A real-time evaluation method for reliability of mechanical equipment of autonomous marine surface ship

Abstract

The application discloses a kind of real-time evaluation method of offshore autonomous water surface ship mechanical equipment reliability, S1: the synthetic data set of ship mechanical equipment operation in line with actual degradation law is constructed, the problem of small sample and data imbalance is solved;S2: the health index value of ship mechanical equipment is evaluated using comprehensive weighting method and a number of processed operating characteristic data;S3: the unsupervised health index prediction model for predicting health index value is constructed, the trained unsupervised health index prediction model is obtained by training unsupervised health index prediction model;S4: the health index value predicted by the trained unsupervised health index prediction model is fitted by Weibull distribution, and the shape parameter and scale parameter of the ship mechanical equipment under Weibull distribution are obtained;S5: based on shape parameter and scale parameter, real-time evaluation of ship mechanical equipment reliability is realized, early fault detection is supported, and the safety of MASS is improved.

Description

Technical Field

[0001] This invention relates to the field of marine mechanical equipment reliability assessment technology, and in particular to a real-time assessment method for the reliability of marine autonomous surface vessels. Background Technology

[0002] With the rapid development of technologies related to Maritime Autonomous Surface Ships (MASS), in order to obtain regulatory approval, industry support, and public recognition, MASSes are required to achieve or exceed the reliable operating time required to complete the same tasks. In maritime transport, reliable ship operation is typically achieved through reactive maintenance (RM) and preventive maintenance (PM). The former involves repairing after a failure occurs, while the latter involves developing regular maintenance plans based on preset time intervals, component lifespan, or operating conditions. Condition-based maintenance (CBM) is also included, which develops maintenance plans by assessing the degradation status of components. In recent years, predictive maintenance (PdM) has optimized maintenance strategies and laid the foundation for in-depth implementation and application.

[0003] During navigation, mechanical equipment serves as the infrastructure ensuring ship operation. Ensuring its reliable operation can effectively reduce unexpected downtime and safety accidents caused by equipment failure. However, current implementations of MASS reliability primarily focus on hull structural integrity, collision avoidance algorithms, and remote control systems, while the reliability of mechanical equipment, a key element of navigation safety, has not received sufficient attention. Filling this gap is crucial for the effective execution of long-term autonomous missions by MASS. The highly manual nature of mechanical equipment makes the maintenance and repair of faulty components on MASS increasingly complex. Since MASS ultimately needs to operate without maintenance personnel, RM (Repair Management) cannot prevent failures, and PM (Maintenance Maintenance) relies on experience-based fixed maintenance cycles that need to be coordinated with planned port calls. Although CBM (Continuous Maintenance Management) is cost-effective, cumulative system degradation may reach a failure threshold, increasing the risk of failure. To ensure the reliable operation of MASS during navigation, PdM (Property Management) has become an effective method for ensuring the reliability of mechanical equipment. However, due to the slow accumulation of fault data for MASS mechanical equipment and the tendency for purely algorithm-generated data to deviate from real-world conditions, resulting in low model reliability, and the lack of fault labels in actual monitoring data, traditional methods struggle to directly quantify the reliability of mechanical equipment. Summary of the Invention

[0004] This invention provides a real-time assessment method for the reliability of mechanical equipment on autonomous surface vessels at sea, in order to overcome the aforementioned technical problems.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] A real-time reliability assessment method for the mechanical equipment of autonomous surface vessels at sea, comprising the following steps:

[0007] S1: Construct a synthetic dataset that conforms to the actual degradation law of ship machinery and equipment operation, the synthetic dataset including several types of processed operation feature data;

[0008] S2: The health index value of ship machinery and equipment is evaluated using a comprehensive weighted method and the aforementioned processed operational characteristic data to establish a health index dataset;

[0009] S3: Construct an unsupervised health index prediction model for predicting health index values, and train the unsupervised health index prediction model based on the health index dataset to obtain the trained unsupervised health index prediction model.

[0010] S4: Fit the health index values ​​predicted by the unsupervised health index prediction model after training to a Weibull distribution, and solve for the shape parameters and scale parameters of the ship's machinery and equipment under the Weibull distribution;

[0011] S5: Real-time assessment of the reliability of ship machinery and equipment is achieved based on the shape parameters and dimensional parameters.

[0012] Furthermore, in S1, the specific steps for constructing a synthetic dataset that conforms to the actual degradation patterns of ship machinery and equipment include:

[0013] S11: Acquire operational data of ship machinery and equipment, wherein the operational data of ship machinery and equipment includes several types of operational characteristic data and degradation state coefficients;

[0014] S12: Select multiple subsets of data from the ship's mechanical equipment operating data according to the required ship speed;

[0015] S13: Combine the variation law of Weibull distribution and the degradation state coefficient to remove one or more operational feature data that do not change over time from the multiple subsets of data to obtain the initial dataset;

[0016] S14: The initial dataset is scaled to the [-1,1] interval using the min-max normalization method;

[0017] S15: The Wasserstein generative adversarial network is used to augment the scaled dataset;

[0018] S16: Perform denormalization on the data augmented dataset to convert it into its true form;

[0019] S17: Set the time period to be analyzed, and select data within that time period from several subsets of the dataset after inverse normalization to construct a synthetic dataset that includes several types of processed runtime feature data.

