Ship power system fault diagnosis method based on control perception affine network

Abstract

The invention relates to the technical field of ship power fault diagnosis, in particular to a ship power system fault diagnosis method based on a control-aware affine network, which comprises the following steps of: constructing a ship diesel engine fuel injection system simulation CAAN model based on the control-aware affine network; the method comprises the following steps of: firstly, fusing a fault signal, embedding a control signal feature into the fault signal through a control perception calibration CAC module so as to enhance the semantic expression capability of the feature, then decoupling the fused feature by a model, respectively extracting a maneuvering invariant feature and a fault severity feature, and in order to ensure the complementarity and discrimination of the two types of features, determining the fault severity of the maneuvering invariant feature and the fault severity of the maneuvering invariant feature and the fault severity. Regularization constraint is introduced into the model to promote differential feature learning, finally, two types of invariant features are spliced and fused to form unified feature representation, and the feature representation is input into a classifier to achieve fault recognition; the method has the advantages of adapting to frequent working condition changes of the marine diesel engine, improving fault early warning reliability and the like.

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 as a common strategy aims to improve the robustness of the model by learning domain-invariant features, so as to realize reliable fault diagnosis under unknown working conditions. Existing domain generalization methods mainly include feature alignment, adversarial training, data augmentation and meta-learning optimization. However, most methods regard working condition changes as external disturbances, and only rely on complex network structure for passive adaptation or compensation, which fails to fully solve the problem of the internal coupling relationship between fault features and working condition features. As shown in Fig. 1(a), when receiving a deceleration instruction, the electronic control unit (ECU) will execute a feedforward control strategy to appropriately extend the injection duration to maintain the stability of the combustion process. Similarly, Figs. 1(b)-(c) show that when the injector is blocked, the injection duration will also be extended to ensure sufficient fuel supply.

[0006] If the diagnostic system is based on the assumption of steady-state or slowly varying working conditions, dynamic adjustment may be misjudged as an injector fault, leading to misdiagnosis. In the early stage of injector blockage, the compensatory extension of injection timing is small, which is easily masked by the fine-tuning effect of load or temperature fluctuations, causing missed detection. Only when the blockage intensifies and the compensation anomaly is obvious, the fault is detected, but by then the optimal maintenance opportunity has been missed. SUMMARY

[0007] The present application aims to solve the problems and deficiencies in the prior art by introducing a control-aware calibration module to effectively embed control signals into fault features, achieving feature mapping based on working condition driving, thereby improving the robustness of the model under complex operating conditions. A ship power system fault diagnosis method based on control-aware affine network is proposed, which maintains semantic consistency through intra-domain and inter-domain collaborative strategies, and enhances cross-domain generalization ability through an adversarial decoupling mechanism.

[0008] The present application achieves the above-mentioned purposes by the following measures: A ship power system fault diagnosis method based on control-aware affine network, characterized in that it comprises the following steps: Step 1: Construct a control-aware affine network (CAAN) model for the fuel injection system of a marine diesel engine. The CAAN model first aligns the fault signals and control signals in time, and then embeds the control signal features into the fault signals through a control-aware calibration (CAC) module to enhance the semantic expression ability of the features. Subsequently, the model decouples the fused features to extract maneuver-invariant features and fault severity features: the former is obtained through a gradient reversal mechanism, and the latter is extracted by a teacher network through knowledge distillation. To ensure the complementarity and discriminability of the two types of features, the model introduces regularization constraints to promote differential feature learning. Finally, the two types of invariant features are fused to form a unified feature representation, which is input into a classifier to realize fault recognition. Step 2: Collect the control signal and the corresponding fuel injection mass flow signal to form a fault data set; set three standard operating conditions of acceleration, steady state and deceleration to construct three-domain data sets, select one domain as the target domain and the other two as the source domain, thereby forming a typical domain generalization task.

[0009] Step 3: For each domain generalization task, the CAAN model proposed is trained, and the fault classification performance under cross-domain conditions is evaluated.

