Ship power system fault diagnosis method based on control-aware affine network

Abstract

The present application relates to the technical field of ship power fault diagnosis, in particular to a ship power system fault diagnosis method based on control-aware affine network, comprising constructing a simulation CAAN model of marine diesel engine fuel injection system based on control-aware affine network, first aligning the fault signal and the control signal in time, and embedding the control signal features into the fault signal through the control-aware calibration CAC module to enhance the semantic expression ability of the features, then decoupling the fused features, and extracting the maneuvering invariant features and the fault severity features respectively, in order to ensure the complementarity and discriminability of the two types of features, the model introduces a regularization constraint to promote differentiated feature learning, finally, the two types of invariant features are spliced and fused to form a unified feature representation, and input into the classifier to realize fault recognition; has the advantages of being able to adapt to the frequent changes of marine diesel engine working conditions, improving the reliability of fault early warning, etc.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ship power fault diagnosis, in particular to a ship power system fault diagnosis method based on control-aware affine network, which can adapt to frequent changes of working conditions of marine diesel engines and improve the reliability of fault early warning. BACKGROUND

[0002] As the core power system of a ship, a diesel engine undertakes the important task of propulsion and auxiliary energy supply, and its operating state directly affects the safety of navigation, energy efficiency and economic benefits. However, marine diesel engines usually work in a complex and variable environment and are subjected to stresses such as high temperature, high pressure and dynamic load for a long time. Under such harsh working conditions, key components are prone to accelerated wear, performance degradation, and even serious failure. These failures not only can cause high downtime losses and transportation delays, but also can lead to serious accidents such as loss of control or fire, which pose a threat to the safety of crew, the integrity of the ship and the marine environment. Therefore, it is of great significance to carry out real-time monitoring and early fault diagnosis to ensure the safety of navigation and improve economic benefits.

[0003] At present, the fault diagnosis methods of marine diesel engines mainly include three types: model-based methods, traditional machine learning-based methods and deep learning-based methods. Among them, the model-based method relies on the analytical model derived from mathematics and physical mechanism, which has strong interpretability and clear physical meaning. However, due to the complex structure, strong nonlinearity and time-varying parameters of marine diesel engine systems, it is difficult to build a complete model that can accurately describe the dynamic behavior of the system. The traditional machine learning method avoids the complex process of physical modeling by introducing expert knowledge for feature engineering. This method usually manually extracts statistical features in the time domain, frequency domain or time-frequency domain from sensor signals and inputs them into a classifier to achieve fault recognition. Although this method can achieve good diagnostic performance without strict model assumptions, its effectiveness is highly dependent on the manually designed features, and the feature extraction process is often time-consuming and dependent on professional experience. In addition, these artificial features generally lack robustness and generalization ability, and their applicability is limited under complex and variable operating conditions.

[0004] In contrast, the deep learning-based method breaks through the traditional diagnosis paradigm and realizes an end-to-end feature learning process, which can automatically extract discriminative fault features from raw high-dimensional sensor data. However, this method usually assumes that the training data and test data follow the same distribution, which is difficult to meet in practical applications. The working conditions of marine diesel engines change frequently, resulting in a significant shift in data distribution, which reduces the generalization ability of the model in unknown environments and significantly degrades the diagnostic performance.

[0005] Domain generalization, a commonly used strategy, aims to improve the robustness of a model by learning domain-invariant features, thereby achieving reliable fault diagnosis under unknown operating conditions. Existing domain generalization methods mainly include feature alignment, adversarial training, data augmentation, and meta-learning optimization. However, most methods treat changes in operating conditions as external disturbances, relying solely on complex network structures for passive adaptation or compensation, failing to fully address the inherent coupling relationship between fault features and operating condition features. As shown in Figure 1(a), when a deceleration command is received, the electronic control unit (ECU) executes a feedforward control strategy, appropriately extending the injection duration to maintain the stability of the combustion process. Similarly, Figures 1(b)–(c) show that injector blockage also leads to an extension of the injection duration, thereby ensuring sufficient fuel supply.

