Ship engine component cross-working-condition life prediction method and system

Abstract

The invention discloses a ship engine component cross-working-condition life prediction method and system, and the method comprises the steps: collecting the operation data of key components of a ship engine in real time, and carrying out the preprocessing of the operation data of each component, so as to guarantee the data synchronism and data validity of the operation data; inputting the preprocessed operation data into the corresponding trained life prediction network model, and outputting a life prediction result corresponding to each component; the life prediction network model comprises a feature extraction module used for extracting time sequence features under different scales, splicing all the time sequence features on a feature dimension, performing feature enhancement on a splicing result based on a Transform model, and outputting a global feature vector after compression of a global average pooling layer; the life prediction module is used for outputting a life prediction result according to the global feature vector; the domain discrimination module is used for outputting a domain discrimination result according to the global feature vector; in training, the feature extraction module and the domain discrimination module carry out adversarial training.

Description

Technical Field

[0001] This invention belongs to the field of cross-condition life prediction technology for marine engine components, specifically relating to a method and system for cross-condition life prediction of marine engine components. Background Technology

[0002] In today's era of deep integration between intelligent manufacturing and the shipping industry, the safe operation and efficient maintenance of marine engines are facing unprecedented challenges. As a core component of the marine propulsion system, the operating status of key engine components (such as pistons and cylinder liners) directly affects the reliability and economy of the ship. In reality, the marine environment and operating conditions faced by ships vary, such as load size, fuel quality differences, speed, and route, causing significant fluctuations in the condition monitoring signals of marine engines under different operating conditions, which in turn leads to significant differences in the distribution of engine life data under different operating conditions.

[0003] However, current mainstream lifespan prediction methods, especially deep learning-based models, largely rely on the premise that "training and testing data are identically distributed," failing to adequately consider the data offset problem between different operating conditions. The dataset with known lifespan labels during training is called the source domain, and the dataset consisting of unlabeled or minimally labeled data generated by ship engine components under new operating conditions is called the target domain. Data offset between different operating conditions can lead to inaccurate model predictions. That is, a ship engine lifespan prediction model trained in the source domain may not perform well when placed in a new production environment (new operating conditions) using data from the target domain. In other words, these models generally suffer from weak generalization ability and unstable predictions when facing lifespan prediction tasks across operating conditions, severely limiting their practicality in real-world ship environments. Therefore, researching an intelligent lifespan prediction method with cross-operating-condition modeling capabilities is crucial. In recent years, domain adaptation and transfer learning techniques have shown great potential in handling the problem of differences in the distribution between the source and target domains. For example: The invention patent "A Method and System for Predicting the Life of Rolling Bearings Across Working Conditions" (Publication No. CN119047116A, Application Date 20240703) applied for by Cen Zilang, Hu Shaolin, Sun Guoxi, et al. proposes a method for predicting the life of rolling bearings across working conditions. Specifically, by collecting vibration signals under different working conditions, and after time-domain, frequency-domain, and time-frequency-domain feature extraction and filtering, a segmented degradation model is constructed using fuzzy C-means clustering as a label. A domain adaptive hybrid model is designed (combining a sequence decomposition MLP feature extractor, a dilated convolution RUL (Remaining Useful Life) predictor, and a domain discriminator). Through maximum mean discrepancy (MMD) and adversarial training, the feature distributions of the source domain and the target domain are aligned, ultimately achieving high-precision prediction of the remaining life with only a small amount of target domain data. While the cross-condition bearing life prediction method proposed in this patent is innovative, it still has several shortcomings: the piecewise degradation model constructed by relying on fuzzy C-means clustering may not accurately reflect the nonlinear characteristics of complex degradation processes; the feature extractor, which uses a fixed-parameter sequence decomposition method, has limited adaptability to non-stationary vibration signals and may lose transient fault features; and using only simple wavelet packets and frequency domain features may miss crucial information about early, subtle bearing faults.

[0004] The invention patent "A Method for Predicting Remaining Useful Life (RUL) for Applications Across Multiple Operating Conditions" (Publication No. CN118171240B, Application Date: 20240412) filed by Lü Yi, Shen Zaichen, Zhou Ningxu, et al., proposes a method for predicting RUL for applications across multiple operating conditions. Specifically, it separates the common features and private features of the source and target domain data, and uses a deep feature adaptation module to align the common features from both global and local distribution perspectives. Furthermore, it utilizes the InfoNCE (Information Noise-Contrastive Estimation Loss) loss function in contrastive learning to maximize the mutual information between the input data and the features, ensuring the representativeness of the features. Finally, a regression prediction module achieves high-precision RUL prediction. However, this invention patent has the following drawbacks: it has high computational complexity, involving joint optimization of multiple modules such as bidirectional LSTM, feature separation, dynamic distribution alignment and contrastive learning, which increases training time and resource consumption; it is highly dependent on the pseudo-label of the target domain, and if the accuracy of the subdomain discriminator is insufficient in the early stage, it will lead to error accumulation; the hyperparameter (such as adaptive coefficient, trade-off coefficient, etc.) adjustment strategy is complicated, and it is difficult to tune parameters in practical applications.