[0020] Furthermore, in S13, the specific steps for removing one or more operational feature data that do not change over time from the multiple subsets of data, based on the variation pattern of the Weibull distribution and the degradation state coefficient, include:

[0021] S131: The initial shape and scale parameters of the Weibull distribution are set through expert experience to ensure that the degradation curve conforms to the physical characteristics of mechanical wear. Combined with the variation law of the Weibull distribution, a degradation curve characterizing the change of the degradation state coefficient over time is obtained, expressed as:

[0022]

[0023] In the formula, λ is the scale parameter, β is the shape parameter, and t is the degradation time;

[0024] S132: Add time series data to several types of operational feature data in the subset based on the degradation curve to obtain a subset with added time series data;

[0025] S133: Remove one or more runtime feature data from the subset of data with added time series according to the set removal rules;

[0026] The elimination rule is as follows: determine whether each type of operational characteristic data remains unchanged over time; if so, eliminate all data for that type of operational characteristic parameter; otherwise, retain that type of operational characteristic data.

[0027] Furthermore, in S2, the formula for evaluating the health index value of ship machinery and equipment using the comprehensive weighting method and the aforementioned processed operational characteristic data is as follows:

[0028]

[0029] Among them, h i (t) represents the health index value of ship machinery equipment i based on multiple operational characteristic data; m is the total number of operational characteristic data; ρ j The weight of the j-th running feature data;

[0030] h ij (t) represents the final health index value of the j-th operational characteristic data of device i. The steps to obtain this value include:

[0031] The initial health index values ​​of the several processed operational feature data are obtained by normalizing the data according to the following formula:

[0032]

[0033] In the formula, X is the initial health index value for the j-th running feature data, with a value range of [0,1]; j For the j-th running feature data sample, X max and X min These are the maximum and minimum values ​​of the j-th running feature data, respectively;

[0034] The initial health index values ​​of all operational characteristic data corresponding to the initial time of ship operation are reset to the maximum health index value L, L = 1;

[0035] Calculate the difference between the initial health index value and the maximum health index value of each operational characteristic data corresponding to the initial time of ship operation, and add the absolute value of the difference to the initial health index value of each operational characteristic data corresponding to other ship operation times besides the initial time of ship operation to obtain the changed health index value of each operational characteristic data.

[0036] Determine whether the health index value of each operational characteristic data after the change is greater than or equal to 1. If it is greater than or equal to 1, then transform the health index value greater than or equal to 1 according to the transformation formula, and take the transformed health index value as the final health index value of the operational characteristic data. The transformation formula is:

[0037] h ij (t) = 2 - h' ij (t) (4)

[0038] In the formula, h' ij (t) represents a health index value greater than or equal to 1.

[0039] Furthermore, in S3, the unsupervised health index prediction model used to predict health index values ​​is a model constructed based on the principal component analysis-long short-term memory network model PCA-LSTM architecture.

[0040] The specific steps for training the unsupervised health index prediction model based on the aforementioned health index dataset include:

[0041] The health index dataset is normalized to eliminate the influence of dimensionality on the calculation results. An original matrix X is then established based on the normalized health index dataset. * ;

[0042] Based on the original matrix X *And determine the covariance matrix P according to formula (5), the formula is:

[0043]

[0044] Where S is the number of samples in the synthetic dataset;

[0045] Based on the covariance matrix P and by solving according to formula (6), the eigenvalues ​​λ are obtained. i and eigenvector α i , is represented as:

[0046] Pα i =λ i α i (6)

[0047] Where, λ i α represents the magnitude of variance along the principal component direction. i The direction of the principal component is indicated. The number of principal components α is calculated according to formula (7), which is:

[0048]

[0049] Where δ is the threshold for the cumulative contribution rate; The cumulative contribution rate of the first k eigenvalues, k <n;

[0050] By combining the covariance matrix P and the number of principal components α, the original matrix X is... * Perform dimensionality reduction to obtain the original matrix after dimensionality reduction;

[0051] The original matrix after dimensionality reduction is input into the Long Short-Term Memory network model to train it, and the trained unsupervised health index prediction model is obtained.

[0052] Furthermore, in S4, the shape and dimensional parameters of the ship's machinery and equipment are obtained by using the maximum likelihood estimation method (MLE).

[0053] Furthermore, in S5, the formula for real-time evaluation of the reliability of ship machinery based on the shape and dimensional parameters is as follows:

[0054]

[0055] Where R represents the reliability assessment result, and t′ represents the sailing time in the synthetic dataset.

