Marine engine lightweight fault diagnosis system oriented to variable working condition environment

Abstract

The invention discloses a marine engine lightweight fault diagnosis system for a variable working condition environment. The marine engine lightweight fault diagnosis system comprises a fault simulation and data acquisition module, a progressive separation module and a double-view weighted confrontation module, constructing a fault data set containing multiple working conditions through fault simulation, and dividing the fault data set into a training set with labels and a to-be-diagnosed sample set; the progressive separation module uses the training set to train a basic model, extracts to-be-diagnosed sample features to carry out joint scoring and sorting, then uses K-Means clustering and a cross-working-condition class structure matching mechanism to adaptively estimate the number of unknown fault samples, and trains a fine classifier; and the double-view weighted confrontation module fuses the output of the auxiliary domain discriminator and the fine classifier, realizes safe and accurate migration of diagnosis knowledge to a to-be-diagnosed sample set through weighted confrontation training, and obtains a final fault diagnosis model. According to the method, the fault diagnosis precision and robustness under variable working conditions and open class environments are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of marine engine fault diagnosis, and particularly relates to a lightweight fault diagnosis system for marine engines in variable working condition environment. BACKGROUND

[0002] In recent years, the wide application of artificial intelligence technology in the industrial field has promoted the development of the shipbuilding industry towards intelligence and unmanned. This development trend has significantly improved the integration and complexity of the ship engine room system, making the abnormal state and potential failure of the equipment in the actual operation process more concealed and difficult to detect, thereby bringing new challenges to the safe operation of the ship. As the core power unit in the ship engine room, the health status of the marine engine is directly related to the operation reliability and stability of the whole ship. Therefore, the state monitoring and intelligent fault diagnosis of the marine engine are of key significance in the development process of intelligent ships.

[0003] Thanks to the continuous progress of sensor technology, the enhancement of data acquisition and communication capabilities, and the leap development of artificial intelligence algorithms, the state recognition and fault diagnosis of marine engines have ushered in new development opportunities. The integrated application of these technologies enables practitioners to more comprehensively and finely grasp the running state of the equipment, realize early identification and predictive maintenance of potential abnormalities. Especially, data-driven methods have shown significant advantages in improving fault detection accuracy and diagnosis efficiency due to their modeling and analysis capabilities for large-scale operation data. Deep learning methods have been widely applied to the state recognition and fault classification tasks of industrial equipment due to their strong non-linear feature extraction and pattern recognition capabilities, becoming an important technical path for intelligent fault diagnosis research. Especially in dealing with high-dimensional, strong noise and non-linear engine operation data, deep learning has shown excellent feature extraction and classification capabilities, providing strong support for complex fault pattern recognition of marine engines.

[0004] Although deep learning driven diagnostic models have achieved remarkable results in closed laboratory conditions, their generalization ability in actual engineering environments still faces severe challenges. The fundamental reason is that such methods usually rely on a large number of high-quality labeled samples and the assumption that the training data and the target data to be diagnosed are consistent in distribution. In the real running scene of marine engines, it is difficult to obtain sufficient fault labeled data due to the high cost of manual labeling and the actual constraints of safe operation, especially in extreme working conditions, early damage or rare fault modes, and the training set is difficult to cover all potential fault types. And the running state of the device is affected by aging, load fluctuation, environmental disturbance and other factors, which easily induces new types of faults that have not been observed. More importantly, the statistical characteristics of the running data of the marine engine will change significantly when it performs different shipping tasks or runs in different sea conditions, loads and working conditions. Although the fault essential characteristics remain consistent at the physical level, their manifestations in the perception data (such as frequency spectrum, measured values, etc.) change due to the difference in working conditions, resulting in obvious data distribution shift of the same type of fault in different working conditions. SUMMARY

[0005] The present application provides a lightweight fault diagnosis system for marine engines in variable working condition environment to overcome the problems of relying on a large number of high-quality labeled samples, training data consistent with the target data to be diagnosed, and data distribution shift due to working condition difference.

[0006] To achieve the above purpose, the technical scheme of the present application is: A lightweight fault diagnosis system for marine engines in variable working condition environment, comprising: a fault simulation and data acquisition module, a gradual separation module and a dual-view weighted adversarial module; The fault simulation and data acquisition module is used to simulate the running state of the LNG ship engine under normal and fault conditions under various working conditions to collect thermodynamic parameter indicators and construct a fault data set; the fault data set is normalized and divided into a labeled training set and a to-be-diagnosed sample set according to the variable working condition diagnosis task; the thermodynamic parameter indicators are temperature, pressure and speed; The gradual separation module includes an information fusion sorting module and an unknown evaluation separation module; The information fusion sorting module is used to train a fault diagnosis model according to the labeled training set, the fault diagnosis model including a feature extractor, a classifier and a neighborhood similarity discriminator; the fault features of the to-be-diagnosed sample are extracted by the trained feature extractor; the joint score of the to-be-diagnosed sample is obtained by combining the trained classifier and the neighborhood similarity discriminator according to the fault features of the to-be-diagnosed sample; and the to-be-diagnosed sample is sorted according to the size of the joint score; The unknown evaluation separation module is used to set the pseudo-class cluster number of a plurality of samples to be diagnosed, and a K-Means clustering method is adopted to obtain the cosine similarity between the pseudo-class cluster center of the sample set to be diagnosed and the known class cluster center in the training set with labels, so as to obtain the final optimal class matching group, and then obtain the clustering structure consistency score under the current pseudo-class cluster number, so as to determine the optimal cluster number and obtain the number of unknown class samples; according to the number of unknown class samples and the sorted samples to be diagnosed, the samples of the unknown class are obtained; and the unknown class samples are used to train the fine classifier; The double-viewpoint weighted confrontation module is used to obtain a weighted confrontation loss function according to the trained fine classifier and the auxiliary field discriminator; the feature extractor is trained for the second time through the weighted confrontation loss function and the cross-entropy loss function, and the classifier is trained for the second time through the cross-entropy loss function, to obtain the second trained feature extractor and the second trained classifier; the sample to be diagnosed is input into the second trained feature extractor for feature extraction, and the obtained features are input into the second trained classifier, and finally the fault diagnosis result of the sample to be diagnosed is obtained.