[0005] The invention patent "Prediction Method for Remaining Service Life of Bearings under Migrating Operating Conditions Based on Contrastive Dual-Adversarial Mechanism" (publication date: CN119647012B, application date: 20241128) filed by Miao Xiaodong, Sun Aoyun, Ji Xinpeng, et al., proposes a method for predicting the remaining service life of bearings under migrating operating conditions based on a contrastive dual-adversarial mechanism. The main steps include: firstly, acquiring and preprocessing bearing vibration signals; secondly, decomposing and reconstructing the signals using the CEEMDAN (Complete Ensemble Empirical Mode Decomposition with Adaptive Noise) method; and finally... The invention employs a Hidden Markov Model (HMM) with Bayesian inference to detect fault points. Secondly, it constructs samples containing temporal information and designs a Dynamic Temporal-Spatial Convolutional Long Short-Term Memory Neural Network (DTSConvLSTM-NN) as a feature extractor, combined with a Multilayer Perceptron (MLP) to predict remaining lifetime. Finally, by comparing a bi-adversarial training framework, it trains the model using source and target domain data, introducing a domain discriminator and a contrastive loss module to extract domain-invariant features, thus improving cross-condition prediction capabilities. However, this invention has the following drawbacks: high computational complexity (CEEMDAN signal decomposition, bi-adversarial training, and the attention mechanism of the dynamic convolutional LSTM all require significant computational resources, limiting real-time applications); reliance on fault point detection accuracy, misjudgments by the HMM directly affect the reliability of subsequent lifetime predictions; and limitations in signal decomposition, as CEEMDAN's processing of non-stationary signals may introduce residual noise, affecting reconstruction quality. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, this invention provides a method for predicting the lifespan of marine engine components across operating conditions. The technical problem to be solved by this invention is achieved through the following technical solution: In a first aspect, embodiments of the present invention provide a method for predicting the lifespan of marine engine components across operating conditions, the method comprising: Real-time acquisition of operational data from key components of the ship's engine, and preprocessing of the operational data for each component to ensure data synchronization and validity; The preprocessed running data is input into the corresponding trained lifetime prediction network model, and the lifetime prediction result for each component is output. Each component corresponds to a trained lifetime prediction network model, and the network structure of the trained lifetime prediction network models is the same. The lifetime prediction network model includes a feature extraction module, a lifetime prediction module, and a domain discrimination module. The feature extraction module extracts temporal features from the input data at different scales, concatenates all temporal features along the feature dimension, and enhances the concatenated result using a Transformer model. The enhanced feature result is then compressed by a global average pooling layer to output a global feature vector. The lifetime prediction module performs predictions based on the global feature vector and outputs the lifetime prediction result. The domain discrimination module performs discrimination based on the global feature vector and outputs the domain discrimination result. During training, the total loss consists of the lifetime prediction task loss and the domain discrimination task loss. Based on this total loss, adversarial training is performed on the feature extraction module and the domain discrimination module in the lifetime prediction network model to obtain the trained lifetime prediction network model.

[0007] In one embodiment of the present invention, the operating data of each component is preprocessed, including: The process for executing the operational data of each component includes: removing abnormal data from the operational data of that component to obtain first processed data; filling in missing values ​​in the first processed data to obtain second processed data; performing smoothing and denoising processing on the second processed data to obtain third processed data; performing Min-Max normalization processing on the third processed data to obtain fourth processed data; performing data synchronization processing on the fourth processed data to obtain fifth processed data; and performing feature filtering processing on the fifth processed data to obtain sixth processed data.

[0008] In one embodiment of the present invention, missing values ​​in the first processed data are filled to obtain second processed data, including: The missing values ​​in the first processed data are filled in using a forward padding method to obtain the second processed data.

[0009] In one embodiment of the present invention, the fourth processed data is subjected to data synchronization processing to obtain fifth processed data, including: The target resampling period is calculated based on the actual sampling frequency of different components in the fourth processed data. Based on the target resampling period, resampling processing is performed on different components in the fourth processed data; The resampled fourth-processed data is time-aligned and merged to obtain the fifth-processed data.

[0010] In one embodiment of the present invention, based on the target resampling period, resampling processing is performed on different components in the fourth processed data, including: For each component in the fourth processed data, the execution process includes: if the actual sampling frequency of the component is greater than or equal to the target resampling period, then the component is downsampled; otherwise, data filling sampling is performed using linear interpolation.

[0011] In one embodiment of the present invention, the feature extraction module includes three feature extraction channels, a concatenation layer, a Transformer model, and a global average pooling layer; wherein, Each feature extraction channel consists of a convolutional layer, a batch normalization layer, an activation function layer, and an LSTM layer connected in sequence. The convolutional layer serves as the input to the corresponding feature extraction channel. The kernel size of the convolutional layer in each feature extraction channel is different, and the activation function used in the activation function layer is different. The input of the splicing layer is connected to the output of the three feature extraction channels respectively. The splicing layer, the Transformer model and the global average pooling layer are connected in sequence. The output of the global average pooling layer is used as the output of the feature extraction module.

[0012] In one embodiment of the present invention, the lifetime prediction module includes a linear transformation layer, an activation function layer, and a linear transformation layer connected in sequence.

[0013] In one embodiment of the present invention, the domain discrimination module includes a GRL layer, a linear transformation layer, a batch normalization layer, an activation function layer, a linear transformation layer, and a discrimination output layer connected in sequence.

[0014] Secondly, embodiments of the present invention provide a cross-condition life prediction system for ship engine components. The system includes a server deployed with a JAVA backend, a first database, a second database, a frontend, and a pre-trained life prediction network model from the cross-condition life prediction method for ship engine components described in the first aspect, corresponding to key ship engine components. The first database is used to store operational data of key components of ship engines; The JAVA backend is used to periodically read the operating data of key components of the ship engine from the first database, preprocess the operating data of the key components to ensure the data synchronization and validity, and input the preprocessed operating data into the corresponding trained life prediction network model. The lifetime prediction network model is used to output corresponding lifetime prediction results based on the preprocessed running data. The second database is used to store life prediction results for key components of ship engines; The front end is used to read and display the life prediction results of key components of the ship engine from the second database.

