A method and system for evaluating the health state of a marine bearing

Abstract

This invention belongs to the field of marine equipment health management technology, and discloses a method and system for assessing the health status of marine bearings. The method includes the following steps: S1, collecting marine bearing operating status data through multi-channel sensors and processing it to obtain a standardized health status sequence; S2, performing soft partitioning on the standardized health status sequence to obtain a soft-label sequence; S3, extracting and fusing time-series features from the soft-label sequence to obtain bidirectional time-series features; S4, based on the bidirectional time-series features, quantifying the prediction uncertainty using a Monte Carlo mechanism and generating a health status prediction probability through a Softmax function to obtain the marine bearing health status prediction result. This invention introduces a Monte Carlo mechanism, enabling the model to output the confidence interval of the prediction result in a noisy environment, effectively identifying low-confidence samples. It not only solves the problem of traditional methods lacking risk assessment basis, but also supports hierarchical operation and maintenance strategies, significantly enhancing the robustness of the system in complex marine environments.

Description

Technical Field

[0001] This invention relates to the field of marine equipment health management technology, and in particular to a method and system for assessing the health status of marine bearings. Background Technology

[0002] The health and performance of a ship's propulsion equipment are crucial for the reliable and safe navigation of a vessel. Marine bearings, as core components of a ship's power transmission system, bear enormous radial and axial loads, and their health directly determines the safety and reliability of the propulsion system. However, unlike ordinary industrial bearings, marine bearings are constantly exposed to the complex noise interference environment of the ocean, meaning that their vibration signals not only include the degradation characteristics of the equipment itself but also inevitably incorporate various high-intensity interference noises.

[0003] (1) Multi-source mechanical noise: Strong vibrations from adjacent components such as the main engine, gearbox, and propeller are transmitted through the hull structure, forming a wideband background noise;

[0004] (2) Fluid noise: The impact of ocean waves, cavitation effect and random pulsating pressure generated by water turbulence cause the signal-to-noise ratio (SNR) of the sensor to drop sharply;

[0005] (3) Time-varying non-stationary characteristics: The dynamic changes in the ship's navigation conditions result in strong non-stationarity and time-varying characteristics in the noise statistics.

[0006] Currently, health assessment methods are mainly divided into two categories: physical model-driven and data-driven. Physical model-driven methods rely on a deep understanding of the equipment's operating mechanism; however, under conditions where strong noise masks subtle early fault characteristics, it is difficult to establish accurate degradation mathematical models, thus limiting their application. While data-driven methods can perform assessments by mining multi-source sensor data, current technologies still have limitations when dealing with high-noise interference environments.

[0007] First, the robustness of feature extraction is insufficient. Most traditional deep learning models are based on deterministic mapping mechanisms, assuming that the input data is clean or that the noise distribution is fixed. When faced with sudden strong impulse noise or non-Gaussian noise in the marine environment, these models are prone to misclassifying noise as fault features, or causing genuine weak fault features to be submerged by noise, resulting in a large number of false positives and false negatives.

[0008] Secondly, there is a lack of ability to quantify prediction uncertainty. Existing methods mostly output a single deterministic health indicator (hard label), ignoring the uncertainty of the model's own perception in high-noise environments. When the input signal is severely contaminated, the model cannot perceive the credibility of its judgment, resulting in the lack of confidence boundaries in the evaluation results, making it difficult to meet the high reliability requirements of intelligent ship operation and maintenance.

[0009] Finally, the ambiguity of the degradation process is compounded by noise interference. The transition of marine bearings from health to failure is a gradual and continuous process, and the state boundaries themselves are ambiguous. Under strong noise interference, these ambiguous boundaries are further disturbed, making it difficult for traditional hard classification methods to distinguish between them, resulting in a significant reduction in the accuracy of health status assessment. Therefore, how to construct an assessment method that can effectively suppress noise interference from complex marine environments, quantify the uncertainty of model predictions, and accurately capture the gradual degradation characteristics of bearings has become a key challenge that urgently needs to be addressed in the field of marine bearing health management. Summary of the Invention

[0010] To address the technical problems of existing marine bearing health assessment methods, such as the difficulty in accurately depicting the gradual degradation process of equipment using hard labeling, the inability of a single model to simultaneously consider temporal local features and global dependencies, poor feature robustness leading to low assessment accuracy in complex marine environments, and the lack of a mechanism for quantifying prediction uncertainty, this invention proposes a method and system for assessing the health status of marine bearings. This method utilizes a multi-level collaborative mechanism, including fuzzy C-means clustering to generate soft labels, temporal convolutional networks to extract local temporal patterns, bidirectional long short-term memory networks to fuse bidirectional temporal features, beluga optimization algorithm parameter optimization, and Monte Carlo Dropout to quantify prediction uncertainty. This constructs a complete health assessment framework for noisy environments, significantly improving the ability to characterize health status and extract temporal features, achieving high-precision and reliable quantification of marine bearing health.

[0011] This invention is achieved through the following technical solution, including the following steps:

[0012] S1. Collect marine bearing operating status data through multi-channel sensors, and obtain a standardized health sequence after preprocessing and feature fusion;

[0013] S2. Perform soft partitioning on the standardized health sequence to obtain a soft label sequence;

[0014] S3. Extract and fuse the temporal features of the soft-label sequence to obtain bidirectional temporal features;

[0015] S4. Based on the bidirectional time series characteristics, the Monte Carlo mechanism is used to quantify the prediction uncertainty and the Softmax function is used to generate the health status prediction probability to obtain the health status prediction result of marine bearings.