[0010] In step 1 of the application, the control perception calibration module (CAC) is used to capture and parameterize the dynamic response characteristics of the healthy fuel injection system under the operating conditions, and the module characterizes the responses through scaling factor And offset factor These parameters together define a dynamic baseline for normalizing the original features to a health status representation independent of the operating conditions; due to the time correlation between samples, directly calculating the scaling parameter And offset parameter May cause instability in the training process, therefore, an incremental adjustment strategy is adopted to update the parameters according to the current sample features while ensuring overall stability: let the input fault signal be , the control signal is In the first Training iteration, the control signal Is processed through multiple linear layers to generate scaling increment And offset increment : (1), Where, when initialized, set And , respectively, represent the scaling factor and the offset factor, during the training process, these two parameters are constantly updated through iteration, wherein And Indicate the And After gradient descent optimization, finally, the input features after scaling and offset operation get the control perception features : (2).

[0011] In step 1 of the present application, the model decouples the fused features, extracts the maneuvering invariant features and the fault severity features respectively, wherein the fault severity features are trained to be highly sensitive to the change of fault severity, and the maneuvering invariant features are trained to filter out all information related to the maneuvering working condition, and only the basic fault mode existing in various working conditions is reserved, and the fault severity feature learning method is specifically as follows: When the fuel injector is blocked, the fuel injection system will compensate by delaying the injection time to ensure that the fuel injection amount remains unchanged, and the Fourier phase spectrum can effectively represent the injection time delay and the structural time sequence characteristics of the fuel injection system; the input fault feature In the time dimension The Fourier transform is performed on the Fourier phase spectrum to obtain: (3), The Fourier phase angle is defined as: (4)。

[0012] The fault severity feature information distillation in step 1 of the present application specifically includes: the teacher network uses the Fourier phase information for classification, so as to learn the discriminative phase features, and after the training is completed, the network has strong representation ability highly related to the classification semantics; in the student model training process, the intermediate representation of the teacher network is aligned with the feature generated by CAAN through formula (5), so as to guide the student model to inherit the Fourier phase-based feature representation: (5)。

[0013] The maneuvering invariant feature learning method in step 1 of the present application specifically includes: for the source domain, the loss function is defined as the classification target based on label prediction: (6), wherein, indicates the input feature, indicates the model prediction result, is the real label, indicates the source domain data set, and indicates that the sample comes from the source domain distribution, and the loss function of the target domain is the adversarial loss of the domain classifier, which can be represented as: (7), wherein, indicates the prediction result of the target domain, a real label of the target domain; During the back propagation, the gradient reversal layer enhances the discriminability between the source domain and the target domain by reversing the gradient of the extracted features, and the calculation process can be represented by the following formula: (8), wherein, denotes all trainable parameters in the feature extractor; By introducing the gradient reversal layer, the feature extractor learns the maneuvering invariant features in the process of the confusion domain classifier. The training process is as follows: (9).

[0014] In order to realize the complete separation of the fault severity feature and the maneuvering invariant feature at the information level, the cosine similarity is introduced, and its definition is shown in formula (10). By minimizing the cosine similarity, the model is guided to learn two independent information channels, so as to ensure that the fault related information and the working condition related information are separated from each other: (10).

[0015] Compared with the prior art, the present application has the following advantages: (1) The control signal is introduced in the fault diagnosis, and the control and diagnosis are realized. The proposed CAC module converts the maneuvering control information into an active calibration mechanism. The real-time control signal is mapped to the affine transformation parameter, which is used to conditionally modulate the feature channel. This strategy dynamically normalizes and corrects the extracted features using physical prior knowledge, not only improving the interpretability of the features, but also laying a foundation for realizing control-diagnosis collaboration.(2) A dual-feature decoupling framework for fuel injector fault diagnosis is constructed. The features calibrated by the control offset are decomposed into two orthogonal components: one is the maneuvering invariant feature obtained by adversarial learning, and the other is the fault severity feature extracted by knowledge distillation. The two features realize collaborative constraint through complementary regularization, and the fusion result constitutes a comprehensive and good generalization ability fault representation.(3) An expert teacher network for fuel injector fault is established. The teacher model is trained based on the Fourier phase spectrum data, which can effectively capture the key features related to blockage in the fuel injection mass flow. Through knowledge distillation, the representation is migrated to the main diagnosis model, providing a reasonable inductive bias for the model, thereby improving its robustness under different operating conditions. BRIEF DESCRIPTION OF DRAWINGS

