Marine environment multi-parameter coupling equipment fault diagnosis method

Abstract

The application discloses a marine environment multi-parameter coupling equipment fault diagnosis method, and belongs to the technical field of marine equipment state monitoring and fault diagnosis. Real-time observation values of multiple environment parameters are collected, a multi-parameter coupling strength index is calculated, a forbidden edge set and a necessary edge set are constructed, a hard constraint is imposed on data-driven causal diagram learning, a directed acyclic graph is output, a hierarchical dynamic Bayesian network including an environment layer, a degradation layer and a fault layer is constructed, and the multi-parameter coupling strength index is introduced into the state transition probability of the degradation layer to realize dynamic modulation; a multi-fault concurrent scene is decoupled and identified, independent concurrent faults and composite faults are distinguished, and component confidence is output, and online continuous learning is realized through a Bayesian conjugate increment method. The application effectively solves key problems such as multi-environment parameter coupling effect quantization, fusion of physical constraints and data driving, and sensor signal quality self-adaptation, and has long-term self-adaptation capability to service environment changes and new fault types.

Description

Technical Field

[0001] This invention belongs to the field of marine equipment condition monitoring and fault diagnosis technology, specifically involving a fault diagnosis method for equipment coupled with multiple parameters of the marine environment. Background Technology

[0002] Marine equipment operates under the combined influence of various environmental parameters, including temperature, salinity, corrosion current density, and structural stress amplitude. These parameters do not affect the material properties and structural integrity of the equipment independently; rather, they interact through multiple physical mechanisms, such as electrochemistry, thermodynamics, and mechanics, resulting in a degradation rate far exceeding the linear summation of the individual parameters' effects. Therefore, accurate fault diagnosis of marine equipment necessitates a thorough consideration of the coupling and interaction effects among these environmental parameters and their synergistic contribution to the degradation process.

[0003] Most existing fault diagnosis methods for marine equipment treat each environmental parameter as an independent influencing factor and model them separately, or focus only on the impact of a single dominant environmental parameter on the degradation process. They fail to establish quantitative characterization methods for the coupling strength between multiple environmental parameters, nor do they incorporate the multi-parameter coupling effect into the dynamic modulation of the degradation state transition probability. This results in diagnostic models lacking the ability to perceive the accelerated degradation phenomenon under the synergistic effect of multiple parameters, easily leading to underestimation of degradation state and high false alarm rates in harsh environments with multi-parameter coupling. Existing data-driven causal discovery methods, when faced with limited and noisy historical operational data of marine equipment, are prone to introducing false causal relationships that do not conform to physical laws due to misjudgments in statistical tests, or omitting experimentally verified true causal connections. The resulting fault diagnosis network topology is inconsistent with the actual physical mechanism, thus affecting inference accuracy and increasing the false alarm rate. Furthermore, sensors in the marine environment are subject to long-term corrosion and fouling, and the data quality of different sensors deteriorates differently over time. Existing methods typically assign equal inference weights to all sensor data, failing to adaptively distinguish between high-quality and low-quality signals. This leads to noise from damaged sensors adversely interfering with the diagnostic results. In terms of fault identification, marine equipment often encounters complex scenarios with multiple concurrent faults during actual service, including concurrent faults caused independently by different physical reasons and composite faults caused by a common causal root. Existing fault diagnosis frameworks mostly target single fault types and lack the ability to systematically decouple multiple concurrent fault scenarios. They struggle to distinguish between independent concurrent and composite faults and cannot provide the contribution intensity of each fault component. Furthermore, marine equipment typically has a service life of several decades. During service, new fault modes may emerge or material aging may cause systematic changes in degradation characteristics. Existing diagnostic models, once offline training is completed, have their parameters fixed and lack the ability for continuous online learning and adaptive updates, resulting in a gradual decline in model accuracy over time. Summary of the Invention

[0004] To address the problems existing in the background art, the present invention provides a method for equipment fault diagnosis based on multi-parameter coupling in a marine environment, comprising the following steps:

[0005] S1. Calculation of Multi-parameter Coupling Strength Index: Collect real-time observation values ​​of multiple environmental parameters during equipment operation, take the corresponding values ​​of each environmental parameter under the rated operating conditions of the equipment as the benchmark reference value, and calculate the multi-parameter coupling strength index MPCI based on the deviation of each environmental parameter from the benchmark reference value and the coupling strength coefficient between each pair of parameters.

[0006] S2, Physically Constrained Cause-Effect Graph Learning: Pre-defined causal directions that are not allowed to exist form a set of prohibited edges, and proven causal relationships form a set of required edges; Using historical equipment operation data as input, conditional independence tests are performed, and constraints are imposed on the graph structure by combining the set of prohibited edges and the set of required edges, and a directed acyclic graph is output.

[0007] S3. Construction of Hierarchical Dynamic Bayesian Network: Based on the directed acyclic graph obtained in step S2, construct a hierarchical dynamic Bayesian network containing an environment layer, a degradation layer, and a fault layer; introduce the MPCI calculated in step S1 into the state transition probability across time steps of the degradation layer, so that the probability of the degradation state jumping to a higher degradation level increases with the increase of MPCI; perform offline parameter learning on the hierarchical dynamic Bayesian network.

[0008] S4. Adaptive weighted fault reasoning: Calculate the mutual information between the observations of each sensor and the fault state. Use the normalized mutual information as the information contribution weight of the corresponding sensor. Substitute the weighted observation likelihood into the hierarchical dynamic Bayesian network for reasoning and output the posterior probability of each fault type. When the highest posterior probability exceeds the preset threshold, output a fault warning.

[0009] S5. Multi-fault concurrent decoupling identification: Filter candidate faults whose posterior probability exceeds the filtering threshold. Under the condition of knowing the current degradation state, calculate the conditional non-independence metric of any two types of candidate faults. Based on the comparison result of the metric and the judgment threshold, the multi-fault scenario is judged as independent concurrent fault or compound fault, and the corresponding diagnostic conclusions are output respectively.

[0010] Furthermore, step S1 specifically includes:

[0011] S11. Determine the environmental parameter vector and reference values: Obtain from the sensor system. Real-time measurements of several environmental parameters , Each parameter type should at least include temperature, salinity, corrosion current density, and structural stress amplitude; record the reference values ​​for each parameter of the equipment under rated operating conditions. The baseline reference value is determined through factory testing and stored in the diagnostic system before equipment deployment;

[0012] S12. Determine the pairwise coupling strength coefficients: In the accelerated degradation test, apply single environmental parameter excitation and dual-parameter combined excitation to the equipment respectively, and measure the degradation rate under each working condition; for any two environmental parameters... and The coupling strength coefficient is calculated using the following formula. :

[0013] ;

[0014] In the formula: For parameters and The coupling strength coefficient; for and The degradation rate was measured while the excitation was applied simultaneously; To apply only The degradation rate measured during excitation; To apply only The degradation rate measured during excitation; This is the baseline degradation rate under rated operating conditions; when the calculated result is less than zero, let... ; This process is repeated for all parameter pairs to form the coupling strength coefficient matrix. ;

[0015] S13. Calculate the multi-parameter coupling strength index: Calculate using the following formula. Moment :

[0016] ;

[0017] In the formula: for The multi-parameter coupling strength index at time t; This represents the total number of environmental parameters. For all satisfied Summing the parameters; The coupling strength coefficient (dimensionless) obtained in step S12. For the first Environmental parameters in The measured value at that moment; For the first The baseline reference value for each environmental parameter; This is the absolute value operator; This indicates no coupling deviation effect. A larger value indicates a stronger synergistic effect of multiple environmental parameters on degradation.

[0018] Furthermore, step S2 specifically includes:

[0019] S21. Constructing a set of physical constraints: Based on the known laws of marine engineering thermodynamics, electrochemistry, and structural mechanics, construct a set of forbidden edges. With the set of necessary edges Forbidden edge set It must contain at least all causal directions from internal equipment parameters to uncontrollable marine environmental parameters; it must be combined with other methods. It contains at least two types of causal relationships that have been experimentally verified: salinity and corrosion current density, and temperature and material strength degradation;

[0020] S22. Conditional Independence Test to Construct the Skeleton: Using historical equipment operation data as a sample, perform a conditional independence test on all node pairs. Edges between node pairs with statistical dependencies are retained, otherwise they are deleted, thus forming the undirected network skeleton.

[0021] S23. Apply physical constraints to determine edge directions: Delete the set of forbidden edges from the undirected skeleton obtained in S22. All edges appearing in the array will be required to form an edge set. Edges are forcibly added according to the specified causal direction. The direction of the remaining undirected edges is determined one by one according to the V-shaped structure criterion and acyclicity constraint. The output is a directed acyclic graph. This serves as the topology for the hierarchical dynamic Bayesian network in step S3.

[0022] Furthermore, step S3 specifically includes:

[0023] S31. Establish a three-layer network node structure: based on a directed acyclic graph. All nodes are assigned to three layers: the environment layer nodes are the observable environmental parameters collected in step S1; the degradation layer nodes are the hidden variables of the equipment degradation state. ,Pick For normal For mild degeneration, For moderate degradation, For severe degradation, there are four discrete levels; the fault layer nodes are observable internal parameters of the equipment, including at least vibration acceleration and structural strain.

