Marine autonomous surface ship power system reliability real-time evaluation method based on diagnosis input

Abstract

The invention discloses a real-time reliability evaluation method for a marine autonomous surface ship power system based on diagnosis input, and the method comprises the steps: obtaining the operation state data of one or more parts, and obtaining a processed data sequence; obtaining a pre-trained fault diagnosis model, and inputting the processed data sequence into the pre-trained fault diagnosis model to obtain a fault diagnosis result; and constructing a dynamic Bayesian network, performing posterior reasoning through the dynamic Bayesian network in combination with the fault diagnosis result, and outputting a reliability evaluation result of the marine autonomous surface ship power system. According to the method, causal reasoning is carried out by using a component-level fault diagnosis result and combining a dynamic Bayesian network, so that local fault information can dynamically influence system-level prediction, the gap between component-level diagnosis and global reliability estimation is bridged, the problem that system-level quantitative analysis cannot be carried out due to system-level fault data missing is solved, and the reliability of the system is improved. And system-level real-time reliability evaluation can be achieved, and safe operation of the autonomous ship is guaranteed.

Description

Technical Field

[0001] This invention relates to the fields of marine engineering and intelligent operation and maintenance technology, and in particular to a real-time reliability assessment method for autonomous surface vessel propulsion systems based on diagnostic inputs. Background Technology

[0002] In the new development pattern of the shipping industry, based on digitalization and aiming at autonomy, autonomous shipping has become one of the most prominent trends in the current shipping industry. Among them, autonomous vessels are a key element and carrier, and the new autonomous vessel control and operation mode has stimulated new safety requirements, necessitating timely reliability assessments to ensure the safety of life at sea, as well as the safety of cargo and the vessel itself.

[0003] As a crucial system for ensuring the safe operation of autonomous vessels, a continuously reliable propulsion system is of paramount importance for guaranteeing their navigational safety. The propulsion system of autonomous vessels is complex and susceptible to various uncertainties. Current research methods can be broadly categorized into two types: the first employs traditional reliability modeling methods to model the propulsion system and assess its safety; the second focuses on fault diagnosis and health monitoring based on big data analysis. Current research on the reliability of autonomous surface vessels (ASMEs) largely focuses on hull structural integrity, collision avoidance algorithms, and remote control systems. However, the propulsion system, as a crucial subsystem ensuring vessel operation during navigation, plays a vital role in reducing unexpected downtime and accidents caused by equipment failure. In practice, insufficient attention is paid to real-time reliability assessment of propulsion systems, and traditional methods struggle to directly quantify their reliability in real time, resulting in weak fault warning capabilities in unmanned scenarios. Furthermore, the slow accumulation of fault data in ASME propulsion systems, coupled with the tendency for algorithm-generated data to deviate from real-world conditions, leads to low reliability of assessment results, hindering system-level reliability evaluation and fault diagnosis. Summary of the Invention This invention provides a real-time reliability assessment method for autonomous surface vessel propulsion systems based on diagnostic inputs, in order to overcome the aforementioned technical problems.

[0004] To achieve the above objectives, the technical solution of the present invention is as follows: A real-time reliability assessment method for autonomous surface vessel propulsion systems based on diagnostic inputs, comprising the following steps: S1. Acquire the operating status data of one or more components in the autonomous surface vessel power system, preprocess the operating status data, and obtain the processed data sequence; S2. Obtain the pre-trained fault diagnosis model and input the processed data sequence into the pre-trained fault diagnosis model to obtain the fault diagnosis result; The pre-trained fault diagnosis model includes a Transformer model, a bidirectional gated recurrent unit, and an output layer. The Transformer model is used to capture the global correlation of all positions in the processed data sequence, generate a feature representation containing global information, and transmit it to the bidirectional gated loop unit. The bidirectional gated loop unit obtains bidirectional temporal dependency information based on the feature representation containing global information; The output layer outputs fault diagnosis results based on the dual inverse timing dependency information; S3. Construct a dynamic Bayesian network, which is used to simulate the time dependence of each component in the autonomous surface vessel propulsion system and the structural relationship between the components. By using the dynamic Bayesian network and combining it with the fault diagnosis results, posterior reasoning is performed to output the reliability assessment results of the autonomous surface vessel propulsion system.