[0007] Further, the process of obtaining the joint score of the sample to be diagnosed includes: S11, training a fault diagnosis model according to the training set with labels, to obtain a trained feature extractor, a trained classifier and a trained neighborhood similarity discriminator; S12, inputting the sample set to be diagnosed into the trained feature extractor to obtain the fault features of the sample to be diagnosed; S13, obtaining the uncertainty measure and the structural similarity measure respectively through the trained classifier and the trained neighborhood similarity discriminator according to the fault features of the sample to be diagnosed; and obtaining the joint score of the sample to be diagnosed according to the uncertainty measure and the structural similarity measure; S14, sorting the sample set to be diagnosed according to the size of the joint score.

[0008] Further, the fault diagnosis model is trained by using the cross-entropy loss and the binary cross-entropy loss, and the total loss expression of the cross-entropy loss and the binary cross-entropy loss is: (1) In the formula, L is the total loss; L ce is the cross-entropy loss; L bce is the binary cross-entropy loss; is the training set with labels; is the training set sample; is the number of labeled samples in the training set; is corresponding label; a number of known failure classes for the labeled training set; i an index of a known class for the labeled training set; a feature extractor; a classifier; a neighborhood similarity discriminator.

[0009] Further, the uncertainty measure is based on a prediction probability output by the trained classifier, and an entropy value of the sample to be diagnosed is calculated according to the prediction probability output by the trained classifier, and the expression is: (2) (3) wherein, the sample to be diagnosed; a failure feature output by the feature extractor; a probability output of the classifier; an entropy value of the sample set to be diagnosed; The structural similarity measure is the maximum value in the probability outputs of each sub-binary classifier of the neighborhood similarity discriminator, and the expression is: (4) wherein, the probability outputs of each sub-binary classifier of the neighborhood similarity discriminator; the maximum value in the probability outputs of each sub-binary classifier; A joint score is obtained through the uncertainty measure and the structural similarity measure, and the expression is: (5) wherein, the maximum value of the entropy; the joint score.

[0010] Further, the specific execution steps of obtaining the unknown class sample include: S21, setting a number of pseudo-class clusters of a plurality of samples to be diagnosed p , using a K-Meams clustering method to cluster the sample set to be diagnosed to obtain pseudo-cluster centers of the sample set to be diagnosed; the number of pseudo-class clusters of the sample to be diagnosed is C s |≤ p ≤1.5| C s |; S22, obtaining a cosine similarity between the pseudo-cluster centers of the sample set to be diagnosed and the known class cluster centers in the labeled training set, and the formula for obtaining the cosine similarity is: (6) wherein, is the known cluster center of each known class of the labeled training set; is the pseudo cluster center of the sample set to be diagnosed; i represents the index of the known class in the labeled training set; j represents the index of the pseudo class in the sample set to be diagnosed; p is the number of pseudo class clusters of the sample to be diagnosed; represents the cosine similarity between the known cluster center of the i-th labeled training set and the pseudo cluster center of the j-th sample set to be diagnosed; i represents the cosine similarity between the known cluster center of the i-th labeled training set and the pseudo cluster center of the j-th sample set to be diagnosed; j represents the cosine similarity between the known cluster center of the i-th labeled training set and the pseudo cluster center of the j-th sample set to be diagnosed; S23, if the cosine similarity between the i-th pseudo cluster center and the j-th known cluster center is the largest, and the cosine similarity between the i-th known cluster center and the j-th pseudo cluster center is the largest, then the matching pair formed by the i-th pseudo cluster center and the j-th known cluster center at this time is added to the optimal class matching group to obtain the final optimal class matching group; j , i , j i obtains the sum of the cosine similarities of all matching pairs in the final optimal class matching group as the clustering structure consistency score under the current pseudo class cluster number; the pseudo class cluster number with the highest clustering structure consistency score is selected as the optimal cluster number, and if there are two or more pseudo class cluster numbers with the highest clustering structure consistency score, the pseudo class cluster number with the smallest numerical value is selected as the optimal cluster number; the number of samples in the sample set to be diagnosed that are successfully matched is obtained according to the optimal cluster number; p S25, according to the number of samples in the sample set to be diagnosed that are successfully matched in the optimal cluster number, the proportion of the total number of samples in the sample set to be diagnosed is calculated, and the proportion of the unmatched samples is estimated; combined with the current batch size, the number of unknown class samples is obtained, and the expression is: (7) wherein, n is the number of unknown class samples; is the number of samples successfully matched in the optimal cluster number; is the class successfully matched in the sample set to be diagnosed; is the current batch size; is the number of samples in the sample set to be diagnosed; is the index of the class successfully matched in the sample set to be diagnosed.​​​​​​​

[0011] Further, a fine classifier is constructed and trained by unknown class samples to obtain a trained fine classifier; wherein the fine classifier is trained by minimizing binary cross-entropy, and the expression of the minimizing binary cross-entropy is: (8) In the formula, is the fine classifier; is the constructed unlabeled known or unknown sample set; is the unlabeled known or unknown sample; is the known class or unknown class pseudo label.

