Ship diesel engine fault diagnosis method and system based on improved kolmogorov-arnold network

Abstract

The application provides a ship diesel engine fault diagnosis method and system based on an improved Kolmogorov-Arnold network, for the first time introduces the KAN architecture into the field of ship diesel engine fault diagnosis, subverts the paradigm of traditional MLP "fixed node activation function", and places a learnable B-spline nonlinear activation function on the network connection edge, so that the model can adaptively fit the highly distorted nonlinear manifold between the thermal parameters and the fault types; at the same time, the L1 sparsification regularization term is innovatively introduced, which forces the network to automatically prune redundant connections during training, and only the key fault propagation path is reserved, thereby realizing high diagnostic accuracy and greatly compressing the parameter amount, and providing a new solution with high precision, light weight, strong robustness and interpretability for intelligent fault diagnosis of ship diesel engines.

Description

Technical Field

[0001] This application belongs to the field of marine diesel engine fault diagnosis technology, and in particular relates to a marine diesel engine fault diagnosis method and system based on an improved Kolmogorov-Arnold network. Background Technology

[0002] As the core equipment of a ship's power system, the operating status of marine diesel engines directly affects the ship's navigation safety, economy, and environmental performance. With the International Maritime Organization (IMO) imposing increasingly stringent environmental regulations on the Energy Efficiency Design Index (EEDI) and the Ship Energy Efficiency Management Program (SEEMP), and with the shipbuilding industry's trend towards digitalization and intelligence, intelligent fault diagnosis technology for marine diesel engines has become a key technical means to ensure the safe and economical operation of ships.

[0003] Existing methods for diagnosing marine diesel engine faults (especially those based on traditional multilayer perceptrons (MLP), backpropagation (BP), and their combined model SOM-BP) suffer from the following drawbacks: 1. Rigid activation mechanism: Traditional neural networks, using fixed activation functions, struggle to efficiently fit the highly nonlinear and complex mapping relationships between diesel engine thermodynamic parameters, necessitating a significant increase in network depth / width, leading to parameter redundancy. 2. Prone to overfitting with small samples: Obtaining fault samples for marine diesel engines is difficult, making traditional large-parameter models highly susceptible to overfitting and exhibiting poor generalization ability. 3. Lack of interpretability due to the "black box" nature of the model: The decision-making process of traditional neural networks is opaque, making it impossible for engine room managers to determine which thermodynamic parameters caused the fault diagnosis, thus hindering engineering trust. 4. Insufficient noise resistance: The actual engine room environment is subject to strong noise interference, and the global matrix multiplication of traditional models (especially MLPs) amplifies the noise layer by layer, resulting in a significant decrease in diagnostic accuracy.

[0004] Therefore, there is an urgent need for a fault diagnosis method for marine diesel engines that combines high precision, low parameter quantity, strong noise resistance, and good interpretability.

[0005] The foregoing statements are for informational purposes only and are not intended to provide background information in connection with this application. Unless otherwise stated herein, the content described in this section is not prior art to the rest of this application. Summary of the Invention

[0006] This invention presents a method and system for diagnosing marine diesel engine faults based on an improved Kolmogorov-Arnold network. It introduces the KAN architecture into the field of marine diesel engine fault diagnosis for the first time, overturning the traditional MLP paradigm of "fixed node activation function." Instead, it places a learnable B-spline nonlinear activation function on the network connection edges, enabling the model to adaptively fit the highly distorted nonlinear manifold between thermodynamic parameters and fault types. Simultaneously, it innovatively introduces an L1 sparsity regularization term, forcing the network to automatically prune redundant connections during training, retaining only key fault propagation paths, thereby achieving high diagnostic accuracy and significantly reducing the number of parameters.

[0007] Furthermore, the model is endowed with excellent noise robustness by the local support properties of B-splines, and the feature importance is quantified by the explicit expression of the side function. This fundamentally solves the engineering pain point of the "black box" nature of traditional neural networks, which is uninterpretable. It provides a new solution for intelligent fault diagnosis of marine diesel engines that combines high precision, lightweight, strong robustness and interpretability.

[0008] According to a first aspect of the embodiments of this application, a method for diagnosing marine diesel engine faults based on an improved Kolmogorov-Arnold network is provided, comprising:

[0009] Collect thermodynamic parameter data during the operation of marine diesel engines, preprocess the thermodynamic parameter data, and obtain a standardized fault feature sample set; A Kolmogorov-Arnold network diagnostic model is constructed, which includes an input layer, at least one hidden layer, and an output layer. Learnable nonlinear activation functions are set on the connection edges between adjacent layers, and the nonlinear activation functions are parameterized using B-spline curves. The nodes of the input layer are used to receive feature vectors from the standardized fault feature sample set, the nodes of the hidden layer are used to sum and aggregate the function values ​​output by each connection edge of the previous layer, and the output of the output layer is the fault classification result. Construct a total loss function to train the Kolmogorov-Arnold network diagnostic model; The thermal parameter data of the marine diesel engine to be diagnosed is preprocessed and then input into the trained Kolmogorov-Arnold network diagnostic model to obtain the fault diagnosis results.

[0010] In some embodiments of this application, the total loss function includes a classification loss term and a sparsity regularization term; the sparsity regularization term is used to apply a norm penalty to the nonlinear activation function on the connection edge, so as to compress the activation function amplitude of the connection edge with a contribution lower than a preset threshold to zero, thereby realizing automatic pruning of the network structure.

[0011] In some embodiments of this application, the preprocessing employs Min-Max normalization to linearly map each feature parameter to the domain of the B-spline curve, ensuring that the input data falls within the effective support grid range of the B-spline basis function. The formula for calculating Min-Max normalization is: ; in, x j Let j be the original value of the j-th feature. and The training set j Minimum and maximum values ​​of the dimensional feature. These are the normalized eigenvalues.

[0012] In some embodiments of this application, the mathematical expression of the B-spline curve is as follows: ; ; in, b ( x ) is the baseline activation function. For B-spline basis functions, spline functions c k For learnable control point coefficients, w b and w s These are the weighting coefficients for the baseline function and the spline function, respectively.

[0013] In some embodiments of this application, the mathematical expression of the total loss function is as follows: ; in, The cross-entropy classification loss term is used. To connect the first l Layer i The node and the first l +1 floor j The L1 norm of the activation function of each node. λ is the regularization coefficient, used to control the degree of sparsity.

[0014] In some embodiments of this application, the network topology of the Kolmogorov-Arnold network diagnostic model is [ N , M , C ]; in, N The number of nodes in the input layer corresponds to the feature dimension of the thermal parameter data. MThis represents the number of hidden layer nodes. C This represents the number of output layer nodes and the corresponding number of fault categories.

[0015] In some embodiments of this application, the thermodynamic parameter data includes the following characteristic parameters: maximum burst pressure, compression pressure, single-cylinder exhaust temperature, turbocharger speed, scavenging box pressure, air cooler outlet temperature, fuel consumption rate, and exhaust manifold temperature.

[0016] In some embodiments of this application, after the Kolmogorov-Arnold network diagnostic model has been trained, a feature importance evaluation step is also included: Calculate the average L1 norm of all connection functions corresponding to each node in the input layer, and define the average L1 norm as the feature importance weight of the corresponding thermodynamic parameter; The thermal parameters are sorted according to their feature importance weights to identify the key sensitive parameters that lead to specific faults, and the feature importance ranking results are output in a visual manner.