[0005] Furthermore, the specific steps for constructing a dynamic Bayesian network include: Acquire failure rate data to characterize the degradation and failure behavior of powertrain components; A Markov transition probability matrix is ​​constructed based on the failure rate data to describe the probability of a component transitioning from a healthy state to a faulty state within discrete time intervals. The prior failure probability of a component is obtained based on expert experience; The nodes of the dynamic Bayesian network are defined as root nodes and non-root nodes. The prior fault probability and the Markov transition probability matrix are used as the node parameters of the root node. The root node is used to determine the probability of each state of the current root node based on the prior fault probability and the Markov transition probability matrix, and transmits it to the non-root nodes. The non-root nodes are used to determine the probability of each state of the current non-root node based on the data transmitted by the root node. The structure of the dynamic Bayesian network is determined by setting a conditional probability table, and the dynamic Bayesian network is finally obtained.

[0006] Furthermore, the specific steps for performing posterior inference using the dynamic Bayesian network and the bearing fault diagnosis results include: The fault diagnosis results are input into the corresponding nodes of the dynamic Bayesian network; We use dynamic Bayesian networks to perform posterior inference to update the state distribution of nodes without input data and calculate the posterior failure probability of each node. The reliability index of the autonomous surface vessel propulsion system is calculated based on the posterior failure probability and is expressed as:

[0007] In the formula: For the reliability function of the autonomous surface vessel propulsion system; The posterior fault probability is obtained based on the posterior fault probability function. Furthermore, the specific steps for preprocessing the aforementioned operational status data include: The running status data is segmented according to the sampling time, and several segments of running status data are normalized. The normalized operating state data is decomposed using the variational mode decomposition method to obtain the processed data sequence.

[0008] In a specific embodiment, the steps for decomposing the operating state data using the variational mode decomposition method to obtain the decomposed data sequence include: 1) Define a constrained variational problem, expressed as:

[0009] in, Indicates the first k One modal function, K This indicates the total number of preset modes. k K , Represents the impulse basis function. Indicates time t Find the partial derivative. f ( t ) represents the running status data. For Hilbert transform kernel; As a penalty factor; 2) Introduce the Lagrange multiplication operator β The constrained variational problem is transformed into an unconstrained variational problem, expressed as:

[0010] in,{ u k}and{ ω k} represent the set of all modes and their corresponding set of center frequencies, respectively; ω k Indicates the first k The center frequency of each mode; 3) Initialization u k , ω k and β Set the number of iterations m =0; 4) Use the following formula to... u k , ω k and β Perform iterative updates, using the following expression:

[0011] In the formula, and They are respectively for and The form after Fourier transform. It is noise tolerance; It is a constant used to represent the precision of discrimination, and ; 5) Determine if the convergence criterion is met. If not, continue iterating and updating as in step 4). If met, stop iterating and output multiple IMF component signals, which is the decomposed data sequence. The convergence criterion is expressed as:

[0012] Furthermore, the specific steps for constructing the Markov transition probability matrix include: Assume that the homogeneous Markov process satisfies the following equation:

[0013] In the formula: for t The state transition probability at time t; For the first i , j One state; State variables of autonomous surface vessel propulsion systems are defined based on homogeneous Markov processes. and state set E , , n The number of states of a component; Based on state set E Draw a state transition diagram and calculate the state transition matrix based on the diagram. The formula for calculating the state transition matrix is ​​as follows:

[0014] In the formula: This is the state transition matrix; for i State j The transition rate of state transitions; Indicates from state E n The transition rate to state E1; Indicates the overall failure rate of the component; Assuming the state transition probability of a component at a given time follows an exponential distribution, the formula for calculating the state transition probability is:

[0015] The Markov state transition probability matrix is ​​calculated based on the state transition matrix and the state transition probabilities, and is expressed as follows:

[0016] in, To set a time.