[0056] Beneficial Effects: This invention constructs a synthetic dataset of ship machinery and equipment operation that conforms to actual degradation patterns, solving the problems of small sample size and data imbalance. A comprehensive weighting method and several processed operational characteristic data are used to evaluate the health index values ​​of ship machinery and equipment, establishing a health index dataset. An unsupervised health index prediction model is constructed to predict health index values ​​in real time. The health index values ​​predicted by the trained unsupervised health index prediction model are fitted with a Weibull distribution, and the shape and scale parameters of the ship machinery and equipment under the Weibull distribution are obtained. This achieves dynamic real-time assessment of health index-reliability, supports early fault detection, and ultimately improves the safety of MASS (Maintenance, Assurance, and Automation). Attached Figure Description

[0057] 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.

[0058] Figure 1 This is a first flowchart of a real-time reliability assessment method for autonomous surface vessel machinery and equipment in this invention.

[0059] Figure 2 This is a second flowchart of a real-time reliability assessment method for autonomous surface vessel machinery and equipment in this invention.

[0060] Figure 3 This is a structural diagram of the Generative Adversarial Network (GAN) mentioned in this invention;

[0061] Figure 4 This is a structural diagram of the turbine in an embodiment of the present invention;

[0062] Figure 5 This is a comparison chart of the degradation curves of the compressor and turbine in an embodiment of the present invention;

[0063] Figure 6 This is a graph showing the changes in the health index of three main ship speeds in an embodiment of the present invention;

[0064] Figure 7 This is a graph showing the weighted health index calculation results in an embodiment of the present invention;

[0065] Figure 8 This is a graph showing the predicted PCA-LSTM health index values ​​in an embodiment of the present invention.

[0066] Figure 9 This is a diagram showing the calculation results of the size parameters and shape parameters in an embodiment of the present invention;

[0067] Figure 10 This is a diagram showing the reliability evaluation results in an embodiment of the present invention. Detailed Implementation

[0068] 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.

[0069] This embodiment provides a real-time assessment method for the reliability of mechanical equipment on autonomous surface vessels, such as... Figure 1 and Figure 2 As shown, the specific steps include:

[0070] S1: Construct a synthetic dataset that conforms to the actual degradation law of ship machinery and equipment operation, the synthetic dataset including several types of processed operation feature data;

[0071] S2: The health index value of ship machinery and equipment is evaluated using a comprehensive weighted method and the aforementioned processed operational characteristic data to establish a health index dataset;

[0072] S3: Construct an unsupervised health index prediction model for predicting health index values, and train the unsupervised health index prediction model based on the health index dataset to obtain the trained unsupervised health index prediction model.

[0073] S4: Fit the health index values ​​predicted by the unsupervised health index prediction model after training to a Weibull distribution, and solve for the shape parameters and scale parameters of the ship's machinery and equipment under the Weibull distribution;

[0074] S5: Real-time assessment of the reliability of ship machinery and equipment is achieved based on the shape parameters and dimensional parameters.

[0075] In a specific embodiment, S1 includes the following steps for constructing a synthetic dataset that conforms to the actual degradation patterns of ship machinery and equipment:

[0076] S11: Obtain the operating data of the ship's mechanical equipment, which includes several types of operating characteristic data and degradation state coefficients, as shown in Table 1;

[0077] S12: Since the obtained ship machinery and equipment operation data does not have a time series, it is necessary to further process the ship machinery and equipment operation data. Specifically, in this embodiment, multiple subsets of data are obtained from the ship machinery and equipment operation data according to the required ship speed.

[0078] S13: Combine the variation pattern of the Weibull distribution and the degradation state coefficient to remove one or more operational feature data that do not change over time from the multiple subsets of data to obtain the initial dataset;

[0079] Specifically, the Weibull distribution is a continuous probability distribution that is widely used in reliability analysis, survival analysis, and failure time modeling due to its flexibility. This distribution is often used to describe the failure time characteristics of mechanical equipment. In this embodiment, the Weibull distribution is combined with the actual degradation law of the equipment. The degradation state coefficient is combined with shape parameters and scale parameters to characterize the changes of several operating characteristic data over time, so as to facilitate subsequent analysis.

[0080] In a specific embodiment, S13, the specific steps for removing one or more operational feature data that do not change over time from the multiple subsets of data by combining the variation law of the Weibull distribution and the degradation state coefficient include:

[0081] S131: The initial shape and scale parameters of the Weibull distribution are set through expert experience to ensure that the degradation curve better conforms to the physical characteristics of mechanical wear. Combined with the variation law of the Weibull distribution, a degradation curve characterizing the change of the degradation state coefficient over time is obtained, expressed as:

[0082]

[0083] In the formula, λ is the scale parameter, β is the shape parameter, and t is the degradation time;

[0084] S132: Add time series data to several types of operational feature data in the subset based on the degradation curve to obtain a subset with added time series data;

[0085] Specifically, based on the changes in the degradation state coefficient, this embodiment selects data within a two-year range and uses degradation curves to fit the changing trends of several operational characteristic data over time, thereby making the added time series more consistent with the initial operational data of the ship's machinery and equipment.