[0017] According to a second aspect of the embodiments of this application, a marine diesel engine fault diagnosis system based on an improved Kolmogorov-Arnold network is provided, comprising: The data acquisition and processing module is used to collect thermodynamic parameter data during the operation of marine diesel engines, preprocess the thermodynamic parameter data, and obtain a standardized fault feature sample set. The model building module is used to construct the Kolmogorov-Arnold network diagnostic model. The Kolmogorov-Arnold network diagnostic model includes an input layer, at least one hidden layer, and an output layer. The connection edges between adjacent layers are set with learnable nonlinear activation functions, which are parameterized using B-spline curves. The nodes of the input layer are used to receive feature vectors from the standardized fault feature sample set, the nodes of the hidden layer are used to sum and aggregate the function values ​​output by each connection edge of the previous layer, and the output of the output layer is the fault classification result. The training module is used to construct the total loss function to train the Kolmogorov-Arnold network diagnostic model. The total loss function includes a classification loss term and a sparsity regularization term. The sparsity regularization term is used to apply norm penalty to the nonlinear activation functions on the connection edges to compress the magnitude of the activation functions of connection edges with a contribution lower than a preset threshold to zero, thereby realizing automatic pruning of the network structure. The fault diagnosis module is used to input the preprocessed thermodynamic parameter data of the marine diesel engine to be diagnosed into the trained Kolmogorov-Arnold network diagnostic model to obtain the fault diagnosis results.

[0018] According to a third aspect of the embodiments of this application, a marine diesel engine fault diagnosis device based on an improved Kolmogorov-Arnold network is provided, comprising: a storage unit for storing executable instructions; and a processing unit for connecting to the storage unit to execute the executable instructions to complete a marine diesel engine fault diagnosis method based on an improved Kolmogorov-Arnold network.

[0019] According to a fourth aspect of the embodiments of this application, a computer-readable storage medium is provided having a computer program stored thereon; the computer program is executed by a processor to implement a method for diagnosing marine diesel engine faults based on an improved Kolmogorov-Arnold network.

[0020] The present application discloses a method and system for diagnosing marine diesel engine faults based on an improved Kolmogorov-Arnold network. The method includes collecting thermodynamic parameter data during the operation of the marine diesel engine, preprocessing the thermodynamic parameter data, and obtaining a standardized fault feature sample set. A Kolmogorov-Arnold network diagnostic model is constructed, comprising an input layer, at least one hidden layer, and an output layer. Learnable nonlinear activation functions are set on the connection edges between adjacent layers, and these nonlinear activation functions are parameterized using B-spline curves. Nodes in the input layer receive feature vectors from a standardized fault feature sample set, while nodes in the hidden layer sum and aggregate the function values ​​output from each connection edge of the previous layer. The output layer outputs the fault classification result. A total loss function is constructed to train the Kolmogorov-Arnold network diagnostic model, including a classification loss term and a sparsity regularization term. The sparsity regularization term applies a norm penalty to the nonlinear activation functions on the connection edges to compress the activation function amplitudes of connection edges with contributions below a preset threshold to zero, achieving automatic pruning of the network structure. The preprocessed thermal parameter data of the marine diesel engine to be diagnosed is then input into the trained Kolmogorov-Arnold network diagnostic model to obtain the fault diagnosis result.

[0021] This application introduces the KAN architecture into the field of marine diesel engine fault diagnosis for the first time, subverting the traditional MLP paradigm of "fixed node activation function". It places a learnable B-spline nonlinear activation function on the network connection edge, enabling the model to adaptively fit the highly distorted nonlinear manifold between thermodynamic parameters and fault types. At the same time, it innovatively introduces an L1 sparsity regularization term, which forces the network to automatically prune redundant connections during training and retain only the key fault propagation path, thereby achieving high diagnostic accuracy and greatly reducing the number of parameters.

[0022] Furthermore, the model is endowed with excellent noise robustness by the local support properties of B-splines, and the feature importance is quantified by the explicit expression of the side function. This fundamentally solves the engineering pain point of the "black box" nature of traditional neural networks, which is uninterpretable. It provides a new solution for intelligent fault diagnosis of marine diesel engines that combines high precision, lightweight, strong robustness and interpretability.

[0023] Compared with the prior art, the present invention has the following beneficial effects: (1) Higher diagnostic accuracy.

[0024] This invention employs a KAN network architecture, leveraging the powerful nonlinear fitting capability of the B-spline deformable activation function to accurately fit the highly distorted nonlinear manifold between diesel engine thermodynamic parameters and fault types. Experiments show that on a test set containing seven operating conditions, the improved KAN model of this invention achieves a diagnostic accuracy of 98.9%, significantly outperforming traditional SVM (92.4%) and MLP (95.1%) algorithms, and also surpassing the standard KAN (97.8%).

[0025] (2) Fewer model parameters.

[0026] This invention achieves automatic network pruning through L1 sparsity regularization, retaining only critical fault propagation paths. Experiments show that the improved KAN has only about 1.2k parameters, which is about 50% less than MLP and about 30% less than the standard KAN. The lightweight model facilitates subsequent deployment on embedded terminals such as ship engine room monitoring and alarm systems (AMS).

[0027] (3) Excellent noise immunity This invention uses B-spline basis functions to parameterize the activation function. B-splines have local support properties, meaning that local noise in the input data only affects the local shape of the spline curve, unlike MLP which amplifies noise layer by layer through fully connected networks. Experiments show that in noisy environments with a signal-to-noise ratio as low as 20dB, the improved KAN model of this invention can still maintain a diagnostic accuracy of over 91%, while MLP only achieves about 82%.

[0028] (4) White-box interpretability.

[0029] This invention utilizes the explicit side function characteristics of KAN networks to achieve a quantitative assessment of feature importance by calculating the average L1 norm of the input layer connection functions. Feature weight analysis accurately reflects the core role of peak combustion pressure and exhaust temperature in fault diagnosis, and is highly consistent with the physical mechanism of diesel engine combustion. This interpretable diagnostic capability is not possessed by existing "black box" models such as CNN, MLP, and SVM.

[0030] (5) Strong generalization ability under small sample conditions.

[0031] This invention effectively avoids the overfitting problem of traditional deep learning models under small sample conditions by constraining model complexity through L1 sparsity regularization. In practical scenarios of marine diesel engine fault diagnosis with limited training samples, the method of this invention has significant application advantages.

[0032] (6) Fast convergence.