[0012] Further, according to the trained fine classifier and the auxiliary domain discriminator, a weighted adversarial loss function is obtained, and the specific execution steps include: S31, the loss function of the auxiliary domain discriminator is used to train the auxiliary domain discriminator; and the expression of the loss function of the auxiliary domain discriminator is as follows: (9) In the formula, D s is the labeled training set; D t is the sample set to be diagnosed; N s is the number of samples in the training set; N t is the number of samples in the sample set to be diagnosed; is the true domain label of the sample ; is the output of the sample in the auxiliary domain discriminator D aux ; S32, the probability of the unknown class of the sample to be diagnosed output by the trained fine classifier is calculated to calculate the class-level weight; Wherein, the expression of the class-level weight is: (10) In the formula, is the class-level weight; is the probability of the unknown class of the sample to be diagnosed output by the fine classifier; is the training set sample; S33, the output of the trained auxiliary domain discriminator is fused with the class-level weight to obtain the dual-view weight of the sample to be diagnosed; and the expression of the dual-view weight of the sample to be diagnosed is: (11) In the formula, The dual-view weight of the sample to be diagnosed; The proportion of each view weight; The output of the auxiliary field discriminator; S34, the dual-view weight of the sample to be diagnosed is integrated into the adversarial loss function of the adversarial training, and a weighted adversarial loss function is obtained; The expression of the weighted adversarial loss function is: (12) In the formula, The weighted adversarial loss; d i The real field label of the sample in the labeled training set; d j The real field label of the sample in the sample to be diagnosed; The output of the sample in the labeled training set in the field discriminator; The output of the sample in the sample to be diagnosed in the field discriminator.

[0013] Further, the expression of the cross-entropy loss function is as follows: (13) (14) In the formula, The cross-entropy loss of the training set sample ; The cross-entropy loss of the potential unknown class sample of the sample to be diagnosed set; The label of the potential unknown class sample of the sample to be diagnosed set.

[0014] Further, the fault diagnosis system also constructs a lightweight discriminative feature extraction module; the lightweight discriminative feature extraction module is used to replace the traditional standard convolution operation with a deep separable convolution by taking MobileNet as a backbone network, and a lightweight discriminative framework is constructed.

[0015] Beneficial effects: the lightweight fault diagnosis system of the marine engine facing variable working condition environment, through the unknown evaluation separation module, adopts an unsupervised clustering matching mechanism to adaptively estimate the number of unknown class samples, and trains a fine classifier to preliminarily distinguish the known and unknown categories. On this basis, a dual-view weighted adversarial module is introduced, the field discriminator and the feature extractor are trained in an antagonistic manner, the field-invariant fault features irrelevant to the working condition are effectively extracted, the error alignment of the unknown class samples is suppressed, the generalization ability and knowledge transfer effect of the diagnosis model in the working condition change scene are significantly improved, the negative transfer problem is fundamentally avoided, and the robustness of the system is greatly enhanced. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating the execution process of the system of the present invention; Figure 2 This is a diagram showing the overall architecture of the system of the present invention; Figure 3 This is a schematic diagram of the standard convolution process in an embodiment of the present invention; Figure 4 This is a schematic diagram of the depth-separable convolution process in an embodiment of the present invention; Figure 5 This is a simulation model of the Wärtsilä9L34DF dual-fuel marine engine in an embodiment of the present invention; Figure 6a This is a comparison of the diagnostic accuracy of the six cross-condition subtasks of the E1 task in this embodiment of the invention. Figure 6b This is a comparison of the diagnostic accuracy of the six cross-condition subtasks of the E2 task in this embodiment of the invention. Figure 6c This is a comparison of the diagnostic accuracy of the six cross-condition subtasks of the E3 task in this embodiment of the invention. Figure 6d This is a comparison of the diagnostic accuracy of the six cross-condition subtasks of the E4 task in this embodiment of the invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] This embodiment provides a lightweight fault diagnosis system for marine engines operating in varying conditions, such as... Figure 1 As shown, it includes: a fault simulation and data acquisition module, a progressive separation module, and a dual-view weighted adversarial module; The fault simulation and data acquisition module is used to simulate the operating status of LNG ship engines under normal and fault conditions under various operating conditions, so as to collect thermodynamic parameters and construct a fault dataset; the fault dataset is normalized and divided into a labeled training set and a sample set to be diagnosed according to the variable operating condition diagnosis task; the thermodynamic parameters are temperature, pressure and speed. The progressive separation module includes an information fusion and sorting module and an unknown evaluation and separation module; The information fusion and ranking module is used to train a fault diagnosis model based on a labeled training set. The fault diagnosis model includes a feature extractor, a classifier, and a neighborhood similarity discriminator. The trained feature extractor extracts fault features of the samples to be diagnosed. Based on the fault features of the samples to be diagnosed, the trained classifier and neighborhood similarity discriminator are combined to obtain a joint score of the samples to be diagnosed. The samples to be diagnosed are ranked according to the size of the joint score. The unknown evaluation and separation module is used to set the number of pseudo-clusters for multiple samples to be diagnosed, and to use the K-Means clustering method to obtain the cosine similarity between the pseudo-cluster centers of the sample set to be diagnosed and the known cluster centers in the labeled training set, so as to obtain the final optimal category matching group, and then obtain the cluster structure consistency score under the current number of pseudo-clusters to determine the optimal number of clusters and obtain the number of unknown class samples; based on the number of unknown class samples and the sorted samples to be diagnosed, the unknown class samples are obtained; and the unknown class samples are used to train the fine classifier. The dual-view weighted adversarial module is used to obtain a weighted adversarial loss function based on the trained fine classifier and auxiliary domain discriminator; to perform a second training on the feature extractor using the weighted adversarial loss function and the cross-entropy loss function, and to perform a second training on the classifier using the cross-entropy loss function, resulting in a second-trained feature extractor and a second-trained classifier; to input the sample to be diagnosed into the second-trained feature extractor for feature extraction, and to input the obtained features into the second-trained classifier, finally obtaining the fault diagnosis result of the sample to be diagnosed.