[0006] If the diagnostic system is built based on assumptions of steady-state or slowly changing operating conditions, it may misjudge dynamic adjustments as injector faults, leading to misdiagnosis. In the early stages of injector clogging, the compensatory extension of the injection timing is small and easily masked by the fine-tuning effect of load or temperature fluctuations, resulting in missed detection. The fault is only detected when the clogging worsens and the compensation becomes abnormally significant, but by then the optimal repair opportunity has been missed. Summary of the Invention

[0007] This invention addresses the shortcomings and deficiencies of existing technologies by proposing a fault diagnosis method for ship propulsion systems based on control-aware affine networks. This method introduces a control-aware calibration module to effectively embed control signals into fault features, thereby achieving feature mapping driven by operating conditions and improving the robustness of the model under complex operating conditions. Furthermore, it maintains semantic consistency through intra-domain and inter-domain collaborative strategies and enhances cross-domain generalization capabilities by combining adversarial decoupling mechanisms.

[0008] This invention achieves its purpose through the following measures:

[0009] A method for fault diagnosis of ship propulsion systems based on control-aware affine networks, characterized by the following steps:

[0010] Step 1: Construct a simulation CAAN model for marine diesel engine fuel injection system based on a control-aware affine network. The CAAN model first aligns the fault signal and control signal in time, and then embeds the control signal features into the fault signal through the control-aware calibration (CAC) module to enhance the semantic expressive power of the features. Subsequently, the model decouples the fused features and extracts the maneuver invariant features and fault severity features respectively: the former is obtained through the gradient inversion mechanism, and the latter is extracted by the teacher network through knowledge distillation. To ensure the complementarity and discriminativeness of the two types of features, the model introduces regularization constraints to promote the learning of differentiated features. Finally, the two types of invariant features are spliced ​​and fused to form a unified feature representation, which is then input into the classifier to achieve fault identification.

[0011] Step 2: Collect control signals and corresponding fuel injection mass flow signals to form a fault dataset; set three standard operating conditions: acceleration, steady state, and deceleration, to construct a three-domain dataset, select one domain as the target domain, and the other two as the source domains, thus forming a typical domain generalization task.

[0012] Step 3: For each domain generalization task, the proposed CAAN model is trained and its fault classification performance under cross-domain conditions is evaluated.

[0013] In step 1 of this invention, the Control Perception Calibration (CAC) module is used to capture and parameterize the dynamic response characteristics of a healthy fuel injection system under maneuvering conditions. This module uses a scaling factor. and offset factor These responses are characterized by a set of parameters that collectively define a dynamic baseline for normalizing the original features into a condition-independent health status representation. Due to the temporal correlation between samples, directly calculating the scaling parameters from the control signals would be problematic. and offset parameters This could lead to instability during the training process. Therefore, an incremental adjustment strategy is adopted to moderately update the parameters based on the current sample features while ensuring overall stability. Let the input fault signal be... The control signal is In the In each training iteration, the control signal Processed through multiple linear layers to generate scaling increments. and offset increment : (1),

[0014] During initialization, the settings are as follows: and , representing the scaling factor and the offset factor, respectively. During training, these two parameters are continuously updated iteratively. and This indicates the result after gradient descent optimization. and Finally, the input features are scaled and offset to obtain the control-aware features. : (2).

[0015] In step 1 of this invention, the model decouples the fused features and extracts maneuver-invariant features and fault severity features respectively. The fault severity features are trained to be highly sensitive to changes in fault severity, while the maneuver-invariant features are trained to filter out all information related to maneuver conditions, retaining only the basic fault modes that exist under various operating conditions. The specific learning method for the fault severity features is as follows:

[0016] When injectors become clogged, the fuel injection system compensates by delaying the injection timing to ensure a constant fuel injection quantity. Fourier phase spectra can effectively characterize this injection timing delay and the structural timing characteristics of the fuel injection system; for input fault characteristics... In the time dimension Performing a Fourier transform on the above, we get:

[0017] (3),

[0018] Fourier phase angle Defined as:

[0019] (4).