[0016] As a further preferred option, in step S1, principal component analysis is used to preprocess and fuse the marine bearing operating status data to obtain a standardized health sequence.

[0017] As a further preferred option, in step S2, fuzzy C-means clustering is used to perform soft partitioning on the standardized health sequence to generate a soft label sequence.

[0018] As a further preferred option, in step S3, a temporal convolutional network and a bidirectional long short-term memory network are used to extract and fuse the temporal features of the soft-label sequence to obtain bidirectional temporal features.

[0019] As a further preferred option, the specific steps of step S4 are as follows:

[0020] S4.1 Perform N forward propagations on the bidirectional temporal features to obtain N prediction results;

[0021] S4.2. Based on the prediction results obtained in step S4.1, calculate their mean and variance;

[0022] S4.3. Based on the results of a single prediction, construct the final prediction result using the following formula:

[0023] ;

[0024] In the formula, p represents the final prediction result; This represents the probability distribution of the nth prediction; This represents the probability that a sample belongs to class c in the nth prediction.

[0025] S4.4 Based on the final prediction results, the prediction entropy and mutual information are obtained as uncertainties.

[0026] S4.5. Based on the bidirectional time series characteristics, generate the health status prediction probability through the Softmax function;

[0027] S4.6 Output the prediction result of the marine bearing health status, which is composed of the prediction entropy and mutual information obtained in step S4.4 and the health status prediction probability obtained in step S4.5.

[0028] As a further preferred option, step S0 is included before step S1: constructing a marine bearing health status dataset, training the MCD-FT-BiLSTM model, and obtaining a trained MCD-FT-BiLSTM model.

[0029] The present invention also provides an assessment system for implementing the marine bearing health status assessment method described herein, comprising:

[0030] The data acquisition and preprocessing module is used to collect and preprocess marine bearing operating status data through multi-channel sensors.

[0031] The health status construction module processes multi-channel sensor data to construct a standardized health status sequence;

[0032] The soft label segmentation module performs soft segmentation on the standardized health score sequence based on fuzzy C-means clustering to generate a soft label sequence.

[0033] The temporal feature extraction module extracts local temporal features of the soft-label sequence through a temporal convolutional network;

[0034] The feature fusion module is used to fuse the forward and backward temporal dependencies of the output features of the temporal convolutional network through a bidirectional long short-term memory network to generate bidirectional temporal features;

[0035] The parameter optimization module is used to adaptively optimize the hyperparameters of the model using the White Whale Optimization Algorithm.

[0036] The health assessment output module is used to quantify the prediction uncertainty through the Monte Carlo mechanism and output the bearing health status assessment results through the Softmax function.

[0037] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0038] 1. Possesses the ability to quantify prediction uncertainty, enhancing decision-making credibility. The evaluation method provided by this invention introduces the Monte Carlo (MCD) mechanism, enabling the model to output the confidence interval of prediction results in noisy environments, effectively identifying low-confidence samples. This not only solves the problem of traditional methods lacking risk assessment basis but also supports hierarchical operation and maintenance strategies, significantly enhancing the system's robustness in complex marine environments.

[0039] 2. Employing a soft-label mechanism to accurately characterize the progressive degradation process. The evaluation method provided by this invention utilizes fuzzy C-means clustering (FCM) to generate soft labels reflecting state membership, overcoming the shortcomings of traditional hard-label partitioning that ignores the ambiguity of state boundaries. This method better conforms to the physical laws of the continuous evolution of bearings from health to failure, significantly improving the precision and accuracy of state assessment.

[0040] 3. The evaluation method provided by this invention constructs a TCN-BiLSTM two-level architecture to achieve multi-scale feature fusion. It combines the advantages of TCN in capturing local temporal fluctuations and BiLSTM in mining long-distance bidirectional dependencies, solving the problem that a single model cannot take into account both local details and global context, and significantly enhancing the comprehensiveness and anti-interference ability of feature extraction.

[0041] 4. The evaluation method provided by this invention introduces the White Whale Optimization (BWO) algorithm to replace inefficient manual parameter tuning. It can automatically search for the optimal hyperparameter combination of TCN and BiLSTM, avoid getting trapped in local optima, and thus improve the generalization performance and convergence efficiency of the model under different working conditions.

[0042] 5. The evaluation method provided by this invention establishes a four-part evaluation system to support differentiated operation and maintenance decisions. It subdivides the bearing status into four levels: healthy, good, degraded, and faulty, thereby achieving fine monitoring of the entire life cycle, effectively avoiding over-maintenance or under-maintenance, and ensuring the safe and stable operation of the ship's power system. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the overall process of the evaluation method of the present invention.

[0044] Figure 2 This is a schematic diagram of the overall framework of the evaluation system in an embodiment of the present invention.

[0045] Figure 3 This is a detailed flowchart illustrating the internal data processing, training optimization, and uncertainty quantification of the MCD-FT-BiLSTM model in the evaluation method of this invention. Detailed Implementation

[0046] The advantages and features of the present invention will be illustrated and explained by the following non-limiting description of preferred embodiments, which are given by way of example only with reference to the accompanying drawings.

