality indicators, and extraction and probability mapping of state variables; its output is: monitoring evidence entries / inspection evidence entries / historical evidence entries oriented towards a set of safety-state propositions. Each evidence entry contains at least a proposition probability vector and a credibility component, and the evidence entries are written into the evidence layer of the evidence graph.

[0333] The mechanism co-inversion module is used to correct mechanism parameters and generate mechanism evidence under the prior constraints of the evidence map. Its inputs include: monitoring evidence entries, prior intervals / distributions of mechanism parameters and engineering physical constraints provided by the evidence map, and pre-stored mechanism models. Its processing includes: constructing an inversion objective function containing observation consistency terms and prior constraint terms and solving for parameter estimates, while outputting an uncertainty interface, including at least the posterior covariance of parameters, residual statistics, and identifiability indices. Its outputs are: corrected mechanism model parameters, key indicators obtained from verification calculations, and their uncertainties, which are then mapped to mechanism evidence entries and written into the evidence map.

[0334] The fusion arbitration inference module is used to fuse multi-source evidence and perform arbitration inference under conflict conditions. Its inputs include: monitoring / inspection / historical evidence entries and mechanistic evidence entries, as well as consistency constraints provided by the evidence graph. Its processing includes: a) In normal mode, constructing a factor graph model with security-state propositions as variable nodes and various types of evidence as factor nodes, quantifying evidence support based on potential functions, and obtaining the primary posterior probability distribution of the security state through inference; b) Conflict detection and triggering: determining whether conflict triggering conditions are met based on data quality indicators, evidence conflict degree, and / or parameter identifiability indicators; c) In arbitration mode, reducing or redistributing evidence weights based on a logical rule model, performing arbitration inference, outputting the corrected posterior probability distribution, and recording the trigger type, classification results, and weight adjustment results, which are then written into the evidence graph. The module's output includes: the primary posterior probability distribution of the security state or the posterior probability distribution after arbitration, and the corresponding explanatory chain, including evidence paths, key rules, and weight contributions.

[0335] The Credible Decision Output Module is used to calibrate the credibility of probabilistic outputs and generate final results for management decision-making. Its inputs include: the probability distribution from the fusion arbitration inference module and a pre-stored historical calibration sample set containing known true security status labels; its processing includes: constructing an inconsistency metric and calculating quantile thresholds, and using conformal prediction to generate a prediction set as confidence intervals; its outputs include: a security status diagnostic conclusion containing confidence intervals, confidence level parameters, an evidence interpretation chain, and traceable trigger / arbitration records.

[0336] In this embodiment, the system can be deployed on the dam management center server or a cloud platform. The evidence graph service can use a graph database or RDF storage, providing a unified graph query interface and evidence writing interface to each module. The fusion inference engine and PSL solver serve as the computational kernel of the fusion arbitration inference module to achieve an adaptive inference architecture of normal factor graph fusion and abnormal logic arbitration. Through modular design, the system not only outputs safety-state diagnostic results but also synchronously outputs evidence paths, weight adjustments, and arbitration basis, meeting the application requirements of traceability and interpretability in engineering projects.

[0337] This invention achieves bidirectional coupling between mechanistic parameter inversion and evidence fusion by constructing a collaborative inversion objective function and introducing evidence consistency constraints and fusion posterior guidance terms. This forms a closed-loop mechanism of inversion-fusion-correction, effectively solving the problem of unidirectional error propagation and inability to correct errors in traditional serial methods, making the diagnostic results more robust. By defining conflict classification variables and four conflict types, and performing classification inference based on feature vectors and probabilistic soft logic rules, the invention achieves automatic location of conflict root causes, providing differentiated handling strategies for different types of conflicts. This avoids the risk of erroneous suppression of real abnormal signals that may result from the one-size-fits-all weighting of conflict evidence in traditional methods. Through a closed-loop update mechanism of evidence reliability weights, when a model mismatch conflict is detected, the parameter recalibration process is automatically triggered, enabling the system to automatically recover from a contaminated state to a self-cleaning state, achieving collaborative optimization of inversion and arbitration.

[0338] Furthermore, by constructing a two-layer adaptive inference architecture combining factor graph fusion and logical arbitration, probabilistic Soft Logic (PSL) arbitration is immediately triggered upon detecting missing data or conflicting evidence. This introduces higher-order engineering rules to dynamically reallocate weights, ensuring that diagnostic conclusions remain consistent with engineering logic even under extreme conditions. A collaborative inversion method under graph constraints is proposed. By explicitly adding physical prior constraints from the evidence graph to the inversion objective function, the parameter optimization process is forced to occur within a physically consistent range, effectively solving the problem of the disconnect between the mechanistic model and actual operating conditions. Conformal prediction technology is introduced. By calculating inconsistency measures and constructing prediction sets, traditional single probability values ​​are transformed into statistically rigorous confidence intervals, providing reliable quantitative support for risk decision-making by engineering managers.

[0339] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A data-mechanism-knowledge jointly driven method for self-diagnosis of dam safety status, characterized in that, include: Acquire operational status data of the dam, extract state variables characterizing the dam's safety status based on a pre-constructed computable evidence map, and map them into monitoring evidence for safety status propositions; Under the prior constraints of the computable evidence map, the pre-stored mechanism model parameters are inverted and corrected using monitoring evidence, and the verification calculation is performed accordingly to obtain the mechanism evidence. By using monitoring evidence and mechanistic evidence as observation factors and integrating them into a pre-configured probabilistic graphical model for joint inference, a primary probability distribution of the safety state is obtained. When the preset conflict triggering conditions are met, the weights of monitoring evidence and mechanism evidence are adjusted using a pre-configured logical rule model, and arbitration inference is performed to obtain the corrected security state posterior probability distribution. Based on a pre-stored historical calibration sample set, the confidence level of the primary probability distribution or the corrected posterior probability distribution of the safety state is calibrated using a conformal prediction method, resulting in a safety state diagnostic conclusion containing a confidence interval.