[0020] The distillation of fault severity feature information in step 1 of this invention specifically includes: the teacher network uses Fourier phase information for classification, thereby learning discriminative phase features. After training, the network possesses a strong representational ability highly correlated with classification semantics; during the training of the student model, the intermediate representation of the teacher network is combined with the features generated by CAAN through equation (5). Alignment is performed to guide student models to inherit Fourier phase-based feature representations:

[0021] (5).

[0022] The machine-invariant feature learning method in step 1 of this invention specifically includes: for the source domain, its loss function... Defined as a classification objective based on label prediction:

[0023] (6),

[0024] in, Indicates input features, This indicates the model's prediction results. For real labels, This represents the source domain dataset, while The loss function for the target domain, representing the sample originating from the source domain distribution, is the adversarial loss of the domain classifier, and can be expressed as:

[0025] (7),

[0026] in, This represents the prediction result for the target domain. The true label for the target domain;

[0027] During backpropagation, the gradient inversion layer enhances the separability between the source and target domains by inverting the gradients of the extracted features. Its calculation process can be represented by the following formula:

[0028] (8),

[0029] in, This represents all trainable parameters in the feature extractor;

[0030] By introducing a gradient inversion layer, the feature extractor learns machine-invariant features during the confusion domain classifier process. Its training process is as follows:

[0031] (9).

[0032] To achieve complete separation of fault severity features and maneuver invariance features at the information level, this invention introduces cosine similarity, defined as shown in equation (10). By minimizing the cosine similarity, the model is guided to learn two independent information channels, thereby ensuring that fault-related information and operating condition-related information are separated from each other. (10).

[0033] Compared with the prior art, the present invention has the following advantages: (1) It introduces control signals into fault diagnosis to achieve the synergy between control and diagnosis. The proposed CAC module transforms the maneuver control information into an active calibration mechanism. The real-time control signal is mapped to affine transformation parameters for conditionally modulating the feature channel. This strategy uses physical prior knowledge to dynamically normalize and correct the extracted features, which not only improves the interpretability of the features but also lays the foundation for achieving control-diagnosis synergy. (2) A dual-feature decoupling framework for injector fault diagnosis is constructed. The features after control offset calibration are decomposed into two orthogonal components: one is the maneuver invariant feature obtained through adversarial learning, and the other is the fault severity feature extracted through knowledge distillation. The two are synergistically constrained through complementary regularization, and their fusion result constitutes a comprehensive fault characterization with good generalization ability. (3) An expert teacher network for injector faults is established. The teacher model is trained based on Fourier phase spectrum data, which can effectively capture key features related to blockage in fuel injection mass flow rate. Through knowledge distillation, this representation is transferred to the main diagnostic model, providing the model with a reasonable inductive bias, thereby improving its robustness under different operating conditions. Attached Figure Description

[0034] Appendix Figure 1 This is the diesel engine injection mass flow rate diagram in this invention, wherein... Figure 1 (a) is a fuel mass flow rate curve during diesel engine deceleration, (b) is a schematic diagram of the diesel engine injector structure, and (c) is a curve showing different degrees of blockage of the diesel engine injector.

[0035] Appendix Figure 2 This is a diagram of the CAAN network structure in this invention.

[0036] Appendix Figure 3 This is a network structure diagram of the CAC module in this invention.

[0037] Appendix Figure 4 This is a flowchart of the Fourier phase distillation learning process in this invention.

[0038] Appendix Figure 5 This is the overall flowchart of the fault diagnosis of the present invention.

[0039] Appendix Figure 6 This is a schematic diagram of the diesel injection system model of the present invention.

[0040] Appendix Figure 7 This is a schematic diagram comparing the accuracy of different models in this invention.

[0041] Appendix Figure 8 This is the confusion matrix of the CAAN model in this invention under three tasks, where (a) is task 1, (b) is task 2, and (c) is task 3.