[0020] Specifically, such as Figure 2 As shown, a fault simulation and data acquisition module is first constructed to simulate the normal and various fault states of LNG marine engines under multiple variable operating conditions. The system collects key thermodynamic parameters such as temperature, pressure and speed to construct a multi-condition fault dataset. The dataset is then normalized and divided into a labeled training set and a sample set to be diagnosed, providing a data foundation for subsequent diagnostic tasks. Secondly, a progressive separation module is constructed, which consists of two core units: information fusion ranking and unknown evaluation separation. The information fusion ranking module trains an initial fault diagnosis model using a labeled training set. This model integrates a feature extractor, a classifier, and a neighborhood similarity discriminator. The trained model extracts the deep fault features of the samples to be diagnosed and calculates their joint score (combining model prediction confidence and neighborhood similarity). Based on the score, all samples to be diagnosed are ranked according to their confidence level. The unknown assessment and separation module sets multiple pseudo-category cluster numbers and uses the K-Means clustering method to calculate the cosine similarity between the pseudo-cluster centers of the sample set to be diagnosed and the cluster centers of the known clusters, thereby determining the optimal number of clusters and the number of unknown class samples. Based on this number and the ranking results, unknown class fault samples are accurately separated from the samples to be diagnosed, and these samples are used to train a fine classifier to enhance the model's recognition of unknown faults. Next, a dual-view weighted adversarial module is constructed. This module introduces an auxiliary domain discriminator that works in conjunction with the trained fine classifier to calculate a weighted adversarial loss function. By combining this loss function with the cross-entropy loss function, the feature extractor is subjected to a second adversarial training, while the classifier is optimized to ensure that the features extracted by the model under varying operating conditions have both fault discriminative power and operating condition invariance. Finally, the samples to be diagnosed are input into the second-trained model to achieve accurate identification of known faults and effective discovery of unknown faults, completing lightweight and high-precision fault diagnosis under varying operating conditions.

[0021] Preferably, the process of obtaining the joint score of the sample to be diagnosed includes: S11. Based on the labeled training set, train the fault diagnosis model to obtain a trained feature extractor, a trained classifier, and a trained neighborhood similarity discriminator. S12. Input the set of samples to be diagnosed into the trained feature extractor to obtain the fault features of the samples to be diagnosed. S13. Based on the fault characteristics of the sample to be diagnosed, uncertainty measure and structural similarity measure are obtained through the trained classifier and the trained neighborhood similarity discriminator, respectively; based on the uncertainty measure and structural similarity measure, the joint score of the sample to be diagnosed is obtained. S14. Sort the sample set to be diagnosed according to the joint score.

[0022] Preferably, the fault diagnosis model is trained using cross-entropy loss and binary cross-entropy loss. The total loss expression for cross-entropy loss and binary cross-entropy loss is as follows: (1) In the formula, L Total loss; Lce Cross-entropy loss; L bce For binary cross-entropy loss; For labeled training set; These are training set samples; The number of labeled samples in the training set; for Corresponding tags; This represents the number of known fault categories in the labeled training set. i For the index of the known classes in the labeled training set; For feature extractors; For classifiers; It serves as a neighborhood similarity discriminator.

[0023] Preferably, the uncertainty measure is based on the predicted probability output by the trained classifier, and the entropy value of the sample to be diagnosed is calculated based on the predicted probability output by the trained classifier, expressed as: (2) (3) In the formula, The sample is to be diagnosed. The fault features output by the feature extractor; This represents the probability output of the classifier. The entropy value of the sample set to be diagnosed; The structural similarity metric is the maximum value among the probability outputs of each sub-binary classifier of the neighborhood similarity discriminator, expressed as: (4) In the formula, This represents the probability output of each sub-binary classifier of the neighborhood similarity discriminator; This represents the maximum probability output of each sub-binary classifier; The joint score is obtained by using the uncertainty measure and the structural similarity measure, expressed as: (5) In the formula, This represents the maximum value of entropy. For combined scores.