2. The method according to claim 1, characterized in that, Computable evidence maps include an ontology layer and an evidence layer; The ontology layer defines the topological relationships between engineering objects, structural parts, safety indicators, disease types, and sets of safety-state propositions; The evidence layer defines the association path between evidence items and the set of security-state propositions, and the edge weights of the association path are composed of probability components and credibility components.

3. The method according to claim 1, characterized in that, Operational status data includes monitoring time-series data, inspection record data, and historical event data; Mapping state variables to monitoring evidence for security-state propositions includes: Amplitude features, trend features, and environmental response features are extracted from the monitoring time series data as state variables. The state variables are then converted into probability vectors for safety state propositions using a preset scoring function. The inspection record data and historical event data are converted into semantic evidence items according to structural parts and preset hierarchical rules, and then written into monitoring evidence in combination with probability vectors.

4. The method according to claim 1, characterized in that, The parameters of the pre-stored mechanistic model are inverted and corrected using monitoring evidence, including: The prior intervals of physical parameters and engineering physical constraints of the mechanism model parameters are extracted from the ontology layer of the computable evidence map. Construct an inversion objective function, which includes an observation consistency term that characterizes the deviation between the mechanism model response and the monitoring evidence, and a priori constraint term that characterizes the deviation of the mechanism model parameters from the prior interval of the physical parameters. The objective function of the inversion is optimized under the condition of satisfying the engineering physical constraints to obtain the parameter estimates of the mechanism model.

5. The method according to claim 4, characterized in that, The method also includes an uncertainty interface that outputs an uncertainty characterization of the inversion confidence level. The uncertainty interface includes the following metrics: The posterior covariance matrix of the parameters, calculated based on the second derivative information of the inverted objective function, is used to characterize the distribution range of the parameter estimates. The residual statistics calculated based on the residual sequence after inversion convergence are used to characterize the explanatory power of the mechanistic model for monitoring evidence. The identifiability index, calculated based on the parameter sensitivity matrix, is used to characterize the invertibility of mechanistic model parameters under current monitoring evidence.

6. The method according to claim 1, characterized in that, The probabilistic graphical model is specifically a factor graphical model; The construction of probabilistic graphical models includes: Establish variable nodes that correspond one-to-one with the safety state proposition, as well as factor nodes that correspond to monitoring evidence and mechanistic evidence, respectively; Based on the edge weights defined in the computable evidence graph, a potential function is constructed to connect variable nodes and factor nodes.

7. The method according to claim 1, characterized in that, The preset conflict triggering conditions include at least one of the following conditions: The data integrity rate calculated based on the running status data is lower than the preset integrity threshold, or the data anomaly rate is higher than the preset anomaly rate threshold. The degree of evidence conflict between monitoring evidence and mechanistic evidence calculated using evidence theory or distance metric functions is higher than the preset conflict threshold.

8. The method according to claim 7, characterized in that, The weights of monitoring evidence and mechanistic evidence are adjusted according to the conflict classification-driven principle, specifically as follows: When the conflict is triggered by the data integrity rate or anomaly rate not meeting the requirements, a linear decay function is constructed based on the proportion of missing or anomaly data to reduce the weight of the monitoring evidence. When the conflict trigger condition is that the degree of conflict of evidence is higher than the conflict threshold, the consistency score of each piece of evidence is calculated based on the logical rule model, and the weights of monitoring evidence and mechanism evidence are redistributed accordingly.

9. The method according to claim 1, characterized in that, The logical rule model adopts a probabilistic soft logic model; The corrected posterior probability distribution of the safety state is obtained, including: Map the truth values ​​of security-state propositions to continuous variables on the closed interval [0, 1]. Based on the Lukasiewicz fuzzy logic operator combined with continuous variables, a weighted set of logic rules containing monitoring evidence and mechanistic evidence is constructed; By minimizing the hinge loss distance of the weighted logic rule set on all security-state propositions using a convex optimization algorithm, the corrected security-state posterior probability distribution is inferred.

10. A data-mechanism-knowledge jointly driven dam safety state self-diagnosis system, characterized in that, include: The holographic perception and evidence mapping module is used to acquire the dam's operational status data, extract state variables that characterize the dam's safety status based on a pre-constructed computable evidence map, and map them into monitoring evidence for safety status propositions. The mechanism co-inversion module is used to invert and correct the pre-stored mechanism model parameters using monitoring evidence under the prior constraints of the computable evidence map, and to perform verification calculations to obtain mechanism evidence. The fusion arbitration inference module is used to input monitoring evidence and mechanistic evidence as observation factors into a pre-configured probabilistic graphical model for joint inference to obtain the primary probability distribution of the safety state; It is used to adjust the weights of monitoring evidence and mechanism evidence and perform arbitration inference when the preset conflict triggering conditions are met, so as to obtain the corrected security state posterior probability distribution. The credible decision output module is used to perform credibility calibration on the primary probability distribution of the safety state or the corrected posterior probability distribution of the safety state based on the pre-stored historical calibration sample set using the conformal prediction method, and obtain a safety state diagnostic conclusion containing a confidence interval.

Images