[0024] S32. Introducing MPCI's setting of degradation layer state transition probabilities: Degradation states are determined by levels. Jump to a higher degradation level ( The transition probability of ) is calculated using the following formula:

[0025] ;

[0026] In the formula: Given the degenerate state at the previous time step: And currently Under certain conditions, the degenerate state jumps to... The transition probability (dimensionless). When MPCI modulation is not considered arrive The base transition probability (dimensionless) is initialized from historical fault data; Coupling effect coefficient ( (dimensionless), determined by maximum likelihood estimation from historical fault data; Calculated for step S1 Time-varying multi-parameter coupling strength exponent (dimensionless); It is a natural exponential function; This is a normalization constant, its value equal to the sum of all transition probabilities in that row, ensuring that the sum of the transition probabilities of all rows is 1; the above amplification only applies to... That is, apply the probability of jumping to a higher degradation level, and then normalize the entire row;

[0027] S33. Expectation-Maximization (EM) Algorithm Parameter Learning: Divide historical running data into training and validation sets, perform EM algorithm iterations on the training set, and calculate the latent variables of the degenerate state in the E-step. The posterior distribution is obtained, and all conditional probability parameters are updated in the M-step based on the E-step results. The two steps are alternately iterated until the log-likelihood increment between two adjacent iterations is less than the convergence threshold. The classification accuracy is evaluated with a validation set. After the validation is passed, the network parameters are fixed, and the hierarchical dynamic Bayesian network with completed parameter learning is output.

[0028] Furthermore, step S4 specifically includes:

[0029] S41. Sliding Window Mutual Information Estimation: Using a sliding time window of preset length, for the first... Observation data sequence of each sensor Fault status labels within the corresponding time period Perform joint frequency statistics and estimate using the mutual information formula. Output the mutual information of each sensor at the current moment. (bit);

[0030] S42. Normalized mutual information weight calculation: Calculate the weight using the following formula. One sensor in Information contribution weight at any given moment :

[0031] ;

[0032] In the formula: For the first One sensor in Information contribution weight at time (dimensionless, satisfying) ); The estimated first step in step S41 Mutual information (bits) of each sensor; For all Summing the values ​​from all sensors; when all sensors... When both are 0, let ;

[0033] S43, Weighted Belief Propagation Reasoning and Early Warning Output: Based on For the exponent with respect to the first Conditional observation likelihood of each sensor After performing power-weighted processing, the overall observation likelihood is obtained as follows:

[0034] ;

[0035] In the formula: For comprehensive observational likelihood (dimensionless); For the first One sensor in Conditional observational likelihood at time (dimensionless); The first result obtained in step S42 The information contribution weight of each sensor (dimensionless); To multiply all sensors together; Substitute the results into a hierarchical dynamic Bayesian network to perform belief propagation inference and output various fault types. The posterior probability; when the highest posterior probability exceeds a preset threshold. When a fault warning is triggered, the corresponding fault type and its confidence level are output.

[0036] Furthermore, step S5 specifically includes:

[0037] S51. Candidate Concurrent Fault Screening: From the posterior probability vector output in step S4, screen for faults whose posterior probabilities exceed the screening threshold. The fault types constitute the candidate set, where If there is only one fault type in the candidate set, the single fault diagnosis conclusion will be output directly.

[0038] S52. Calculation of Conditional Non-Independence Metric: For any two types of faults in the candidate set... and Given the current degradation state Under the given conditions, calculate the conditional non-independence measure using the following formula. :

[0039] ;

[0040] In the formula: For fault and In a known degenerate state A measure of conditional non-independence under given conditions (dimensionless). Known under conditions and Simultaneous joint conditional probability (dimensionless). , They are known respectively under conditions , The conditional probability of each occurrence (dimensionless); This is the absolute value operator; where By including degenerate state nodes in a Bayesian network After instantiation, the data is obtained using a connection tree algorithm or Gibbs sampling.

[0041] S53, Fault Component Output: When Less than the judgment threshold At that time, the judgment and For independent concurrent faults, each fault outputs its own diagnostic conclusion and confidence level independently; when Not less than When the fault is determined to be a compound fault, a non-negative least squares (NNLS) optimization model is established: Let Let matrix be the feature vector extracted from the sensor signal at the current time. Each column is a standard sensor response template vector corresponding to a single fault. Given the fault component coefficient vector to be solved, solve... ,constraint , with coefficient vector The magnitude of each element serves as the diagnostic confidence level for each fault component. The output lists the components of the composite fault and the confidence level of each component.

[0042] Furthermore, the present invention also includes step S6 of online incremental learning: accumulating confirmed labeled fault samples; when the accumulated number reaches a preset batch size, updating the conditional probability parameters of the hierarchical dynamic Bayesian network using a Bayesian conjugate incremental approach, and monitoring the change in parameter distribution before and after two adjacent updates to determine whether the current batch update has converged; specifically including:

[0043] S61. Accumulation of Labeled Samples and Batch Triggering: Continuously receives labeled samples whose fault type has been confirmed on-site by maintenance personnel, and temporarily stores them in a sample buffer; when the number of samples in the buffer reaches the preset batch size... Parameter updates are triggered on demand; when a new fault type appears in the sample buffer that has never been seen in the training set, the parameters are reduced. To accelerate initial adaptation;

[0044] S62, Dirichlet parameter incremental update: for each type of fault in the network. The corresponding conditional probability parameters are incrementally updated using the following formula:

[0045] ;

[0046] In the formula: For the updated version Positive integer pseudo-counts of the Dirichlet distribution corresponding to the fault class; The positive integer pseudo-count of the Dirichlet distribution before the update; This is the first labeled sample in this batch. The actual number of times each type of fault occurred; during system initialization. Taking 1 corresponds to no prior information;

[0047] S63, KL divergence convergence monitoring and write-back: After each parameter update, calculate the KL divergence between the conditional probability distributions before and after the update; when the KL divergence is less than the convergence threshold, determine that the update of this batch has converged, write the updated conditional probability parameters back to the hierarchical dynamic Bayesian network, clear the sample buffer of the current batch, and enter the next batch to wait.

[0048] The beneficial effects achieved by this invention are as follows:

[0049] This invention proposes a quantification method for a multi-parameter coupling strength index. By collecting real-time observations of multiple environmental parameters and combining them with pairwise coupling strength coefficients calibrated through accelerated degradation tests, the synergistic promoting effect of various environmental parameters such as temperature, salinity, corrosion current density, and structural stress amplitude on equipment degradation is quantified into a single scalar index that can be calculated in real time. The index is introduced into the cross-time step state transition probability of the degradation layer of the hierarchical dynamic Bayesian network. Through an exponential function modulation mechanism, the probability of the degradation state jumping to a higher degradation level increases monotonically with the increase of coupling strength. This enables the diagnostic model to dynamically reflect the accelerating effect of the cross-coupling of multiple environmental parameters on the equipment degradation rate. This overcomes the problem of large degradation state estimation bias and high fault false alarm rate caused by treating each environmental parameter as an independent factor in existing methods, and significantly improves the fault diagnosis accuracy under harsh multi-parameter coupling environments.

[0050] This invention proposes a physically constrained causal graph learning method. Based on data-driven conditional independence testing, it embeds known physical laws from marine engineering thermodynamics, electrochemistry, and structural mechanics as hard constraints in the causal graph learning process, in the form of forbidden edge sets and required edge sets. This ensures that the learned directed acyclic graph conforms to statistical data, effectively avoiding the problem of false causal edges introduced by purely data-driven methods due to limited samples and noise interference. This reduces the negative impact of causal structure errors on the inference quality of Bayesian networks, thereby reducing the false alarm rate. Simultaneously, this invention proposes an adaptive sensor weighting mechanism based on mutual information. Using the mutual information between each sensor observation and the fault state within a sliding time window as the information contribution weight, it performs power-weighted processing on the conditional observation likelihood. This allows sensors with higher mutual information to contribute more to the inference results, while the contributions of sensors affected by corrosion, resulting in signal drift or increased noise, are automatically suppressed, improving the robustness of the diagnostic system to low-quality sensor data.

[0051] This invention proposes a multi-fault concurrent decoupling identification method. By calculating the conditional non-independence metric between candidate faults under known degradation conditions, it classifies multi-fault scenarios into two categories: independent concurrent faults and composite faults. For composite faults, a non-negative least squares optimization model is used to separate the contribution intensity of each fault component, outputting a complete combined diagnostic conclusion and the confidence level of each component. This method overcomes the limitation of traditional single-fault diagnostic frameworks in multi-fault concurrent scenarios, which can only identify the primary fault while missing secondary faults, thus improving the diagnostic completeness and accuracy in complex fault scenarios.

[0052] This invention proposes an online learning method based on Bayesian conjugate increments. It uses accumulated labeled fault samples to perform lightweight incremental updates on the conditional probability parameters of a hierarchical dynamic Bayesian network. Each batch of updates only requires adding pseudo-counts, resulting in extremely low computational cost and eliminating the need for downtime for full model retraining. Convergence is determined by monitoring the KL divergence between the conditional probability distributions before and after the update, and timely warnings are issued when a systematic shift in fault modes is detected. This enables the diagnostic system to continuously improve diagnostic accuracy as service time increases and new fault modes accumulate, providing long-term adaptability to changes in the service environment and new fault types. Attached Figure Description

[0053] Figure 1 This is a bar chart comparing the fault diagnosis accuracy of Examples 1, 2, and 3 with Comparative Examples 1, 2, 3, and 4.

[0054] Figure 2 This is a comparison curve of the impact of MPCI on the degradation state transition probability of Examples 1, 2, and 3 and Comparative Example 1.

[0055] Figure 3 The charts show the comparison of false alarm rate and false alarm rate between Examples 1, 2, and 3 and Comparative Examples 1, 2, 3, and 4, where (a) is a bar chart comparing false alarm rate (MDR) and (b) is a bar chart comparing false alarm rate (FAR).

[0056] Figure 4 These are the online incremental learning KL divergence convergence curves for Examples 1, 2, and 3.

[0057] Figure 5 This is a schematic diagram of the multi-parameter coupling strength index calculation process of the present invention.