[0024] Preferably, the specific execution steps for obtaining unknown class samples include: S21. Set the pseudo-class clustering number for multiple samples to be diagnosed. p The K-Meams clustering method was used to cluster the sample set to be diagnosed, obtaining pseudo-cluster centers of the sample set; the number of pseudo-clusters of the sample set to be diagnosed was | C s |≤p ≤1.5| C s |; S22. Obtain the cosine similarity between the pseudo-cluster centers of the sample set to be diagnosed and the known cluster centers in the labeled training set. The formula for obtaining the cosine similarity is: (6) In the formula, The cluster centers of each known class in the labeled training set; These serve as pseudo-cluster centers for the sample set to be diagnosed. i Represents the index of a known class in the labeled training set; j An index representing the pseudo-class in the set of samples to be diagnosed; p The pseudo-class cluster number of the sample to be diagnosed; Indicates the first i The known cluster centers of the training set with the nth label and the nth label j Cosine similarity between pseudo-cluster centers of a set of samples to be diagnosed; S23. If after traversing all known cluster centers, the... j The pseudo-cluster centers and the first , The cosine similarity is maximized among the known cluster centers. Simultaneously, after traversing all pseudo-cluster centers, the th... i The known cluster centers and the first , If the cosine similarity between the pseudo-cluster centers is maximized, then the i-th cluster center will be... j The pseudo-cluster centers and the first i The matching pairs formed by the known cluster centers are added to the optimal category matching group to obtain the final optimal category matching group; Specifically, considering the one-to-one category correspondence between working conditions, in order to improve the robustness of structure matching, optimal category matching is performed independently from two perspectives: From the perspective of the labeled training set, based on cosine similarity, a pseudo-cluster center that is most similar in the sample set to be diagnosed is matched for each known cluster center; From the perspective of the sample set to be diagnosed, based on cosine similarity, each pseudo-cluster center is matched with a known cluster center that is most similar in the training set; Extract the items with consistent category pairs from the matching results of the two perspectives, and use them as the final optimal category matching group; The sum of the cosine similarities of all matching pairs in the final optimal class matching group is used as the current pseudo-class cluster number. pThe cluster structure consistency score is used as the optimal number of clusters. If there are two or more pseudo-category clusters with the highest cluster structure consistency scores, the pseudo-category cluster with the smallest cluster number is selected as the optimal number of clusters. Based on the optimal number of clusters, the number of samples in the set of samples to be diagnosed that have been successfully matched is obtained. S25. Based on the number of successfully matched samples in the diagnostic set within the optimal clustering number, calculate the ratio of this number to the total number of samples in the diagnostic set, and estimate the proportion of unmatched samples; combined with the current batch size, obtain the number of samples in the unknown class, expressed as: (7) In the formula, n The number of samples in the unknown class; This represents the number of samples that were successfully matched in the optimal cluster. The categories that were successfully matched in the set of samples to be diagnosed; This is the current batch size; The number of samples in the sample set to be diagnosed; An index of the categories that were successfully matched in the sample set to be diagnosed.

[0025] Preferably, n samples are selected from the samples with the lowest joint scores as potential unknown class samples, and n samples are selected from the samples with the highest scores as known class samples, which are used as the training set for the fine classifier; the fine classifier is constructed and trained using the fine classifier training set to obtain the trained fine classifier. The training of the fine classifier is optimized by minimizing the binary cross-loss, and the expression for minimizing the binary cross-loss is as follows: (8) In the formula, For fine classifiers; For the constructed unlabeled set of known or unknown samples; These are unlabeled known or unknown samples; Pseudo-labels for known or unknown classes.

[0026] Specifically, this implementation introduces the DANN framework to effectively transfer diagnostic knowledge obtained from the labeled training set to the sample set to be diagnosed. The DANN framework imitates the architecture of the GAN network and adds a domain discriminator on the basis of the traditional feature extractor and classifier structure. It achieves cross-condition invariant features through the adversarial relationship between the feature extractor and the domain discriminator. Specifically, the feature extractor outputs a feature that makes it impossible for the domain discriminator to determine the source of the sample (labeled training set or sample set to be diagnosed), while the domain discriminator tries its best to identify which sample set the input feature comes from. The two work against each other and train continuously until they reach Nash equilibrium. During training, the adversarial loss function is used for optimization. After training is complete, the feature extractor can output a sample feature that the domain discriminator cannot determine its source. This implicitly finds the sample feature that is invariant across operating conditions. To further constrain this feature, supervised training is performed on the feature output by the feature extractor using the training set during each training session to ensure that the final cross-operating condition invariant feature has the discriminative feature of fault detection.

[0027] Preferably, in this embodiment, an auxiliary domain discriminator is introduced and trained; based on the trained fine classifier and the auxiliary domain discriminator, a weighted adversarial loss function is obtained. The specific execution steps include: S31. The auxiliary domain discriminator is trained using its loss function; the expression for the loss function of the auxiliary domain discriminator is as follows: (9) In the formula, D s The training set is labeled; D t The sample set to be diagnosed; N s The number of samples in the training set; N t The number of samples in the sample set to be diagnosed; For the sample Real-world domain labels; For the sample In auxiliary domain discriminator D aux The output; S32. Calculate the class-level weights by training a fine classifier to output the probability of the unknown category of the sample to be diagnosed. The expression for class-level weights is as follows: (10) In the formula, Class-level weights; The probability of the unknown category output by the fine classifier; These are training set samples; S33. The output of the trained auxiliary domain discriminator is fused with the class-level weights to obtain the dual-view weights of the sample to be diagnosed; the expression for the dual-view weights of the sample to be diagnosed is: (11) In the formula, The weights for the two perspectives of the sample to be diagnosed; The weighting percentage for each perspective; For the output of the auxiliary domain discriminator; S34. Incorporate the dual-view weights of the sample to be diagnosed into the adversarial loss function of the adversarial training to obtain the weighted adversarial loss function; The expression for the weighted adversarial loss function is: (12) In the formula, To weight the losses; d i The true domain labels for the samples in the labeled training set; d j The true domain labels of the samples in the sample set to be diagnosed; The output of the domain discriminator for labeled samples in the training set; This is the output of the neighborhood discriminator for the samples in the sample set to be diagnosed.

[0028] Preferably, the expression for the cross-entropy loss function is as follows: (13) (14) In the formula, For training set samples Cross-entropy loss; Potentially unknown class samples in the sample set to be diagnosed Cross-entropy loss; Potentially unknown class samples in the sample set to be diagnosed The tag.