[0017] Beneficial Effects: This invention constructs a dynamic Bayesian network and uses it to simulate the time-dependent relationships of each component in the autonomous surface vessel (ASME) propulsion system, as well as the structural relationships between components. Fault diagnosis results obtained through a pre-trained fault diagnosis model are input into the dynamic Bayesian network. Based on these results, the network performs posterior inference and ultimately outputs a reliability assessment result for the ASME propulsion system. This invention uses component-level fault diagnosis results combined with causal inference via a dynamic Bayesian network, enabling local fault information to dynamically influence system-level predictions. This bridges the gap between component-level diagnosis and global reliability estimation, solves the problem of insufficient system-level fault data for quantitative analysis, and achieves real-time system-level reliability assessment. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of the real-time reliability assessment method for autonomous surface vessel propulsion systems based on diagnostic input in this invention; Figure 2 This is a schematic diagram of the dynamic Bayesian network in an embodiment of the present invention; Figure 3 This is a Markov state transition diagram of a typical unrepairable system in an embodiment of the present invention; Figure 4 This is a schematic diagram of the Transformer model in an embodiment of the present invention; Figure 5 This is a schematic diagram of the bidirectional gated loop unit in an embodiment of the present invention; Figure 6This is a connection diagram of the parallel hybrid power system in an embodiment of the present invention; Figure 7 This is an example diagram illustrating the dynamic failure of a component in an embodiment of the present invention; Figure 8 This is a schematic diagram of the dynamic Bayesian network of the diesel-electric hybrid system in an embodiment of the present invention; Figure 9 These are initial vibration signal diagrams for ten operating conditions in embodiments of the present invention; Figure 10 This is a diagram showing the VMD signal decomposition results in an embodiment of the present invention; Figure 11 This is a graph showing the accuracy and loss curves of the training and validation sets in this embodiment of the invention. Figure 12 This is a fault diagnosis confusion matrix diagram in an embodiment of the present invention; Figure 13 This is a diagram showing the fault diagnosis results in an embodiment of the present invention; Figure 14 This is a t-SNE visualization based on the original data test set in an embodiment of the present invention; Figure 15 This is a visualization of t-SNE based on extracted features in an embodiment of the present invention; Figure 16 This is a schematic diagram of system-level reliability curves in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] This embodiment provides a real-time reliability assessment method for the propulsion system of autonomous surface vessels based on diagnostic input, such as... Figure 1 As shown, the specific steps include: S1. Acquire the operating status data of one or more components in the autonomous surface vessel power system, preprocess the operating status data, and obtain the processed data sequence; S2. Obtain the pre-trained fault diagnosis model and input the processed data sequence into the pre-trained fault diagnosis model to obtain the fault diagnosis result; The pre-trained fault diagnosis model includes a Transformer model, a bidirectional gated recurrent unit (BiGRU), and an output layer; The Transformer model is used to capture the global correlation of all positions in the processed data sequence, generate a feature representation containing global information, and transmit it to the bidirectional gated loop unit. The bidirectional gated recurrent unit obtains bidirectional temporal dependency information based on the feature representation containing global information. The network structure of the bidirectional gated recurrent unit is as follows: Figure 5 As shown; The output layer outputs fault diagnosis results based on the dual inverse timing dependency information; Specifically, the Transformer model is a deep learning model based on a multi-head self-attention mechanism, mainly composed of an encoder and a decoder, and its structure is as follows: Figure 4 As shown. The encoder-decoder structure of the Transformer model can capture long-term dependencies in a sequence, resulting in higher accuracy and performance in time series prediction. The Transformer model can weight the importance of each element in the input data sequence, learning the relative importance of each element in the sequence relative to other elements. The calculation formula is as follows:

[0022] In the formula, Q The problem matrix, K The key matrix, V These are value matrices, and all three matrices are obtained by performing linear transformations on the input. d k For matrix Q , K The vector dimension.

[0023] Specifically, the Transformer model, based on a multi-head attention mechanism, decomposes a single attention into several matrices that yield multiple sets of attention weights. An additional weight matrix is ​​then introduced, concatenating these matrices and performing a linear transformation to obtain the final representation. Through this multi-head attention mechanism, each attention head can acquire local features of the data sequence, helping the model capture richer information and enhancing its expressive power. Simultaneously, it can reduce the impact of inaccurate attention weights at certain locations on the model's prediction accuracy, enhancing its robustness. The calculation formula is as follows:

[0024] In the formula, W O This is the weight matrix. , , These are the trainable parameter matrices; S3. Construct a dynamic Bayesian network (DBN), which is used to simulate the time dependence of each component in the autonomous surface vessel propulsion system and the structural relationship between the components; perform posterior reasoning through the dynamic Bayesian network and the fault diagnosis results, and then output the reliability assessment results of the autonomous surface vessel propulsion system.