[0058] Figure 6 This is a schematic diagram of the workflow of the online incremental learning module of the present invention.

[0059] Figure 7 This is a schematic diagram of the hierarchical dynamic Bayesian network structure of the present invention. Detailed Implementation

[0060] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. In addition, the forms of the various structures described in the following embodiments are merely illustrative. The present invention is not limited to the structures described in the following embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] This invention addresses the complex fault problems caused by the combined effects of multiple environmental parameters such as temperature, salinity, corrosion, and stress on marine equipment. It proposes a fault diagnosis method that integrates physical constraint causal graph learning, hierarchical dynamic Bayesian inference, adaptive sensor weighting, and online Bayesian incremental updates, and constructs a corresponding diagnostic system based on this method. The diagnostic method of this invention is executed sequentially according to steps S1 to S6, as follows:

[0062] Step S1: Calculation of the Multi-Parameter Coupling Intensity Index (MPCI); The goal of Step S1 is to quantify the synergistic promoting effect of multiple environmental parameters on equipment degradation into a scalar index that can be calculated in real time, namely, MPCI. Environmental parameters such as temperature, salinity, corrosion current density, and structural stress amplitude experienced by marine equipment in actual service do not act independently on materials and structures, but rather through cross-coupling via multiple physical mechanisms including electrochemical, thermodynamic, and mechanical mechanisms, resulting in a degradation rate far exceeding the sum of the individual effects of each parameter. (Refer to...) Figure 5 Step S1 includes the following three sub-steps.

[0063] S11 determines the environmental parameter vector and the baseline reference value. The diagnostic system obtains data from various sensors. Environmental parameters at the current moment The real-time measurement value is denoted as ,in The minimum value should be 4, and each parameter type should at least cover temperature, salinity, corrosion current density, and structural stress amplitude. Meanwhile, the system's baseline reference values ​​for each parameter should be determined through factory rated operating condition tests before the equipment leaves the factory. The parameters are then stored in the parameter library of the diagnostic system. The reference value represents the typical marine environmental conditions under which the equipment operates normally. The greater the deviation of each environmental parameter from the reference value, the more significant the difference between the equipment's environment and its rated operating conditions. S11 outputs the real-time vector of environmental parameters and the vector of reference values.

[0064] S12 is the coupling strength coefficient that determines the relationship between any two environmental parameters. Since the degradation process of marine equipment involves complex multiphysics coupling, any two environmental parameters... and The interaction effect on equipment degradation rate cannot be directly and accurately derived from theoretical formulas; therefore, accelerated degradation tests are used to quantitatively determine the coupling strength of each parameter pair. In the test, a single parameter is applied to the equipment. Excitation, single parameter Incentives and A dual-parameter combined excitation and a rated operating condition reference excitation were used to record the steady-state degradation rate of the equipment under each operating condition. Coupling strength coefficient. Calculate using the following formula:

[0065] ;

[0066] In the formula: For parameters and The coupling strength coefficient (dimensionless). for and The degradation rate was measured while the excitation was applied simultaneously; To apply only The degradation rate measured during excitation; To apply only The degradation rate measured during excitation; This is the baseline degradation rate under rated operating conditions; all degradation rates have the same dimension; when the calculation result is negative, let... .

[0067] The numerator of the above formula reflects the additional degradation increment of the two-parameter joint excitation relative to the sum of the two single-parameter excitations, which is attributed to the coupling interaction between the two parameters; the denominator is normalized by the baseline degradation rate, making These become dimensionless coefficients, facilitating comparisons between parameter pairs with different dimensions. When This indicates that there is no interaction effect between the two parameters at the degradation level; A larger value indicates a stronger coupling effect. After conducting experiments on all parameter pairs sequentially, a [database / structure] is formed. The coupling strength coefficient matrix is ​​obtained from the first-order coupling strength matrix, and S12 outputs this coefficient matrix.

[0068] It should be noted that the aforementioned coupling strength coefficient matrix is ​​calibrated and stored before the equipment leaves the factory. During the long-term service of the equipment, if the accumulated fault data indicates that the initially measured coupling coefficient can no longer accurately reflect the current material state, the coupling strength coefficient matrix can be periodically fine-tuned by using the long-term accumulated fault data and environmental parameter records through reverse inference to maintain the long-term accuracy of the model.

[0069] S13 is calculated based on S11 and S12. Moment The calculation formula is as follows:

[0070] ;

[0071] In the formula: for The multi-parameter coupling strength exponent at time t (dimensionless). This represents the total number of environmental parameters. The normalization coefficient (dimensionless) is the average of the parameters over the total number. For all satisfied The parameters are used to perform summation; Obtained from S12 and The coupling strength coefficient (dimensionless). For the first Environmental parameters in The measured value at that moment; For the first The baseline reference value for each environmental parameter; This is the absolute value operator; This indicates that all parameters are near the reference value and there is no coupling deviation effect; The larger the value, the stronger the synergistic effect of the current combination of environmental parameters on equipment degradation.

[0072] The above formula uses the absolute value symbol. The physical basis is as follows: For the marine equipment metal corrosion and structural degradation application scenario targeted by this invention, regardless of the direction of deviation of environmental parameters such as temperature and salinity (above or below the benchmark reference value), they will all accelerate the material degradation process through different physical mechanisms. For example, excessively high temperatures will accelerate corrosion by increasing the electrochemical reaction rate, while excessively low temperatures may indirectly promote structural damage by altering the brittle-ductile transition characteristics of the material. Therefore, using absolute values ​​to quantify the magnitude of parameter deviation can reasonably reflect the unidirectional promoting characteristic of the deviation direction on the coupled degradation effect in this application scenario.

[0073] To ensure the coupling effect coefficient in subsequent step S3 Numerical stability of parameter estimation is recommended in [the context of]... Before being passed to step S3, normalization is performed. The normalization method can be maximum-minimum normalization (mapping the maximum and minimum MPCI values ​​from historical data to a specific value). (interval) or hyperbolic tangent function compression ( This ensures that the normalized MPCI value range remains bounded, thereby avoiding The item overflowed due to an excessively large MPCI value. S13 output. The real-time value is fed into step S3 in real time and is used to dynamically modulate the degenerate state transition probability of the hierarchical dynamic Bayesian network.

[0074] Step S2: Physically Constrained Causal Graph Learning; The goal of Step S2 is to automatically discover the causal structure between environmental parameters, degradation states, and fault types from historical operational data. Simultaneously, known physical laws in the field of marine engineering are embedded as hard constraints into the learning process, thereby obtaining a directed acyclic graph (DAG) that conforms to both statistical data patterns and physical mechanisms. Step S2 consists of three sub-steps.

[0075] S21 involves constructing a set of physical constraints. Before performing data-driven causal learning, two types of constraint sets are manually compiled based on known principles in thermodynamics, electrochemistry, and structural mechanics. The first type is the set of forbidden edges. The first category includes all physically impermissible causal directions. For example, vibration signals inside equipment cannot be a causal source affecting the salinity of the surrounding seawater. Directions pointing from observable internal parameters to uncontrollable marine environmental parameters are all included in the prohibited edge set. The second category is the mandatory edge set. This includes causal relationships confirmed by electrochemical corrosion experiments and metal fatigue fracture theory, such as increased salinity leading to increased corrosion current density and the relationship between salinity and corrosion current density, as well as increased temperature causing decreased yield strength and material strength degradation. S21 Output Forbidden Edge Set and must set of edges The configuration details are completed by domain experts based on the specific equipment type before equipment deployment and remain fixed during the equipment's service life.

[0076] S22 uses historical equipment operation data as input to perform graph skeleton learning based on conditional independence (CI) tests. The CI test is a statistical method used to determine whether a statistical dependency exists between two variables given several other variables. This invention performs CI tests on each variable node pair. Commonly used methods include correlation tests based on Fisher-Z transform, suitable for continuous Gaussian variables, and tests based on chi-square distribution, suitable for discrete variables.

[0077] Before performing the CI test, the historical data is first tested for normality. If the data does not meet the Gaussian distribution assumption, such as ocean temperature or salinity data showing multimodal distribution or heavy-tailed characteristics, non-parametric independence test methods are used, such as kernel function-based independence test or distance correlation test, to ensure the statistical validity of graph skeleton learning and avoid erroneous deletion or retention of edges due to the failure to meet the distribution assumption.

[0078] The historical data used in S22 should be no less than 6 months of continuous equipment operation records to ensure that the sample covers a variety of typical marine environmental conditions. Undirected edges are retained between node pairs with statistical dependencies, while edges are deleted between node pairs without statistical dependencies, ultimately forming an undirected network skeleton. S22 outputs the undirected network skeleton for S23 to further determine the direction of the edges.

[0079] S23 applies physical constraints and determines the direction of edges based on the undirected skeleton obtained in S22, outputting a directed acyclic graph. The specific operation consists of three steps: First, remove the set of forbidden edges from the undirected skeleton. All edges appearing in the array; then the set of necessary edges. Edges in the specified direction are forcibly added to the skeleton; for the remaining undirected edges whose directions are not yet determined, their directions are determined one by one according to the V-shaped structure criterion and acyclic constraints. The V-shaped structure criterion refers to: if nodes With nodes All with nodes Connected and and If there is no direct connection between them, then CI testing does not support it. and In a given In the case of independent time, the direction is determined as ,Right now These are nodes with shared effects. The acyclicity constraint requires that no new directed loops be introduced during the direction determination process. After the above process is completed, a directed acyclic graph is output. The data is stored as a structured file and loaded into step S3.