[0015] The beneficial effects of this invention are: This invention proposes a cross-condition life prediction method for ship engine components. It designs life prediction network models with identical network structures for key ship engine components. By introducing multi-scale feature extraction, attention mechanisms, and adversarial training, the life prediction network models enhance their adaptability to different data distributions under various operating conditions. This achieves feature alignment and accurate life prediction for key ship engine components across operating conditions, significantly improving the model's generalization ability and stability in complex environments. The proposed method employs adversarial training during training, effectively utilizing historical operational data with life labels and limited operational data without life labels under new operating conditions to establish a robust life prediction network model. This provides reliable remaining service life predictions for key ship engine components, supports predictive maintenance decisions, avoids unplanned downtime, improves navigation safety, and reduces maintenance costs. The proposed method can effectively grasp the health status of key ship engine components, provide early warnings of potential faults, reduce the risk of unplanned maintenance, extend component service life, and improve ship operational safety and maintenance management efficiency, demonstrating significant engineering application value and economic benefits.

[0016] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a method for predicting the lifespan of a ship engine component across operating conditions, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of piston operation data provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the lifetime prediction network model provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of a cross-condition life prediction system for ship engine components provided in an embodiment of the present invention. Detailed Implementation

[0018] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0019] Firstly, please see Figure 1 This invention provides a method for predicting the lifespan of ship engine components across operating conditions, specifically including the following steps: S10. Real-time acquisition of operational data of key components of the ship's engine, and preprocessing of the operational data of each component to ensure data synchronization and validity.

[0020] Marine engines consist of several key components, such as cylinder liners, cylinder heads, pistons, cylinder head exhaust valves, cylinder head starting valves, fuel injectors, and high-pressure fuel pumps. For each component, its corresponding operating parameters can be collected separately. At the same time, the operating parameters of other related components that affect its operating status should also be considered. Taking the piston as an example... Figure 2 As shown, the collected operating data includes not only physical variables (such as temperature, pressure, vibration, and stress), but also other engine operating parameters, such as workload, crankshaft speed, and coolant temperature. These parameter values ​​collected at different times together constitute the operating data of this component.

[0021] The collected data can be timestamped and stored in the database for subsequent data analysis and processing.

[0022] Furthermore, since the meanings, formats, and types of the various operating parameters in the collected operating data vary significantly, data preprocessing is required to standardize the data. This embodiment of the invention preprocesses the operating data for each component, including: performing the following process for each component's operating data: removing abnormal data from the component's operating data to obtain first processed data; filling in missing values ​​in the first processed data to obtain second processed data; performing smoothing and denoising processing on the second processed data to obtain third processed data; performing Min-Max normalization processing on the third processed data to obtain fourth processed data; performing data synchronization processing on the fourth processed data to obtain fifth processed data; and performing feature filtering processing on the fifth processed data to obtain sixth processed data. More specifically: For outlier removal: In lifespan prediction tasks, the collected operational data of ship engine components often contain outliers or noise points caused by factors such as sensor failure, communication interference, and human error, or some operational data may exceed the variation range, resulting in unreasonable extreme values. These outliers can interfere with degradation trend modeling and even guide the model to learn incorrect patterns, thus severely affecting the accuracy and stability of lifespan prediction. Therefore, it is necessary to remove these outliers. For unreasonable extreme values, it is necessary to filter them according to the numerical range specified in the sensor's parameter description, directly removing extreme values ​​and setting them as null values. This allows for the preparation of the first processed data after outlier removal.

[0023] For handling missing values: During equipment operation, missing values ​​(NaN or null values ​​after outlier removal) often appear in the operating data due to sensor failures, communication interruptions, data acquisition delays, etc. Missing values ​​can lead to model training failure or decreased prediction performance, and in severe cases, can mask the degradation trend. For missing values ​​generated after outlier removal, this embodiment of the invention uses forward imputation, that is, using the previous valid observation value to fill the current missing value. In other words, forward imputation is used to fill the missing values ​​in the first processed data to obtain the second processed data. If the missing value appears at the beginning of the first processed data, the first valid value immediately following it is used for backward imputation.

[0024] For smoothing and denoising: In lifetime prediction tasks, equipment operating data is easily affected by high-frequency noise, electromagnetic interference, and sampling errors during the acquisition process, resulting in significant fluctuations and noise. Unprocessed raw signals can interfere with the extraction of degradation trends, making it difficult for the model to accurately learn the lifetime evolution process. Therefore, this invention proposes smoothing the second-processed data by taking the average value within a fixed time window to obtain the third-processed data. The fixed time window is a hyperparameter that can be dynamically set according to the data sampling frequency and noise characteristics. For data with a sampling interval of 1 second, the window size can be set to 10 seconds. The basic principle is that the fixed time window should be large enough to filter out noise, but not so large as to smooth out the true degradation trend changes.

[0025] For Min-Max normalization: a uniform feature scale is applied, linearly scaling the value of each feature to the [0, 1] interval. The normalization formula is as follows: ; in, The normalized values ​​of the original data. It is the original data value of a certain feature. It is the minimum value within that feature. It is the minimum value within that feature.