[0025] Specifically, this embodiment utilizes a fault diagnosis model to perform real-time fault diagnosis on the power system and its equipment, and inputs the fault diagnosis results into a dynamic Bayesian network. Upon receiving the diagnosis results, the dynamic Bayesian model updates the reliability status in real time, providing real-time decision-making basis for autonomous vessel scheduling, degradation control, and maintenance decisions, ensuring the safe operation of autonomous vessels. Simultaneously, posterior inference is performed using the dynamic Bayesian network to reshape the probability distribution of relevant components. Through this diagnosis-based causal inference mechanism, local fault information can dynamically influence system-level predictions, bridging the gap between component-level diagnosis and global reliability estimation. The method proposed in this embodiment not only advances the theoretical integration between fault diagnosis models and causal inference but also provides a practical tool for intelligent health monitoring of autonomous maritime systems. Specifically, this embodiment uses VMD-Transformer-BiGRU to perform fault diagnosis based on real-time observation data. The diagnosis result, i.e., the current state of the equipment, is used as evidence in a dynamic Bayesian network constructed based on failure rate data to calculate the system's reliability index. As time changes, the observation data also changes, and the fault diagnosis results change accordingly. The fault probability obtained after inputting into the dynamic Bayesian network can be updated in real time, thereby achieving real-time reliability assessment.

[0026] In a specific embodiment, the specific steps for preprocessing the operating status data include: The running status data is segmented according to the sampling time, and several segments of running status data are normalized. The normalized running state data is decomposed using the Variational Mode Decomposition (VMD) method to obtain the processed data sequence.

[0027] Specifically, Variational Mode Decomposition (VMD) is an adaptive signal decomposition method used to extract multi-scale features from fault signals, decomposing the signal into a series of sub-signals (modes) with specific sparsity, where these modes have limited bandwidth in the frequency domain. In a specific embodiment, the specific steps for decomposing the operating state data using the variational mode decomposition method to obtain the decomposed data sequence include: 1) Define a constrained variational problem, expressed as:

[0028] in, Indicates the first k Individual Modal Functions (IMFs) K This indicates the total number of preset modes. k K , Represents the impulse basis function. Indicates time t Find the partial derivative. f ( t ) represents the running status data. For Hilbert transform kernel; As a penalty factor; 2) Introduce the Lagrange multiplication operator β The constrained variational problem is transformed into an unconstrained variational problem, expressed as:

[0029] in,{ u k}and{ ω k} represent the set of all modes and their corresponding set of center frequencies, respectively; ω k Indicates the first k The center frequency of each mode; 3) Initialization u k , ω k and β Set the number of iterations m =0; 4) Use the following formula to... u k , ω k and β Perform iterative updates, using the following expression:

[0030] In the formula, and They are respectively for and The form after Fourier transform. It is noise tolerance; It is a constant used to represent the precision of discrimination, and ; 5) Determine if the convergence criterion is met. If not, continue iterating and updating as in step 4). If met, stop iterating and output multiple IMF component signals, which is the decomposed data sequence. The convergence criterion is expressed as:

[0031] Specifically, Bayesian networks (BNs) are theoretical models based on Bayesian methods. They have been extensively studied and are considered one of the most effective models for expressing and reasoning about uncertainty, widely used in system reliability modeling, reasoning, and diagnosis. These networks are based on a graphical model where nodes represent random variables and directed edges represent probabilistic causal relationships between events. Bayesian networks offer significant advantages in reliability modeling. They leverage conditional independence to reduce the difficulty of information collection and solution, possess polymorphic representation capabilities to effectively express multiple node states and the strength of causal relationships, and support bidirectional reasoning for fault assessment and prediction, identifying system weaknesses. For static BNs, the first step is to determine the system's variables and their dependencies, i.e., the conditional probability tables. The network structure consists of nodes and directed edges; nodes represent variables, and directed edges represent dependencies between variables. The conditional probability tables for fault influencing factor nodes and fault symptom nodes are crucial for establishing a static Bayesian network model. Each node must be associated with a conditional probability table, which is based on the probability of previous information and past experience. The conditional probability of the root node is determined by real-world data, which is the probability of the node being in each state. The conditional probability of a child node is the probability of the node being in each state, which is the probability of different levels of failure, considering all the conditions that affect it.

[0032] Specifically, the DBN in this embodiment is a probability distribution model that integrates the BN structure and time series principles. DBN comprises two main parts: a statically structured BN and a time model used to describe dynamic changes. DBN forms a probability distribution model for processing information by adding a time dimension to the static BN.

[0033] Specifically, based on A typical DBN can be defined, such as Figure 2 As shown, For the initial Bayesian network, it defines the prior distribution of BN. ; For a transition network, let represent the transition probability between every two time slices. The state transition probabilities between two time slices are expressed as follows:

[0034] in, express i The conditional probability of a node, where N represents the number of nodes.