[0033] The deformable B-spline activation function of the KAN network in this invention can rapidly capture the nonlinear mapping relationship between diesel engine thermodynamic parameters and fault modes. Experiments show that within the first 15 training epochs, the model loss value decreases exponentially, and the accuracy rapidly climbs to over 90%, with training efficiency far exceeding that of traditional MLP networks. Attached Figure Description

[0034] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 The diagram illustrates the steps of a marine diesel engine fault diagnosis method based on an improved Kolmogorov-Arnold network according to an embodiment of this application. Figure 2 The diagram shows a comparison of the MLP and KAN structures according to embodiments of this application; Figure 3 The diagram shows a schematic representation of the B-spline function fitting principle according to an embodiment of this application. Figure 4 The diagram shows an architecture of a diesel engine fault diagnosis model based on an improved KAN according to an embodiment of this application; Figure 5 The figure shows the loss function and accuracy variation curves of the improved KAN model according to the embodiments of this application on the training set and the validation set; Figure 6 The diagram shows the confusion matrix of the improved KAN on the test set according to an embodiment of this application; Figure 7 The diagram shows the importance weight distribution of feature parameters according to embodiments of this application; Figure 8 The figure shows the accuracy variation trend under different signal-to-noise ratios according to embodiments of this application; Figure 9 The diagram shows a structural schematic of a marine diesel engine fault diagnosis system based on an improved Kolmogorov-Arnold network according to an embodiment of this application; Figure 10The diagram shows a structural schematic of a marine diesel engine fault diagnosis device based on an improved Kolmogorov-Arnold network according to an embodiment of this application. Detailed Implementation

[0035] Regarding this application, currently, there are several main technical approaches for diagnosing marine diesel engine faults: I. Fault diagnosis method based on mechanism model.

[0036] This method establishes the thermodynamic differential equations of a diesel engine to construct a mathematical model of its ideal operating process, and uses the residuals between the model output and actual measurements for fault diagnosis. However, the combustion process of a diesel engine involves complex nonlinear processes such as turbulent flow and chemical reaction kinetics, making it extremely difficult to establish a high-precision mathematical model. Furthermore, the model is overly sensitive to boundary conditions, and the mechanistic model will gradually become ineffective due to performance degradation caused by long-term operation of the diesel engine.

[0037] II. Fault Diagnosis Methods Based on Expert Systems This method summarizes the diagnostic experience of senior chief engineers into reasoning rules and constructs a knowledge base for fault reasoning. However, expert systems suffer from a "bottleneck" problem in knowledge acquisition, making it difficult to cover all unknown fault combinations and lacking self-learning capabilities, thus failing to adapt to newly emerging fault modes.

[0038] III. Data-Driven Fault Diagnosis Methods With the large-scale application of sensor technology, data-driven intelligent fault diagnosis methods have become a research hotspot in recent years. Existing technologies mainly include the following approaches: (I) Diagnostic methods based on traditional neural networks: A method and system for diagnosing thermal faults of marine diesel engines based on CBAM-CNN. This method uses convolutional neural networks (CNN) to extract fault features and introduces channel attention mechanism and spatial attention mechanism (CBAM) to optimize the feature map. However, this method has the following shortcomings: (1) The number of parameters in the CNN network is large. The four convolutional layers combined with four CBAM modules result in excessive model complexity and high requirements for hardware resources; (2) CNN is essentially still a "black box" model and cannot explain the specific contribution of each input thermal parameter to the diagnostic results. It is difficult for marine engineers to trust the diagnostic conclusions; (3) This method mainly relies on a large number of labeled samples for training. In the actual scenario where it is difficult to obtain marine diesel engine fault samples, the generalization ability is limited.

[0039] A method and apparatus for diagnosing thermal faults in marine diesel engines based on PCA-BO-SVM are proposed. This method uses principal component analysis (PCA) for feature dimensionality reduction, supports vector machines (SVM) for classification, and Bayesian optimization (BO) to optimize the SVM parameters. However, this method has the following shortcomings: (1) SVM is a shallow learning model, and its kernel function is difficult to select, resulting in limited processing capabilities for large-scale data; (2) PCA is a linear dimensionality reduction method, which cannot effectively capture the nonlinear correlation between the thermal parameters of the diesel engine; (3) The SVM model also lacks interpretability and cannot intuitively show the degree of influence of each parameter on fault judgment.

[0040] A method and system for monitoring and early warning of marine diesel engines based on data processing. This method identifies the combination of related parameters by calculating the correlation coefficient between parameter types, extracts key monitoring parameters based on the contribution rate formula, and constructs a fault prediction model. However, this method still has the following shortcomings: (1) Its contribution rate calculation is based on linear statistical methods, which cannot capture the nonlinear interaction effect between parameters; (2) This method focuses on the monitoring and early warning process at the system architecture level and does not involve a deep learning network architecture that can adaptively fit complex nonlinear mappings.

[0041] (ii) Diagnostic methods based on multilayer perceptron (MLP).

[0042] Traditional backpropagation (BP) neural networks and their improved models (such as SOM-BP) are the most widely used deep learning methods in diesel engine fault diagnosis. However, MLP and its variants suffer from the following structural defects: (1) Activation function rigidity: The neurons of MLP use fixed activation functions (such as Sigmoid, Tanh or ReLU), and the shape of the activation function remains unchanged regardless of the changes in training data. The network can only combine and stack these fixed curves by adjusting linear weights to approximate the real fault distribution, resulting in low fitting efficiency.

[0043] (2) Parameter redundancy and overfitting: In order to fit the highly distorted nonlinear manifold between the thermodynamic parameters of the diesel engine and the fault type, the MLP needs to significantly increase the network depth or width, resulting in an exponential increase in the weight parameters to be optimized. Under the "small sample" problem where it is difficult to obtain fault samples of marine diesel engines and the amount of typical fault data is limited, an excessive number of parameters can easily lead to overfitting.

[0044] (3) Poor interpretability: The complex weight connections within the MLP make the model decision-making process like a "black box", making it impossible to trace which abnormal fluctuation of thermal parameter led to the fault judgment, making it difficult to gain the trust of the engine management personnel.

[0045] In summary, there is an urgent need in the existing technology for a fault diagnosis method for marine diesel engines that combines high precision, low parameter quantity, strong noise resistance, and good interpretability.

[0046] Therefore, this invention provides a fault diagnosis method for marine diesel engines based on an improved Kolmogorov-Arnold network (KAN), aiming to solve the following technical problems: (1) Solve the problems of low fitting efficiency and redundant network parameters caused by fixed activation functions in traditional MLP and its variants; (2) To address the problem that existing diagnostic models are prone to overfitting and have poor generalization ability under small sample conditions; (3) Solve the problem of uninterpretable diagnostic results caused by the "black box" characteristics of existing deep neural network models; (4) Solve the problem that the diagnostic accuracy of the existing model drops significantly in the strong noise environment of the cabin.

[0047] KAN (Kolmogorov-Arnold Network): A novel neural network architecture based on the Kolmogorov-Arnold representation theorem.

[0048] This application presents a method and system for diagnosing marine diesel engine faults based on an improved Kolmogorov-Arnold network. It introduces the KAN architecture into the field of marine diesel engine fault diagnosis for the first time, overturning the traditional MLP paradigm of "fixed node activation function." Instead, it places a learnable B-spline nonlinear activation function on the network connection edges, enabling the model to adaptively fit the highly distorted nonlinear manifold between thermodynamic parameters and fault types. Simultaneously, it innovatively introduces an L1 sparsity regularization term, forcing the network to automatically prune redundant connections during training, retaining only key fault propagation paths, thereby achieving high diagnostic accuracy and significantly reducing the number of parameters.

[0049] Furthermore, the model is endowed with excellent noise robustness by the local support properties of B-splines, and the feature importance is quantified by the explicit expression of the side function. This fundamentally solves the engineering pain point of the "black box" nature of traditional neural networks, which is uninterpretable. It provides a new solution for intelligent fault diagnosis of marine diesel engines that combines high precision, lightweight, strong robustness and interpretability.