[0042] AppendixFigure 9 This is a schematic diagram illustrating the impact of the number of control signals on model accuracy in an embodiment of the present invention, wherein... Figure 9 (a) shows the accuracy rate, and (b) shows the F1 score.

[0043] Appendix Figure 10 This is a schematic diagram of t-SEN results under different control signal quantities according to an embodiment of the present invention, wherein... Figure 10 (a), (b), and (c) involve three control signals, (d), (e), and (f) involve two control signals, and (g), (h), and (i) involve one control signal. Detailed Implementation

[0044] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0045] To address the cross-domain performance degradation caused by loss of control perception during maneuvers, this invention proposes a domain generalization method combining control signals and fault signals, and applies it to fault diagnosis of marine diesel engine injection systems, as shown in the attached figure. Figure 5 As shown, the main steps include:

[0046] (1) Construct a simulation model of the marine diesel engine fuel injection system and integrate the control unit; synchronously collect control signals and corresponding fuel injection mass flow signals to form a fault dataset;

[0047] (2) Set three standard operating conditions: acceleration, steady state and deceleration, and construct a three-domain dataset. Select one domain as the target domain and the other two as the source domains to form a typical domain generalization task.

[0048] (3) For each domain generalization task, the CAAN model is trained and its fault classification performance under cross-domain conditions is evaluated.

[0049] As shown in Figure 2, the CAAN model first aligns the fault signal and control signal in time, and then embeds the control signal features into the fault signal through the CAC module to enhance the semantic expressiveness of the features. Subsequently, the model decouples the fused features, extracting maneuver-invariant features and fault severity features separately: the former is obtained through gradient inversion, while the latter is extracted by the teacher network through knowledge distillation. To ensure the complementarity and discriminative power of the two types of features, the model introduces regularization constraints to promote the learning of differentiated features. Finally, the two types of invariant features are concatenated and fused to form a unified feature representation, which is then input into the classifier to achieve fault identification.

[0050] Since control signals themselves do not directly contain fault information, using them together with fault signals as feature input to the diagnostic model may lead to the model learning incorrect correlation patterns under conditions of insufficient or unevenly distributed data. Therefore, this invention proposes a Control Awareness Calibration (CAC) module to capture and parameterize the dynamic response characteristics of a healthy fuel injection system under maneuvering conditions. This module uses a scaling factor... and offset factor These responses are characterized. This set of parameters collectively defines a dynamic baseline used to normalize the raw features into a condition-independent health status representation.

[0051] The structure of the CAC module is as follows: Figure 3 As shown, due to the temporal correlation between samples, directly calculating the scaling parameter based on the control signal would be problematic. and offset parameters This could lead to instability during the training process. Therefore, this invention employs an incremental adjustment strategy, which updates the parameters appropriately based on the current sample features while ensuring overall stability. Let the input fault signal be... The control signal is In the first In each training iteration, the control signal Processed through multiple linear layers to generate scaling increments. and offset increment :

[0052] (1),

[0053] During initialization, the settings are as follows: and , where and represent the scaling factor and offset factor, respectively. During training, these two parameters are continuously updated iteratively. and This indicates the result after gradient descent optimization. and Ultimately, the input features are scaled and offset to obtain the control-aware features. : (2).

[0054] Given that existing research has shown that simple cross-domain feature alignment methods struggle to obtain highly discriminative feature representations in domain generalization tasks—although such methods can reduce distributional differences between different domains, they may also weaken the discriminative power of features and limit their diversity—a more effective strategy is to perform feature modeling simultaneously at both intra-domain and cross-domain levels. In this example, the fault severity feature is trained to be highly sensitive to changes in fault severity; the maneuver-invariant feature is trained to filter out all information related to maneuver conditions, retaining only the basic fault modes that exist under all operating conditions.