[0047] like Figure 1 and Figure 3 As shown, this invention provides a method for assessing the health status of marine bearings. It constructs a Monte Carlo-Fuzzy C-Means-Temporal Convolutional Network-Bi-directional Long Short-Term Memory (MCD-FT-BiLSTM) model, which integrates Monte Carlo Dropout (MCD), Fuzzy C-Means clustering (FCM), Temporal Convolutional Network (TCN), and Bi-directional Long Short-Term Memory (BiLSTM). This model is suitable for monitoring the health status and intelligent operation and maintenance of core components of marine power transmission systems in complex marine environments. The implementation entity of the assessment method provided by this invention can be a marine bearing health status assessment system. Depending on different requirements, the order of steps in the assessment method can be changed, and some steps can be omitted. The specific steps of the assessment method provided by this invention are as follows:

[0048] Step S0: Construct a marine bearing health status dataset and train the MCD-FT-BiLSTM model to obtain the trained MCD-FT-BiLSTM model.

[0049] Step S0.1: Construct a marine bearing health status dataset. The dataset includes multiple sets of health status soft label sequences and corresponding health status labels. The preferred dataset of this invention comes from the XJTU rolling bearing accelerated life test dataset of Xi'an Jiaotong University. This dataset contains 15 bearing vibration signals throughout their entire life cycle from a healthy state to complete failure.

[0050] Step S0.2: Divide the marine bearing health status dataset into a training set, a validation set, and a test set in a ratio of 7:2:1;

[0051] Step S0.3: Initialize the parameters of the MCD-FT-BiLSTM model;

[0052] Step S0.4: Apply artificial Gaussian white noise to the training set data. The noise intensity is controlled by the signal-to-noise ratio (SNR) parameter to construct a training sample set under multiple noise levels and enhance the model's adaptability to noise interference.

[0053] Step S0.5: The Beluga Whale Optimization (BWO) algorithm is used to adaptively optimize the key hyperparameters of the MCD-FT-BiLSTM model to obtain the optimized parameter combination, thereby obtaining the optimized MCD-FT-BiLSTM model, including the dilation rate d of the temporal convolutional network TCN, the kernel size K, the number of kernels, and the number of hidden units of the bidirectional long short-term memory network BiLSTM.

[0054] The specific steps of the Beluga optimization algorithm are as follows:

[0055] Step S0.5.1: Initialize the beluga whale population using the following formula:

[0056] ;

[0057] In the formula, Let be the i1th beluga whale individual, i.e., the i1th hyperparameter combination; LB and UB are the lower and upper bounds of the parameter search, respectively; rand(0,1) represents a random number in the interval [0,1].

[0058] Step S0.5.2: Take the accuracy of the validation set as the fitness value, and then update the position using the following formula:

[0059] ;

[0060] ;

[0061] In the formula, This is the current optimal solution; , and Both are coefficient vectors; Represents the combination of hyperparameters at time t; This represents the combination of hyperparameters at time t+1;

[0062] Step S0.5.3: Employ spiral update to enhance local fine-grained search, as shown in the following formula:

[0063] ;

[0064] In the formula, b is a constant that defines the spiral shape; l is a random number between [-1, 1].

[0065] Step S0.5.4: After reaching the required number of iterations, output the current optimal solution, which is the optimized parameter combination, to obtain the optimized MCD-FT-BiLSTM model.

[0066] The expansion rate of TCN was optimized using the white whale optimization algorithm. kernel size Key hyperparameters such as the number of convolutional kernels and the number of hidden units in BiLSTM are optimized to obtain the optimized parameter combination.

[0067] By using the BWO optimization step, adaptive optimization of key hyperparameters of the model is achieved, solving the technical problems of low efficiency and easy getting trapped in local optima in traditional methods of manual parameter tuning.

[0068] Step S0.6: Input the training set into the optimized MCD-FT-BiLSTM model and output the health status prediction results;

[0069] Step S0.7: Calculate the training loss using the classification cross-entropy loss function, and update the network parameters using the backpropagation algorithm;

[0070] Step S0.8: Introduce label smoothing technology during training to alleviate overfitting. The label smoothing coefficient is set to 0.1.

[0071] Step S0.9: Monitor model performance using the validation set. When the accuracy on the validation set no longer improves for 5 consecutive training cycles, or when the change in the training loss value is less than the preset threshold of 0.001, stop training, complete the model building work, save the model weight file, and obtain the trained MCD-FT-BiLSTM model.

[0072] Step S0.10: Model testing. Predict and evaluate the test set, output fault labels and model performance indicators, and verify the robustness of the model under different signal-to-noise ratios (SNR={40,30,20,10}).

[0073] By introducing noise enhancement mechanisms and label smoothing techniques during the training phase, the generalization ability and robustness of the MCD-FT-BiLSTM model under noisy environments are significantly improved, solving the technical problem of unstable performance of traditional methods in complex marine environments.

[0074] By clarifying the specific training steps of the MCD-FT-BiLSTM model, efficient and stable training and optimization of the model were achieved, ensuring that the model can fully learn the complete degradation process of marine bearings from health to failure under noisy conditions, and providing reliable technical support for practical engineering applications.

[0075] Step S1: Collect marine bearing operating status data through multi-channel sensors, and obtain a standardized health sequence after preprocessing and feature fusion;

[0076] By acquiring multi-channel sensor time-series monitoring data of marine bearing operation status, real-world operational data of marine bearings in complex marine environments is collected, providing a raw data foundation for subsequent health assessments. This ensures that the assessment method is based on actual monitoring signals and enhances the model's practicality and data-driven capabilities.