[0050] To make the technical solutions and advantages of the embodiments of this application clearer, the exemplary embodiments of this application will be described in further detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not an exhaustive list of all embodiments. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other.

[0051] Example 1 Figure 1 The diagram illustrates the steps of a marine diesel engine fault diagnosis method based on an improved Kolmogorov-Arnold network according to an embodiment of this application.

[0052] like Figure 1 As shown, a fault diagnosis method for marine diesel engines based on an improved Kolmogorov-Arnold network includes the following steps: S1: Collect thermodynamic parameter data during the operation of the ship's diesel engine, preprocess the thermodynamic parameter data, and obtain a standardized fault feature sample set.

[0053] Preferably, the thermodynamic parameter data includes the following eight parameters: maximum burst pressure. P max Compression pressure P comp Single-cylinder exhaust temperature T exh turbocharger speed n tc scavenging chamber pressure P scav air cooler outlet temperature T air Fuel consumption rate b e and exhaust manifold temperature T man .

[0054] In S1, the preprocessing uses Min-Max normalization to linearly map each feature parameter to the domain of the B-spline curve to ensure that the input data falls within the effective support grid range of the B-spline basis function; preferably, each thermodynamic parameter is linearly mapped to the [-1,1] interval.

[0055] The formula for calculating Min-Max normalization is: ; in, x j Let j be the original value of the j-th feature. and The training set j Minimum and maximum values ​​of the dimensional feature. These are the normalized eigenvalues.

[0056] S2: Construct a Kolmogorov-Arnold network diagnostic model. The Kolmogorov-Arnold network diagnostic model includes an input layer, at least one hidden layer, and an output layer. Learnable nonlinear activation functions are set on the connection edges between adjacent layers. The nonlinear activation functions are parameterized using B-spline curves. The nodes of the input layer are used to receive feature vectors from the standardized fault feature sample set. The nodes of the hidden layer are used to sum and aggregate the function values ​​output by each connection edge of the previous layer. The output of the output layer is the fault classification result.

[0057] B-Spline (Basis Spline): A parameterized activation function used in KAN networks. It consists of a series of basis functions and control point coefficients. During training, by adjusting the control point coefficients, the B-spline curve can adaptively deform into arbitrary shapes, thereby accurately fitting the nonlinear relationship between input features and faults.

[0058] Preferably, the mathematical expression of the B-spline curve is as follows: ; ; in, b ( x ) is the baseline activation function. For B-spline basis functions, spline functions c k For learnable control point coefficients, w b and w s These are the weighting coefficients for the baseline function and the spline function, respectively.

[0059] In the preferred implementation, the B-spline is a cubic uniform B-spline, the number of spline grids is set to 5, and the domain is [-1,1]; control point coefficients ck Initialize to 0, making the initial activation function a linear function. ( x )= x .

[0060] In a preferred embodiment, the network topology of the Kolmogorov-Arnold network diagnostic model is [ N , M , C ]; in, N The number of nodes in the input layer corresponds to the feature dimension of the thermal parameter data. M This represents the number of hidden layer nodes. C This represents the number of output layer nodes and the corresponding number of fault categories.

[0061] More preferably, the network topology is [8,5, C The input layer contains 8 nodes, each corresponding to one of the 8 thermal parameters mentioned above; the hidden layer contains 5 nodes.

[0062] S3: Construct the total loss function to train the Kolmogorov-Arnold network diagnostic model; Preferably, the total loss function includes a classification loss term and a sparsity regularization term; the sparsity regularization term is used to apply norm penalty to the nonlinear activation function on the connection edge, so as to compress the magnitude of the activation function of the connection edge with a contribution lower than a preset threshold to zero, thereby realizing automatic pruning of the network structure.

[0063] Preferably, the sparsification regularization term is a weighted sum of the L1 norms of the B-spline activation functions on each layer's connecting edges; by minimizing the total loss function, the amplitude of the B-spline activation function on connecting edges whose contribution to fault diagnosis is less than the threshold is compressed to zero, thereby achieving adaptive pruning of the network structure.

[0064] Preferably, the mathematical expression of the total loss function is as follows: ; in, The cross-entropy classification loss term is used. To connect the first l Layer i The node and the first l +1 floor j The L1 norm of the activation function of each node. λ is the regularization coefficient, used to control the degree of sparsity.

[0065] During model training, the AdamW optimizer was used to train the Kolmogorov-Arnold network diagnostic model, with an initial learning rate set to 1×10⁻⁶. 3 It also incorporates a cosine annealing strategy to decay the learning rate.

[0066] S4: The thermal parameter data of the marine diesel engine to be diagnosed is preprocessed and then input into the trained Kolmogorov-Arnold network diagnostic model to obtain the fault diagnosis results.

[0067] Further optimization of the implementation process includes a feature importance evaluation step after the Kolmogorov-Arnold network diagnostic model has been trained: Calculate the average L1 norm of all connection functions corresponding to each node in the input layer, and define the average L1 norm as the feature importance weight of the corresponding thermodynamic parameter; The thermal parameters are sorted according to their feature importance weights to identify the key sensitive parameters that lead to specific faults, and the feature importance ranking results are output in a visual manner.

[0068] Visualization methods include displaying the characteristic importance weights of each thermodynamic parameter in the form of bar charts, and / or displaying the nonlinear activation function shape of each connection edge in the form of line graphs.

[0069] In other preferred embodiments, S1 collects thermal parameter data of the marine diesel engine, including collecting thermal parameter data at the following typical load points: 50%, 75%, 85% and 100% load points; The final fault category includes one or more of the following operating conditions: normal condition, excessive fuel injection advance angle, insufficient fuel injection advance angle, poor injector atomization, air leakage in the intake valve, and air leakage in the exhaust valve.

[0070] The following detailed implementation process further illustrates the process of applying the improved Kolmogorov-Arnold (KAN) network to marine diesel engine fault diagnosis in this application embodiment.

[0071] Figure 2 The diagram shows a comparison of the MLP and KAN structures according to embodiments of this application.

[0072] like Figure 2 As shown, the KAN network transfers learnable nonlinear activation functions (parameterized as B-splines) to the network's "connection edges," while the neurons simply perform simple additive aggregation operations. KAN uses flexible splines to fit curves.

[0073] MLP attempts to piece together a curve using countless straight lines. Under the premise of achieving the same fitting accuracy, KAN usually requires far fewer parameters than MLP. Furthermore, since its activation function is an explicit spline curve, users can intuitively understand the inference logic of the model by visualizing the function shape of each layer's edge.

[0074] The improved Kolmogorov-Arnold (KAN) network in this application mainly consists of a cascaded input layer, a hidden layer, and an output layer.

[0075] Input layer: Responsible for receiving the preprocessed diesel engine thermodynamic parameter feature vector. The number of neurons in this layer is consistent with the dimension of the input features (e.g., maximum burst pressure). P max Compression pressure P comp(etc.), mainly serving the function of signal transmission, and not performing calculations.

[0076] Hidden layer: Located between the input layer and the output layer, it is the core of feature extraction in the network.