[0055] The fault severity feature learning method is Fourier phase distillation learning: when the injector becomes clogged, the injection system compensates by delaying the injection timing to ensure that the fuel injection quantity remains constant; the Fourier phase spectrum can effectively characterize this injection timing delay and the structural timing characteristics of the injection system; for the input fault features In the time dimension Performing a Fourier transform on the above, we get:

[0056] (3),

[0057] Fourier phase angle Defined as: (4),

[0058] Fault Information Distillation: To avoid repetitive Fourier transform calculations and achieve efficient end-to-end reasoning during the prediction phase, this example proposes a compact knowledge distillation framework, as shown in Figure 4. The teacher network uses Fourier phase information for classification, thereby learning discriminative phase features. After training, the network possesses strong representational capabilities highly correlated with classification semantics.

[0059] Intermediate representations of the teacher network during student model training. Features generated by equation (5) and CAAN Alignment is performed to guide student models to inherit feature representations based on Fourier phase.

[0060] (5).

[0061] In this example, the maneuver-invariant feature learning method is as follows: In the complex operating scenarios of marine diesel engines, the differences in data distribution between different maneuvering conditions exhibit highly complex and nonlinear characteristics. Therefore, moment-matching-based global alignment methods (such as MMD or CORAL) struggle to effectively capture and align the high-order or fine-grained structures within these distributions. In contrast, Domain Adversarial Networks (DANNs), through adversarial training mechanisms, are better able to handle such complex nonlinear distribution variations.

[0062] For the source domain, its loss function Defined as a classification objective based on label prediction:

[0063] (6),

[0064] in, Indicates input features, This indicates the model's prediction results. For real labels, This represents the source domain dataset, while This indicates that the sample comes from the source domain distribution.

[0065] The loss function for the target domain is the adversarial loss of the domain classifier, which can be expressed as:

[0066] (7),

[0067] in, This represents the prediction result for the target domain. The true label for the target domain.

[0068] During backpropagation, the gradient inversion layer enhances the separability between the source and target domains by inverting the gradients of the extracted features. Its calculation process can be represented by the following formula:

[0069] (8),

[0070] in, This represents all trainable parameters in the feature extractor.

[0071] By introducing a gradient inversion layer, the feature extractor learns machine-invariant features during the confusion domain classifier process. Its training process is as follows:

[0072] (9),

[0073] To achieve complete separation of fault severity features and maneuver invariance features at the information level, this example introduces cosine similarity, defined in equation (10). By minimizing the cosine similarity, the model is guided to learn two independent information channels, thereby ensuring the separation of fault-related information and operational condition-related information. This strategy can extract purer and more reliable decoupled features:

[0074] (10).

[0075] The experimental data in this example comes from a simulation model of a diesel engine fuel injection system, such as... Figure 6 As shown, the system structure includes a camshaft, a lift pump, four in-line injection pumps, and four injectors. Fault signals correspond to the fuel mass flow rate of each injector during the injection process. To simulate injector clogging, different fault states are set by changing the nozzle orifice diameter; the specific parameters are shown in Table 1. The system's dynamic control signals include crankshaft angular velocity. Cam speed and cam position These three variables together determine the opening and closing characteristics of the fuel injection valve, thus directly affecting the fuel mass flow rate during the injection process. .

[0076] Table 1 Injector clogging fault settings

[0077]

[0078] To construct a domain generalization experimental environment, this example establishes a dataset containing three operating states: acceleration (C1, engine speed increases from 926 rpm to 955 rpm), steady-state (C2, engine speed maintained at 955 rpm), and deceleration (C3, engine speed decreases from 955 rpm to 926 rpm). The data sampling frequency is 10 kHz, and 2,500 samples are collected for each fault type under each operating state. Based on this dataset, several cross-domain transfer learning tasks are designed to evaluate the model's generalization ability under different operating states. Detailed configurations for each task are shown in Table 2.

[0079] Table 2 Task Configuration

[0080]

[0081] The CAAN model presented in this example is implemented using the deep learning framework PyTorch 2.6.0, developed by Meta, which supports GPU-accelerated computing. The experimental environment was configured with an Intel Core i9-13900HX processor and an NVIDIA GeForce RTX 4060 graphics card.