[0077] In some specific embodiments, the multi-channel sensor timing monitoring data of the marine bearing includes vibration signals, temperature signals, and pressure signals. The signal sampling frequency is 25.6 kHz, the sampling interval is 1 min, and the duration of each sampling is 1.28 s. Among them, the vibration signal is collected by using two PCB352C33 unidirectional accelerometers, which are respectively mounted vertically and horizontally on the test bearing housing.

[0078] By specifically defining the types and acquisition parameters of multi-channel sensor data, comprehensive monitoring of the operating status of marine bearings is achieved, enhancing the representativeness and completeness of the input data and improving the model's ability to capture bearing degradation characteristics in noise-interference environments.

[0079] In some specific embodiments, Principal Component Analysis (PCA) combined with a weighted distance mechanism and an exponential decay function is used to preprocess and fuse the original data to obtain a standardized health sequence. The original data is the marine bearing operating status data collected by multi-channel sensors. The specific steps are as follows:

[0080] Step S1.1: Principal Component Analysis (PCA) is used to reduce the dimensionality of the original data to obtain the comprehensive feature sequence, as shown in the following formula:

[0081] ;

[0082] In the formula, This represents the sensor data for the p-channel; Indicates the kth 主 The weight matrix of each principal component; This represents the matrix after dimensionality reduction; T represents the sample point length of a single sequence.

[0083] Step S1.2: Based on the comprehensive feature sequence, calculate the weighted distance between the i-th health feature and other features at each time point, as shown in the following formula:

[0084] ;

[0085] In the formula, This represents the weighted distance between the i-th health feature and other features at time t; and represents the values ​​of the i-th and j-th health features at time t, respectively; k represents the total number of equipment health indicators;

[0086] Step S1.3: Construct the weighting coefficients using an exponential decay function, as shown in the following formula:

[0087] ;

[0088] In the formula, Indicates the weighting coefficient; represents the attenuation coefficient, which is the weighted distance between the i-th health feature and other features at time t obtained from step S1.2; exp represents the exponential operation with base e.

[0089] Step S1.4: Based on the weighting coefficients obtained in Step S1.3, weighted fusion is performed on each health feature to obtain a standardized health index. The standardized health index is then arranged in time sequence to obtain a standardized health sequence, as shown in the following formula:

[0090] ;

[0091] In the formula, This represents the health status value of the equipment at time t, i.e., the standardized health status index. This represents the value of the i-th health characteristic at time t.

[0092] Through the health status construction steps, the fusion processing of multi-source sensor data for marine bearings was realized, generating a single-variable health status sequence that can accurately reflect the degradation state of the equipment, thus solving the technical problems of multi-source data fusion and quantitative assessment of health status.

[0093] Step S2: Perform soft partitioning on the standardized health sequence obtained in step S1 to generate a soft label sequence containing the probability distribution of the four states: "healthy-good-degraded-faulty".

[0094] By preprocessing multi-channel sensor time-series monitoring data to obtain univariate health status sequences and health status soft labels, high-dimensional complex sensor data can be transformed into quantitative indicators that reflect the degradation status of equipment. This reduces data dimensional redundancy and noise interference, and improves the training efficiency and evaluation accuracy of subsequent deep learning models.

[0095] In some specific embodiments, this invention uses the root mean square (RMS) value of horizontal vibration signals to construct bearing health curves. The three-stage variation characteristics of the RMS value of bearing vibration signals with the number of samples—slow growth, fluctuating rise, and sharp increase—correspond to the stable operation, gradual degradation, and sudden failure processes of the bearing, respectively. The health value constructed based on RMS is precisely divided into four intervals: healthy (HI=0.75-1.0), good (HI=0.5-0.75), degraded (HI=0.25-0.5), and failed (HI=0.0-0.25). This four-part evaluation system achieves a detailed characterization of the entire process of marine bearings from health to failure, providing accurate basis for ship operation and maintenance decisions.

[0096] In some specific embodiments, this invention employs Fuzzy C-Means (FCM) clustering to softly partition the standardized health status sequence, generating a soft label sequence. The FCM clustering process includes: inputting the standardized health status sequence, data preprocessing and parameter initialization, constructing a fuzzy membership matrix, calculating cluster centers, iterative optimization, convergence judgment, and finally outputting the soft label matrix. In this invention, the number of clusters c=4, corresponding to four state levels: healthy, good, deteriorated, and faulty. The specific steps are as follows:

[0097] Step S2.1: Construct the FCM objective function, as shown in the following formula:

[0098] ;

[0099] In the formula, The objective function value is represented by U; the fuzzy membership matrix is ​​represented by the formula: R represents the set of real numbers; T represents the length of the sample points in a single sequence; This represents the membership degree of the i0th sample to the j0th cluster center; Represents the set of cluster centers; This represents the fuzzy weighting index, used to control the degree of fuzziness in clustering; Denotes the Euclidean norm; represents the health value of the i0th sample at time t, i.e., the standardized health index, which is an element in the standardized health sequence; c represents the number of clusters. Let j represent the j0th cluster center; T represents the length of the sample points in a single sequence.

[0100] Step S2.2: Construct the cluster center update formula as follows:

[0101] ;

[0102] Step S2.3: Construct the membership update formula as follows:

[0103] ;

[0104] In the formula, This represents the k0th cluster center.