[0077] Output layer: Outputs the final diagnostic results. These are typically processed using a Softmax function to obtain the probability distributions for each fault category.

[0078] In this embodiment, a learnable nonlinear function is placed on the “edge” of the network.

[0079] KAN defines a network as a stack of linear layers. In a KAN layer, the computation between the input and output is no longer matrix multiplication, but the summation of nonlinear functions.

[0080] For the The first layer Each neuron receives its input from all neurons in the previous layer after passing through a learnable function. The sum after processing: ; here, It is the connection of the first Layer The node and the first Layer A one-dimensional function with nodes.

[0081] In order to make the above one-dimensional function With strong fitting capabilities, KAN uses B-splines for parameterization.

[0082] Specifically, each activation function is defined as a baseline function. spline functions The linear combination of is given by the following formula: ; ; in, It is the baseline activation function (usually SiLU) used to capture the overall trend; It is a B-spline basis function; These are learnable control point coefficients.

[0083] Figure 3 The diagram shows a schematic diagram of the B-spline function fitting principle according to an embodiment of this application.

[0084] like Figure 3 As shown, by dynamically adjusting the control coefficients during training... , It can adaptively deform into any desired shape. This mechanism fundamentally solves the problem of "rigidity" in MLP activation functions.

[0085] Next, considering that marine diesel engine monitoring data typically contains a certain degree of redundancy, this application improves the KAN model by introducing sparsity regularization to enhance its noise resistance and interpretability.

[0086] To remove redundant connections that do not contribute to the diagnostic results, this embodiment introduces an L1 norm penalty term into the loss function of KAN, forcing the network to perform sparse learning.

[0087] Improved total loss function Defined as: ; in, This is the cross-entropy loss function, used to ensure classification accuracy; The L1 norm of the activation function (i.e., the average magnitude of the function over its domain). This is the regularization coefficient, used to control the sparsity.

[0088] By minimizing this loss function, the corresponding connection function for input features that are not important for fault diagnosis is... The amplitude will be automatically compressed to zero (approaching a horizontal line). This means that the network has achieved automatic pruning, retaining only the critical fault propagation paths (i.e., retaining the most sensitive thermal parameters), which greatly improves the interpretability and running efficiency of the model.

[0089] Figure 4 The diagram shows an architecture diagram of a diesel engine fault diagnosis model based on an improved KAN according to an embodiment of this application.

[0090] like Figure 4 As shown, the specific architecture of the diagnostic model actually constructed in this application embodiment to meet the actual monitoring needs of a certain type of marine diesel engine is as follows: Input Layer: This layer has 8 nodes, corresponding to normalized thermal parameters. wait.

[0091] KAN Hidden Layers: (This likely refers to a specific implementation or method) The structural configuration.

[0092] The first layer contains 8 input nodes and 5 hidden nodes, used for feature fusion and nonlinear mapping. Inter-layer connections use cubic B-splines with a grid size of 5 to balance fitting accuracy and computational cost.

[0093] Output Layer and Classification Header: Includes... Each node represents a fault category, such as: normal, valve leakage, injector blockage, etc. Finally, the result is fed into a Softmax layer to output a probability vector.

[0094] The diesel engine fault diagnosis model based on the improved KAN in this application not only inherits the powerful nonlinear fitting ability of KAN, but also achieves automatic feature selection and network lightweighting, taking into account the multi-parameter coupling characteristics of diesel engines.

[0095] I. Diesel engine fault simulation and data acquisition.

[0096] Regarding the construction of experimental objects and simulation platforms: Given that marine diesel engines are large and complex thermodynamic devices, conducting destructive failure experiments in actual operation is extremely costly and poses safety hazards, making it difficult to obtain complete failure sample data throughout their entire lifespan. Therefore, this application uses AVL BOOST to establish a zero-dimensional / one-dimensional thermodynamic cycle simulation model of a certain type of marine low-speed two-stroke diesel engine, using this as the source for generating failure data.

[0097] In calibrating the simulation model parameters, this application uses the W 6S60ME-C diesel engine as the research object. Based on the testbed data and technical specifications of this engine model, the main geometric dimensions (cylinder bore, stroke, connecting rod length), valve timing, and turbocharger diagram of the simulation model are refined.

[0098] To ensure model fidelity, steady-state verification was performed under calibration conditions (100% load) and partial load conditions. The verification results show that the key thermodynamic parameters output by the simulation model (maximum burst pressure) are within acceptable limits. Exhaust temperature The relative errors between the model and the bench test values ​​were all controlled within 3%. This indicates that the model can accurately reproduce the thermodynamic working process of a real diesel engine and is effective as a data source for fault diagnosis.

[0099] Next, we will construct the failure mode settings and sample set.

[0100] Diesel engine thermal system failures are multi-source and coupled, mainly concentrated in the fuel injection system, intake and exhaust system, and cylinder combustion chamber components. Based on statistical analysis of actual engineering cases and prior expert knowledge, this application's embodiments define one normal state and six typical single fault modes as diagnostic targets.

[0101] Different types of faults are simulated by introducing degradation biases in physical parameters into the simulation model. The physical meaning of each fault mode and the parameter adjustment strategy are shown in Table 1 below.

[0102]

[0103] Table 1 Diesel Engine Fault Modes and Parameter Settings Regarding the selection of feature parameters, feature selection directly determines the upper limit of the diagnostic model. Based on the thermodynamic process mechanism analysis of diesel engines, this embodiment selects eight thermodynamic parameters that are most sensitive to the above-mentioned faults and are easily obtained through shipborne sensors to form the input feature vector. .

[0104] 1. Maximum burst pressure ( ): A core indicator reflecting combustion power, extremely sensitive to fuel timing and airtightness failures.

[0105] 2. Compression pressure ( ): Directly reflects the cylinder sealing performance (such as piston ring wear).

[0106] 3. Single-cylinder exhaust temperature ( ): A direct reflection of combustion quality and afterburning phenomenon.

[0107] 4. Turbocharger speed ( ): Reflects the energy balance state of intake and exhaust.

[0108] 5. Scavenging air box pressure ( ): Characterizes the filling capacity of the intake system.

[0109] 6. Air cooler outlet temperature ( ): Key parameters affecting intake air density and combustion start point.

[0110] 7. Fuel consumption rate ( ): Overall economic indicators, reflecting the decline in system efficiency.

[0111] 8. Exhaust manifold temperature ( ): Reflects the overall exhaust energy level of the machine.

[0112] Next, let's discuss sample generation and partitioning.

[0113] To cover the full operating characteristics of the diesel engine, simulations were conducted at four typical load points (50%, 75%, 85%, 100%) for the seven states (F0~F6) mentioned above.

[0114] Total sample size: 200 data sets were collected for each fault mode, totaling [number missing]. Group sample.

[0115] Dataset partitioning: Stratified sampling was used to divide the dataset into a training set (980 sets), a validation set (280 sets), and a test set (140 sets) in a ratio of 7:2:1 to ensure that the distribution of various faults is even in the subsets.

[0116] II. Data Preprocessing.

[0117] The raw thermal parameters are characterized by inconsistent dimensions and large numerical ranges (e.g.) MPa, (r / min). If this order-of-magnitude difference is not addressed, it will not only lead to slow gradient descent convergence but also severely affect the effective domain of the B-spline basis functions in the KAN network, causing local fitting failure. Therefore, data preprocessing is a crucial step in the embodiments of this application.