[0082] To ensure the reliability of generalization during the knowledge distillation process, the teacher network was pre-trained for 100 epochs before training CAAN. The pre-training used the Adam optimization algorithm with a learning rate of 0.001 and a batch size of 64. Detailed hyperparameter configurations for this network are shown in Table 3.

[0083] Table 3. Structural details of the teacher network

[0084]

[0085] The core objective of the CAAN model is to achieve the co-learning of fault severity features and maneuver invariance features. This objective is achieved through joint constraints of distillation loss and domain difference loss, with adjustment factors set. =1.0, to balance the learning process of the two types of features. To achieve proper alignment of the domain classifiers, the gradient inversion coefficient is set to . =0.1. The model training process includes four loss functions: classification loss (CrossEntropyLoss), domain discriminant loss (BCELoss), knowledge distillation loss (MSELoss), and domain feature difference loss (Cosine Similarity). Training uses the Adam optimizer with an initial learning rate of 0.001 and a batch size of 64.

[0086] To verify the domain generalization effectiveness of the proposed model in multiple scenarios, this example selects CORAL, DANN, MLDG, and GroupDRO as baseline methods for comparative experiments. Performance comparisons were conducted on three scenario datasets: Task 1, Task 2, and Task 3. The corresponding results are shown below. Figure 7 As shown.

[0087] The evaluation metrics used were accuracy (ACC) and F1 score, and detailed comparison results are listed in Table 4. The confusion matrix of the CAAN model under three different speed conditions is shown below. Figure 8 As shown.

[0088] In Task 1, the model demonstrated high classification accuracy under the given conditions, with most samples correctly classified into their respective categories. The diagonal elements of the confusion matrix were almost equal to the total number of samples in each category, indicating high model accuracy with only a small number of misclassifications between adjacent categories in some healthy states.

[0089] In Task 2, the model's cross-domain generalization ability significantly decreased compared to Task 1. Consistent with the previous results, most misclassifications still occurred between categories representing near-healthy states, but their frequency increased. The performance degradation was particularly pronounced in samples representing moderate health levels, where classification confusion was more evident.

[0090] In Task 3, the model's cross-domain generalization ability dropped to its lowest point. The confusion matrix showed a significant increase in off-diagonal elements, indicating that the model had considerable difficulty distinguishing between similar fault levels, resulting in a marked decrease in classification reliability.

[0091] Table 4 Performance Comparison of Different Task Models

[0092]

[0093] To evaluate the dual-feature decoupling mechanism and the role of the CAC module in CAAN, a series of ablation experiments were conducted. The analysis considered three model configurations: a baseline model using only machine-invariant features (MIF), a variant model extracting both features simultaneously (MIF+FSF), and an enhanced model incorporating the CAC module (MIF+FSF+CAC). Experiments were performed on three domain generalization tasks, and the results were averaged to ensure consistency and reduce random fluctuations. The corresponding evaluation results are listed in Table 5.

[0094] With the configuration combining dual features and control signals, the proposed model achieves average accuracies of 98.83%, 98.14%, and 97.65% in Tasks 1, 2, and 3, respectively, with corresponding F1 scores of 0.988, 0.981, and 0.976, demonstrating the best performance among all configurations. In contrast, when the control signal is removed, the F1 scores decrease to 0.955, 0.951, and 0.956, representing reductions of 3.34%, 3.06%, and 2.05%, respectively. These results indicate that although the control signal does not directly contain fault label information, the affine transformation mechanism in the CAC module enables adaptive alignment of domain features, thereby significantly improving the model's diagnostic accuracy and cross-domain generalization ability.

[0095] When the model relies solely on maneuver-invariant features, its average accuracies on Task 1, Task 2, and Task 3 are 94.96%, 94.07%, and 94.39%, respectively, with corresponding F1 scores of 0.949, 0.941, and 0.944. Compared to the complete model that fuses control signals, the overall performance of this configuration is significantly lower. This result indicates that although the introduction of dual features improves cross-domain recognition capabilities to some extent, the model still struggles to differentiate fault severity in the target domain, suggesting that the effect of using feature decoupling mechanisms alone is limited.