[0105] Step S2.4: Construct the convergence criterion for FCM, as shown in the following formula:

[0106] ;

[0107] In the formula, This indicates the preset convergence threshold; Indicates the first The membership matrix of the next iteration; Indicates the first The membership matrix of the next iteration;

[0108] Step S2.5: Iteratively solve the objective function in step S2.1 using the cluster center update formula constructed in step S2.2 and the membership update formula constructed in step S2.3 until the convergence criterion in step S2.4 is satisfied, and output the soft label sequence. ,in ,and .

[0109] This invention achieves soft clustering partitioning by minimizing the objective function, solves the objective function through iterative optimization, and realizes accurate characterization of the progressive degradation process of marine bearings through the FCM soft label partitioning step, thus solving the technical problem of traditional hard label partitioning methods ignoring the state fuzzy boundary.

[0110] Step S3: Extract and fuse the temporal features of the soft-label sequence obtained in step S2 to obtain bidirectional temporal features.

[0111] In some specific embodiments, the present invention employs a temporal convolutional network (TCN) and a bidirectional long short-term memory network (BiLSTM) to extract and fuse temporal features from soft-label sequences, thereby obtaining bidirectional temporal features. For example... Figure 3 As shown in the figure, the temporal feature extraction module incorporates a temporal convolutional network (TCN), which sequentially includes causal convolutional layers, dilated convolutional layers, and residual connection structures. The residual connection part specifically includes weight normalization, a 1×1 convolution operation, and a ReLU activation function. Causal convolution ensures temporal causality, dilated convolution expands the receptive field by introducing a dilation rate, and the residual structure effectively alleviates the gradient vanishing problem in deep networks, achieving accurate capture of local fluctuation patterns in healthy state soft-label sequences. The specific steps are as follows:

[0112] S3.1 The causal convolutional layers in the Temporal Convolutional Network (TCN) obtain temporal causal convolutional features based on the soft label sequence obtained in step S2, as shown in the following formula:

[0113] ;

[0114] In the formula, y 1t represents the output at time t, and represents the feature vector composed of the causal convolutional features at time t; This represents the input sequence at time t-k0, i.e., the values ​​in the soft-label sequence; represents the convolution kernel; K represents the size of the convolution kernel.

[0115] S3.2. The dilated convolutional layers in the Temporal Convolutional Network (TCN) introduce a dilation rate to expand the receptive field, as shown in the following formula:

[0116] ;

[0117] In the formula, y 2t represents the output at time t, represents the dilated feature vector at time t, i.e., the dilated convolution feature; d represents the dilation rate.

[0118] S3.3 In the Temporal Convolutional Network (TCN), the residual connection layer performs residual connection operations based on the dilated convolution features to obtain temporal local features, as shown in the following formula:

[0119] ;

[0120] In the formula, o is the output of the residual connection layer, and is the element that makes up the temporal local features; Activation is the ReLU activation function; This represents a 1×1 convolution kernel parameter matrix; denoted as the residual function; x represents the input feature, i.e., the dilated convolution feature obtained in step S3.2.

[0121] S3.4. The forward LSTM in the Bidirectional Long Short-Term Memory (BiLSTM) network processes the temporal local features output by the temporal convolutional network (TCN) in a forward time series manner to obtain the forward hidden state, as shown in the following formula:

[0122] ;

[0123] In the formula, This represents the forward hidden state at time t; Represents the forward hidden state at time t-1; x t This represents the input at time t, that is, the element at time t in the local features of the time sequence;

[0124] S3.5. In the Bidirectional Long Short-Term Memory (BiLSTM) network, the backward LSTM processes the temporal local features of the output of the temporal convolutional network (TCN) in a forward time series manner to obtain the backward hidden state, as shown in the following formula:

[0125] ;

[0126] In the formula, This represents the backward hidden state at time t; Represents the backward hidden state at time t-1; x t This represents the input at time t, that is, the element at time t in the local features of the time sequence;

[0127] Step S3.6: Perform bidirectional output fusion on the forward hidden state obtained in step S3.4 and the backward hidden state obtained in step S3.5 to obtain the memory unit output, as shown in the following formula:

[0128] ;

[0129] In the formula, h t This represents the output of the memory unit at time t; Indicates a vector concatenation operation;

[0130] Step S3.7: The memory unit output obtained in step S3.6 is sequentially processed through forget gate calculation, input gate calculation, candidate state calculation, cell state update, output gate calculation, and hidden state calculation to obtain bidirectional temporal features, as shown in the following formula:

[0131] ;

[0132] ;

[0133] ;

[0134] ;

[0135] ;

[0136] ;

[0137] In the formula, The activation value for the forget gate; For activation functions; This represents the weight vector of the forget gate; Represents convolution operation; h t-1 x is the output of the memory unit at time t-1; t for The input to the time-series memory unit is the element at time t in the local features of the time sequence; For the bias of the forget gate; i t This is the activation value of the input gate; The weight vector of the input gate; The bias is used to control the input gate; These are candidate states for memory cells; The weight vector represents the candidate states of the input gate. The bias is used to set the candidate states of the input gate; for The state of the memory unit at all times; This is the activation value of the output gate; Let be the weight vector of the output gate. For the output gate bias; h 1t This represents the output of the hidden state at time t, i.e., the bidirectional temporal characteristic.

[0138] Step S3.8: Map the bidirectional temporal features to the original output scores of various states through a fully connected layer, as shown in the following formula:

[0139]

[0140] In the formula, This represents the state score vector at time t; This represents the weight matrix of the fully connected layer, with dimension 1. (in Bidirectional time series characteristics (4 represents the number of states in the four categories of health, good, deterioration, and failure as defined in this invention). This represents the bias vector.