[0016] attached Figure 1 is the diesel engine injection mass flow diagram in the present application, wherein Figure 1Fig. 1 is a diagram of fuel mass flow rate curve in diesel engine deceleration, Fig. 2 is a schematic diagram of diesel engine nozzle structure, and Fig. 3 is a diagram of different clogging degrees of diesel engine nozzle.

[0017] Fig. 4 is a schematic diagram of CAAN network structure in the application. Figure 2

[0018] Fig. 5 is a schematic diagram of CAC module network structure in the application. Figure 3

[0019] Fig. 6 is a flowchart of Fourier phase distillation learning in the application. Figure 4

[0020] Fig. 7 is a flowchart of overall fault diagnosis in the application. Figure 5

[0021] Fig. 8 is a schematic diagram of diesel injection system model in the application. Figure 6

[0022] Fig. 9 is a schematic diagram of accuracy rate comparison of different models in the application. Figure 7

[0023] Fig. 10 is a confusion matrix of CAAN model in three tasks in the application, wherein (a) is task 1, (b) is task 2, and (c) is task 3. Figure 8

[0024] Fig. 11 is a schematic diagram of influence of control signal quantity on model accuracy in the embodiment of the application, wherein Figure 9 Fig. 12 is a schematic diagram of accuracy rate, and Fig. 13 is a schematic diagram of F1 score. Figure 9

[0025] Fig. 14 is a schematic diagram of t-SEN results of the embodiment of the application under different control signal quantities, wherein Figure 10 (a), (b) and (c) in Fig. 14 involve three control signals, (d), (e) and (f) involve two control signals, and (g), (h) and (i) involve one control signal. Figure 10 DETAILED DESCRIPTION The application will be further described below in combination with the drawings and embodiments.

[0026] To cope with the cross-domain performance degradation problem caused by loss of control perception in the maneuvering process, the application proposes a domain generalization method combining control signals and fault signals, and applies it to fault diagnosis of a marine diesel engine injection system, as shown in Fig. 1, which mainly includes the following steps:

[0027] Figure 5 ​​​​​​​​​​(1) Construct a simulation model of the fuel injection system of a marine diesel engine, and integrate the control unit; simultaneously collect the control signals and the corresponding fuel injection mass flow signals to form a fault data set; (2) Set three standard operating conditions, namely acceleration, steady state, and deceleration, to construct three-domain data sets, select one domain as the target domain, and the other two as the source domains, thereby forming a typical domain generalization task; (3) For each domain generalization task, the CAAN model is trained, and the fault classification performance under cross-domain conditions is evaluated.

[0028] The CAAN model is shown in Figure 2. The network first aligns the fault signals and the control signals in time, and then embeds the control signal features into the fault signals through the CAC module to enhance the semantic expression ability of the features. Subsequently, the model decouples the fused features, extracts the maneuvering invariant features and the fault severity features: the former is obtained through the gradient reversal mechanism, and the latter is extracted by the teacher network through knowledge distillation. To ensure the complementarity and discriminability of the two types of features, the model introduces regularization constraints to promote differential feature learning. Finally, the two types of invariant features are spliced and fused to form a unified feature representation, which is input into the classifier to realize fault recognition.