[0118] Preferably, Min-Max normalization is used for data preprocessing.

[0119] To accommodate the B-spline basis functions in KAN networks, they are typically defined in a standard interval (e.g., or Regarding the characteristics within the range, this application embodiment uses Min-Max normalization to linearly map all features to... The interval. For the first... 3D features The conversion formula is: ; in, and These are the maximum and minimum values ​​of the feature in the training set, respectively. Normalization ensures that all input variables fall within the effective range of the spline grid, thereby activating enough basis functions for nonlinear fitting.

[0120] Regarding noise injection and robustness enhancement: The actual ship engine room environment is complex, and sensor data is often affected by electromagnetic interference or vibration noise. To verify the robustness of the improved KAN model under non-ideal conditions, this application's embodiments constructed a noisy dataset during the testing phase. Specifically, Gaussian white noise with a signal-to-noise ratio (SNR) of 30 dB was superimposed on the original normalized data to simulate measurement uncertainties under real-world operating conditions.

[0121] III. Experimental Results Analysis and Model Validation.

[0122] Based on the diesel engine fault simulation dataset and experimental environment constructed above, the proposed "improved KAN diagnostic model with sparse regularization" was comprehensively tested. The experiment aimed to verify the superiority of the model over traditional methods from four dimensions: convergence characteristics, diagnostic accuracy, model interpretability, and noise robustness.

[0123] Figure 5 The figure shows the loss function and accuracy variation curves of the improved KAN model according to an embodiment of this application on the training set and the validation set.

[0124] For the analysis of model training convergence characteristics. For example... Figure 5 As shown, the curves illustrating the changes in loss and accuracy of the improved KAN model on the training and validation sets are presented.

[0125] The experimental parameters were set as follows: maximum number of iterations (Epochs) was set to 100, batch size was set to 32, the optimizer was AdamW, and the initial learning rate was set to... It also incorporates a cosine annealing strategy to decay the learning rate.

[0126] like Figure 5 As shown, it can be seen that (1) fast fitting ability: within the first 15 epochs, the model loss value decreases exponentially, and the accuracy quickly rises to over 90%. This indicates that the deformable B-spline activation function in the KAN network can capture the nonlinear mapping relationship between the diesel engine's thermodynamic parameters and the fault mode extremely quickly.

[0127] (2) No overfitting: As training progresses, the accuracy of the validation set increases in tandem with that of the training set, eventually stabilizing at around 98.5%, with the difference consistently kept within 0.5%. This demonstrates that the introduced L1 sparsity regularization term effectively constrains the model complexity and avoids the overfitting problem common in traditional neural networks.

[0128] IV. To objectively evaluate model performance, this application's embodiments select three representative benchmark algorithms for horizontal comparison: SVM (Support Vector Machine): Employs the RBF kernel function and represents a traditional machine learning method.

[0129] MLP (Multilayer Perceptron): Topology is This represents a classic deep learning method.

[0130] Standard KAN: The original KAN model without the addition of a sparsity regularization term. Improved KAN (method of this application embodiment): Added sparsity regularization and automatic pruning mechanism.

[0131] The performance metrics of each model on the test set (data not used in training) are shown in Table 2 below.

[0132]

[0133] Table 2 Comparison of performance indicators of different fault diagnosis models KAN vs. MLP: The improved KAN achieves an accuracy increase of approximately 3.8% compared to MLP. This confirms that the linear combination of fixed activation functions in MLP struggles to perfectly fit the complex fault boundaries of diesel engines, while KAN's adaptive spline function can achieve higher-dimensional feature segmentation with fewer neurons.

[0134] Compared to the standard KAN, the improved model proposed in this application reduces the number of parameters by approximately 30% while maintaining extremely high accuracy. This demonstrates that the sparsity regularization successfully eliminates redundant connections in the network, achieving a "lightweight" model.

[0135] V. In order to explore the model's ability to distinguish specific fault categories, Figure 6 The confusion matrix of the improved KAN on the test set is given.

[0136] like Figure 6 As shown, the classification confusion matrix of the improved model on the test set is displayed. The shades of gray in the figure intuitively reflect the density of the sample distribution. The dark areas on the diagonal represent the number of samples correctly predicted by the model, while the light-colored blocks off-diagonal represent misclassified samples.

[0137] Quantitative analysis of the confusion matrix leads to the following conclusions: 1. Excellent overall recognition performance.

[0138] Out of a total of 300 test samples, the model correctly classified 294 samples, achieving an overall accuracy of 98.00%. The vast majority of samples accurately fell on the main diagonal of the matrix, demonstrating the model's effectiveness in multi-class classification tasks.

[0139] 2. The recognition rate for specific categories reaches 100%.

[0140] Observation of the matrix shows that the identification effect of the "normal state" and "piston ring wear" faults is the best. The rows and columns containing these faults are all white (with a value of 0) except for the diagonal lines, indicating that the characteristic fingerprints of these two types of operating conditions are very unique, and the model can achieve 100% accurate identification and separation.

[0141] 3. Analysis of the mechanism of misjudgment (analysis of difficulties).

[0142] The few misclassifications shown in the figure are mainly concentrated among fault categories with similar physical characteristics, exhibiting the characteristic of "intra-class confusion": Valve leakage confusion: This is the main source of misclassification. Specifically, 3 "exhaust valve leakage" samples were misclassified as "intake valve leakage", and 1 "intake valve leakage" sample was misclassified in reverse.

[0143] From the perspective of diesel engine thermodynamic cycle mechanism, whether it is leakage in the intake valve or the exhaust valve, the direct consequence is a decrease in cylinder sealing, leading to a decrease in compression pressure. ) and peak burst pressure ( The compression and expansion lines of both are highly similar in trend, resulting in a slight overlap in the decision boundaries in the feature space. Two other "injection advance" samples were misclassified as "injection lag." This is mainly because both are timing faults in the fuel system, causing a shift in the combustion phase, resulting in slightly lower feature discrimination on some critical samples.

[0144] Despite the very minor confounding issues mentioned above, it is worth noting that the model successfully distinguished between "valve-related faults" (air leaks) and "fuel-related faults" (injection anomalies), without any misclassification across major categories. This demonstrates that the features extracted by the improved algorithm possess strong physical interpretability and robustness, meeting the needs of practical engineering diagnostics.

[0145] VI. For model interpretability and feature importance analysis, this is the core advantage of KAN that distinguishes it from "black box" MLP.

[0146] Because this embodiment introduces a sparsity penalty in the loss function, the model will automatically decouple irrelevant features during training. The weights are pushed to zero. By statistically analyzing the average amplitude (L1 norm) of the connection functions of each node in the input layer, the importance of each thermal parameter for fault diagnosis is inferred.

[0147] Figure 7 The diagram shows the importance weight distribution of feature parameters according to an embodiment of this application.

[0148] like Figure 7 As shown, the importance weight distribution of the eight feature parameters of the input layer is displayed after the model has been trained.