[0026] For data synchronization: During navigation, ships are affected by factors such as sea conditions (e.g., waves, ocean currents) and navigation tasks (e.g., acceleration, deceleration, turning). Data collection from different monitoring systems (e.g., vibration sensors, temperature sensors, speed sensors) may have inconsistent time bases. For example, when a ship encounters strong waves and rough seas, some sensors may experience temporary malfunctions or communication delays, leading to data timestamp discrepancies. Alternatively, although different sensors may have the same sampling frequency setting, slight differences in sampling periods can occur due to fluctuations in the ship's electrical system. Over time, this can cause multi-source data to become "out of sync" in the time dimension. Without targeted data synchronization processing for specific ship scenarios, subsequent time-series-based feature extraction and lifetime prediction will suffer from "parameter correspondence errors" due to the chaotic time base, severely impacting model accuracy.

[0027] Data synchronization aims to unify the sampling frequencies of sensors on different components of a ship's engine, preventing inconsistencies in time granularity from causing data misalignment from multiple sources and impacting subsequent feature fusion and model training. To address this issue, this invention proposes data synchronization processing for the fourth-processed data to obtain the fifth-processed data. This includes: calculating a target resampling period based on the actual sampling frequencies of different components in the fourth-processed data; resampling different components in the fourth-processed data based on the target resampling period; and merging the resampled fourth-processed data in time to obtain the fifth-processed data. Resampling different components in the fourth-processed data based on the target resampling period includes: for each component in the fourth-processed data, the process includes: if the actual sampling frequency of the component is greater than or equal to the target resampling period, downsampling is performed on the component; otherwise, linear interpolation is used for data filling sampling. More specifically: (1) Determine the target resampling period: Analyze the actual sampling frequency of all components in the fourth processing data, select a suitable sampling frequency from all time sampling frequencies to calculate the target resampling period, and construct a unified time axis. For example, in this embodiment of the invention, the target resampling period is calculated by selecting the sampling frequency range where most sampling frequency distributions are located, or by taking the median value, or by taking the average value.

[0028] (2) Resampling: Resampling is performed on each component in the fourth processing data. For components with an actual sampling frequency greater than or equal to the target resampling period, downsampling is performed by taking the average of all data within one target resampling period. For components with an actual sampling frequency less than the target resampling period, data padding is performed by linear interpolation, i.e., the two data points with the lower actual sampling frequency are used for data padding. and (corresponding time point is) and Interpolation between ) (corresponding time point) The following formula can be used: ; Available from The time value is taken once every other target resampling period.

[0029] (3) Time alignment and merging: After the target resampling period is determined and resampling is performed, all component data have been synchronized to a time axis. Then, all data are merged according to time points to form a multivariate aligned dataset with a complete time series structure.

[0030] For feature selection: The operational data of ship engine components often has a high dimensionality, containing numerous sensor variables of various types, many of which are irrelevant to the equipment's lifespan prediction or degradation process. These redundant or irrelevant features not only increase the complexity of model training, leading to overfitting, but also mask key degradation information, reducing prediction accuracy. To address this, this invention proposes using the Pearson correlation coefficient for feature selection, retaining features highly correlated with the remaining lifespan of the ship engine components while removing redundant features. Here, the threshold for the Pearson correlation coefficient can be set empirically to 0.5, i.e., retaining features with strong correlation.

[0031] The embodiments of the present invention, through a series of processes including removing abnormal data, handling missing values, smoothing and denoising, Min-Max normalization, data synchronization, and feature filtering, can ensure the data synchronization and validity of the collected operating data of ship engine components, thereby improving the accuracy and stability of life prediction.

[0032] S20. Input the preprocessed running data into the corresponding trained lifetime prediction network model, and output the lifetime prediction result for each component. Each component corresponds to a trained lifetime prediction network model, and the network structures of the trained lifetime prediction network models are identical. The lifetime prediction network model includes a feature extraction module, a lifetime prediction module, and a domain discrimination module. The feature extraction module is used to extract the temporal features of the input data at different scales, concatenate all temporal features along the feature dimension, and perform feature enhancement on the concatenated result based on the Transformer model. The feature enhancement result is compressed by a global average pooling layer and outputs a global feature vector. The lifetime prediction module is used to make predictions based on the global feature vector and output the lifetime prediction result. The domain discrimination module is used to make discriminations based on the global feature vector and output the domain discrimination result. During training, the total loss consists of the lifetime prediction task loss and the domain discrimination task loss. Based on the total loss, the feature extraction module and the domain discrimination module in the lifetime prediction network model are subjected to adversarial training to obtain the trained lifetime prediction network model.

[0033] To achieve more accurate lifespan prediction across operating conditions, this invention proposes a lifespan prediction network model, as follows: Figure 3 As shown, the system includes a feature extraction module, a lifetime prediction module, and a domain discrimination module. The feature extraction module comprises three feature extraction channels, a concatenation layer, a Transformer model, and a global average pooling layer. Each feature extraction channel consists of a convolutional layer, a batch normalization layer, an activation function layer, and an LSTM layer connected in sequence, with the convolutional layer serving as the input to the corresponding feature extraction channel. The kernel size of the convolutional layers in each feature extraction channel is different, and the activation function layers use different activation functions. The input of the concatenation layer is connected to the outputs of the three feature extraction channels. The concatenation layer, Transformer model, and global average pooling layer are connected in sequence, with the output of the global average pooling layer serving as the output of the feature extraction module. The lifetime prediction module comprises a linear transformation layer, an activation function layer, and a linear transformation layer connected in sequence. The domain discrimination module comprises a GRL layer, a linear transformation layer, a batch normalization layer, an activation function layer, a linear transformation layer, and a discrimination output layer connected in sequence.