[0141] By employing the temporal convolutional network (TCN) feature extraction step, the system accurately captures local fluctuation patterns in the health status soft-label sequence, solving the technical problem that existing methods struggle to capture local changes in the health status of ship equipment. Through the synergistic effect of the forward LSTM, backward LSTM, and bidirectional output fusion of the bidirectional long short-term memory network (BiLSTM) feature fusion module, the system achieves a comprehensive understanding of the forward and backward context of the health status time-series data, effectively capturing long-distance temporal dependencies. This solves the modeling challenges of nonlinearity, time-varying nature, and uncertainty in the state evolution of marine bearings in complex operating environments, and enhances the model's ability to model long-term dependencies in noisy environments.

[0142] Step S4: Based on the bidirectional time series features obtained in step S3, the prediction uncertainty is quantified using the Monte Carlo (MCD) mechanism and the health status prediction probability is generated through the Softmax function to obtain the health status prediction result of the marine bearing.

[0143] Step S4.1: Perform N forward propagations on the bidirectional temporal features to obtain N prediction results, as shown in the following formula:

[0144] ;

[0145] In the formula, This represents the prediction result of the nth forward propagation; This represents the subset of parameters generated by the randomness of Dropout during the nth forward propagation; This represents the input, namely the bidirectional temporal features obtained in step S3;

[0146] Step S4.2: Based on the prediction results obtained in step S4.1, calculate its mean and variance, using the following formulas:

[0147] ;

[0148] ;

[0149] In the formula, This represents the predicted mean. This represents the variance of the prediction; the larger the variance, the higher the uncertainty of the model's prediction for that sample.

[0150] Step S4.3: Based on the single prediction result, construct the final prediction result, i.e., the average probability distribution, as shown in the following formula:

[0151] ;

[0152] In the formula, p represents the average probability distribution, i.e., the final prediction result; This represents the probability distribution of the nth prediction; This represents the probability that a sample belongs to class c in the nth prediction.

[0153] Step S4.4: Based on the final prediction results, obtain the prediction entropy and mutual information as uncertainties, as shown in the following formulas:

[0154] ;

[0155] ;

[0156] In the formula, This indicates cognitive uncertainty; Indicates randomness and uncertainty; Represents mutual information; This represents the average predicted probability that a sample belongs to class c.

[0157] Step S4.5: Based on the bidirectional time-series features obtained in step S3, generate the health status prediction probability using the Softmax function, as shown in the following formula:

[0158] ;

[0159] In the formula, The vector obtained in step S3.8 represents The j0th element in the equation is the unnormalized output score of the model for the j0th class of states. This score reflects the confidence level of the model in believing that the sample belongs to a certain class under the current parameters. The vector obtained in step S3.8 represents The k0th element in the array; c represents the number of categories, which is 4 in this invention; y t The predicted state category at time t is obtained from step S3.

[0160] Step S4.6: Output the prediction result of the marine bearing health status, which is composed of the prediction entropy and mutual information obtained in step S4.4 and the health status prediction probability obtained in step S4.5.

[0161] By quantifying the uncertainty of the Monte Carlo (MCD) mechanism and utilizing the output of the Softmax function, a multi-class probabilistic assessment and confidence quantification of health status are achieved. This provides rich auxiliary information for marine bearing operation and maintenance decisions, helping to formulate differentiated and refined maintenance strategies. Particularly in noisy environments, it effectively distinguishes between high-confidence and low-confidence samples. The Softmax function output enables a multi-class probabilistic assessment of health status, providing rich auxiliary information for marine bearing operation and maintenance decisions and helping to formulate differentiated and refined maintenance strategies.

[0162] In some specific embodiments, the present invention also provides a system for performing the marine bearing health status assessment method described herein, comprising:

[0163] The data acquisition and preprocessing module is used for the acquisition and cleaning of multi-source sensor data;

[0164] A health score construction module is used to generate univariate health score sequences using principal component analysis (PCA) and weighted distance mechanisms.

[0165] The soft label segmentation module is used to generate probabilistic soft labels that reflect gradual degradation based on the FCM algorithm.

[0166] The temporal feature extraction and fusion module is used to extract multi-scale spatiotemporal features through the TCN-BiLSTM two-level architecture;

[0167] The parameter optimization module is used to achieve uncertainty quantification by using the Beluga Whale Optimization (BWO) algorithm for adaptive optimization and combining it with the Monte Carlo (MCD) mechanism.

[0168] like Figure 2 As shown, in some preferred embodiments, the system used to perform the marine bearing health status assessment method of the present invention is further refined, and the MCD-FT-BiLSTM model includes...

[0169] The data acquisition and preprocessing module is used to collect and preprocess marine bearing operating status data through multi-channel sensors.

[0170] The health status construction module processes multi-channel sensor data to construct a standardized health status sequence;

[0171] The soft label partitioning module performs soft partitioning on the standardized health sequence based on the FCM algorithm, generating a soft label sequence containing the membership probabilities of four states.

[0172] The temporal feature extraction module extracts local temporal features of the soft-label sequence through a TCN network, wherein the TCN includes causal convolution, dilated convolution and residual connection structures;

[0173] The feature fusion module is used to fuse the forward and backward temporal dependencies of the TCN output features through the BiLSTM network to generate bidirectional temporal features;

[0174] The parameter optimization module is used to adaptively optimize the hyperparameters of the TCN-BiLSTM model using the BWO algorithm.