[0149] (maximum burst pressure) and The weight of exhaust temperature is significantly higher than that of other parameters. This is highly consistent with the principles of diesel engine combustion—most combustion system and valve malfunctions are directly reflected in drastic fluctuations in these two parameters. (scavenging temperature) and The weight of (lubricating oil pressure) is automatically compressed to near 0 by the model. This is consistent with the physical fact that in the thermal fault scenario set up in this experiment, the lubricating oil system is usually not directly affected.

[0150] The improved KAN not only provides diagnostic results but also "tells" which parameters it uses to make the judgment. This interpretability is of great reference value for the actual troubleshooting work of ship engineers.

[0151] VII. Model noise robustness test.

[0152] The electromagnetic interference in actual ship engine room environments is complex, and sensor data is often accompanied by noise. To verify robustness, Gaussian white noise of different intensities (signal-to-noise ratio SNR decreasing from 40dB to 10dB) was injected into the test set data to test the accuracy decay of each model.

[0153] Figure 8 The figure shows the accuracy variation trend under different signal-to-noise ratios according to embodiments of this application.

[0154] like Figure 8 As shown, experimental results indicate that when the signal-to-noise ratio drops to 20 dB (in a noisy environment), the accuracy of MLP falls sharply to 82%, while the improved KAN still maintains above 91%. MLP relies on global matrix multiplication, and noise is amplified layer by layer along the fully connected network; while the B-spline basis functions used by KAN have local support characteristics, and local noise has a limited impact on the overall shape of the function, thus exhibiting stronger anti-interference ability.

[0155] In summary, this application addresses the fact that marine diesel engines are the core of a ship's power system, and their operational status directly affects shipping safety and economic efficiency. To overcome the bottlenecks of traditional data-driven diagnostic methods (such as MLP and SVM), including rigid activation mechanisms, redundant network structures, and poor interpretability, this application, based on in-depth research into cutting-edge deep learning theories, proposes the aforementioned diesel engine fault diagnosis model based on sparse regularized improved KAN (Kolmogorov-Arnold Networks).

[0156] This application's embodiment is the first to introduce the KAN architecture into the field of marine diesel engine fault diagnosis, breaking through the architectural limitations of traditional multilayer perceptrons (MLPs) and backpropagation neural networks with their "linear weights + fixed node activation functions." This invention utilizes the characteristics of the Kolmogorov-Arnold network (KAN) to transfer learnable nonlinear activation functions (B-spline curves) to the network's connection "edges." This architecture is better suited to the highly nonlinear and strongly coupled manifold characteristics of diesel engine thermodynamic parameters, achieving higher fitting accuracy and convergence speed than traditional deep learning models even with small sample data.

[0157] The algorithm improvement of the embodiment scheme in this application is based on the automatic pruning mechanism of L1 sparsity regularization. In view of the computational redundancy problem existing in standard KAN, an L1 norm penalty term of the activation function is innovatively introduced into the loss function.

[0158] This forces the model to automatically compress the weights of connection functions that contribute little to fault detection to zero during training (i.e., automatic pruning). This achieves both network lightweighting (reducing the number of parameters by about 30%) and prevention of overfitting, significantly improving the model's generalization ability under partial load or variable operating conditions.

[0159] The functional breakthrough of the embodiment of this application realizes interpretable diagnosis from "black box" to "white box", solving the pain point that traditional neural networks (such as CNN and MLP) cannot explain the basis for fault judgment.

[0160] By utilizing the explicit spline function characteristics of the KAN network, and by statistically analyzing the average amplitude (L1 norm) of the input layer connection function, various thermodynamic parameters (such as...) are quantitatively calculated. , The importance weights of input features and fault categories can be intuitively displayed, providing marine engineering managers with diagnostic basis that conforms to physical mechanisms (such as clearly indicating which parameters' abnormalities led to the fault determination).

[0161] The robustness of the proposed solution is enhanced by a noise-resistant design based on the local support properties of B-splines. Utilizing the "local support" mathematical property of B-splines, local noise in the input data only affects the local shape of the spline curve, unlike in MLPs where noise spreads through fully connected layers to the entire network. This allows the model to maintain a diagnostic accuracy of over 90% even in noisy environments with a signal-to-noise ratio as low as 20dB.

[0162] The marine diesel engine fault diagnosis method based on the improved Kolmogorov-Arnold network provided in this application can be widely applied to condition monitoring and fault diagnosis of various marine diesel engines. It can be deployed as software in a marine engine room monitoring and alarm system (AMS) or edge computing gateway to perform online analysis and fault identification of real-time collected thermal parameter data. It can also serve as a core algorithm module of a ship's intelligent operation and maintenance platform, providing fault warnings, diagnostic reasoning, and maintenance decision support for marine engineering managers, and has broad prospects for industrial application.

[0163] Example 2 This application also provides a marine diesel engine fault diagnosis system based on an improved Kolmogorov-Arnold network. For details not disclosed in this embodiment of the marine diesel engine fault diagnosis system based on an improved Kolmogorov-Arnold network, please refer to the specific implementation of the marine diesel engine fault diagnosis method based on an improved Kolmogorov-Arnold network in other embodiments.

[0164] Figure 9 The diagram shows a structural schematic of a marine diesel engine fault diagnosis system based on an improved Kolmogorov-Arnold network according to an embodiment of this application.

[0165] like Figure 9 As shown, this application provides a marine diesel engine fault diagnosis system based on an improved Kolmogorov-Arnold network, comprising: The data acquisition and processing module 10 is used to collect thermodynamic parameter data during the operation of the marine diesel engine, preprocess the thermodynamic parameter data, and obtain a standardized fault feature sample set. The model building module 20 is used to build the Kolmogorov-Arnold network diagnostic model. The Kolmogorov-Arnold network diagnostic model includes an input layer, at least one hidden layer, and an output layer. The connection edges between adjacent layers are set with learnable nonlinear activation functions, which are parameterized using B-spline curves. The nodes of the input layer are used to receive feature vectors from the standardized fault feature sample set, the nodes of the hidden layer are used to sum and aggregate the function values ​​output by each connection edge of the previous layer, and the output of the output layer is the fault classification result. Training module 30 is used to construct the total loss function to train the Kolmogorov-Arnold network diagnostic model. The total loss function includes a classification loss term and a sparsity regularization term. The sparsity regularization term is used to apply norm penalty to the nonlinear activation function on the connection edge to compress the magnitude of the activation function of the connection edge with a contribution lower than a preset threshold to zero, thereby realizing automatic pruning of the network structure. The fault diagnosis module 40 is used to input the preprocessed thermodynamic parameter data of the marine diesel engine to be diagnosed into the trained Kolmogorov-Arnold network diagnostic model to obtain the fault diagnosis results.

[0166] This application presents a marine diesel engine fault diagnosis system based on an improved Kolmogorov-Arnold network. For the first time, the KAN architecture is introduced into the field of marine diesel engine fault diagnosis, overturning the traditional MLP paradigm of "fixed node activation function." A learnable B-spline nonlinear activation function is placed on the network connection edges, enabling the model to adaptively fit the highly distorted nonlinear manifold between thermodynamic parameters and fault types. Simultaneously, an innovative L1 sparsity regularization term is introduced, forcing the network to automatically prune redundant connections during training, retaining only key fault propagation paths, thereby achieving high diagnostic accuracy and significantly reducing the number of parameters.