[0035] In a specific embodiment, considering the various fault modes and functional dependencies existing in the hybrid power system, a dynamic Bayesian network is constructed to capture the causal relationships and time-dependent transitions between key components and the system. The specific steps for constructing the dynamic Bayesian network include: Acquire failure rate data to characterize the degradation and failure behavior of powertrain components; A Markov transition probability matrix is ​​constructed based on the failure rate data to describe the probability of a component transitioning from a healthy state to a faulty state within discrete time intervals. The prior failure probability of a component is obtained based on expert experience; The nodes of the dynamic Bayesian network are defined as root nodes and non-root nodes. The prior fault probability and the Markov transition probability matrix are used as the node parameters of the root node. The root node is used to determine the probability of each state of the current root node based on the prior fault probability and the Markov transition probability matrix, and transmits it to the non-root nodes. The non-root nodes are used to determine the probability of each state of the current non-root node based on the data transmitted by the root node. Specifically, this embodiment uses the Markov transition probability matrix as one of the node parameters of the root node to characterize the change of the Markov transition probability over time, such as the increased probability of failure due to wear and tear.

[0036] The structure of the dynamic Bayesian network is determined by setting a conditional probability table, and the dynamic Bayesian network is finally obtained.

[0037] Specifically, this embodiment combines Markov chains to construct a dynamic Bayesian network for an autonomous marine propulsion system, enabling dynamic assessment of the failure probability and reliability of the propulsion system and its equipment. In this specific embodiment, the steps for constructing the Markov transition probability matrix include: Assume that the homogeneous Markov process satisfies the following equation:

[0038] In the formula: for t The state transition probability at time t; For the first i , j One state; State variables of autonomous surface vessel propulsion systems are defined based on homogeneous Markov processes. and state set E , , n The number of states of a component; Based on state set ETo draw a state transition diagram, the defined states must be sufficient to distinguish the different states of the system. The state transition matrix is ​​then calculated based on the state transition diagram. The formula for solving the state transition matrix is ​​as follows:

[0039] In the formula: This is the state transition matrix; for i State j The transition rate of state transitions; Indicates from state E n The transition rate to state E1; Indicates the overall failure rate of the component; Assuming the state transition probability of a component at a given time follows an exponential distribution, the formula for calculating the state transition probability is:

[0040] The Markov state transition probability matrix is ​​calculated based on the state transition matrix and the state transition probabilities, and is expressed as follows:

[0041] in, To set a time.

[0042] In a specific embodiment, the specific steps of using the dynamic Bayesian network and combining it with the bearing fault diagnosis results for posterior inference include: The fault diagnosis results are input into the corresponding nodes of the dynamic Bayesian network; Dynamic Bayesian networks perform posterior inference to update the state distribution of nodes without input data and calculate the posterior failure probability of nodes; The reliability index of the autonomous surface vessel propulsion system is calculated based on the posterior failure probability and is expressed as:

[0043] In the formula: For the reliability function of the autonomous surface vessel propulsion system; The posterior fault probability is obtained based on the posterior fault probability function. Specifically, in practice, a dynamically updated reliability curve can be plotted based on the reliability index to more intuitively reflect the latest diagnostic information and the probability changes over time.

[0044] To validate the proposed method, this embodiment presents a case study of a representative autonomous surface vessel propulsion system. Specifically, diesel-electric hybrid power systems are increasingly used in shipbuilding due to their advantages such as operational flexibility, fuel efficiency, and reduced emissions. In particular, hybrid power systems typically involve multiple subsystems, such as diesel generators, electric motors, power converters, and propulsion shafts, which are tightly coupled both functionally and temporally. These characteristics make hybrid systems an ideal test platform for evaluating the effectiveness of causal models of dynamic diagnostic information, as they require both time-dependent reasoning and the ability to incorporate local fault observation. Figure 6 This is a simplified diagram of a parallel hybrid power system.

[0045] Due to dataset limitations, this embodiment combines publicly available diagnostic datasets with data from literature and related reports to verify the feasibility of the proposed method. Faulty components and their failure rates are shown in Table 1. Table 1 Component Failure Rate

[0046] This embodiment assumes that each basic component of the autonomous surface vessel's propulsion system exists in four states: normal operation (E1), partial failure (E2), severe failure (E3), and complete failure (E4). That is, the root node of the DBN has four states: normal operation, partial failure, severe failure, and complete failure, and its state transition rate follows a Markov process. Maintenance transforms a faulty state into a normal state. E 1. Its state transition diagram is as follows: Figure 3 As shown; E1 represents a completely healthy state, E2 represents a slight decrease in functional output due to wear or fatigue, E3 represents a significant decrease in functional output, and E4 represents a state where no functional output can be provided. Other non-root nodes of DBN only have two states: normal operation and complete failure. Compared to most two states (normal and fault), this embodiment adds two intermediate states, which can more accurately measure the component status.