[0029] Preferably, this fault diagnosis system also constructs a lightweight discriminative feature extraction module; the lightweight discriminative feature extraction module is used to construct a lightweight discriminative framework by using MobileNet as the backbone network and replacing the traditional standard convolution operation with depthwise separable convolution.

[0030] Specifically, MobileNet uses depthwise separable convolution instead of traditional standard convolution operations; like Figure 3 As shown, traditional standard convolution operations need to simultaneously perform spatial feature extraction and channel-dimensional feature fusion; the computational cost and parameter count of standard convolution can be expressed as: (15) (16) In the formula, This represents the computational cost of standard convolution. The number of parameters for a standard convolution; The kernel size; The feature output size; Input the number of channels; This represents the number of output channels.

[0031] The depthwise separable convolution used in this embodiment breaks down the traditional standard convolution operation into two operations: depthwise convolution and pointwise convolution. Depthwise convolution performs spatial convolution on each input channel independently, and the number of output channels is the same as the number of input channels. Pointwise convolution uses a 1×1 convolution kernel to fuse the output of depthwise convolution to obtain a fused output. like Figure 4 As shown, the computational cost and parameter count of depthwise separable convolution can be expressed as: (17) (18) In the formula, The computational cost of depthwise separable convolution; This represents the computational cost of the depthwise convolution process. This represents the computational cost of the pointwise convolution process. The number of parameters for depthwise separation convolution; The number of parameters in the depthwise convolution process; The number of parameters in the pointwise convolution process; In the diagram, Depthwise Conv represents depthwise convolution; Pointwise Conv represents pointwise convolution.

[0032] In this embodiment, a ratio method is used to compare and analyze the standard convolution with depthwise separable kernels. The expression is as follows: (19) (20) By comparing formulas (17) and (18), depthwise separable convolution has lower computational cost and fewer parameters than standard convolution, which significantly reduces deployment costs on resource-constrained embedded platforms.

[0033] In a specific embodiment, using the Wärtsilä 9L34DF dual-fuel marine diesel engine mounted on an LNG vessel as the experimental object, an AvlBoost simulation model of the engine was constructed. The simulation model is as follows: Figure 5 As shown; Four typical fault types were simulated using a simulation model: intercooler water-side blockage (F1), intercooler gas-side blockage (F2), compressor failure (F3), and turbine failure (F4). Specific fault simulation details are shown in Table 1. Table 1. Engine fault setting parameter range

[0034] This embodiment simulates faults under three typical load conditions (100%, 75%, and 50%), covering a variety of operating states and fault types, thereby enhancing the diversity and adaptability of the data and constructing a representative fault dataset.

[0035] Specifically, this embodiment designs an evaluation system for variable operating conditions and open-category scenarios to address the issues of incomplete labels and data distribution drift. The incomplete label problem arises because the training set may be missing some fault categories due to limitations in sample collection conditions under actual operating conditions, causing the model to face potential interference from unknown categories during the diagnostic process. The data distribution drift problem arises because the statistical distribution of test samples often differs significantly from that of training samples due to factors such as load fluctuations and environmental disturbances during actual operation, increasing the difficulty of generalization in diagnosis.

[0036] The evaluation system in this embodiment comprehensively considers two typical real-world problems in the task design: incomplete labels in the training sample set and distribution drift of the sample set to be diagnosed.

[0037] In order to simulate the situation where training data collection is limited under actual working conditions, one or more fault categories are randomly missing in the training sample set, and unknown fault types that may be encountered in the diagnosis process are introduced to reflect the diagnostic uncertainty caused by incomplete label information. The sample set to be diagnosed is uniformly selected from unlabeled data that have significant differences in operating status compared with the training sample set, in order to simulate the distribution shift caused by factors such as load fluctuations and environmental disturbances, and to reflect the challenges of different working conditions in real scenarios.

[0038] In the designed tasks, each task is configured with six typical operating condition change paths (100%→75%, 75%→50%, 75%→100%, 50%→75%, 100%→50%, 50%→100%) to comprehensively cover the state change trends that may occur in actual operation; and 1000 sets of samples are collected under each operating condition to ensure a balanced distribution of various types of samples; the detailed configuration of each task is shown in Table 2. Table 2. Variable Operating Condition Diagnostic Task Design

[0039] In the table, u represents the category of unknown faults that occur in the training conditions due to incomplete data collection.

[0040] In a specific embodiment, four representative diagnostic methods are selected as comparison methods for the system of this embodiment; the four diagnostic methods are Domain Adversarial Neural Network (DANN), State Transition Algorithm (STA), Improved Weighted Statistical Spherical Decoding Algorithm (IWOSDA), and Theory Guided Progressive Transfer Learning Network (TPTLN). The following three evaluation metrics are introduced to comprehensively evaluate the fault diagnosis performance of the system (PSWAN) in this embodiment: 1. Overall Accuracy: The average diagnostic accuracy of all samples to be diagnosed, expressed as Acc. * ; 2. Shared Class Accuracy: The diagnostic accuracy among samples belonging to known classes in the training set, denoted as Acc. k ; 3. Unknown Class Accuracy: The accuracy rate of identifying samples of unknown class among the samples to be diagnosed, expressed as Acc. un .

[0041] The comparison results are shown in Table 3. It can be concluded that the system in this embodiment achieved the best diagnostic effect in all tasks: Table 3. Comparison under different tasks (E1~E4).