[0086] S133: Remove one or more runtime feature data from the subset of data with added time series according to the set removal rules;

[0087] The elimination rule is as follows: determine whether each type of operational characteristic data remains unchanged over time; if so, eliminate all data for that type of operational characteristic parameter; otherwise, retain that type of operational characteristic data.

[0088] S14: The initial dataset is scaled to the [-1, 1] interval using the min-max normalization method, as shown in the formula:

[0089]

[0090] In the formula, target_data is the initial dataset, data_min is the minimum value in the initial dataset, data_max is the maximum value in the initial dataset, and normalized_data is the data after normalization.

[0091] S15: Use Wasserstein Generative Adversarial Network (WGAN) to perform data augmentation on the scaled dataset;

[0092] Specifically, Generative Adversarial Networks (GANs), as a widely used data augmentation technique, can effectively alleviate the data imbalance problem by generating synthetic samples. This adversarial architecture comprises two neural network components performing minimax optimization: a generator (G) generates artificial samples by capturing latent feature distributions, and a discriminator (D) distinguishes real data from generated data through a learnable decision boundary. Through adversarial training and iterative optimization, a Nash equilibrium between the two is sought. The optimization of GANs fundamentally relies on the strategic construction of the adversarial objective function. According to the original GAN ​​framework, the adversarial training process constitutes a two-player game theory scenario, and its minimax optimization objective can be mathematically expressed as:

[0093]

[0094] In the formula, P data (x) represents the true data distribution, P z (z) represents the prior distribution of the latent variable. The discriminator dynamically adjusts the decision boundary through the parameter function D(x). When the output value exceeds 0.5, it indicates a higher confidence level in the authenticity of the sample. G(z) represents the sample data generated by the generator network through random noise.

[0095] During GAN model training, problems such as gradient vanishing and model collapse may be encountered, and coordinating the training levels of the generator and discriminator network parameters often consumes a significant amount of time. Therefore, this embodiment employs WGAN, which can alleviate the gradient vanishing problem caused by the inconsistency between the distribution of real samples and generated data in low-dimensional manifolds. WGAN is based on improvements to the GAN architecture, including:

[0096] 1) Remove the Sigmoid activation function at the end of the discriminator in the Generative Adversarial Network (GAN) to achieve unbounded gradient propagation, enhance the network's sensitivity to distribution differences, and force the Lipschitz continuity constraint to be satisfied through backpropagation weight clipping to prevent parameter divergence and improve the stability of generated data.

[0097] 2) The Wasserstein distance loss function is used to reconstruct the adversarial target loss function instead of the log loss function in the generative adversarial network (GAN), which can more effectively capture the relationship between distributions;

[0098] 3) During training, the Stochastic Gradient Descent (SGD) and Root Mean Square Propagation (RMSProp) optimizer are used to replace the momentum-based Adam algorithm in Generative Adversarial Networks (GANs), which reduces the random fluctuations in parameter updates and makes the adversarial relationship between the discriminator and the generator more stable.

[0099] Specifically, the Wasserstein distance in WGAN is defined as follows:

[0100]

[0101] In the formula, p r For the true data distribution, p g Represents the distribution of generated data, the set of joint probability distributions ∏(p r ,p g p r and p g Let W(p) be a marginal distribution, where γ is one of the joint probability distributions. r ,p g ) represents the lower bound of the expectation of γ, that is, the requirement that the true data distribution p r The sample x and the generated data distribution p g The distance of sample y ||xy|| needs to be adapted to p r and p g .

[0102] In a specific embodiment, the adversarial target loss function reconstructed using Wasserstein distance is:

[0103]

[0104] In the formula, f w The discriminator's final output layer maintains linearity, and w is the bias peak of the corresponding neuron; This represents the expected value of the discriminator's output for the real sample; This represents the expected value of the discriminator's output for the generated samples.

[0105] S16: Perform denormalization on the data augmented dataset to convert it into its true form;

[0106] S17: Set the time period to be analyzed, and select data within that time period from several subsets of the dataset after inverse normalization to construct a synthetic dataset that includes several types of processed runtime feature data.

[0107] Specifically, in the process of constructing a synthetic dataset, it is necessary to combine physical knowledge, match it according to the actual ship speed during navigation, and reasonably divide the time proportion of different datasets.

[0108] Specifically, MASS (Maneuvering Assault System) suffers from slow data accumulation of mechanical equipment failures, and data generated purely by algorithms (such as traditional GANs) is prone to deviating from real-world operating conditions, resulting in low model reliability. This embodiment employs WGAN for data augmentation and integrates physical knowledge (such as prior Weibull distribution parameters) to form a dataset of mechanical equipment operation over a single flight, ensuring it conforms to actual degradation patterns. This provides a dataset synthesis method for research on mechanical equipment reliability and solves the problems of small sample sizes and data imbalance.