[0034] Specifically: (1) Feature extraction module 1) Employing three channels (three types of convolutional kernels) to extract temporal features at different scales. Marine engines are complex dynamic systems, and the degradation process of their components leaves multi-scale feature traces in time-series data. Operational data is mixed with data from different time scales, such as high-frequency features from early bearing wear and low-frequency features from the slow temperature rise caused by lubricant performance degradation. A single-size convolutional kernel can only capture local temporal patterns at a fixed scale and cannot comprehensively cover the multi-stage features of component degradation. Therefore, in this embodiment of the invention, the feature extraction module uses three-size convolutional kernels to focus on temporal information at three different scales, avoiding the one-sidedness of single-scale features and reducing prediction bias caused by missing local information.

[0035] 2) Long Short-Term Memory (LSTM) layer captures long-range dependencies during the degradation process. The degradation of ship engine components is a long-term, time-dependent process (e.g., the accelerated wear of bearings is closely related to the vibration peak and cumulative temperature increase in the previous few hours). Traditional recurrent neural networks (RNNs) suffer from the problem of "long-term dependency forgetting" (the further back in time historical information is, the more severe the gradient decay during propagation), and they cannot capture the key correlation of "historical state influencing current state" during the degradation process.

[0036] This invention introduces an LSTM layer, which selectively retains important historical features (such as outliers from last week) and forgets irrelevant noise (such as instantaneous fluctuations) through a gating mechanism of "input gate, forget gate, and output gate," thereby accurately capturing the cumulative effect of the degradation process.

[0037] 3) Feature concatenation preserves multi-scale feature semantics. Features extracted from multiple scales and channels are heterogeneous (different dimensions and semantics). Direct superposition can lead to confusion of feature information (e.g., the weights of high-frequency vibration features and low-frequency temperature features are averaged). Therefore, embodiments of the present invention need to splice heterogeneous features according to dimensions to fully preserve the independent semantics of each feature, such as high-frequency vibration features and low-frequency temperature features.

[0038] 4) Long-distance dependencies between features in the Transformer layer and differences in attention weights The degradation of ship engine components is the result of the synergistic effect of multiple features (e.g., bearing failure is not only related to its own vibration, but also strongly correlated with lubricating oil temperature and speed fluctuations). While LSTM can capture long-range dependencies of a single sequence, it is difficult to model the "correlation relationship between different feature dimensions". Therefore, this embodiment of the invention further introduces the Transformer model. The self-attention mechanism of the Transformer model can calculate the correlation weight between any two features (or time-series nodes), and through "multi-head attention + positional encoding", it can more efficiently model the dependencies in ultra-long time series (such as monitoring data over several months), improving the globality of feature fusion.

[0039] The goal of the feature extraction module in this embodiment of the invention is to generate domain-invariant features, meaning that data from different domains are similarly distributed in the feature space, maximizing the error of the domain discriminator. After preprocessing, the running data is copied into three copies and passed through three parallel feature extraction channels. Each channel's convolutional layer (CNN) uses convolutional kernels of different sizes to extract temporal features at different levels. Batch normalization is then used to normalize the convolutional output, accelerating convergence. Nonlinear transformations are introduced through the activation functions of the three convolutional layer channels: ReLU, LeakyReLU, and GELU, respectively, increasing representational power. A Long Short-Term Memory (LSTM) layer supplements the modeling of long-range temporal dependencies. Then, the temporal features extracted from the three channels are concatenated according to feature dimensions. The Transformer model further fuses the concatenated multi-scale temporal features, modeling long-distance dependencies and attention weight differences between them, enhancing feature representation capabilities. Then, a global average pooling layer (GAP) is used to aggregate the enhanced features output by the Transformer model: the average of all feature values ​​in the sequence length dimension is calculated, and the sequence features of each sample are compressed into a global feature vector.

[0040] (2) Lifetime Prediction Module: This module performs lifetime prediction based on the global feature vector output by the feature extraction module. The global feature vector output by the feature extraction module is input into the linear transformation layer of the lifetime prediction section, then passes through an activation function layer (ReLU), and is then input into another linear transformation layer to obtain the lifetime prediction value.

[0041] (3) Domain discrimination module: Minimizes the error in domain discrimination, enabling the model to distinguish between the source and target domains, and performs adversarial training against the feature extraction module. The global feature vector output by the feature extraction module passes through a gradient inversion layer (GRL), which is equivalent to an identity mapping during forward propagation. , It is the input of the GRL layer. This is the output of the GRL layer. During backpropagation, the GRL multiplies the gradient by a negative number. Reverse, that is , It is a domain-based assessment of task loss. These are the adversarial weights of GRL, which control the intensity of adversarial training. ∈(0 1), The direction of the backward update of the feature extraction module is the direction of "confusion of the domains," meaning that the GRL layer plays a role in adversarial training during backpropagation. Then, it passes through a linear transformation layer, batch normalization layer, ReLU activation function layer, another linear transformation layer, and finally a discriminative output layer (LogSoftmax) to obtain the domain discrimination result.

[0042] More specifically: Suppose the running data needs further processing into a tensor dataset of (64, 50, F), meaning each batch contains 64 samples, each sample is a sequence containing 50 consecutive time steps, and F is the remaining feature dimension after preprocessing. The dataset labels are the true RUL values. Then, the input data is copied three times and fed into three parallel feature extraction channels. The specific structure and parameters of the lifetime prediction network model are as follows: In the first feature extraction channel (CNNLSTMBranch1): CNNs employ one-dimensional convolutional layers: using 3×3 kernels for temporal convolution operations. The input channel count is the original feature dimension F, and the output channel count is 32. Padding (padding=1) maintains the sequence length.