[0080] Step S3: Construction of a hierarchical dynamic Bayesian network; the goal of step S3 is to construct a network based on a directed acyclic graph. A hierarchical dynamic Bayesian network (DBN) is constructed, and the MPCI calculated in step S1 is introduced into the state transition probability of the degradation layer during the parameter learning phase. This enables the network to dynamically reflect the modulation effect of environmental parameter coupling on the equipment degradation rate. Figure 7 As shown, step S3 includes three sub-steps.

[0081] S31 establishes a three-layer network node structure based on a directed acyclic graph. All nodes are distributed across three layers. The first layer is the environment layer, where nodes represent observable environmental parameters directly measured by sensors; the second layer is the degradation layer, where nodes represent latent variables related to the equipment's degradation state. ,Pick (normal), (Mild degeneration) (Moderate degeneration) (Severe degradation) consists of four discrete levels; the third layer is the fault layer, where nodes represent observable internal equipment parameters, such as vibration acceleration and structural strain, driven by the degradation layer nodes. This three-layer structure allows the network to introduce an explicit intermediate layer of degradation states between environmental inputs and internal sensor observations, thereby enabling the inference of implicit degradation processes. S31 determines the set of nodes in the network.

[0082] S32 is the setting of the degradation layer state transition probability introduced by MPCI. Assume the equipment degradation state is in... Always at the level The current MPCI value is ,but The degradation state can be jumped to a higher degradation level at any time. ( The conditional transition probability of ) is calculated using the following formula:

[0083] ;

[0084] In the formula: Given the degenerate state at the previous time step: And currently Under certain conditions, the degenerate state jumps to... The transition probability (dimensionless). When MPCI modulation is not considered arrive The basic transition probability (dimensionless) is initialized from historical fault data; Coupling effect coefficient ( (dimensionless), determined by maximum likelihood estimation from historical fault data; Calculated for step S1 Time-varying multi-parameter coupling strength exponent (dimensionless); It is a natural exponential function; This is a normalization constant, its value equal to the sum of all transition probabilities in that row, ensuring that the sum of the transition probabilities of all rows is 1; the above amplification only applies to... That is, the probability of jumping to a higher degradation level is applied, while the probability of state retention and change to a lower level is not introduced into MPCI modulation, and the entire line is renormalized.

[0085] The above formula correlates MPCI with state transition probability through an exponential function: the larger the MPCI value, the higher the probability of state transition. The value increases accordingly, thus amplifying the probability of equipment transitioning from a degraded state to a higher level. Parameter The positive constraint ensures the physical monotonicity that the larger the MPCI, the stronger the degradation acceleration effect.

[0086] Regarding the coupling effect coefficient The following details need to be specifically explained in the estimate:

[0087] 1) The value of must match the numerical scale of MPCI. It is recommended that the value be used in the estimation. First, normalize the MPCI according to the method described at the end of S13 to ensure that the MPCI value range is within a reasonable range and avoid... The item caused a numerical overflow.

[0088] 2) When using the EM algorithm for parameter learning, for The estimate should be based on a reasonable prior distribution (such as the gamma distribution). hyperparameters and (This can be set based on historical experience) or determined through cross-validation. The search range is determined to ensure the numerical stability and convergence of the optimization process.

[0089] 3) The final estimate should be validated by the classification accuracy on the validation set.

[0090] S33 utilizes the Expectation-Maximization (EM) algorithm for offline learning of the network's conditional probability parameters. The EM algorithm is an iterative optimization method for estimating parameters in probabilistic models with latent variables. Its core idea is to calculate the posterior distribution of the latent variables using the current parameter estimates in the E-step, and then maximize the log-likelihood of the observed data based on this posterior distribution in the M-step. These two steps are iterated alternately until convergence. In this invention, the degenerate state... As latent variables, only environmental parameters and internal sensor observations are available in the historical data. The historical data is divided into training and validation sets. The EM algorithm is iterated using the training set, stopping when the log-likelihood increment between two consecutive iterations is less than a convergence threshold. The classification accuracy is evaluated using the validation set, and once the accuracy threshold is met, the network parameters are fixed. The fixed parameters include the basic transition probability matrix. Coupling influence coefficient And the conditional probability table for each node in the fault layer. S33 outputs the hierarchical dynamic Bayesian network with learned parameters, which is then used for online inference in step S4.

[0091] Step S4: Adaptive Weighted Fault Inference; The goal of Step S4 is to use the network obtained in Step S3 to perform real-time probabilistic inference for each fault type, and to differentiate sensor signals of different qualities through a mutual information adaptive weighting mechanism, thereby improving the robustness of diagnosis to low-quality sensor data. Mutual information (MI) quantifies the statistical dependence between two random variables. The larger the mutual information, the higher the predictive value of one variable for the other. Therefore, using mutual information to measure the contribution of each sensor to fault diagnosis is supported by information theory. Step S4 includes three sub-steps.

[0092] S41 estimates the mutual information between each sensor and the fault state using a sliding time window. The sensor is evaluated using a sliding time window of a preset length. Joint frequency statistics are performed on the observed sequences and the corresponding fault status labels for each time period to estimate the frequency. Momentary mutual information (bit). The length of the sliding time window should be set reasonably according to the sensor sampling rate. When a sensor affected by corrosion experiences signal drift or increased noise, the sensor's... The value decreases accordingly, thus automatically reducing its inference contribution in subsequent weighting steps. S41 outputs the mutual information estimate of each sensor at the current time. .

[0093] S42 normalizes the mutual information of each sensor to obtain the information contribution weights. Normalization is performed using the following formula:

[0094] ;

[0095] In the formula: For the first One sensor in Information contribution weight at time step (dimensionless, satisfying that the sum of all weights equals 1); The estimated first step in step S41 Individual sensor mutual information (bit); For all Summing of the sensors; The total number of sensors; when all sensors When all weights are 0, set all weights to be equal, that is... This is to ensure numerical stability.

[0096] S43 substitutes the weighted observation likelihood into a hierarchical dynamic Bayesian network to perform inference, outputs the posterior probability of each fault type, and triggers an early warning when necessary. Let the... One sensor in The conditional observation likelihood at time is Then, the comprehensive observation likelihood is taken as the conditional likelihood of each sensor. In the form of a product of exponents:

[0097] ;

[0098] In the formula: For comprehensive observational likelihood (dimensionless); For the first One sensor in Conditional observational likelihood at time (dimensionless); The first result obtained in step S42 The information contribution weight of each sensor (dimensionless); The result is a product of all sensors; sensors with higher mutual information contribute more to the overall likelihood, while the contributions of sensors with lower mutual information are suppressed.

[0099] The aforementioned power-weighted likelihood, within the framework of information theory and robust Bayesian inference, is equivalent to minimizing the weighted logarithmic loss function, where The log-likelihood contribution of each sensor was scaled proportionally to its information content. It is important to note that... When the weights are not all 1, it is no longer a normalized probability density in the strict sense; in actual reasoning, Substituting the evidence factor into Bayes' formula, its normalization is guaranteed by the partition function in the posterior probability calculation, and does not affect the validity of the inference result. Substitute the network into forward and backward reasoning, and update each fault type according to Bayes' theorem. posterior probability When the highest posterior probability exceeds a preset threshold. When the threshold is reached, the system outputs an early warning, providing the fault type and posterior probability confidence interval; if the threshold is not reached, it outputs a pending observation status prompt. S43 outputs the posterior probability vector for each fault type and passes it to step S5.

[0100] Step S5: Multi-fault concurrent decoupling identification; The goal of step S5 is to analyze the multi-fault concurrent scenario based on the posterior probability vector output from step S4, distinguish between independent concurrent faults and compound faults, and provide corresponding diagnostic conclusions for each. Step S5 consists of 3 sub-steps.

[0101] S51 is the candidate concurrent fault screening. The posterior probability is extracted from the posterior probability vector output by S4. Exceeding the filtering threshold All fault types constitute the candidate concurrent fault set. Filtering threshold. Less than the warning threshold in S4 Its function is to broadly include potentially coexisting fault candidates with a low probability threshold, avoiding the omission of truly existing secondary faults due to probability dispersion. If there is only one fault type in the candidate set after screening, the single fault diagnosis conclusion is directly output, and subsequent sub-steps do not need to be executed. S51 outputs the candidate concurrent fault set.

[0102] S52 is a measure of conditional non-independence between any two types of faults in the candidate set. and Given the current degradation state Under the given conditions, calculate the conditional non-independence measure using the following formula. :

[0103] ;

[0104] In the formula: For fault and In a known degenerate state A measure of conditional non-independence under given conditions (dimensionless). Known under conditions and Simultaneous joint conditional probability (dimensionless). Known under conditions The single-class conditional probability of occurrence (dimensionless); Known under conditions The single-class conditional probability of occurrence (dimensionless); This is the absolute value operator.

[0105] Regarding joint conditional probability Efficient computation method: In practical implementation, degenerate state nodes are represented in the Bayesian network. The network is instantiated to the currently inferred degradation level, and then the connection tree algorithm is performed on the network for exact inference to obtain the joint posterior probability distribution of the two faulty nodes. When the network is large, approximate inference methods such as Gibbs sampling can also be used. Gibbs sampling strikes a balance between acceptable accuracy and computational overhead to meet the latency requirements of real-time inference.

[0106] The above formula is derived from the definition of conditional independence in probability theory, that is, if and In a given If the conditions are independent, then the absolute value of the difference between the two is 0. However, in actual data, statistical errors will not be exactly zero; therefore, a preset threshold is used. Distinguish between the conditions of independence and independence. S52 calculates each fault pair in the candidate set one by one. .

[0107] S53 outputs a diagnostic conclusion based on the results of S52. When Less than the judgment threshold At that time, the judgment and Under the current degenerate state, conditions are independent, meaning both are caused by independent physical causes, and the system outputs independently. and The diagnostic conclusions and their respective posterior probability confidence levels. When Not less than When the fault is identified as a composite fault, it means that the two types of faults exhibit a common causal root or a mutually reinforcing relationship.