[0109] In a specific embodiment, in S2, the formula for evaluating the health index value of ship machinery and equipment using a comprehensive weighted method and the aforementioned processed operational characteristic data is as follows:

[0110]

[0111] Among them, h i (t) represents the health index value of ship machinery equipment i based on multiple operational characteristic data; m is the total number of operational characteristic data; ρ j The weight of the j-th operational feature data reflects the impact of this operational feature data on the equipment status, 0≤ρ j ≤1, and

[0112] h ij (t) represents the final health index value of the j-th operational characteristic data of device i. The steps to obtain this value include:

[0113] The initial health index values ​​of the several processed operational feature data are obtained by normalizing the data according to the following formula:

[0114]

[0115] In the formula, X is the initial health index value for the j-th running feature data, with a value range of [0,1]; j For the j-th running feature data sample, Xmax and X min These are the maximum and minimum values ​​of the j-th running feature data, respectively;

[0116] The initial health index values ​​of all operational characteristic data corresponding to the initial time of ship operation are reset to the maximum health index value L, L = 1;

[0117] Calculate the difference between the initial health index value and the maximum health index value of each operational characteristic data corresponding to the initial time of ship operation, and add the absolute value of the difference to the initial health index value of each operational characteristic data corresponding to other ship operation times besides the initial time of ship operation to obtain the changed health index value of each operational characteristic data.

[0118] Determine whether the health index value of each operational characteristic data after the change is greater than or equal to 1. If it is greater than or equal to 1, then transform the health index value greater than or equal to 1 according to the transformation formula, and take the transformed health index value as the final health index value of the operational characteristic data. The transformation formula is:

[0119] h ij (t) = 2 - h' ij (t) (8)

[0120] In the formula, h' ij (t) represents a health index value greater than or equal to 1.

[0121] In a specific embodiment, in S3, the unsupervised health index prediction model used to predict the health index value is a model constructed based on the principal component analysis-long short-term memory network model PCA-LSTM architecture.

[0122] The specific steps for training the unsupervised health index prediction model based on the aforementioned health index dataset include:

[0123] The health index dataset is normalized to eliminate the influence of dimensionality on the calculation results. An original matrix X is then established based on the normalized health index dataset. * ;

[0124] Based on the original matrix X * And determine the covariance matrix P according to formula (9), the formula is:

[0125]

[0126] Where S is the number of samples in the synthetic dataset;

[0127] Based on the covariance matrix P, the eigenvalues ​​λ are obtained by solving according to formula (10). i and eigenvector α i, is represented as:

[0128] Pα i =λ i α i (10)

[0129] Where, λ i α represents the magnitude of variance along the principal component direction. i The direction of the principal components is indicated. The number of principal components α is calculated according to formula (11) to retain the main variance of the data. The formula is as follows:

[0130]

[0131] Where δ is the threshold for the cumulative contribution rate, which is usually set to 85%; The cumulative contribution rate of the first k eigenvalues ​​(k <n);

[0132] By combining the covariance matrix P and the number of principal components α, the original matrix X is... * Perform dimensionality reduction to obtain the original matrix after dimensionality reduction;

[0133] The original matrix after dimensionality reduction is input into the Long Short-Term Memory network model to train it, and the trained unsupervised health index prediction model is obtained.

[0134] Specifically, PCA is a linear dimensionality reduction technique that maps high-dimensional data to a low-dimensional space through linear projection. This involves projecting the original features onto the dimension with the largest variance in the captured data, ensuring that the low-dimensional representation retains as much of the original information as possible. This embodiment uses PCA to transform the original matrix X... * Dimensionality reduction and training the Long Short-Term Memory (LSTM) network model with the original dimensionality-reduced matrix can improve the training accuracy of LSTM. Specifically, LSTM demonstrates strong advantages in handling time series prediction and nonlinear mapping problems. LSTM incorporates memory units to remember past information and adds three gate structures—input gate, output gate, and forget gate—to control the transmission of historical information. This solves the gradient vanishing and exploding problems caused by the chain-like connection of network units in traditional recurrent neural networks, effectively improving learning time.

[0135] Specifically, in this embodiment, the following evaluation metrics are used to compare the predictive performance of the PCA-LSTM model, including:

[0136] (1) Mean Squared Error (MSE) is the ratio of the sum of squared deviations between observed and true values ​​to the number of observations. It sensitively reflects the magnitude of the model's prediction error, but is greatly affected by outliers. Because MSE squares the error, larger errors receive greater weight. It is commonly used as the loss function in linear regression models, optimizing model parameters by minimizing MSE. The formula for MSE is:

[0137]

[0138] Among them, y i For the true value, is the predicted value, and n is the number of observations.

[0139] (2) Mean Absolute Error (MAE) is similar to MSE, but uses absolute error. It is more robust to outliers than MSE because it does not amplify these outliers through squared errors. MAE can more intuitively reflect prediction errors in some practical applications. It is suitable for situations requiring direct and interpretable error measurement. The formula for MAE is:

[0140]

[0141] (3)R 2 R measures the proportion of true label variance that the model fails to capture. Its value ranges from 0 to 1, with values ​​closer to 1 indicating a better model fit. It is widely used in statistical modeling, especially crucial in evaluating model interpretability. 2 The formula is:

[0142]

[0143] in, This is the mean of the true values.