[0043] Batch normalization layer: Normalizes the 32 channels output by the convolutional layer.

[0044] Activation function layer: ReLU activation function is used.

[0045] LSTM layer: It is a single-layer structure with 32 hidden units. It receives the 32-dimensional feature sequence output by the convolution and outputs the hidden state at all time steps with a shape of (64, 50, 32).

[0046] In the second feature extraction channel (CNNLSTMBranch2): CNN uses one-dimensional convolutional layers: it uses a 5×5 kernel (padding=2) and has 32 output channels.

[0047] Batch normalization layer: Normalizes the 32 channels output by the convolutional layer.

[0048] Activation function layer: LeakyReLU activation function is used.

[0049] LSTM layer: It is a single-layer structure with 64 hidden units. It receives the 32-dimensional feature sequence from the convolution output and outputs the hidden state at all time steps. The output shape is (64, 50, 64).

[0050] In the third feature extraction channel (CNNLSTMBranch3): CNN uses one-dimensional convolutional layers: it uses a 7×7 kernel (padding=3) and has 32 output channels.

[0051] Batch normalization layer: Normalizes the 32 channels output by the convolutional layer.

[0052] Activation function layer: GELU activation function is used.

[0053] LSTM layer: It is a single-layer structure with 128 hidden units. It receives the 32-dimensional feature sequence from the convolution output and outputs the hidden state at all time steps. The output shape is (64, 50, 128).

[0054] The feature vectors output from the three branches are concatenated along the feature dimension to form a fused feature vector with a total dimension of 224 (i.e., 32+64+128) and a shape of (64, 224).

[0055] Subsequently, the following key fusion and processing steps are performed: Multi-scale temporal feature concatenation: The temporal features output from the three parallel feature extraction channels are concatenated along their feature dimensions to form a fused multi-scale spatiotemporal feature sequence with a shape of (64, 50, 224), i.e., 32 + 64 + 128 = 224 dimensions. This concatenated feature sequence fully preserves information from 50 time steps, with features at each time step extracted by convolutional kernels of three different scales.

[0056] Sequence dependency modeling: The feature sequence of shape (64, 50, 224) is directly input into the Transformer model. The Transformer model consists of two identical stacked encoding layers. The core parameters of each Transformer encoding layer are configured as follows: Feature dimension: 224.

[0057] Attention count: 4.

[0058] The hidden layer dimension of the feedforward network is 256.

[0059] Activation function: ReLU.

[0060] Dropout rate: 0.1.

[0061] The Transformer uses a self-attention mechanism to model the long-distance dependencies and global context information across these 50 time steps, outputting an enhanced feature sequence of the same shape (64, 50, 224). A global average pooling strategy is then employed to aggregate the features of the entire sequence output by the Transformer model: averaging all feature values ​​along the sequence length dimension (i.e., the 50 time steps), compressing the sequence features of each sample into a global feature vector with an output shape of (64, 224). This global feature vector is then input into the lifetime prediction module and the domain discrimination module, respectively.

[0062] The lifetime prediction module consists of two or three layers of network: The first layer is a linear transformation layer that maps the input features from 224 dimensions to 64 dimensions.

[0063] The second layer is the activation function layer, which uses the ReLU activation function and introduces a nonlinear transformation.

[0064] The third layer is a linear transformation layer that maps the 64-dimensional features to 1 dimension, directly outputting the final lifetime prediction value.

[0065] The domain discrimination module consists of a six-layer network: The first layer: The GRL layer performs an identity mapping during forward propagation and multiplies the gradient by a factor during back propagation. Reverse the behavior and dynamically adjust the adversarial weights. It started at 0 and gradually increased to 1.

[0066] The second layer is a linear transformation layer that maps the input features from 224 dimensions to 100 dimensions.

[0067] The third layer: batch normalization layer, which standardizes the 100-dimensional features.

[0068] The fourth layer is the activation function layer, which uses the ReLU activation function and introduces a nonlinear transformation.

[0069] The fifth layer is a linear transformation layer that maps 100-dimensional features to 2-dimensional features.

[0070] The sixth layer is the discriminant output layer, which uses the LogSoftmax function to output the log probability distribution of the two domains.

[0071] For key components of marine engines, life prediction network models were designed separately, and the network structures of the life prediction models were identical, such as... Figure 3 As shown.

[0072] Furthermore, the same training method is used for each component of the ship engine. The lifetime prediction network model in this embodiment of the invention adopts a domain adversarial training approach, aiming to reduce performance differences between different data distributions (i.e., source and target domains) through adversarial learning. Its core idea is to ensure that while completing the main task (lifetime regression prediction), the model cannot distinguish which source / target domain (data from different operating conditions) the data comes from, thereby learning domain-invariant feature representations.