[0047] Specifically, due to Since this information is difficult to obtain from the data, this embodiment obtains it through an assumption. Taking a four-state transition as an example, such as Figure 7 As shown, - The failure rate of a component transitioning from one state to another is determined by the following formula:

[0048] Specifically, taking an oil pump as an example, when the same power is provided to the oil pump, E1 indicates that the entire process is operating at the rated efficiency, E2 indicates a slight decrease in efficiency (such as reduced pressure, resulting in a slight reduction in oil intake or discharge, and a reduction in the amount of fuel supplied), E3 indicates a significant decrease in efficiency (such as a significant reduction in pressure, resulting in a significant reduction in oil intake or discharge, and a significant reduction in the amount of fuel supplied), and E4 indicates that the oil pump cannot intake or discharge oil.

[0049] Due to the autonomous nature of autonomous vessels, this embodiment only considers the failure rate of components and ignores their maintenance rate. According to the reliability principle, the failure rate distribution of each component in the entire system is an exponential distribution, and the formula for calculating the state transition probability is shown in Table 2; Table 2 State transition probabilities of a certain component

[0050] In this embodiment, each time slice of system state change is set to 1 day. A cross-time slice conditional probability table is constructed for each node. The failure rate between each state is substituted into Table 2 to calculate the state transition probability of each root node, as shown in Table 3. Table 3. Root Node Time Slice State Transition Probabilities

[0051]

[0052] In real-world maritime operations, conducting thorough system-level checks before each voyage is impractical. Therefore, this embodiment allows components to enter the inference window (dynamic Bayesian network) in a fault state (E2-E4), improving the universality of this method for autonomous vessels. The prior probability distribution reflects the likelihood of potential or undetected faults; the prior state distribution of the root node in this embodiment is shown in Table 4. Table 4 Prior probabilities of root nodes in dynamic Bayesian networks

[0053] Specifically, the key to the DBN model structure lies in determining the conditional probability table. This embodiment uses AND and OR gates to determine the logic structure. The nodes using OR gates are only the main / standby fuel pump and generator; if the main fuel pump / generator fails, the standby fuel pump / generator is used. Therefore, the conditional probability table has special characteristics. Examples of CPTs for AND and OR gates are shown in Tables 5 and 6, respectively. Table 5. Examples of OR gate CPT (Y indicates that the parent node is in this state, N indicates that the parent node is not in this state)

[0054] Table 6. Examples of AND gate CPT

[0055] The dynamic Bayesian structure established in this embodiment is as follows: Figure 8 As shown, intermediate events include bearings, diesel engine machinery, fuel system, fuel pump, diesel engine system, electromagnetic circuitry, power generation system, electric system, and hybrid power system.

[0056] In this embodiment, a bearing dataset from Case Western Reserve University (CWRU) is used for fault diagnosis. The bearing states include rolling element, inner ring, outer ring faults, and normal operation states, with a fault sampling frequency of 12 kHz. Using nine different fault types and ten normal operation states from the drive-end sampling dataset as the research object, the feasibility of applying intelligent diagnostic algorithms to vibration in autonomous ship electric propulsion or diesel machinery is verified. Figure 9 As shown, vibration signals for ten operating conditions were acquired. To ensure that all categories had the same number of sampling points, the original dataset was truncated, retaining 119,808 sampling points for each category. Furthermore, considering the sample size and fault cycle, the signal length was set to 1024, and the sliding window overlap rate was 0.5. The samples were then segmented to obtain a 2330*1025 sample dataset, with the last column representing the classification label. The training, validation, and test sets were then divided in a 7:2:1 ratio, as shown in Table 8. To improve signal quality and reduce noise sensitivity, VMD was used to decompose the original signal. While the VMD method is adaptive in signal decomposition, the total number of modes K affects the accuracy of the decomposition results. An excessively large K value may cause mode aliasing, leading to over-decomposition of the signal. In this embodiment, the center frequency method was used to determine the K value. The original sequence was decomposed starting from K=1, and the K value was continuously increased to obtain the center frequency of each mode corresponding to different K values. Observe the center frequency of the last mode corresponding to K. When two adjacent center frequencies are close, the signal is considered to be over-decomposed, and the previous K value is the optimal K value. This embodiment uses one of the signals under normal operating conditions as an example, and the center frequency calculation results are shown in Table 7; Table 7 Center Frequency Results

[0057] As shown in the table, the optimal K value is 4, and the decomposition results are as follows: Figure 10 As shown.