[0144] In a specific embodiment, in S4, the health index values ​​predicted by the trained unsupervised health index prediction model are fitted with a Weibull distribution. The maximum likelihood estimation (MLE) method is used to solve for the shape and scale parameters of the ship's machinery under the Weibull distribution. The formulas include:

[0145]

[0146] Where L represents the likelihood function, x j Let λ represent the j-th health index value, β be the scale parameter, and N be the total number of health indices used for calculation.

[0147] The formula for calculating the maximum likelihood condition is as follows:

[0148]

[0149] To simplify equation (15), the maximum likelihood condition is expressed as:

[0150]

[0151] Specifically, this embodiment evaluates the reliability of mechanical equipment by fitting the Weibull distribution, uses the maximum likelihood estimation method to solve the shape and scale parameters under the Weibull distribution in real time, and realizes real-time reliability assessment of ship mechanical equipment based on the dynamic changes of the shape and scale parameters.

[0152] In a specific embodiment, in S5, the formula for real-time evaluation of the reliability of ship machinery based on the shape parameters and dimensional parameters is as follows:

[0153]

[0154] Where R represents the reliability assessment result, and t′ represents the sailing time in the synthetic dataset.

[0155] Specifically, actual monitoring data lacks fault labels, making it difficult for traditional methods to directly quantify the reliability of mechanical equipment. This embodiment enhances real-world data by combining physical knowledge and the WGAN algorithm to construct a health index dataset. PCA is used to reduce the dimensionality of the health index dataset, extracting key features and combining LSTM to accurately capture long-term dependencies, enabling real-time prediction of health index values ​​and improving prediction accuracy. The predicted health index values ​​are fitted with a Weibull distribution, and shape and scale parameters are obtained in real-time using MLE (Mean Scale), directly establishing a mathematical correlation between health status and the reliability of ship machinery. This more accurately characterizes the equipment degradation process and reliability trends. Real-time assessment of ship machinery reliability can be achieved using real-time calculations of shape and scale parameters without fault-labeled data. This real-time reliability assessment allows for timely early fault detection, enabling preventative measures to avoid accidents, improving system safety, and providing real-time decision support for equipment maintenance management.

[0156] Specifically, in this embodiment, the Weibull distribution parameters are updated every hour using MLE, and combined with real-time health index feedback, to achieve dynamic calibration of the reliability assessment model.

[0157] To effectively evaluate the real-time reliability of MASS mechanical equipment, this embodiment proposes a real-time reliability assessment method for MASS mechanical equipment. Combining physical knowledge, the degradation coefficient is fitted to a Weibull distribution to reflect actual operating conditions. The limitations of actual monitoring data are overcome by adding time-series features to the state data, and WGAN is used to enhance data richness. A 720-hour cruise state monitoring dataset is extracted for case verification.

[0158] Given the future development of MASS, ensuring high reliability and power redundancy of the propulsion system, minimizing human intervention, and enabling remote control are crucial. Furthermore, prioritizing low emissions and environmental sustainability is also key. For intelligent ships, battery-powered and hybrid propulsion systems are considered the optimal solutions. Therefore, in this embodiment, a dataset is generated using an advanced gas turbine (GT) propulsion system simulator installed on a naval frigate equipped with a combined diesel-electric and gas turbine (CODLAG) propulsion system. The data is sourced from the University of California, Irvine (UCI) repository. The degradation state coefficients represent the degradation of the turbine and compressor. The GT compressor degradation state coefficient and the GT turbine degradation state coefficient are used as features of this model, and other features of the GT components are used as targets for state prediction. The structure of the GT is as follows... Figure 4 As shown, its specific operational characteristics are shown in Feature 1-16 in Table 1.

[0159] Table 1

[0160]

[0161]

[0162] In these operational characteristic data, lp (characteristic 1) shows a linear relationship with ship speed, while T1, P1, and Pexh remain constant throughout the dataset. Therefore, these four operational characteristics were excluded from the analysis, leaving the dataset covering nine different ship speeds: 3, 6, 9, 12, 15, 18, 21, 24, and 27 knots. Since the condition of other gas turbine components varies with different ship speeds, the condition monitoring data for each speed was analyzed separately. In this embodiment, the results are weighted according to the duration ratio of each speed in actual operation. Specifically, after consultation with experts on the operating speed range of ships at sea and analysis of the dataset, infrequently occurring speed data were excluded, retaining only the three most frequently used speeds, with the corresponding ratios shown in Table 2.

[0163] Table 2. Percentage of ship speed time

[0164] Ship speed (knots) 12 15 18 time(%) 20％ 60％ 20％

[0165] In this embodiment, the simulation period for the dataset is set to two years. The CDSC and TDSC values ​​decrease from 1 to 0.95 and 0.975 respectively over the two years. Other features correspond to the monitoring status at each time point and are matched with the decay coefficient values. In this embodiment, a Weibull distribution is used to fit two degradation state coefficients (features 17 and 18) to incorporate them into the time series data. The shape and scale parameters are calculated using equation (1), and plotted as shown below. Figure 5The degradation curves for the turbines and compressors are shown. Each ship speed dataset contains 1326 data points. To alleviate data imbalance and improve the accuracy of training results, this embodiment uses the WGAN algorithm for data augmentation, expanding the dataset to 175,200 data points.