[0073] First, the operational data required for training the lifespan prediction network model is divided into source domain data and target domain data. Source domain data consists of complete, lifespan-labeled operational data of ship engine components under a specific operating condition; the source domain is the primary data source relied upon for model training. Target domain data is unlabeled or minimally labeled data from a "new environment," the domain to which the model ultimately adapts and generalizes. Corresponding operational data is collected in both the source and target domains as training data for the model. During training, the total loss consists of the lifespan prediction task loss and the domain discrimination task loss. Based on this total loss, adversarial training is performed on the feature extraction module and the domain discrimination module of the lifespan prediction network model. (1) The formula for updating the model parameters of the feature extraction model is:

[0074] in, These are the model parameters for the feature extraction module. It's the learning rate. It is a loss in lifespan prediction tasks. It is a domain-based assessment of task loss. These are the adversarial weights of GRL. Among them,

[0075] in, This indicates the number of source domain samples in the source domain sample set. Indicates the first sample in the source domain sample set The true RUL value of each source domain sample Indicates the first sample in the source domain sample set The predicted RUL value of each source domain sample output by the lifetime prediction module.

[0076]

[0077] in, This indicates the number of target domain samples in the target domain sample set. Indicates the first sample in the target domain sample set Domain labels for each target domain sample (0 = source domain, 1 = target domain). Indicates the first sample in the target domain sample set The probability of a target domain sample output by the domain discrimination module.

[0078] The formula for updating model parameters from the feature extraction module It is evident that part of the optimization direction for the feature extraction module is aimed at increasing the loss of the domain discrimination module. This is the result of the gradient inversion layer.

[0079] (2) The model parameter update formula for the domain discrimination module is as follows:

[0080] in, These are the model parameters of the domain discrimination module. The optimization direction of the model parameters of the domain discrimination module is to minimize the loss of the domain discrimination task.

[0081] (3) Model parameter update formula for the lifetime prediction module:

[0082] in, These are the model parameters for the lifetime prediction module. The optimization direction for the model parameters of the lifetime prediction module is to minimize the loss of the lifetime prediction task.

[0083] During the training phase, operational data from ship engine components is first extracted and preprocessed (same as the preprocessing process in S10). Then, source domain data with lifespan labels and target domain data without lifespan labels are input into the lifespan prediction model for training. The feature extraction module extracts domain-invariant features from the source and target domain data, aligning the data distribution between the two domains. The lifespan prediction module trains on lifespan prediction using data from the source domain based on the output of the feature extraction module. The domain discrimination module determines the domain from which the data originates based on the output of the feature extraction module. The domain-invariant features generated by the feature extraction module prevent the domain discrimination module from distinguishing the source of the input data. During backpropagation, the GRL layer enables the feature extraction and domain discrimination modules to be trained adversarially, allowing the feature extraction module to learn more domain-invariant features. This reduces the distribution differences between the source and target domains, improving the model's generalization ability and the accuracy of remaining lifespan prediction in cross-domain scenarios. Finally, hyperparameter tuning optimizes the model performance, resulting in the output of the trained lifespan prediction model.

[0084] During the training process, this invention utilizes a large amount of labeled and complete operational data of key marine engine components as source domain data. When facing new operating conditions (including different operating states and navigation environments), it uses a small amount of unlabeled operational data of key marine engine components as target domain data to train the life prediction network model. This enables the achievement of high life prediction accuracy and is particularly suitable for assessing the remaining life of key marine engine components such as cylinder heads, cylinder liners, and pistons. It also demonstrates good generalization ability and practical value under varying operating conditions.

[0085] In summary, the cross-condition life prediction method for ship engine components proposed in this invention designs life prediction network models with identical network structures for key ship engine components. These models enhance their adaptability to different operating condition data distributions by introducing multi-scale feature extraction, attention mechanisms, and adversarial training. This achieves feature alignment and accurate life prediction for key ship engine components across operating conditions, significantly improving the model's generalization ability and stability in complex environments. The proposed method employs adversarial training during training, effectively utilizing historical operating data with life labels and limited unlabeled operating data under new conditions to establish a robust life prediction network model. This provides reliable remaining service life predictions for key ship engine components, supports predictive maintenance decisions, avoids unplanned downtime, improves navigation safety, and reduces maintenance costs. The proposed method effectively monitors the health status of key ship engine components, provides early warnings of potential faults, reduces the risk of unplanned maintenance, extends component lifespan, and improves ship operational safety and maintenance management efficiency, demonstrating significant engineering application value and economic benefits.

[0086] Secondly, embodiments of the present invention provide a cross-condition life prediction system for marine engine components, the system as follows: Figure 4 As shown, the system includes a server deployed with a JAVA backend, a first database, a second database, a frontend, and a pre-trained life prediction network model for the cross-condition life prediction method for key components of the ship engine described in the first aspect. The first database stores operational data of the key components of the ship engine. The JAVA backend periodically reads the operational data from the first database, preprocesses the read data to ensure data synchronization and validity, and inputs the pre-processed data into the corresponding pre-trained life prediction network model. The life prediction network model outputs the corresponding life prediction result based on the pre-processed operational data. The second database stores the life prediction results of the key components of the ship engine. The frontend reads the life prediction results from the second database and displays them. More specifically: A well-trained lifetime prediction network model can be deployed to a production environment. This can be achieved by establishing, for example... Figure 4 The system described above implements the deployment of cross-condition life prediction functionality for ships. The first database stores operational data (such as vibration, temperature, and pressure) from ship engine components for use by the backend and model. A MySQL database can be used for this first database. The Java backend periodically reads operational data from the first database, preprocesses it, and then sends the preprocessed data to the life prediction network model via a POST request. Upon receiving the life prediction results from the network model, it stores them in a second database. When the frontend requests life prediction data for a component, the Java backend retrieves the relevant prediction results from the second cached database and returns them to the frontend. The life prediction network model, deployed on a server using Flask, receives POST requests from the Java backend, performs component life prediction tasks, and returns the prediction results to the frontend. The second database caches the life prediction results; the Java backend prioritizes querying the second database to reduce frequent calls to the first database or the life prediction network model. The second database can have a set validity period for cached data to achieve periodic updates. A Redis database can be used for this second database. The frontend is responsible for displaying the life prediction results.