[0108] For component separation of complex faults, a non-negative least squares (NNLS) optimization model is used for solution: first, feature vectors are extracted from the raw signals of each sensor at the current time. (such as frequency domain features, time domain statistics, etc.); secondly, construct a template matrix. , Each column corresponds to a standard response template vector for a single fault in the sensor's feature space. The standard template can be obtained by statistically averaging the features of historical single fault samples or generated through simulation. Then, the following optimization model is established:

[0109] Constraints: ;

[0110] In the formula Let be the vector of fault component coefficients to be solved. The Middle The size of the i-th element represents the value of the n-th element. The contribution strength of a single type of fault in current observations. The NNLS problem can be solved efficiently using the Active Set Method or the Projected Gradient Method. The resulting coefficient vector... The fault types corresponding to each non-zero element constitute the components of the composite fault, and the normalized coefficients serve as the diagnostic confidence levels for the corresponding components. S53 outputs complete multi-fault diagnostic conclusions, which are then fed into the system's early warning output module.

[0111] Step S6: Online Incremental Learning; The goal of Step S6 is to continuously update the conditional probability parameters of the hierarchical dynamic Bayesian network online using labeled fault samples verified in the field. This allows the diagnostic system to continuously improve its diagnostic accuracy as service time increases and new fault modes accumulate, without requiring downtime for full model retraining. Step S6 employs the Bayesian conjugate update method, utilizing the conjugate property of the Dirichlet and Categorical distributions to achieve lightweight incremental adjustments to the conditional probability table parameters. (Refer to...) Figure 6 Step S6 consists of 3 sub-steps.

[0112] S61 involves the accumulation and batch triggering of labeled samples. During routine equipment maintenance, whenever maintenance personnel confirm the specific type of a fault event on-site, the sensor observation vector at that moment and the confirmed fault type label together form a labeled sample, which is then stored in the system's sample buffer. When the number of accumulated samples in the buffer reaches a preset batch size... At that time, an online parameter update process is triggered. The settings need to balance the timeliness of updates with statistical stability. When a new fault type sample appears in the sample buffer that has never appeared in the training set, the number of samples required to trigger an update for that type should be reduced to speed up the initial adaptation to the new fault mode. S61 outputs the trigger signal and the set of labeled samples for the current batch.

[0113] S62 represents the incremental update of the Dirichlet distribution parameters. In Bayesian statistics, the Dirichlet distribution is the conjugate prior of the Categorical distribution parameters, and its parameter vector consists of pseudo-counts corresponding to each category. For each type of fault in the network... The corresponding conditional probability parameters are incrementally updated using the following formula:

[0114] ;

[0115] In the formula: For the updated version Positive integer pseudo-counts of the Dirichlet distribution corresponding to the fault class; The positive integer pseudo-count of the Dirichlet distribution before the update; This is the first labeled sample in this batch. The actual number of occurrences of this type of fault is a non-negative integer; This refers to the fault category number; each fault type is identified during system initialization. Setting it to 1 corresponds to no prior information, indicating that no preference is preset for each fault type.

[0116] After completing the pseudo-count update, the first Class failure in a degenerate state The expected conditional probability estimate is calculated using the following formula:

[0117] ;

[0118] In the formula: For the updated version Class failure in a degenerate state Expected conditional probability estimate (dimensionless, range) ); For the updated version pseudo-counts of Dirichlet distribution for class-related faults; This is used to sum the updated pseudo-counts for all fault categories.

[0119] The above update method is mathematically equivalent to using the current parameters as priors, the new batch of samples as observations, and deriving the posterior distribution according to Bayes' theorem, ensuring that the parameter estimates after each update are always within an effective probability distribution space. Since historical observations are encoded in the cumulative value of pseudo-counts, old knowledge is not simply overwritten during updates, and the model forgetting effect is effectively suppressed.

[0120] It should be noted that the aforementioned Dirichlet incremental update mainly targets the conditional probability parameters of the fault layer. To further maintain the long-term accuracy of the dynamic characteristics of the degradation layer, the update scope of online learning can be extended as needed to the degradation layer state transition probability parameters, i.e., the basic transition probability matrix in step S32. and coupling influence coefficient The degradation layer transition probability matrix also has a Dirichlet-Multinomial conjugate structure. It can be updated incrementally using the soft count of degradation states output by the E-step of the EM algorithm; or a lightweight online adjustment can be achieved by using the recursive Bayesian estimation method to cope with the situation where the actual degradation rate of equipment changes systematically due to material aging.

[0121] S63 monitors and determines the convergence of updates. After each S62 step, the KL divergence between the conditional probability distributions before and after the update is calculated to measure the magnitude of the update. KL divergence is a standard measure of the difference between two probability distributions in information theory; a value closer to 0 indicates a closer similarity in parameter distributions before and after the update, meaning the network parameters tend to be stable. When the KL divergence is less than a preset convergence threshold, the update batch is considered converged. After recording the update log, the current batch's sample buffer is cleared, awaiting the next batch of samples. If the KL divergence consistently exceeds a large value in several consecutive batches, it indicates a systematic drift of the recent fault mode. The system issues a model drift warning to maintenance personnel, prompting them to supplement historical data or re-execute offline training to adjust the model's basic structure. The specific values ​​of the convergence threshold and drift warning threshold are flexibly configured by the user based on equipment type and maintenance strategy. S63 enables the diagnostic system to continuously and adaptively update to new fault modes.

[0122] The following describes the fault diagnosis system that implements the above method. The system consists of seven functional modules, which work together through signal connections, data buses, and parameter read / write interfaces to continuously provide real-time fault diagnosis services throughout the entire service life of the equipment.

[0123] The sensor acquisition module is installed in the monitored equipment and its surrounding marine environment, and consists of three parts: an environmental sensor group, an internal state sensor group, and a signal conditioning and conversion unit. The environmental sensor group is deployed on the outer surface of the equipment and in the surrounding seawater area, responsible for collecting environmental parameters such as temperature, salinity, and corrosion current density. The internal state sensor group is installed in the critical load-bearing structural parts of the equipment, responsible for collecting internal operating status signals such as vibration acceleration and structural strain. The signal conditioning and conversion unit performs filtering, amplification, and analog-to-digital conversion on all sensor analog signals, and simultaneously transmits the digital data to the coupling strength calculation module and the network inference module via a fieldbus in a point-to-multipoint manner, realizing real-time distribution of sensor data.

[0124] The signal input terminal of the coupling strength calculation module is connected to the environmental parameter output terminal of the sensor acquisition module, continuously receiving real-time digital measurement values ​​of various environmental parameters. The module internally stores the coupling strength coefficient matrix determined through accelerated degradation testing in step S12 and the baseline reference values ​​of each environmental parameter recorded in step S11. Both are calibrated and permanently written before the equipment leaves the factory and can be updated as needed during equipment service. The coupling strength calculation module calculates in real-time according to the formula in step S13 at a fixed calculation cycle. The calculation results are continuously output to the MPCI input of the network inference module, providing a real-time basis for the dynamic modulation of the degradation state transition probability.

[0125] The data input terminal of the causal graph learning module is connected to the historical data storage unit, while the physical constraint input terminal receives a set of prohibited edges and a set of required edges pre-configured by domain experts. This module performs offline learning once before the equipment is officially put into operation, completing the directed acyclic graph according to step S2. Structural learning will It is stored as a structured file and loaded by the network inference module during initialization. During equipment operation, the causal graph learning module is in standby mode, and only when the user issues a relearning command through the human-computer interaction interface, or when the online learning module detects a continuous model drift warning, will the learning process be restarted and updated. document.

[0126] The network inference module uses the directed acyclic graph output by the causal graph learning module. For the topology, the conditional probability table is initialized with the offline parameter learning results from step S3. During equipment operation, this module performs three operations sequentially within each inference cycle: first, it reads the current... The value is adjusted according to the formula in step S32 to change the transition probability from the degradation layer to a higher degradation level. Then, the adjusted transition probability is multiplied by the degradation state belief vector stored in the previous time step to complete the prediction update of the degradation state. Finally, the current internal state sensor data is read from the sensor acquisition module, the weighted observation likelihood is calculated according to step S4, Bayesian inference is performed to complete the observation correction, the degradation state belief vector is updated, and the posterior probability vector of each fault type is output and sent to the fault diagnosis module. The network inference module is equipped with a parameter read / write interface. This interface allows the online learning module to write new conditional probability parameters after completing the update in S62. The interface has a built-in mutex lock protection to ensure that the parameter writing operation and the ongoing inference calculation do not occur simultaneously.

[0127] The input of the fault diagnosis module is connected to the posterior probability output of the network inference module. Internally, it includes a candidate fault screening unit, a conditional non-independence discrimination unit, and a diagnosis conclusion output unit. The candidate fault screening unit extracts a set of candidate faults from the posterior probability vector according to the logic in step S51; the conditional non-independence discrimination unit calculates the results for each candidate fault pair according to the formula in step S52. The system then transmits the judgment results of each fault pair to the diagnostic conclusion output unit. Following the logic of step S53, the diagnostic conclusion output unit outputs independent diagnostic conclusions for each independent concurrent fault, and for composite faults, it separates the fault components using the built-in NNLS solver and outputs a combined diagnostic conclusion. Simultaneously, the diagnostic conclusions are written to the fault history database. The output of the fault diagnosis module is connected to the early warning output module.