[0166] Specifically, to obtain accurate prediction results, this embodiment uses the PCA-LSTM model to predict the state of various GT components based on the degradation of the turbine and compressor. However, due to the lack of fault labels in the dataset, and to accurately reflect the actual degradation and real-time status of the equipment, this embodiment divides the calculation of the health index value into two parts: the first part calculates the real-time health index value using formula (6), and the relevant results are as follows. Figure 6 and Figure 7 As shown in the figure. The second part uses the PCA-LSTM model to predict the health index value of the device, and the relevant results are as follows. Figure 8 As shown in Table 2, to improve prediction accuracy, model parameters are adjusted promptly based on prediction errors. To maintain consistency with ship operation patterns, this embodiment merges data from different ship speeds according to their respective time proportions during operation. The data is recombined and synthesized based on the time proportion of each ship speed in the actual voyage (12 knots 20%, 15 knots 60%, 18 knots 20%) to ensure the generated data reflects the actual operation pattern. Furthermore, to simulate real operating conditions encountered during actual navigation, the new dataset allows selection of time data for any given time period. To simplify dataset construction, a continuous 720-hour time period was selected. Since different ship speeds imply different operating conditions, the differences between certain features are significant. To address this issue, the dataset is normalized using the min-max method before merging. Figure 6 and Figure 7 The health index values ​​are displayed for three different ship speeds, as well as the health index values ​​from a synthetic dataset. Figure 6 As shown, the health index trended downwards over time, dropping below 0.8 after 17,520 hours (two years) without maintenance. Figure 7 As shown, the newly synthesized dataset shows an overall decline around hour 144. This can be attributed to the fact that devices running at section 12 were in a more stable state compared to the data from sections 15 and 18. After normalization, the health status of section 12 was more favorable at hour 3000, which introduces unavoidable differences when using the merged dataset for prediction. The health index results would likely be higher if calculated using actual operational data.

[0167] To achieve real-time and accurate prediction of the health index, the PCA-LSTM method was chosen, combining the dimensionality reduction capability of PCA with the time series prediction capability of the LSTM network. PCA retains the five most important features, and the LSTM part inputs these dimensionality-reduced features to predict the health index at each time step. The predicted health index is compared with the actual values ​​to evaluate the model's performance. The predicted health index is shown below. Figure 8 As shown. During training, MSE, MAE, and R are used. 2 To evaluate the prediction accuracy of various components. The results are shown in Table 3, where R... 2 A value of 0.98 indicates that the model explains 98% of the variance of the health index, demonstrating high predictive accuracy.

[0168] Table 3 Average values ​​of the measurement parameters

[0169] parameter MSE MAE <![CDATA[R 2 ]]> value 6.74E-05 6.50E-3 9.80E-1

[0170] To further assess the real-time reliability of the propulsion system, the predicted health index value is combined with the Weibull distribution, and the system reliability is estimated in real time using MLE. Based on the prediction results, the two-parameter results are calculated using formulas (15) to (17). Figure 9 As shown. The final assessment results of the ship's mechanical equipment reliability are as follows. Figure 10 As shown, the system reliability dropped to 0.959 at 720 hours. According to the reliability definition, this propulsion system has a 95.9% probability of reliable operation during the voyage (30 days) without human intervention. This verifies the effectiveness of this method in the field of MASS mechanical equipment and provides support for filling research gaps in existing technologies.