[0087] As the system embodiment of the second aspect is basically similar to the method embodiment of the first aspect, the description is relatively simple, and relevant details can be found in the description of the method embodiment of the first aspect.

[0088] In the description of this invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0089] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the specification and accompanying drawings, will understand and implement other variations of the disclosed embodiments in carrying out the claimed invention. In the specification, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality. While certain measures are described in different embodiments, this does not mean that these measures cannot be combined to produce good results.

[0090] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for predicting the lifespan of marine engine components across operating conditions, characterized in that, The method includes: Real-time acquisition of operational data from key components of the ship's engine, and preprocessing of the operational data for each component to ensure data synchronization and validity; The preprocessed running data is input into the corresponding trained lifetime prediction network model, and the lifetime prediction result for each component is output. Each component corresponds to a trained lifetime prediction network model, and the network structure of the trained lifetime prediction network models is the same. The lifetime prediction network model includes a feature extraction module, a lifetime prediction module, and a domain discrimination module. The feature extraction module extracts temporal features from the input data at different scales, concatenates all temporal features along the feature dimension, and enhances the concatenated result using a Transformer model. The enhanced feature result is then compressed by a global average pooling layer to output a global feature vector. The lifetime prediction module performs predictions based on the global feature vector and outputs the lifetime prediction result. The domain discrimination module performs discrimination based on the global feature vector and outputs the domain discrimination result. During training, the total loss consists of the lifetime prediction task loss and the domain discrimination task loss. Based on this total loss, adversarial training is performed on the feature extraction module and the domain discrimination module in the lifetime prediction network model to obtain the trained lifetime prediction network model.

2. The method for predicting the lifespan of marine engine components across operating conditions according to claim 1, characterized in that, The operational data for each component is preprocessed, including: The process for executing the operational data of each component includes: removing abnormal data from the operational data of that component to obtain first processed data; filling in missing values ​​in the first processed data to obtain second processed data; performing smoothing and denoising processing on the second processed data to obtain third processed data; performing Min-Max normalization processing on the third processed data to obtain fourth processed data; performing data synchronization processing on the fourth processed data to obtain fifth processed data; and performing feature filtering processing on the fifth processed data to obtain sixth processed data.

3. The method for predicting the lifespan of marine engine components across operating conditions according to claim 2, characterized in that, Missing values ​​in the first processed data are filled in to obtain the second processed data, including: The missing values ​​in the first processed data are filled in using a forward padding method to obtain the second processed data.

4. The method for predicting the lifespan of marine engine components across operating conditions according to claim 2, characterized in that, The fourth processed data is subjected to data synchronization processing to obtain the fifth processed data, including: The target resampling period is calculated based on the actual sampling frequency of different components in the fourth processed data. Based on the target resampling period, resampling processing is performed on different components in the fourth processed data; The resampled fourth-processed data is time-aligned and merged to obtain the fifth-processed data.

5. The method for predicting the lifespan of marine engine components across operating conditions according to claim 4, characterized in that, Based on the target resampling period, resampling processing is performed on different components in the fourth processed data, including: For each component in the fourth processed data, the execution process includes: if the actual sampling frequency of the component is greater than or equal to the target resampling period, then the component is downsampled; otherwise, data filling sampling is performed using linear interpolation.

6. The method for predicting the lifespan of marine engine components across operating conditions according to claim 1, characterized in that, The feature extraction module includes three feature extraction channels, a splicing layer, a Transformer model, and a global average pooling layer; in, Each feature extraction channel consists of a convolutional layer, a batch normalization layer, an activation function layer, and an LSTM layer connected in sequence, with the convolutional layer serving as the input to the corresponding feature extraction channel; The kernel size of the convolutional layer in each feature extraction channel is different, and the activation function used in the activation function layer is different; The input of the splicing layer is connected to the output of the three feature extraction channels respectively. The splicing layer, the Transformer model and the global average pooling layer are connected in sequence. The output of the global average pooling layer is used as the output of the feature extraction module.

7. The method for predicting the lifespan of marine engine components across operating conditions according to claim 1, characterized in that, The lifetime prediction module includes a linear transformation layer, an activation function layer, and a linear transformation layer connected in sequence.

8. The method for predicting the lifespan of marine engine components across operating conditions according to claim 1, characterized in that, The domain discrimination module includes a GRL layer, a linear transformation layer, a batch normalization layer, an activation function layer, a linear transformation layer, and a discrimination output layer connected in sequence.

9. A cross-condition life prediction system for marine engine components, characterized in that, The system includes a server equipped with a JAVA backend, a first database, a second database, a frontend, and a pre-trained life prediction network model for the cross-condition life prediction method for ship engine components according to any one of claims 1-8, for key ship engine components; wherein... The first database is used to store operational data of key components of ship engines; The JAVA backend is used to periodically read the operating data of key components of the ship engine from the first database, preprocess the operating data of the key components to ensure the data synchronization and validity, and input the preprocessed operating data into the corresponding trained life prediction network model. The lifetime prediction network model is used to output corresponding lifetime prediction results based on the preprocessed running data. The second database is used to store life prediction results for key components of ship engines; The front end is used to read and display the life prediction results of key components of the ship engine from the second database.

Images