[0042] Figure 6a , Figure 6b , Figure 6c , Figure 6d The diagnostic performance of the system (PSWAN) and the comparison model in six varying operating conditions, from E1 to E4, is shown in radar form. It can be clearly observed that the radar map area of ​​the system (PSWAN) in this embodiment is the largest, which further verifies that the system (PSWAN) in this embodiment achieves an excellent balance between fine transfer of shared classes and recognition of fixed classes, and has stronger fault diagnosis performance.

[0043] The present invention has the following beneficial effects: This invention discloses a lightweight fault diagnosis system for marine engines operating under varying conditions. Through an unknown assessment and separation module, an unsupervised clustering matching mechanism is used to adaptively estimate the number of unknown class samples, and a fine-grained classifier is trained to initially distinguish between known and unknown categories. Building upon this, a dual-view weighted adversarial module is introduced. Through adversarial training of the domain discriminator and feature extractor, domain-invariant fault features independent of operating conditions are effectively extracted, suppressing misalignment of unknown class samples. This significantly improves the generalization ability and knowledge transfer effect of the diagnostic model under varying operating conditions, fundamentally avoiding negative transfer problems and greatly enhancing the system's robustness.

[0044] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A lightweight fault diagnosis system for marine engines in variable operating conditions, characterized by, The application relates to a fault diagnosis method based on a double-view weighted adversarial network. The fault simulation and data acquisition module is used for simulating the running state of an LNG ship engine under normal and fault states under various working conditions to collect thermodynamic parameter indexes and construct a fault data set. The fault data set is normalized and divided into a labeled training set and a to-be-diagnosed sample set according to a variable working condition diagnosis task; the thermodynamic parameter indexes are temperature, pressure and rotating speed; The progressive separation module comprises an information fusion sorting module and an unknown evaluation separation module. The information fusion sorting module is used for training a fault diagnosis model according to the labeled training set, wherein the fault diagnosis model comprises a feature extractor, a classifier and a neighborhood similarity discriminator; the fault feature of the to-be-diagnosed sample is extracted through the trained feature extractor; The joint score of the to-be-diagnosed sample is obtained by combining the trained classifier and the neighborhood similarity discriminator according to the fault feature of the to-be-diagnosed sample; and the to-be-diagnosed sample is sorted according to the size of the joint score. The unknown evaluation separation module is used for setting the pseudo-class cluster number of the to-be-diagnosed sample, adopting a K-Means clustering method to obtain the cosine similarity between the pseudo-class center of the to-be-diagnosed sample set and the known-class cluster center in the labeled training set, obtaining the final optimal class matching group, and then obtaining the clustering structure consistency score under the current pseudo-class cluster number to determine the optimal cluster number and obtain the unknown-class sample number; The sample of the unknown class is obtained according to the unknown-class sample number and the sorted to-be-diagnosed sample; The fine classifier is trained by using the sample of the unknown class; The double-view weighted adversarial module is used for obtaining a weighted adversarial loss function according to the trained fine classifier and an auxiliary field discriminator; the feature extractor is secondly trained through the weighted adversarial loss function and a cross-entropy loss function, and the classifier is secondly trained through the cross-entropy loss function, so that the secondly trained feature extractor and the secondly trained classifier are obtained; the to-be-diagnosed sample is input into the secondly trained feature extractor for feature extraction, and the obtained feature is input into the secondly trained classifier, so that the fault diagnosis result of the to-be-diagnosed sample is finally obtained. The process of obtaining the joint score of the to-be-diagnosed sample comprises the following steps:

2. The lightweight fault diagnosis system for marine engine facing variable working condition environment according to claim 1, characterized in that, S11, training a fault diagnosis model according to the labeled training set to obtain a trained feature extractor, a trained classifier and a trained neighborhood similarity discriminator; S12, inputting the to-be-diagnosed sample set into the trained feature extractor to obtain the fault feature of the to-be-diagnosed sample; S13, obtaining an uncertainty measure and a structural similarity measure through the trained classifier and the trained neighborhood similarity discriminator according to the fault feature of the to-be-diagnosed sample; and obtaining the joint score of the to-be-diagnosed sample according to the uncertainty measure and the structural similarity measure; S14, sorting the to-be-diagnosed sample set according to the size of the joint score. The fault diagnosis model is trained by using a cross-entropy loss and a binary cross-entropy loss, and the total loss expression of the cross-entropy loss and the binary cross-entropy loss is 3. The lightweight fault diagnosis system for marine engine facing variable working condition environment according to claim 2, characterized in that, ​ （1） wherein, L is the total loss; L ce is the cross-entropy loss; L bce is the binary cross-entropy loss; is the labeled training set; is a training set sample; is the number of labeled samples in the training set; is is the corresponding label; is the number of known failure classes in the labeled training set; i is the index of the known class in the labeled training set; is a feature extractor; is a classifier; is a neighborhood similarity discriminator.

4. The lightweight fault diagnosis system for marine engine facing variable working condition environment according to claim 2, characterized in that, The uncertainty measure is a prediction probability output based on the trained classifier, and the entropy value of the sample to be diagnosed is calculated according to the prediction probability output of the trained classifier, and the expression is: （2） （3） wherein is the sample to be diagnosed; is the fault feature output by the feature extractor; is the probability output of the classifier; is the entropy value of the sample set to be diagnosed; The structural similarity measure is the maximum value in the probability output of each sub-binary classifier of the neighborhood similarity discriminator, and the expression is: （4） wherein is the probability output of each sub-binary classifier of the neighborhood similarity discriminator; is the maximum value of the probability outputs of each sub-binary classifier. Through the uncertainty measure and the structural similarity measure, a joint score is obtained, and the expression is: （5） wherein is the maximum of the entropy; is the joint score.