[0029] Since the control signal does not directly contain fault information, inputting it together with the fault signal as features into the diagnostic model may cause the model to learn incorrect association patterns in the case of insufficient data or uneven distribution. Therefore, the present application proposes a control-aware calibration module (CAC) to capture and parameterize the dynamic response characteristics of the healthy fuel injection system under maneuvering conditions. The module characterizes these responses through scaling factors and offset factors . These parameters together define a dynamic baseline that normalizes the original features to a healthy state representation independent of the operating condition.

[0030] The structure of the CAC module is shown in Figure 3 . Since there is a time correlation between samples, directly calculating the scaling parameters and offset parameters from the control signal may cause instability in the training process. Therefore, the present application adopts an incremental adjustment strategy to update the parameters based on the current sample features while ensuring overall stability. Let the input fault signal be , and the control signal be . In the th training iteration, the control signal is processed through multiple linear layers to generate the scaling increment and the offset increment : (1), where and are the scaling factor and offset factor, respectively. During the training process, these two parameters are updated iteratively, where and denote the optimized and after gradient descent. and and and Finally, the input feature is scaled and offset to obtain the control-aware feature : (2).

[0031] Given that previous studies have shown that simple cross-domain feature alignment methods are difficult to obtain strong discriminative feature representations in domain generalization tasks, although such methods can reduce the distribution difference between different domains, they may also weaken the discriminability of features and limit their diversity. Therefore, a more effective strategy is to model features both within and across domains: in this case, 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 the maneuver operating condition, leaving only the basic fault patterns that exist under all operating conditions.

[0032] where the fault severity feature learning method is Fourier phase distillation learning: when the fuel injector is blocked, the fuel injection system will compensate by delaying the injection time to ensure that the fuel injection amount remains constant; the Fourier phase spectrum can effectively represent this injection timing delay and the structural timing features of the fuel injection system; the input fault feature is Fourier transformed in the time dimension to obtain: (3), the Fourier phase angle is defined as: (4), Fault information distillation: to avoid repeated Fourier transform calculations during the prediction phase and achieve efficient end-to-end inference, 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 is complete, the network has strong representation capabilities that are highly relevant to the classification semantics.

[0033] During the student model training process, the intermediate representation of the teacher network is generated by the features of the CAAN and equation (5) ​​Aligning to guide the student model to inherit the Fourier phase-based feature representation.

[0034] (5).

[0035] The motor-invariant feature learning method in this example is as follows: in the complex operating scenarios of a marine diesel engine, the data distribution difference between different motor conditions has a highly complex and nonlinear feature. Therefore, the global alignment method based on matrix matching (such as MMD or CORAL) is difficult to effectively capture and align the high-order or fine-grained structure in these distributions. In contrast, the domain adversarial network (DANN) can better cope with such complex nonlinear distribution changes through an adversarial training mechanism.

[0036] For the source domain, the loss function is defined as the classification target based on label prediction: (6), wherein, represents the input feature, represents the model prediction result, is the true label, represents the source domain dataset, and indicates that the sample comes from the source domain distribution.

[0037] The loss function of the target domain is the adversarial loss of the domain classifier, which can be represented as: (7), wherein, represents the prediction result of the target domain, is the true label of the target domain.

[0038] In the backpropagation process, the gradient reversal layer enhances the separability between the source domain and the target domain by performing a reverse operation on the gradient of the extracted feature, and its calculation process can be represented by the following formula: (8), wherein, represents all trainable parameters in the feature extractor.

[0039] By introducing the gradient reversal layer, the feature extractor learns the motor-invariant features in the process of confusing the domain classifier. Its training process is as follows: (9), To achieve complete separation of fault severity features and maneuvering invariance features at the information level, a cosine similarity is introduced in this case, as defined in equation (10). By minimizing the cosine similarity, the model is guided to learn two independent information channels, ensuring that fault-related information and working condition-related information are separated from each other. This strategy can extract more pure and reliable decoupled features: (10).