[0058] Table 8 Dataset Size

[0059] Specifically, the processed data sequence to be input into the fault diagnosis model is first reshaped into shape fragments (batch_size, 64, 64) to be compatible with the Transformer encoder. The Transformer module consists of two encoding layers, each with 4 attention heads and 128 hidden dimensions. The BiGRU stack contains two layers with 32 and 64 hidden units respectively. Each layer operates in bidirectional mode. The output of the last BiGRU layer at the last time step is passed to the fully connected Softmax layer to generate a 10-dimensional probability vector corresponding to 10 predefined fault categories.

[0060] Specifically, in this embodiment, the batch size used during training is 32, the learning rate is 0.0003, and the training lasts for 50 epochs. The cross-entropy loss function is used, and the Adam optimizer is employed to optimize the model parameters. The training process includes training and validation loops, recording accuracy and loss at each epoch. The model with the highest validation accuracy is saved as the final diagnostic model. Figure 11 The training dynamics at different stages are illustrated, showing the trends in training / validation accuracy and loss. The final fault diagnosis model achieved near 100% maximum validation accuracy. After 20 epochs, the accuracy on the training set exceeded 99%, and the loss curve showed no overfitting or underfitting, indicating that the fault diagnosis model performed exceptionally well on this data. This embodiment uses the confusion matrix to detect the diagnostic performance of the trained model on the test set, such as... Figure 13 As shown in the figure, C1-C10 represent different operating conditions. The columns represent the actual operating conditions of the fault, and the rows represent the predictions of the fault diagnosis model. The figure shows that the accuracy on the test set is close to 100%, accurately diagnosing the category of the fault signal. The predicted fault labels are compared with the true labels in the test set, as shown... Figure 13 As shown, the high consistency between the predicted and actual labels indicates that the model has strong generalization ability in sequential diagnosis scenarios and can adapt to time-series fault detection tasks. These results demonstrate that the VMD-Transformer-BiGRU model proposed in this embodiment can accurately and consistently classify fault types under different conditions. The predicted fault labels are then mapped to the corresponding component state categories and input into a dynamic Bayesian network, which helps in real-time reliability inference.

[0061] Furthermore, to visualize the effectiveness of the feature representations learned by the model, this embodiment applies t-distributed random neighborhood embedding (t-SNE) to both the original input features and the final hidden features extracted by the trained model. For example... Figure 14-15As shown, there is significant overlap between categories in the original data, while the extracted features form well-separated clusters, indicating that the model has a strong category recognition capability.

[0062] Specifically, after the fault diagnosis model completes fault classification, it maps the identified fault types to the corresponding health states of bearing sub-components in the dynamic Bayesian network. The total time is set according to the time required to complete one voyage (30 days). Specifically, in the first 10 days, the inner ring, outer ring, and rolling elements are all diagnosed as being in state E1. In the 10th day, the inner ring and rolling elements remain in state E1, while the outer ring is in state E2. The diagnostic results of the first 10 days are injected into the DBN as observed evidence to predict reliability changes after 10 days, and posterior inference is performed to obtain system-level reliability, such as... Figure 16 As shown, after setting the DBN evidence on day 10, the reliability of the diesel engine machinery and diesel engine system, which are directly related to the bearings, changed significantly, decreasing from 0.975 and 0.979 to 0.730 and 0.733, respectively. This decrease verifies the effectiveness of the method proposed in this embodiment in capturing fault propagation dynamics. On day 30, the overall reliability of the hybrid propulsion system will decrease to 0.986, providing a reference for proactively scheduling maintenance or making informed mission abort decisions after completing the navigation mission.

[0063] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A real-time reliability assessment method for the propulsion system of autonomous surface vessels based on diagnostic input, characterized in that, The specific steps include: S1. Acquire the operating status data of one or more components in the autonomous surface vessel power system, preprocess the operating status data, and obtain the processed data sequence; S2. Obtain the pre-trained fault diagnosis model and input the processed data sequence into the pre-trained fault diagnosis model to obtain the fault diagnosis result; The pre-trained fault diagnosis model includes a Transformer model, a bidirectional gated recurrent unit, and an output layer. The Transformer model is used to capture the global correlation of all positions in the processed data sequence, generate a feature representation containing global information, and transmit it to the bidirectional gated loop unit. The bidirectional gated loop unit obtains bidirectional temporal dependency information based on the feature representation containing global information; The output layer outputs fault diagnosis results based on the dual inverse timing dependency information; S3. Construct a dynamic Bayesian network, which is used to simulate the time dependence of each component in the autonomous surface vessel propulsion system and the structural relationship between the components. By using the dynamic Bayesian network and combining it with the fault diagnosis results, posterior reasoning is performed to output the reliability assessment results of the autonomous surface vessel propulsion system.