[0171] 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 evaluation method for the reliability of mechanical equipment on autonomous surface vessels, characterized in that, The specific steps include: S1: Construct a synthetic dataset that conforms to the actual degradation law of ship machinery and equipment operation, the synthetic dataset including several types of processed operation feature data; In S1, the specific steps for constructing a synthetic dataset that conforms to the actual degradation patterns of ship machinery and equipment include: S11: Acquire operational data of ship machinery and equipment, wherein the operational data of ship machinery and equipment includes several types of operational characteristic data and degradation state coefficients; S12: Select multiple subsets of data from the ship's mechanical equipment operating data according to the required ship speed; S13: Combine the variation law of Weibull distribution and the degradation state coefficient to remove one or more operational feature data that do not change over time from the multiple subsets of data to obtain the initial dataset; S14: The initial dataset is scaled to the [-1,1] interval using the min-max normalization method; S15: The Wasserstein generative adversarial network is used to augment the scaled dataset; S16: Perform denormalization on the data augmented dataset to convert it into its true form; S17: Set the time period to be analyzed, and select the data within that time period from several subsets of the dataset after inverse normalization to construct a synthetic dataset that includes several types of processed runtime feature data. S2: The health index value of ship machinery and equipment is evaluated using a comprehensive weighted method and the aforementioned processed operational characteristic data to establish a health index dataset; S3: Construct an unsupervised health index prediction model for predicting health index values, and train the unsupervised health index prediction model based on the health index dataset to obtain the trained unsupervised health index prediction model. In S3, the unsupervised health index prediction model used to predict health index values ​​is a model constructed based on the principal component analysis-long short-term memory network model PCA-LSTM architecture. The specific steps for training the unsupervised health index prediction model based on the aforementioned health index dataset include: The health index dataset is normalized to eliminate the influence of dimensionality on the calculation results, and the original matrix is ​​established based on the normalized health index dataset. X * ; Based on the original matrix X * And determine the covariance matrix according to formula (5). P The formula is: in, S This represents the number of samples in the synthetic dataset. Based on covariance matrix P The eigenvalues ​​are obtained by solving formula (6). λ i and eigenvectors α i , is represented as: in, λ i This indicates the magnitude of variance along the principal component direction. α i Indicate the direction of the principal components, and calculate the number of principal components according to formula (7). α The formula is: in, δ The threshold for cumulative contribution rate; For the front k The cumulative contribution rate of each eigenvalue k < n ; By combining the covariance matrix P and the number of principal components α For the original matrix X * Perform dimensionality reduction to obtain the original matrix after dimensionality reduction; The original matrix after dimensionality reduction is input into the Long Short-Term Memory network model to train it, and the trained unsupervised health index prediction model is obtained. S4: Fit the health index values ​​predicted by the unsupervised health index prediction model after training to a Weibull distribution, and solve for the shape parameters and scale parameters of the ship's machinery and equipment under the Weibull distribution; S5: Real-time assessment of the reliability of ship machinery and equipment is achieved based on the shape parameters and dimensional parameters.

2. The real-time reliability assessment method for the mechanical equipment of autonomous surface vessels according to claim 1, characterized in that, In S13, the specific steps for removing one or more operational feature data that do not change over time from the multiple subsets of data, based on the variation pattern of the Weibull distribution and the degradation state coefficient, include: S131: The initial shape and scale parameters of the Weibull distribution are set through expert experience to ensure that the degradation curve conforms to the physical characteristics of mechanical wear. Combined with the variation law of the Weibull distribution, a degradation curve characterizing the change of the degradation state coefficient over time is obtained, expressed as: In the formula, λ For scale parameters, β For shape parameters; t For degradation time; S132: Add time series data to several types of operational feature data in the subset based on the degradation curve to obtain a subset with added time series data; S133: Remove one or more runtime feature data from the subset of data with added time series according to the set removal rules; The elimination rule is as follows: determine whether each type of operational characteristic data remains unchanged over time; if so, eliminate all data for that type of operational characteristic parameter; otherwise, retain that type of operational characteristic data.

3. The real-time reliability assessment method for the mechanical equipment of autonomous surface vessels according to claim 2, characterized in that, In S2, the formula for evaluating the health index value of ship machinery and equipment using the comprehensive weighted method and the aforementioned processed operational characteristic data is as follows: in, h i ( t ) for marine machinery and equipment i Health index values ​​obtained based on multiple operational characteristic data; m The total number of feature data points; ρ j For the first j The weights of each running feature data; h ij ( t ) for equipment i The j The steps to obtain the final health index value of each running feature data include: The initial health index values ​​of the several processed operational feature data are obtained by normalizing the data according to the following formula: In the formula, For the first j The initial health index value of each running feature data is in the range of [0,1]. For the first j A data sample of operational feature data, X max and X min The first j The maximum and minimum values ​​of each running feature data; The initial health index values ​​of each operational characteristic data corresponding to the initial time of ship operation are reset to the maximum health index value L, L=1; Calculate the difference between the initial health index value and the maximum health index value of each operational characteristic data corresponding to the initial time of ship operation, and add the absolute value of the difference to the initial health index value of each operational characteristic data corresponding to other ship operation times besides the initial time of ship operation to obtain the changed health index value of each operational characteristic data. Determine whether the health index value of each operational characteristic data after the change is greater than or equal to 1. If it is greater than or equal to 1, then transform the health index value greater than or equal to 1 according to the transformation formula, and take the transformed health index value as the final health index value of the operational characteristic data. The transformation formula is: h ij ( t ) =2 h’ ij ( t ) （4） In the formula, h’ ij ( t () is a health index value greater than or equal to 1.

4. The real-time reliability assessment method for the mechanical equipment of autonomous surface vessels according to claim 3, characterized in that, In S4, the shape and dimensional parameters of the ship's machinery and equipment are obtained by using the maximum likelihood estimation method (MLE).

5. The real-time reliability assessment method for the mechanical equipment of autonomous surface vessels according to claim 4, characterized in that, In S5, the formula for real-time evaluation of the reliability of ship machinery and equipment based on the shape parameters and dimensional parameters is as follows: in, R This indicates the reliability assessment results. This represents the sailing time in the synthetic dataset.

Images