[0128] The online learning module's labeled data input terminal receives labeled fault samples entered through an external human-computer interaction interface. Internally, it consists of a sample buffer and a parameter update unit. The sample buffer stores labeled samples at different times, categorized by fault type, and monitors whether the accumulated number reaches the batch size. Upon reaching the target value, a trigger signal is sent to the parameter update unit, along with the current batch sample set. Upon receiving the trigger signal, the parameter update unit reads the Dirichlet distribution pseudo-counts for each fault category through the parameter read / write interface of the network inference module. It then performs an additive accumulation update according to the formula in step S62, calculates the new expected conditional probability, and calculates the KL divergence before and after the update according to step S63. It then determines whether the update has converged and pushes the historical KL divergence curve to the early warning output module for maintenance personnel to view. After convergence, the new parameters are written back to the network inference module, and the current batch buffer is cleared.

[0129] The input of the early warning output module is connected to the fault diagnosis module, and it also receives the KL divergence monitoring signal pushed by the online learning module. When the highest posterior probability output by the fault diagnosis module exceeds a preset threshold... At that time, the early warning output module issues fault warnings to maintenance personnel through a human-computer interaction interface using graphics and sound, displaying the fault type identifier, posterior probability value, confidence interval, and directed acyclic graph. The visualization results help maintenance personnel quickly locate the root cause of faults. When the online learning module reports that the KL divergence continuously exceeds the drift warning line, the early warning output module simultaneously displays model drift warning information, prompting maintenance personnel to pay attention to the systemic changes in recent fault modes. The early warning output module also provides a query function for historical early warning records. All early warning events are written to the fault history database, indexed by timestamps and device identifiers, for subsequent trend analysis and maintenance decisions.

[0130] Example 1: This example uses the BOP (Blowout Preventer) assembly of a deep-water semi-submersible drilling platform as the diagnostic object. The platform is deployed in the northern part of the South China Sea at a depth of about 1,500 m. It is subjected to the coupled effects of multiple environmental parameters such as high temperature, high salinity, strong corrosion and alternating stress all year round, which is suitable for verifying the effectiveness of the whole process method of the present invention.

[0131] Step S1: Calculation of multi-parameter coupling strength index: Execute step S11 to collect data from the sensor system. Real-time measured values ​​of several environmental parameters, the parameter types being seawater temperature. ,salinity Corrosion current density and structural stress amplitude The reference values ​​for each parameter under rated operating conditions are as follows: ℃ ‰、 mA / cm2 , MPa, the above benchmark reference value was determined through factory rated operating condition tests and stored in the diagnostic system parameter library before platform deployment.

[0132] Perform step S12, applying single-parameter excitation and dual-parameter combined excitation to the critical sealing components of the BOP in the accelerated degradation test, measuring the degradation rate under each operating condition, and applying the formula... Calculate the coupling strength coefficient of each parameter to form Coupling strength coefficient matrix; where temperature-salinity coupling strength coefficient Temperature-corrosion current density Temperature-stress amplitude Salinity-corrosion current density Salinity-stress amplitude Corrosion current density - stress amplitude .

[0133] Perform step S13, according to the formula Real-time computing The multi-parameter coupling strength exponent at time step 1 is mapped to the MPCI using a maximum-minimum normalization method. The interval is then passed to step S3.

[0134] Step S2: Physical constraint cause-effect graph learning: Execute step S21 to construct a set of forbidden edges based on the laws of marine engineering electrochemistry and structural mechanics. Includes 8 prohibited directions and required edges, such as vibration signal and salinity, structural strain and temperature. This includes two essential directions: salinity and corrosion current density, and temperature and material strength degradation. Step S22 involves using approximately 70,000 records (sampled at 5-minute intervals) of historical data from the platform over the past 8 months as input. A conditional independence test based on the Fisher-Z transform is then performed on all node pairs to assess the significance level of this test. This forms the undirected network skeleton. Step S23 is then executed to delete elements from the undirected skeleton. Edges in, forced addition The remaining edge directions are determined according to the V-shaped structure criterion and acyclic constraints, and a directed acyclic graph is output. .

[0135] Step S3: Construction of Hierarchical Dynamic Bayesian Network: Execute step S31, based on... Nodes are assigned to the four observable environmental parameter nodes in the environmental layer and the latent variables of the degradation state in the degradation layer. ,Pick normal, Mild degeneration, Moderate degeneration The system has four discrete levels of severe degradation and two observable nodes: vibration acceleration at the fault layer and structural strain. Step S32 is executed to set the coupling influence coefficient. Determined from historical fault data through maximum likelihood estimation, according to the formula Calculate the state transition probability of the degenerate layer across time steps. Execute step S33, divide the historical data into a training set and a validation set in a 7:3 ratio, and iteratively learn the network parameters using the EM algorithm, setting the convergence threshold to... After the network parameters were fixed, the classification accuracy of the validation set reached 91.3%.

[0136] Step S4 Adaptive Weighted Fault Reasoning: Execute step S41, with the sliding window length... Each sampling point estimates the mutual information between each sensor and the fault state. Execute step S42, according to the formula. Calculate the contribution weight of normalized information from each sensor.

[0137] Perform step S43 to weight the observation likelihood. Substitute the hierarchical dynamic Bayesian network to perform belief propagation inference, and set the warning threshold to... When the highest posterior probability exceeds Output fault warnings in real time.

[0138] Step S5: Multi-fault concurrent decoupling identification: Execute step S51 to filter thresholds. ( Candidate faults are selected from the posterior probability vector. Step S52 is executed, given the known degradation state. Under the condition, according to the formula Calculate the conditional non-independence measure and determine the threshold. Execute step S53, based on and The comparison results determine whether the fault is independent or concurrent. For composite faults, the confidence scores of each component are solved using the NNLS model.

[0139] Step S6 Online Incremental Learning: Execute step S61 and set the batch size. Execute step S62, according to the Dirichlet conjugate increment formula. Update the conditional probability parameters. Execute step S63 to determine whether the update has converged using the KL divergence convergence threshold of 0.01. If converged, write back the network parameters.

[0140] Example 2: This example uses a certain offshore wind turbine monopile foundation structure as the diagnostic object, which is deployed in the East China Sea at a water depth of about 40 m. It is subjected to the dual effects of alternating wave loads and marine atmospheric corrosion. The difference between this example and Example 1 is that the environmental parameter dimensions and the values ​​of key parameters are different.

[0141] Collect in step S1 One environmental parameter was added to the four parameters in Example 1, along with dissolved oxygen concentration. Benchmark reference value mg / L and wave load frequency Benchmark reference value Hz; the coupling strength coefficient matrix is ​​extended to The order, where the dissolved oxygen-corrosion current density coupling coefficient is... Wave load-stress amplitude coupling coefficient The coupling coefficients of the other newly added parameters are between 0.15 and 0.48.

[0142] Step S2 involves collecting historical data covering 12 months, totaling approximately 105,000 records; this must be done by edge aggregation. Increased dissolved oxygen → corrosion current density causality. Coupling effect coefficient in step S3. The degradation layer state transition probability is more sensitive to MPCI than in Example 1. The warning threshold in step S4... In step S5, the screening threshold is used. Determine the threshold Batch size in step S6 The KL divergence convergence threshold is 0.01. The execution flow of the remaining steps is the same as in Example 1.

[0143] Example 3: This example uses a submarine oil and gas pipeline as the diagnostic object. It is laid on the seabed of Bohai Bay and is about 120 km long. It is affected by both seabed sediment cover and ocean current erosion. The difference between this example and Examples 1 and 2 is the application scenario and parameter configuration.

[0144] Collect in step S1 One environmental parameter is added to the four parameters in Example 1, along with ocean current velocity. Benchmark reference value m / s; Temperature-salinity coupling coefficient Ocean current velocity-corrosion current density coupling coefficient Ocean current velocity-stress amplitude coupling coefficient Step S2 involves historical data covering 10 months, approximately 90,000 records. Step S3 involves the coupling influence coefficient. The degradation layer state transition probability is less sensitive to MPCI than in Examples 1 and 2. The warning threshold in step S4... In step S5, the screening threshold is used. Determine the threshold Batch size in step S6 The KL divergence convergence threshold is 0.008. The execution flow of the remaining steps is the same as in Example 1.

[0145] Comparative Example 1: Removing Multi-parameter Coupling Comparative Example 1 uses the same diagnostic objects, datasets, and network structure as Example 1, with the only difference being that in setting the degradation layer state transition probability in step S3, the coupling influence coefficient is set to... That is, the degradation state transition probability is not modulated by MPCI and always uses the basic transition probability. This is equivalent to not performing the multi-parameter coupling strength index calculation in step S1. The degradation layer only performs Markov transitions based on the degradation state at the previous moment, without considering the synergistic coupling effect of multiple environmental parameters. The execution flow and parameter settings of the remaining steps S2, S4, S5, and S6 are exactly the same as in Example 1.

[0146] Comparative Example 2: No physically constrained cause-effect graph was used; Comparative Example 2 used the same diagnostic objects and dataset as Example 1, the only difference being that: a set of prohibited edges was not constructed in step S2. and must set of edges , that is to say , Causal graph learning relies entirely on data-driven conditional independence test results to determine the graph structure, without imposing any domain physical constraints. The execution flow and parameter settings of the remaining steps S1, S3, S4, S5, and S6 are exactly the same as in Example 1.

[0147] Comparative Example 3: No adaptive sensor weighting was used; Comparative Example 3 used the same diagnostic objects, datasets, and network structures as Example 1, the only difference being that: in step S4, the mutual information estimation and normalized weight calculation in steps S41 and S42 were not performed, but instead, uniform weights were assigned to all sensors. The total number of sensors is represented by the weighted geometric average of the conditional likelihoods of each sensor, which is the overall observation likelihood degenerated into the weighted geometric average of the conditional likelihoods of each sensor. The execution flow and parameter settings of the remaining steps S1, S2, S3, S5, and S6 are exactly the same as in Example 1.