[0040] The experimental data of this case comes from a simulation model of a diesel engine fuel injection system, as shown in Figure 6 , the system structure includes camshaft, lift pump, four in-line fuel pumps and four injectors, and the fault signal corresponds to the fuel mass flow of each injector during injection. To simulate the nozzle clogging fault, different fault states are set by changing the nozzle aperture, and the specific parameters are shown in Table 1. The dynamic control signals of the system include the crankshaft angular velocity , camshaft speed and cam position . These three variables together determine the opening and closing characteristics of the fuel injection valve, thereby directly affecting the fuel mass flow during injection.

[0041] Table 1 Nozzle clogging fault settings

[0042] To build a domain generalization experimental environment, this case establishes a dataset containing three operating states: acceleration condition (C1, speed increases from 926 rpm to 955 rpm), steady state condition (C2, speed maintains at 955 rpm) and deceleration condition (C3, speed decreases from 955 rpm to 926 rpm). The data sampling frequency is 10 kHz, and 2,500 samples of each fault are collected under each operating condition. Based on this dataset, multiple cross-domain transfer learning tasks are designed to evaluate the model's generalization ability under different operating conditions. The detailed configuration of each task is shown in Table 2.

[0043] Table 2 Task configuration

[0044]

[0045] The CAAN model proposed in this case is implemented based on the deep learning framework PyTorch 2.6.0, which is developed by Meta and supports GPU accelerated computing. The experimental environment is configured with an Intel Core i9-13900HX processor and an NVIDIA GeForce RTX 4060 graphics card.

[0046] To ensure the reliability of generalization in the process of knowledge distillation, the teacher network is first pre-trained for 100 epochs before training the CAAN. The pre-training uses the Adam optimization algorithm with a learning rate of 0.001 and a batch size of 64. The detailed hyperparameter configuration of the network is shown in Table 3.

[0047] Table 3 Structure details of the teacher network

[0048] The core goal of the CAAN model is to achieve the collaborative learning of fault severity features and maneuvering invariant features. This goal is achieved through the joint constraint of distillation loss and domain difference loss, and the adjustment factor = 1.0 is set to balance the learning process of the two types of features. To achieve moderate alignment of the domain classifier, the gradient reversal coefficient is set to = 0.1. The model training process includes four types of loss functions: classification loss (CrossEntropyLoss), domain discrimination loss (BCELoss), knowledge distillation loss (MSELoss), and domain feature difference loss (Cosine Similarity). The training uses the Adam optimizer with an initial learning rate of 0.001 and a batch size of 64.

[0049] 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 comparison is conducted on three scenario datasets, namely Task 1, Task 2, and Task 3, and the corresponding results are shown in Figure 7 .

[0050] The evaluation indicators are accuracy (Accuracy, ACC) and F1 score, and the detailed comparison results are shown in Table 4. The confusion matrix of the CAAN model under three different speed conditions is shown in Figure 8 .

[0051] In Task 1, the model shows high classification accuracy under the given working conditions, and most samples can be correctly classified into the corresponding categories. The diagonal elements of the confusion matrix are almost equal to the total number of samples in each category, indicating that the model has high accuracy, and only a small amount of misclassification occurs between adjacent categories of some healthy states.

[0052] In Task 2, compared with Task 1, the cross-domain generalization ability of the model has decreased significantly. Consistent with the previous results, most misclassifications still occur between categories close to the healthy state, but the frequency has increased. The performance decline is particularly significant in samples representing moderate health levels, and the classification confusion phenomenon is more obvious.

[0053] In Task 3, the cross-domain generalization ability of the model is at the lowest. The off-diagonal elements in the confusion matrix increase significantly, indicating that the model has great difficulty in distinguishing these similar fault levels, and the classification reliability is significantly weakened.

[0054] Table 4 Performance comparison of different task models

[0055]

[0056] To evaluate the role of the dual-feature decoupling mechanism and the CAC module in CAAN, a series of ablation experiments were conducted. Three model configurations were considered for analysis: a baseline model that only uses maneuver-invariant features (MIF), a variant model that extracts dual features (MIF+FSF), and an enhanced model that integrates the CAC module (MIF+FSF+CAC). Experiments were conducted 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.