[0096] Table 5 Ablation Experiment Results

[0097]

[0098] To further evaluate the impact of the number of control signals on model performance, this example designs a series of comparative experiments to analyze the role of control signal dimensions in domain generalization. Although control signals do not directly convey fault label information, they can characterize the system's operating state, thus providing contextual information for dual-feature decoupling and feature generalization. Therefore, this example sets up three experimental schemes, using one-dimensional, two-dimensional, and three-dimensional control signals respectively. In all configurations, the network structure and training parameters remain consistent. The performance of the experiments is evaluated on tasks 1 to 3, and the classification accuracy results and corresponding t-SNE visualizations are shown in [link to relevant documentation]. Figure 9 and Figure 10 .

[0099] The results show that introducing control signals can continuously improve model performance, and the accuracy increases with the number of control signals. When using only a single control signal, the model's accuracy on Task 1, Task 2, and Task 3 is 95.93%, 95.69%, and 96.18%, respectively, showing only a slight improvement compared to the model without control signals. At this stage, t-SNE shows that the class cluster boundaries are blurred, with significant overlap and limited inter-class separability. After introducing two control signals, the accuracy increases to 97.64%, 97.40%, and 97.32%, respectively, with a significant improvement in inter-class separation. The boundaries of classes F2-F4 are clearer, and intra-class samples are more tightly clustered, indicating that the model has learned more discriminative and robust features. However, a small amount of confusion still exists between normal operating conditions and low-level faults. When using all three control signals, the model performs optimally, with almost complete separation of classes in the feature space, clear cluster boundaries, and obvious cluster structure, fully demonstrating the model's stronger generalization ability and classification performance.

[0100] To overcome the performance degradation in fault diagnosis caused by data distribution variations during the maneuvering operation of marine diesel engines, this paper proposes a novel CAAN-based method. This method emphasizes actively utilizing operating condition information to enhance feature robustness. First, the CAC module applies a dynamic affine transformation to the latent features based on real-time control signals, thereby calibrating and decoupling feature biases caused by adverse operating condition changes. Subsequently, a dual-feature decoupling framework categorizes the calibrated features into two classes: fault severity features and maneuver-invariant features. The maneuver-invariant features are obtained through adversarial training, while the fault severity features undergo knowledge distillation guided by an expert teacher network. By fusing these two orthogonal and complementary feature classes, CAAN achieves high-precision and high-reliability classification of injector clogging faults.

[0101] Experimental results show that, compared with existing multi-domain generalization methods, the proposed CAAN model achieves significantly higher accuracy and better overall performance in various cross-condition diagnostic tasks. Its advantages are mainly reflected in two aspects. First, the CAC module effectively alleviates the inherent coupling between fault-related features and maneuver-related features, enabling the model to better adapt to unknown operating environments. Second, the mechanism of dual-feature decoupling combined with complementary learning ensures that the fault representation obtained by the model not only possesses good cross-domain generalization ability but also maintains strong class discriminativeness, thus achieving an excellent balance between generalization and accuracy.

Claims

1. A method for fault diagnosis of ship propulsion systems based on control-aware affine networks, characterized in that, Includes the following steps: Step 1: Construct a simulation CAAN model for marine diesel engine fuel injection system based on a control-aware affine network. The CAAN model first aligns the fault signal and control signal in time, and then embeds the control signal features into the fault signal through the control-aware calibration (CAC) module to enhance the semantic expressive power of the features. Subsequently, the model decouples the fused features and extracts the maneuver invariant features and fault severity features respectively: the former is obtained through the gradient inversion mechanism, and the latter is extracted by the teacher network through knowledge distillation. To ensure the complementarity and discriminativeness of the two types of features, the model introduces regularization constraints to promote the learning of differentiated features. Finally, the two types of invariant features are spliced ​​and fused to form a unified feature representation, which is then input into the classifier to achieve fault identification. Step 2: Collect control signals and corresponding fuel injection mass flow signals to form a fault dataset; set three standard operating conditions: acceleration, steady state, and deceleration, to construct a three-domain dataset, select one domain as the target domain, and the other two as the source domains, thus forming a typical domain generalization task. Step 3: For each domain generalization task, the proposed CAAN model is trained and its fault classification performance under cross-domain conditions is evaluated.