[0148] Comparative Example 4: No online incremental learning was used. Comparative Example 4 used the same diagnostic subjects, datasets, and network structures as Example 1, with the only difference being that: the online incremental learning in step S6 was not performed; the conditional probability parameters of the hierarchical dynamic Bayesian network were fixed after offline training and were no longer updated; the static parameters obtained from the initial training were used throughout the entire service process. The execution flow and parameter settings of the remaining steps S1, S2, S3, S4, and S5 were exactly the same as in Example 1.

[0149] Experiment Example 1: Comparison Experiment of Fault Diagnosis Accuracy; This experiment uses the methods of Examples 1, 2, and 3, and Comparative Examples 1, 2, 3, and 4 to evaluate fault diagnosis performance on their respective test datasets. The test dataset is the last 20% of the historical running data of each example / comparative example, extracted chronologically, and is not used in any training process.

[0150] The evaluation metric used is fault diagnosis accuracy, defined as the percentage of correctly diagnosed fault samples out of the total number of test samples. Each method runs the complete diagnostic process on the test set, and the matching of the fault type output by each diagnosis with the actual labeled fault type is statistically analyzed.

[0151] Experimental results are as follows Figure 1 As shown. The fault diagnosis accuracy of Example 1 is 93.7%, Example 2 is 94.2%, and Example 3 is 92.8%; Comparative Example 1 with MPCI coupling removed is 81.5%, Comparative Example 2 with physical constraints removed is 86.3%, Comparative Example 3 with adaptive weighting removed is 88.1%, and Comparative Example 4 with online learning removed is 85.6%.

[0152] from Figure 1 As can be seen, the fault diagnosis accuracy of the three embodiments is above 92%, significantly higher than that of the four comparative examples. Comparative example 1 has the lowest accuracy at 81.5%, indicating that after removing the multi-parameter coupling strength index, the degradation state transition probability cannot reflect the synergistic effect of environmental parameters, leading to a large deviation in degradation state estimation and severely affecting the diagnosis accuracy. Comparative example 2 has an accuracy of 86.3%, indicating that data-driven causal graph learning after removing physical constraints easily introduces false causal edges, reducing the quality of network inference. Comparative example 3 has an accuracy of 88.1%, indicating that uniform weights cannot effectively suppress the interference of low-quality sensor noise. Comparative example 4 has an accuracy of 85.6%, indicating that without online learning, the model parameters cannot adapt to the degradation mode drift during service, demonstrating that the technical means in steps S1 to S6 of this invention are all indispensable to the diagnosis accuracy.

[0153] Experimental Example 2: The effect of MPCI on the probability of degenerate state transition; This experimental example adopts the methods described in Example 1. Example 2 Example 3 Comparative Example 1 The method examines how the degradation state changes from 0 to 1.0 as the normalized MPCI value varies. Mild degeneration Transition probability of moderate degradation jump The changing pattern of the basic transition probability. According to the formula Calculate the transition probability for each MPCI value, where the normalization constant is... Ensure that the sum of the transition probabilities for the same row is 1.

[0154] Experimental results are as follows Figure 2 As shown. When MPCI=0, the transition probability of all four methods is equal to the base value of 0.15; as MPCI increases, the transition probabilities of Examples 1, 2, and 3 all show a monotonically increasing trend, with Example 2 showing the highest probability. The fastest growth was observed, with a transition probability of approximately 0.36 at MPCI=1.0; Example 1 Secondly, when MPCI=1.0, it is approximately 0.29; Example 3 The slowest growth was approximately 0.24 when MPCI=1.0; Comparative Example 1 The transition probability remains at a constant level of 0.15 and does not change with MPCI.

[0155] from Figure 2 It can be seen that the present invention introduces an exponential modulation mechanism of MPCI on the state transition probability of the degradation layer through step S32, which enables the probability of the degradation state to jump to a higher level to increase monotonically with the increase of the coupling strength of environmental parameters, which is in line with the physical law of accelerated degradation of marine equipment in a harsh environment with multi-parameter coupling. The larger the value, the more sensitive the transition probability is to MPCI, which is suitable for corrosion-sensitive equipment such as the wind power foundation structure in Example 2. When the value is small, the response is relatively smooth, which is suitable for equipment with a relatively slow degradation process, such as the subsea pipeline in Example 3. Comparative Example 1 does not respond to MPCI changes at all and cannot capture the physical effect of multi-parameter coupling on degradation acceleration, indicating that steps S1 and S3, which introduce MPCI into the degradation state transition probability, have clear physical significance and engineering effectiveness.

[0156] Experiment Example 3: Comparison of Missed Detection Rate and False Alarm Rate; This experiment uses the same method and test dataset as Experiment Example 1 to evaluate the Missed Detection Rate (MDR) and False Alarm Rate (FAR) of each method. The missed detection rate is defined as the proportion of samples that actually experience a fault but are not detected by the system; the false alarm rate is defined as the proportion of samples that receive a fault warning from the system but do not actually experience a fault.

[0157] Experimental results are as follows Figure 3As shown. The false negative rate of Example 1 was 3.2% and the false positive rate was 4.5%; the false negative rate of Example 2 was 2.8% and the false positive rate was 5.1%; the false negative rate of Example 3 was 3.6% and the false positive rate was 3.9%; the false negative rate of Comparative Example 1 was 12.4% and the false positive rate was 8.7%; the false negative rate of Comparative Example 2 was 8.9% and the false positive rate was 10.2%; the false negative rate of Comparative Example 3 was 7.5% and the false positive rate was 9.8%; and the false negative rate of Comparative Example 4 was 9.1% and the false positive rate was 7.3%.

[0158] from Figure 3 As can be seen, the false negative rate of all three embodiments is below 4%, and the false positive rate is below 6%, which is far superior to the comparative examples. Comparative Example 1 has a false negative rate as high as 12.4% because the degradation state estimation is too low without considering MPCI, causing the system to still judge it as a low degradation state when the actual degradation has reached a high level, thus missing a large number of fault events. Comparative Example 2 has a false positive rate as high as 10.2% because the introduction of false causal edges in the pure data-driven causal graph leads to too many erroneous fault associations in the network inference. Comparative Example 3 has both high false negative and false positive rates because uniform weighting makes noisy sensors contribute equally to high-quality sensors, reducing inference accuracy and failing to meet the stringent requirements for low false negative and low false positive rates in marine equipment safety monitoring.

[0159] Experiment Example 4: Online Incremental Learning Convergence Experiment; This experiment example adopts the methods described in Example 1. Example 2 and Example 3 In step S6 of the online incremental learning method, the trend of KL divergence after each parameter update is examined as labeled sample batches accumulate. KL divergence, as defined in step S63, measures the difference between the conditional probability distributions before and after the update. Comparative Example 4 does not perform online learning and does not generate KL divergence data, serving as a baseline. During the experiment, 20 batches of labeled samples are accumulated sequentially. Each batch triggers a Dirichlet parameter incremental update step S62, and the KL divergence is calculated immediately after the update.

[0160] Experimental results are as follows Figure 4 As shown. In Example 1, the KL divergence was 0.148 after the first batch update, and it monotonically decreased with the increase of batches, dropping to 0.009 by the 8th batch, and then stabilizing below the convergence threshold of 0.01; in Example 2, the KL divergence was 0.182 in the first batch, and it decreased due to the batch size. The smaller batch size and relatively richer information per batch resulted in a faster convergence speed than Example 1, decreasing to 0.008 by the 6th batch. Example 3 had a KL divergence of 0.123 in the 1st batch, which decreased to 0.007 by the 7th batch. All three examples achieved convergence within 10 batches, verifying the convergence of the Dirichlet conjugate incremental update mechanism in step S62.

[0161] from Figure 4 As can be seen, the online incremental learning method in step S6 of this invention utilizes the conjugate property of the Dirichlet and Categorical distributions. Each batch update only requires adding the pseudo-counts, resulting in extremely low computational cost. With the continuous accumulation of labeled samples, the network parameters gradually stabilize, and the KL divergence monotonically decreases below the convergence threshold, proving the numerical convergence of the online update. Batch size The smaller the value, the greater the proportion of new information carried per batch, and the fewer batches are required for convergence, but the variance of parameter estimation is slightly larger; The larger the value, the smoother the convergence, but the more batches are required. Users can configure it flexibly according to the equipment maintenance frequency. Comparative Example 4 does not perform online learning; its network parameters remain at the initial offline training values, making it unable to adapt to new fault modes and degradation mode drift that occur during service.