2. The real-time reliability assessment method for autonomous surface vessel propulsion systems based on diagnostic input as described in claim 1, characterized in that, The specific steps for constructing a dynamic Bayesian network include: Acquire failure rate data to characterize the degradation and failure behavior of powertrain components; A Markov transition probability matrix is ​​constructed based on the failure rate data to describe the probability of a component transitioning from a healthy state to a faulty state within discrete time intervals. The prior failure probability of a component is obtained based on expert experience; The nodes of the dynamic Bayesian network are defined as root nodes and non-root nodes. The prior fault probability and the Markov transition probability matrix are used as the node parameters of the root node. The root node is used to determine the probability of each state of the current root node based on the prior fault probability and the Markov transition probability matrix, and transmits it to the non-root nodes. The non-root nodes are used to determine the probability of each state of the current non-root node based on the data transmitted by the root node. The structure of the dynamic Bayesian network is determined by setting a conditional probability table, and the dynamic Bayesian network is finally obtained.

3. The real-time reliability assessment method for autonomous surface vessel propulsion systems based on diagnostic input as described in claim 2, characterized in that, The specific steps for performing posterior inference using the dynamic Bayesian network and the bearing fault diagnosis results include: The fault diagnosis results are input into the corresponding nodes of the dynamic Bayesian network; We use dynamic Bayesian networks to perform posterior inference to update the state distribution of nodes without input data and calculate the posterior failure probability of each node. The reliability index of the autonomous surface vessel propulsion system is calculated based on the posterior failure probability and is expressed as: In the formula: For the reliability function of the autonomous surface vessel propulsion system; The posterior fault probability is obtained based on the posterior fault probability function.

4. The real-time reliability assessment method for autonomous surface vessel propulsion systems based on diagnostic input as described in claim 1, characterized in that, The specific steps for preprocessing the aforementioned operational status data include: The running status data is segmented according to the sampling time, and several segments of running status data are normalized. The normalized operating state data is decomposed using the variational mode decomposition method to obtain the processed data sequence.

5. The real-time reliability assessment method for autonomous surface vessel propulsion systems based on diagnostic input as described in claim 4, characterized in that, The specific steps for decomposing the operating state data using the variational mode decomposition method to obtain the decomposed data sequence include: 1) Define a constrained variational problem, expressed as: in, Indicates the first k One modal function, K This indicates the total number of preset modes. k K , Represents the impulse basis function. Indicates time t Find the partial derivative. f ( t ) represents the running status data. For Hilbert transform kernel; As a penalty factor; 2) Introduce the Lagrange multiplication operator β The constrained variational problem is transformed into an unconstrained variational problem, expressed as: in,{ u k }and{ ω k } represent the set of all modes and their corresponding set of center frequencies, respectively; ω k Indicates the first k The center frequency of each mode; 3) Initialization u k , ω k and β Set the number of iterations m =0; 4) Use the following formula to... u k , ω k and β Perform iterative updates, using the following expression: In the formula, and They are respectively for and The form after Fourier transform. It is noise tolerance; It is a constant used to represent the precision of discrimination, and ; 5) Determine if the convergence criterion is met. If not, continue iterating and updating as in step 4). If met, stop iterating and output multiple IMF component signals, which is the decomposed data sequence. The convergence criterion is expressed as: 。 6. The real-time reliability assessment method for autonomous surface vessel propulsion systems based on diagnostic input according to claim 2, characterized in that, The specific steps for constructing the Markov transition probability matrix include: Assume that the homogeneous Markov process satisfies the following equation: In the formula: for t The state transition probability at time t; For the first i , j One state; State variables of autonomous surface vessel propulsion systems are defined based on homogeneous Markov processes. and state set E , , n The number of states of a component; Based on state set E Draw a state transition diagram and calculate the state transition matrix based on the diagram. The formula for calculating the state transition matrix is ​​as follows: In the formula: This is the state transition matrix; for i State j The transition rate of state transitions; Indicates from state E n The transition rate to state E1; Indicates the overall failure rate of the component; Assuming the state transition probability of a component at a given time follows an exponential distribution, the formula for calculating the state transition probability is: The Markov state transition probability matrix is ​​calculated based on the state transition matrix and the state transition probabilities, and is expressed as follows: in, To set a time.

Images