[0175] The health assessment output module is used to quantify the prediction uncertainty through the Monte Carlo (MCD) mechanism, and output the bearing health status assessment results through the Softmax function, generating an assessment report containing four levels: healthy, good, degraded, and faulty.

[0176] By inputting the health status soft label sequence into the trained MCD-FT-BiLSTM model to output the health status assessment result, the accurate identification of the health status of marine bearings using a multi-level feature extraction structure is achieved. This can efficiently generate high-precision health assessment results and reduce human judgment errors, thereby improving the overall diagnostic performance and practicality of the system. In particular, it exhibits good robustness in noisy environments.

[0177] In addition to the above embodiments, the present invention may have other implementation methods. All technical solutions formed by equivalent substitution or equivalent transformation fall within the protection scope claimed by the present invention.

Claims

1. A method for assessing the health status of marine bearings, characterized in that: Includes the following steps: S1. Collect marine bearing operating status data through multi-channel sensors, and obtain a standardized health sequence after preprocessing and feature fusion; S2. Perform soft partitioning on the standardized health sequence to obtain a soft label sequence; S3. Extract and fuse the temporal features of the soft-label sequence to obtain bidirectional temporal features; S4. Based on the bidirectional time series characteristics, the Monte Carlo mechanism is used to quantify the prediction uncertainty and the Softmax function is used to generate the health status prediction probability to obtain the health status prediction result of marine bearings.

2. The method for assessing the health status of marine bearings according to claim 1, characterized in that: In step S1, principal component analysis is used to preprocess and fuse the marine bearing operating status data to obtain a standardized health sequence. The specific steps are as follows: S1.1 Principal component analysis is used to reduce the dimensionality of the marine bearing operating status data to obtain a comprehensive feature sequence; S1.

2. Based on the comprehensive feature sequence, calculate the weighted distance between the i-th health feature and other features at each time point; S1.

3. Constructing weighting coefficients using an exponential decay function: S1.

4. Based on the weighting coefficients obtained in step S1.3, the health characteristics are weighted and fused to obtain standardized health indicators. The standardized health indicators are then arranged in chronological order to obtain a standardized health sequence.

3. The method for assessing the health status of marine bearings according to claim 2, characterized in that: In step S2, fuzzy C-means clustering is used to perform soft partitioning on the standardized health sequence to generate a soft label sequence. The specific steps are as follows: S2.1 Construct the FCM objective function, as shown in the following formula: ； In the formula, U represents the objective function value; U represents the fuzzy membership matrix. This represents the membership degree of the i0th sample to the j0th cluster center; The cluster center set is represented by m; the fuzzy weighting index is represented by m. Denotes the Euclidean norm; This represents the standardized health index of the i0th sample at time t; c represents the number of cluster categories; This represents the j0th cluster center; T represents the length of the sample points in a single sequence. S2.2 Construct the cluster center update formula as follows: ； S2.3 Construct the membership update formula as follows: ； In the formula, This represents the k0th cluster center; S2.4 Construct the convergence criterion for FCM, the formula is as follows: ； In the formula, This indicates the preset convergence threshold; Indicates the first The membership matrix of the next iteration; Indicates the first The membership matrix of the next iteration; S2.

5. Iteratively solve the objective function in step S2.1 using the cluster center update formula constructed in step S2.2 and the membership update formula constructed in step S2.3 until the convergence criterion in step S2.4 is satisfied, and output the soft label sequence.

4. The method for assessing the health status of marine bearings according to claim 3, characterized in that: In step S3, a temporal convolutional network and a bidirectional long short-term memory network are used to extract and fuse temporal features of the soft-label sequence to obtain bidirectional temporal features. The specific steps are as follows: S3.

1. Based on the soft-label sequence, obtain the temporal causal convolutional features, as shown in the following formula: ； In the formula, y 1t This represents the output at time t; This represents the input sequence at time t-k0; K represents the convolution kernel; K represents the size of the convolution kernel. S3.2, The dilated convolutional layer introduces a dilation rate, as shown in the following formula: ； In the formula, y 2t d represents the dilated convolution feature at time t; d represents the dilation rate. S3.3 The residual connection layer performs residual connection operations based on the dilated convolution features to obtain temporal local features, as shown in the following formula: ； In the formula, o is the output of the residual connection layer; Activation is the ReLU activation function; This represents a 1×1 convolution kernel parameter matrix; Represents the residual function; x represents the dilated convolution feature; S3.

4. The forward LSTM processes the temporal local features in the forward direction of the time series to obtain the forward hidden state; S3.

5. Backward LSTM processes local temporal features in the forward direction of the time series to obtain the backward hidden state; S3.

6. Perform bidirectional output fusion on the forward hidden state and the backward hidden state to obtain the memory unit output; S3.

7. The output of the memory unit is sequentially processed through forget gate calculation, input gate calculation, candidate state calculation, cell state update, output gate calculation, and hidden state calculation to obtain bidirectional temporal features, as shown in the following formula: ； ； ； ； ； ； In the formula, The activation value for the forget gate; For activation functions; This represents the weight vector of the forget gate; Represents convolution operation; h t-1 x is the output of the memory unit at time t-1; t for Input to the time-of-flight memory unit; For the bias of the forget gate; i t This is the activation value of the input gate; The weight vector of the input gate; The bias is used to control the input gate; These are candidate states for memory cells; The weight vector represents the candidate states of the input gate. The bias is used to set the candidate states of the input gate; for The state of the memory unit at all times; This is the activation value of the output gate; Let be the weight vector of the output gate. For the output gate bias; h 1t This represents the output of the hidden state at time t; S3.