5. The lightweight fault diagnosis system for marine engine facing variable working condition environment according to claim 1, characterized in that, The specific implementation steps of obtaining the unknown class sample include: S21, set the pseudo-class cluster number of a plurality of samples to be diagnosed p , a K-Meams clustering method is used to cluster the sample set to be diagnosed to obtain a pseudo-cluster center of the sample set to be diagnosed; the pseudo-cluster number of the sample to be diagnosed is C s |≤ p ≤1.5| C s |; S22, obtaining the cosine similarity between the pseudo clustering center of the sample set to be diagnosed and the known class clustering center in the labeled training set, and the formula for obtaining the cosine similarity is: （6） wherein, is the known cluster center of each known class of the labeled training set; is the pseudo cluster center of the sample set to be diagnosed; i represents the index of the known class in the labeled training set; j represents the index of the pseudo class in the sample set to be diagnosed; p is the number of pseudo class clusters of the sample to be diagnosed; represents the cosine similarity between the known cluster center of the i th labeled training set and the pseudo cluster center of the j th sample set to be diagnosed; S23, if the cosine similarity between the first pseudo cluster center and the first known cluster center is maximum, while the cosine similarity between the first known cluster center and the second pseudo cluster center is maximum, then the matching pair formed by the first pseudo cluster center and the first known cluster center at this time is added to the optimal class matching group to obtain the final optimal class matching group. j , i , j i ​​​​​​ S24, obtaining the sum of cosine similarities of all matching pairs in the final optimal category matching group as the current pseudo-category cluster number p the cluster structure consistency score of the pseudo-category cluster number with the highest cluster structure consistency score as the optimal cluster number, if there are two or more pseudo-category cluster numbers with the highest cluster structure consistency score, selecting the pseudo-category cluster number with the smallest numerical value as the optimal cluster number; and obtaining the number of samples in the successfully matched sample set in the sample set to be diagnosed according to the optimal cluster number; S25, according to the number of samples in the sample set to be diagnosed that are successfully matched in the optimal cluster number, the proportion of the total number of samples in the sample set to be diagnosed is calculated, and the proportion of the unmatched samples is estimated; combined with the size of the current batch, the number of unknown class samples is obtained, and the expression is: （7） wherein, n is the number of samples of unknown classes; is the number of samples of unknown classes that are successfully matched in the optimal cluster number; is the class that is successfully matched in the sample set to be diagnosed; is the current batch size; is the number of samples in the sample set to be diagnosed; is the index of the class that is successfully matched in the sample set to be diagnosed.

6. The lightweight fault diagnosis system for marine engine facing variable working condition environment according to claim 1, characterized in that, A fine classifier is constructed and trained by unknown class samples to obtain a trained fine classifier; wherein the fine classifier is trained by minimizing the binary cross-entropy loss, and the expression of the binary cross-entropy loss is: （8） wherein is a fine classifier; is a constructed unlabeled known or unknown sample set; is an unlabeled known or unknown sample; is a known class or unknown class pseudo label.

7. The lightweight fault diagnosis system for marine engine facing variable working condition environment according to claim 1, characterized in that, According to the trained fine classifier and the auxiliary field discriminator, a weighted adversarial loss function is obtained, and the specific implementation steps include: S31, the loss function of the auxiliary field discriminator is used to train the auxiliary field discriminator; the expression of the loss function of the auxiliary field discriminator is as follows: （9） wherein, D s is a labeled training set; D t is a set of samples to be diagnosed; N s is the number of samples in the training set; N t is the number of samples in the set of samples to be diagnosed; is a sample with a true domain label; is a sample with an output from the auxiliary domain discriminator D aux ; S32, the probability of the unknown class of the sample to be diagnosed output by the trained fine classifier is calculated to calculate the class-level weight; The expression of the class-level weight is: （10） wherein, is a class-level weight; is a probability that the unknown class of the sample to be diagnosed is output by the fine classifier output; is a training set sample; S33, the output of the trained auxiliary field discriminator is fused with the class-level weight to obtain the dual-view weight of the sample to be diagnosed; the expression of the dual-view weight of the sample to be diagnosed is: （11） In the formula, is a double-view weight of the sample to be diagnosed; is a proportion of each view weight; is an output of the auxiliary field discriminator; S34, the dual-view weight of the sample to be diagnosed is integrated into the adversarial loss function of the adversarial training to obtain a weighted adversarial loss function; The expression of the weighted adversarial loss function is: （12） wherein is a weighted adversarial loss; d i is a true domain label of a sample in the labeled training set; d j is a true domain label of a sample in the set of samples to be diagnosed; is an output of the domain discriminator for a sample in the labeled training set; is an output of the domain discriminator for a sample in the set of samples to be diagnosed.

8. The lightweight fault diagnosis system for marine engine facing variable working condition environment according to claim 1, characterized in that, The expression of the cross-entropy loss function is as follows: （13） （14） wherein, is the cross-entropy loss for the training set samples ; is the cross-entropy loss for the potential unknown class samples of the set of samples to be diagnosed ; is the label for the potential unknown class samples of the set of samples to be diagnosed .

9. The lightweight fault diagnosis system for marine engine facing variable working condition environment according to claim 1, characterized in that, The fault diagnosis system also constructs a lightweight discriminative feature extraction module; the lightweight discriminative feature extraction module is used to replace the traditional standard convolution operation with a deep separable convolution by taking MobileNet as the backbone network to construct a lightweight discriminative framework.

Images