[0162] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for diagnosing equipment faults involving multi-parameter coupling in a marine environment, characterized in that, Includes the following steps: S1. Calculation of Multi-parameter Coupling Strength Index: Collect real-time observation values ​​of multiple environmental parameters during equipment operation, take the corresponding values ​​of each environmental parameter under the rated operating conditions of the equipment as the benchmark reference value, and calculate the multi-parameter coupling strength index MPCI based on the deviation of each environmental parameter from the benchmark reference value and the coupling strength coefficient between each pair of parameters. S2, Learning physical constraint cause-effect graphs; Pre-defined causal directions that are not allowed to exist constitute the forbidden edge set, and proven causal relationships constitute the necessary edge set; using historical equipment operation data as input, conditional independence tests are performed, and constraints are imposed on the graph structure by combining the forbidden edge set and the necessary edge set, outputting a directed acyclic graph; S3. Construction of hierarchical dynamic Bayesian network: Based on the directed acyclic graph obtained in step S2, construct a hierarchical dynamic Bayesian network containing an environment layer, a degradation layer, and a fault layer; introduce the MPCI calculated in step S1 into the state transition probability across time steps of the degradation layer, so that the probability of the degradation state jumping to a higher degradation level increases with the increase of MPCI. Offline parameter learning for hierarchical dynamic Bayesian networks; Step S3 specifically includes: S31. Establish a three-layer network node structure: based on a directed acyclic graph. All nodes are assigned to three layers: the environment layer nodes are the observable environmental parameters collected in step S1; the degradation layer nodes are the hidden variables of the equipment degradation state. ,Pick For normal For mild degeneration, For moderate degradation, For severe degradation, there are four discrete levels; the fault layer nodes are observable internal parameters of the equipment, including at least vibration acceleration and structural strain. S32. Introducing MPCI's setting of degradation layer state transition probabilities: Degradation states are determined by levels. Jump to a higher degradation level ( The transition probability of ) is calculated using the following formula: ； In the formula: Given the degenerate state at the previous time step: And currently Under certain conditions, the degenerate state jumps to... The transition probability; When MPCI modulation is not considered arrive The base transition probability is initialized from historical fault data; The coupling effect coefficient is determined by maximum likelihood estimation from historical fault data; Calculated for step S1 Time-varying multi-parameter coupling strength index; It is a natural exponential function; This is a normalization constant, its value equal to the sum of all transition probabilities in that row, ensuring that the sum of the transition probabilities of all rows is 1; the above amplification only applies to... That is, apply the probability of jumping to a higher degradation level, and then normalize the entire row; S33. Expectation-Maximization (EM) Algorithm Parameter Learning: Divide historical running data into training and validation sets, perform EM algorithm iterations on the training set, and calculate the latent variables of the degenerate state in the E-step. The posterior distribution is obtained, and all conditional probability parameters are updated in the M step based on the E step results. The two steps are alternately iterated until the log-likelihood increment between two adjacent iterations is less than the convergence threshold. The classification accuracy is evaluated with the validation set. After the validation is passed, the network parameters are fixed and the hierarchical dynamic Bayesian network with completed parameter learning is output. S4. Adaptive weighted fault reasoning: Calculate the mutual information between the observations of each sensor and the fault state, use the normalized mutual information as the information contribution weight of the corresponding sensor, substitute the weighted observation likelihood into the hierarchical dynamic Bayesian network for reasoning, and output the posterior probability of each fault type; when the highest posterior probability exceeds the preset threshold, output a fault warning. S5. Multi-fault concurrent decoupling identification: Screen candidate faults whose posterior probability exceeds the screening threshold. Under the condition of knowing the current degradation state, calculate the conditional non-independence metric of any two types of candidate faults. Based on the comparison result of the metric and the judgment threshold, the multi-fault scenario is judged as independent concurrent fault or compound fault, and the corresponding diagnostic conclusions are output respectively.

2. The method according to claim 1, characterized in that, Step S1 specifically includes: S11. Determine the environmental parameter vector and reference values: Obtain from the sensor system. Real-time measurements of several environmental parameters , Each parameter type should at least include temperature, salinity, corrosion current density, and structural stress amplitude; record the reference values ​​for each parameter of the equipment under rated operating conditions. The baseline reference value is determined through factory testing and stored in the diagnostic system before equipment deployment; S12. Determine the pairwise coupling strength coefficients: In the accelerated degradation test, apply single environmental parameter excitation and dual-parameter combined excitation to the equipment respectively, and measure the degradation rate under each working condition; for any two environmental parameters... and The coupling strength coefficient is calculated using the following formula. : ； In the formula: For parameters and The coupling strength coefficient; for and The degradation rate was measured while the excitation was applied simultaneously; To apply only The degradation rate measured during excitation; To apply only The degradation rate measured during excitation; This is the baseline degradation rate under rated operating conditions; when the calculated result is less than zero, let... ; This process is repeated for all parameter pairs to form the coupling strength coefficient matrix. ; S13. Calculate the multi-parameter coupling strength index: Calculate using the following formula. Moment : ； In the formula: for The multi-parameter coupling strength index at time t; This represents the total number of environmental parameters. For all satisfied Summing the parameters; The coupling strength coefficient obtained in step S12; For the first Environmental parameters in The measured value at that moment; For the first The baseline reference value for each environmental parameter; This is the absolute value operator; This indicates no coupling deviation effect. A larger value indicates a stronger synergistic effect of multiple environmental parameters on degradation.

3. The method according to claim 1, characterized in that, Step S2 specifically includes: S21. Constructing a set of physical constraints: Based on the known laws of marine engineering thermodynamics, electrochemistry, and structural mechanics, construct a set of forbidden edges. With the set of necessary edges Forbidden edge set It must contain at least all causal directions from internal equipment parameters to uncontrollable marine environmental parameters; it must be combined with other methods. It contains at least two types of causal relationships that have been experimentally verified: salinity and corrosion current density, and temperature and material strength degradation; S22. Conditional Independence Test to Construct the Skeleton: Using historical equipment operation data as a sample, perform a conditional independence test on all node pairs. Edges between node pairs with statistical dependencies are retained, otherwise they are deleted, thus forming the undirected network skeleton. S23. Apply physical constraints to determine edge directions: Delete the set of forbidden edges from the undirected skeleton obtained in S22. All edges appearing in the array will be required to form an edge set. Edges are forcibly added according to the specified causal direction. The direction of the remaining undirected edges is determined one by one according to the V-shaped structure criterion and acyclicity constraint. The output is a directed acyclic graph. This serves as the topology of the hierarchical dynamic Bayesian network in step S3.

4. The method according to claim 1, characterized in that, Step S4 specifically includes: S41. Sliding Window Mutual Information Estimation: Using a sliding time window of preset length, for the first... Observation data sequence of each sensor Fault status labels within the corresponding time period Perform joint frequency statistics and estimate using the mutual information formula. Output the mutual information of each sensor at the current moment. ; S42. Normalized mutual information weight calculation: Calculate the weight using the following formula. One sensor in Information contribution weight at any given moment : ； In the formula: For the first One sensor in Weight of information contribution at any given moment; The estimated first step in step S41 Mutual information between sensors; For all Summing the values ​​from all sensors; when all sensors... When both are 0, let ; S43, Weighted Belief Propagation Reasoning and Early Warning Output: Based on For the exponent with respect to the first Conditional observation likelihood of each sensor After performing power-weighted processing, the overall observation likelihood is obtained as follows: ； In the formula: For comprehensive observational likelihood; For the first One sensor in Conditional observational likelihood at time; The first result obtained in step S42 The information contribution weight of each sensor; To multiply all sensors together; Substitute the results into a hierarchical dynamic Bayesian network to perform belief propagation inference and output various fault types. The posterior probability; when the highest posterior probability exceeds a preset threshold. When a fault warning is triggered, the corresponding fault type and its confidence level are output.

5. The method according to claim 1, characterized in that, Step S5 specifically includes: S51. Candidate Concurrent Fault Screening: From the posterior probability vector output in step S4, screen for faults whose posterior probabilities exceed the screening threshold. The fault types constitute the candidate set, where If there is only one fault type in the candidate set, the single fault diagnosis conclusion will be output directly. S52. Calculation of Conditional Non-Independence Metric: For any two types of faults in the candidate set... and Given the current degradation state Under the given conditions, calculate the conditional non-independence measure using the following formula. : ； In the formula: For fault and In a known degenerate state A measure of conditional non-independence under given conditions; Known under conditions and The joint conditional probability of simultaneous occurrence; , They are known respectively under conditions , The conditional probability of each occurring; This is the absolute value operator; where By including degenerate state nodes in a Bayesian network After instantiation, the data is obtained using a connection tree algorithm or Gibbs sampling. S53, Fault Component Output: When Less than the judgment threshold At that time, the judgment and For independent concurrent faults, each fault outputs its own diagnostic conclusion and confidence level independently; when Not less than When the fault is determined to be a compound fault, a non-negative least squares optimization model is established: Let Let matrix be the feature vector extracted from the sensor signal at the current time. Each column is a standard sensor response template vector corresponding to a single fault. Given the fault component coefficient vector to be solved, solve... ,constraint , with coefficient vector The magnitude of each element serves as the diagnostic confidence level for each fault component. The output lists the components of the composite fault and the confidence level of each component.

6. The method according to claim 1, characterized in that, It also includes step S6 of online incremental learning: accumulating confirmed labeled fault samples; when the accumulated number reaches the preset batch size, updating the conditional probability parameters of the hierarchical dynamic Bayesian network using a Bayesian conjugate incremental approach, and monitoring the change in parameter distribution before and after two adjacent updates to determine whether the current batch update has converged; specifically including: S61. Accumulation of Labeled Samples and Batch Triggering: Continuously receives labeled samples whose fault type has been confirmed on-site by maintenance personnel, and temporarily stores them in a sample buffer; when the number of samples in the buffer reaches the preset batch size... Parameter updates are triggered on demand; when a new fault type appears in the sample buffer that has never been seen in the training set, the parameters are reduced. To accelerate initial adaptation; S62, Dirichlet parameter incremental update: for each type of fault in the network. The corresponding conditional probability parameters are incrementally updated using the following formula: ； In the formula: For the updated version Positive integer pseudo-counts of the Dirichlet distribution corresponding to the fault class; The positive integer pseudo-count of the Dirichlet distribution before the update; This is the first labeled sample in this batch. The actual number of times each type of fault occurred; during system initialization. Taking 1 corresponds to no prior information; S63, KL divergence convergence monitoring and write-back: After each parameter update, calculate the KL divergence between the conditional probability distributions before and after the update; when the KL divergence is less than the convergence threshold, determine that the update of this batch has converged, write the updated conditional probability parameters back to the hierarchical dynamic Bayesian network, clear the sample buffer of the current batch, and enter the next batch to wait.

Images