8. The bidirectional temporal features are mapped to the original output scores of various states through a fully connected layer, as shown in the following formula: ； In the formula, This represents the state score vector at time t; This represents the weight matrix of the fully connected layer; This represents the bias vector.

5. The method for assessing the health status of marine bearings according to claim 4, characterized in that: The specific steps of step S4 are as follows: S4.1 Perform N forward propagations on the bidirectional temporal features to obtain N prediction results; S4.

2. Based on the prediction results obtained in step S4.1, calculate their mean and variance; S4.

3. Based on the results of a single prediction, construct the final prediction result using the following formula: ； In the formula, p represents the final prediction result; This represents the probability distribution of the nth prediction; This represents the probability that a sample belongs to class c in the nth prediction. S4.

4. Based on the final prediction results, the prediction entropy and mutual information are obtained as uncertainties, as shown in the following formulas: ； ； In the formula, This indicates cognitive uncertainty; Indicates randomness and uncertainty; Represents mutual information; This represents the average predicted probability that a sample belongs to class c. S4.

5. Based on the bidirectional time series characteristics, the health status prediction probability is generated using the Softmax function, as shown in the following formula: ； In the formula, The vector obtained in step S3.8 represents The j0th element in; The vector obtained in step S3.8 represents The k0th element in the array; c represents the number of categories; y t The predicted state category at time t is obtained from step S3; S4.6 Output the prediction result of the marine bearing health status, which is composed of the prediction entropy and mutual information obtained in step S4.4 and the health status prediction probability obtained in step S4.

5.

6. The method for assessing the health status of marine bearings according to claim 5, characterized in that: Before step S1, step S0 is included: constructing a marine bearing health status dataset, training the MCD-FT-BiLSTM model, and obtaining the trained MCD-FT-BiLSTM model. The specific steps are as follows: Step S0.1: Construct a marine bearing health status dataset; Step S0.2: Divide the dataset into a training set, a validation set, and a test set; Step S0.3: Initialize the parameters of the MCD-FT-BiLSTM model; Step S0.4: Construct a training sample set under multiple noise levels; Step S0.5: Adaptively optimize the key hyperparameters of the MCD-FT-BiLSTM model using the White Whale Optimization Algorithm to obtain the optimized parameter combination, thereby obtaining the optimized MCD-FT-BiLSTM model. Step S0.6: Input the training set into the optimized MCD-FT-BiLSTM model and output the health status prediction results; Step S0.7: Calculate the training loss using the classification cross-entropy loss function, and update the network parameters using the backpropagation algorithm; Step S0.8: Introduce label smoothing technology during training to alleviate overfitting. Step S0.9: Monitor model performance using the validation set. When the accuracy on the validation set no longer improves for 5 consecutive training cycles, or when the change in training loss value is less than the preset threshold of 0.001, stop training and obtain the trained MCD-FT-BiLSTM model.

7. The method for assessing the health status of marine bearings according to claim 6, characterized in that: The specific steps of the beluga optimization algorithm in step S0.5 are as follows: S0.5.1 Initialize the beluga whale population, using the following formula: ； In the formula, Let be the i-th hyperparameter combination; LB and UB are the lower and upper bounds of the parameter search, respectively; rand(0,1) represents a random number in the interval [0,1]. S0.5.

2. Take the accuracy of the validation set as the fitness value, and then update the position, as shown in the following formula: ； ； In the formula, This is the current optimal solution; , and Both are coefficient vectors; Represents the combination of hyperparameters at time t; This represents the combination of hyperparameters at time t+1; S0.5.

3. Employ spiral update to enhance local fine-grained search, as shown in the following formula: ； In the formula, b is a constant that defines the spiral shape; l is a random number between [-1, 1]. S0.5.

4. After reaching the required number of iterations, output the optimized parameter combination to obtain the optimized MCD-FT-BiLSTM model.

8. The method for assessing the health status of marine bearings according to claim 6, characterized in that: In step S0.2, the ratio of the training set, validation set, and test set is 7:2:

1.

9. The method for assessing the health status of marine bearings according to claim 6, characterized in that: The multi-channel sensor in step S1 includes an accelerometer for detecting vertical and horizontal vibration signals, a temperature sensor for detecting temperature, and a pressure sensor for detecting pressure.

10. An assessment system for implementing the health status assessment method for marine bearings according to any one of claims 6 to 9, characterized in that: include The data acquisition and preprocessing module is used to collect and preprocess marine bearing operating status data through multi-channel sensors. The health status construction module processes multi-channel sensor data to construct a standardized health status sequence; The soft label segmentation module performs soft segmentation on the standardized health score sequence based on fuzzy C-means clustering to generate a soft label sequence. The temporal feature extraction module extracts local temporal features of the soft-label sequence through a temporal convolutional network; The feature fusion module is used to fuse the forward and backward temporal dependencies of the output features of the temporal convolutional network through a bidirectional long short-term memory network to generate bidirectional temporal features; The parameter optimization module is used to adaptively optimize the hyperparameters of the model using the White Whale Optimization Algorithm. The health assessment output module is used to quantify the prediction uncertainty through the Monte Carlo mechanism and output the bearing health status assessment results through the Softmax function.

Images