2. The method for fault diagnosis of ship propulsion system based on control-aware affine network according to claim 1, characterized in that, In step 1, the control perception calibration module (CAC) is used to capture and parameterize the dynamic response characteristics of a healthy fuel injection system under maneuvering conditions. This module uses a scaling factor... and offset factor These responses are characterized by a set of parameters that collectively define a dynamic baseline for normalizing the original features into a condition-independent health status representation. Due to the temporal correlation between samples, directly calculating the scaling parameters from the control signals would be problematic. and offset parameters This could lead to instability during the training process. Therefore, an incremental adjustment strategy is adopted to moderately update the parameters based on the current sample features while ensuring overall stability. Let the input fault signal be... The control signal is In the In each training iteration, the control signal Processed through multiple linear layers to generate scaling increments. and offset increment : (1)， During initialization, the settings are as follows: and , representing the scaling factor and the offset factor, respectively. During training, these two parameters are continuously updated iteratively. and This indicates the result after gradient descent optimization. and Finally, the input features are scaled and offset to obtain the control-aware features. : (2).

3. The method for fault diagnosis of ship propulsion system based on control-aware affine network according to claim 1, characterized in that... In step 1, the model decouples the fused features and extracts maneuver-invariant features and fault severity features separately. The fault severity features are trained to be highly sensitive to changes in fault severity, while the maneuver-invariant features are trained to filter out all information related to maneuver conditions, retaining only the basic fault modes that exist under all operating conditions. The specific learning method for the fault severity features is as follows: When the injector becomes clogged, the fuel injection system compensates by delaying the injection timing to ensure that the fuel injection quantity remains constant. The Fourier phase spectrum can effectively characterize this injection timing delay and the structural timing characteristics of the fuel injection system. Input fault characteristics In the time dimension Performing a Fourier transform on the above, we get: (3)， Fourier phase angle Defined as: (4)。 4. The method for fault diagnosis of ship propulsion system based on control-aware affine network according to claim 1, characterized in that... Step 1, the distillation of fault severity feature information, specifically includes: the teacher network using Fourier phase information for classification, thereby learning discriminative phase features. After training, this network possesses a strong representational ability highly correlated with classification semantics; during the student model training process, the intermediate representations of the teacher network... Features generated by equation (5) and CAAN Alignment is performed to guide student models to inherit Fourier phase-based feature representations: (5)。 5. The method for fault diagnosis of ship propulsion system based on control-aware affine network according to claim 1, characterized in that... The machine-invariant feature learning method in step 1 specifically includes: for the source domain, its loss function... Defined as a classification objective based on label prediction: (6)， in, Indicates input features, This indicates the model's prediction results. For real labels, This represents the source domain dataset, while The loss function for the target domain, representing the sample originating from the source domain distribution, is the adversarial loss of the domain classifier, and can be expressed as: (7)， in, This represents the prediction result for the target domain, where represents the true label for the target domain. During backpropagation, the gradient inversion layer enhances the separability between the source and target domains by inverting the gradients of the extracted features. Its calculation process can be represented by the following formula: (8)， in, This represents all trainable parameters in the feature extractor; By introducing a gradient inversion layer, the feature extractor learns machine-invariant features during the confusion domain classifier process. The training process is as follows: (9)。 6. The method for fault diagnosis of ship propulsion system based on control-aware affine network according to claim 1, characterized in that... To achieve complete separation of fault severity features and maneuver invariance features at the information level, cosine similarity is introduced, defined as shown in equation (10). By minimizing the cosine similarity, the model is guided to learn two independent information channels, thereby ensuring that fault-related information and operating condition-related information are separated from each other. (10)。

Images