[0057] In the configuration that combines dual features and control signals, the proposed model achieved an average accuracy 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 was removed, the F1 scores dropped to 0.955, 0.951, and 0.956, with decreases of 3.34%, 3.06%, and 2.05%, respectively. These results indicate that although the control signal does not directly contain fault label information, it can achieve adaptive alignment of domain features through the affine transformation mechanism in the CAC module, thereby significantly improving the model's diagnostic accuracy and cross-domain generalization ability.

[0058] When the model relies only on maneuver-invariant features, its average accuracy on Tasks 1, 2, and 3 is 94.96%, 94.07%, and 94.39%, respectively, with corresponding F1 scores of 0.949, 0.941, and 0.944. Compared with the complete model that integrates control signals, the overall performance of this configuration is significantly reduced. This result indicates that although the introduction of dual features improves cross-domain recognition ability to some extent, the model still has difficulty in distinguishing fault severity in the target domain, indicating that the effect of using the feature decoupling mechanism alone is limited.

[0059] Table 5 Ablation experiment results

[0060]

[0061] To further evaluate the impact of the number of control signals on the model performance, a series of comparative experiments are designed in this example to analyze the role of control signal dimension in domain generalization. Although the control signals do not directly convey fault label information, they can represent the operating state of the system, thereby providing contextual information for dual-feature decoupling and feature generalization. Therefore, three experimental schemes are set up, using one-dimensional, two-dimensional, and three-dimensional control signals, respectively. In all configurations, the network structure and training parameters remain the same. The experiments are performed on tasks 1 to 3 for performance evaluation, and the classification accuracy results and corresponding t-SNE visualizations are shown in Figure 9 and Figure 10 .

[0062] The results show that the introduction of control signals can continuously improve the model performance, and with the increase in the number of control signals, the accuracy shows an increasing trend. When only a single control signal is used, the accuracy of the model on tasks 1, 2, and 3 is 95.93%, 95.69%, and 96.18%, respectively, with only a slight improvement compared to the model without control signals. In this stage, t-SNE shows that the class cluster boundaries are fuzzy, with obvious overlap, and the class separability is limited. After introducing two control signals, the accuracy is improved to 97.64%, 97.40%, and 97.32%, respectively, and the class separation degree is significantly improved. The boundaries of classes F2-F4 are more clear, and the intra-class samples are more closely clustered, indicating that the model has learned more discriminative and robust features. However, there are still a small number of confusions between normal operating conditions and low-level faults. When all three control signals are used, the model performance reaches the optimal, and the classes in the feature space are almost completely separated, with clear clustering boundaries and obvious cluster structure, fully reflecting the stronger generalization ability and classification performance of the model.

[0063] To overcome the problem of performance degradation in fault diagnosis of marine diesel engines due to changes in data distribution during maneuvering operation, a new method based on CAAN is proposed in this example, which emphasizes the active use of operating condition information to enhance the robustness of features. First, the CAC module applies a dynamic affine transformation to the potential features based on real-time control signals, thereby calibrating and decoupling the feature bias caused by adverse operating condition changes. Subsequently, the dual-feature decoupling framework divides the calibrated features into fault severity features and maneuvering invariant features. The maneuvering invariant features are obtained through adversarial training, while the fault severity features are distilled under the guidance of the expert teacher network. By fusing these two types of orthogonal and complementary features, CAAN can achieve high-precision and high-reliability classification of fuel injector clogging faults.

[0064] Experimental results show that, compared with existing multiple domain generalization methods, the proposed CAAN model achieves significantly higher accuracy and better overall performance in multiple cross-condition diagnosis tasks. Its advantages mainly lie 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 dual-feature decoupling combined with complementary learning enables the model to obtain fault representations that not only have good cross-domain generalization ability but also maintain strong class distinguishability, thereby 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