[0167] Furthermore, the model is endowed with excellent noise robustness by the local support properties of B-splines, and the feature importance is quantified by the explicit expression of the side function. This fundamentally solves the engineering pain point of the "black box" nature of traditional neural networks, which is uninterpretable. It provides a new solution for intelligent fault diagnosis of marine diesel engines that combines high precision, lightweight, strong robustness and interpretability.

[0168] Example 3 This embodiment provides a marine diesel engine fault diagnosis device based on an improved Kolmogorov-Arnold network. For details not disclosed in this embodiment, please refer to the specific implementation details of the marine diesel engine fault diagnosis device or method based on an improved Kolmogorov-Arnold network in other embodiments.

[0169] Figure 10 The diagram shows a structural schematic of a marine diesel engine fault diagnosis device based on an improved Kolmogorov-Arnold network according to an embodiment of this application.

[0170] like Figure 10 As shown, the marine diesel engine fault diagnosis device 410 based on the improved Kolmogorov-Arnold network includes: a storage unit 402 for storing executable instructions; and a processing unit 401 for connecting to the storage unit 402 to execute the executable instructions to complete the marine diesel engine fault diagnosis method based on the improved Kolmogorov-Arnold network.

[0171] Those skilled in the art will understand that the illustration Figure 10 This is merely an example of a marine diesel engine fault diagnosis device 410 based on an improved Kolmogorov-Arnold network, and does not constitute a limitation on the marine diesel engine fault diagnosis device 410 based on an improved Kolmogorov-Arnold network. It may include more or fewer components than shown, or combine certain components, or different components. For example, the marine diesel engine fault diagnosis device 410 based on an improved Kolmogorov-Arnold network may also include input / output devices, network access devices, buses, etc.

[0172] The processing unit 401 (Central Processing Unit, CPU) can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor, or the processing unit 401 can be any conventional processor. The processing unit 401 is the control center of the marine diesel engine fault diagnosis device 410 based on the improved Kolmogorov-Arnold network, connecting all parts of the device 410 using various interfaces and lines.

[0173] Storage unit 402 can be used to store computer-readable instructions. Processing unit 401 implements various functions of the marine diesel engine fault diagnosis device 410 based on the improved Kolmogorov-Arnold network by running or executing the computer-readable instructions or modules stored in storage unit 402 and calling data stored in storage unit 402. Storage unit 402 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the marine diesel engine fault diagnosis device 410 based on the improved Kolmogorov-Arnold network, etc. In addition, storage unit 402 may include hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, read-only memory (ROM), random access memory (RAM), or other non-volatile / volatile storage devices.

[0174] If the integrated module of the marine diesel engine fault diagnosis device 410 based on the improved Kolmogorov-Arnold network is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by instructing related hardware through computer-readable instructions. These computer-readable instructions can be stored in a computer-readable storage medium, and when executed by a processor, they can implement the steps of the various method embodiments described above.

[0175] Example 4 This embodiment provides a computer-readable storage medium having a computer program stored thereon; the computer program is executed by a processor to implement a marine diesel engine fault diagnosis system or method based on an improved Kolmogorov-Arnold network as described in other embodiments.

[0176] In summary, the marine diesel engine fault diagnosis device and storage medium based on the improved Kolmogorov-Arnold network proposed in this application introduce the KAN architecture into the field of marine diesel engine fault diagnosis for the first time. This overturns the traditional MLP paradigm of "fixed node activation function" by placing a learnable B-spline nonlinear activation function on the network connection edges, enabling the model to adaptively fit the highly distorted nonlinear manifold between thermodynamic parameters and fault types. At the same time, it innovatively introduces an L1 sparsity regularization term, which forces the network to automatically prune redundant connections during training, retaining only the key fault propagation paths, thereby achieving high diagnostic accuracy and greatly reducing the number of parameters.

[0177] Furthermore, the model is endowed with excellent noise robustness by the local support properties of B-splines, and the feature importance is quantified by the explicit expression of the side function. This fundamentally solves the engineering pain point of the "black box" nature of traditional neural networks, which is uninterpretable. It provides a new solution for intelligent fault diagnosis of marine diesel engines that combines high precision, lightweight, strong robustness and interpretability.

[0178] Those skilled in the art will understand that the terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The singular forms “a,” “an,” and “the” as used in this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in embodiments of this application refers to and includes any or all possible combinations of one or more of the associated listed items.

[0179] It should be understood that although the terms first, second, third, etc., may be used in this invention to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first information may also be referred to as second information without departing from the scope of this invention, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0180] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0181] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for fault diagnosis of marine diesel engines based on an improved Kolmogorov-Arnold network, characterized in that, include: Collect thermodynamic parameter data during the operation of the marine diesel engine, preprocess the thermodynamic parameter data, and obtain a standardized fault feature sample set; A Kolmogorov-Arnold network diagnostic model is constructed, comprising an input layer, at least one hidden layer, and an output layer. Learnable nonlinear activation functions are set on the connection edges between adjacent layers, and these nonlinear activation functions are parameterized using B-spline curves. The nodes in the input layer are used to receive feature vectors from the standardized fault feature sample set. The nodes in the hidden layer are used to sum and aggregate the function values ​​output by each connection edge of the previous layer. The output of the output layer is the fault classification result. The total loss function is constructed to train the Kolmogorov-Arnold network diagnostic model; The thermal parameter data of the marine diesel engine to be diagnosed is preprocessed and then input into the trained Kolmogorov-Arnold network diagnostic model to obtain the fault diagnosis results.

2. The method according to claim 1, characterized in that, The total loss function includes a classification loss term and a sparsity regularization term; the sparsity regularization term is used to apply a norm penalty to the nonlinear activation function on the connection edge, so as to compress the magnitude of the activation function of the connection edge with a contribution lower than a preset threshold to zero, thereby realizing automatic pruning of the network structure.

3. The method according to claim 1, characterized in that, The preprocessing employs Min-Max normalization to linearly map each feature parameter to the domain of the B-spline curve, ensuring that the input data falls within the effective support grid range of the B-spline basis function. The formula for calculating the Min-Max normalization is: ； in, x j Let j be the original value of the j-th feature. and The training set j Minimum and maximum values ​​of the dimensional feature. These are the normalized eigenvalues.

4. The method according to claim 1, characterized in that, The mathematical expression of the B-spline curve is as follows: ； ； in, b ( x ) is the baseline activation function. For B-spline basis functions, spline functions c k For learnable control point coefficients, w b and w s These are the weighting coefficients for the baseline function and the spline function, respectively.

5. The method according to claim 2, characterized in that, The mathematical expression of the total loss function is as follows: ； in, The cross-entropy classification loss term is used. To connect the first l Layer i The node and the first l +1 floor j The L1 norm of the activation function of each node. λ is the regularization coefficient, used to control the degree of sparsity.

6. The method according to claim 1, characterized in that, The network topology of the Kolmogorov-Arnold network diagnostic model is [ N , M , C ]; in, N The number of input layer nodes corresponds to the feature dimension of the thermal parameter data; M This represents the number of hidden layer nodes. C This represents the number of output layer nodes and the corresponding number of fault categories.

7. The method according to claim 1, characterized in that, The thermodynamic parameters include the following characteristic parameters: maximum burst pressure, compression pressure, single-cylinder exhaust temperature, turbocharger speed, scavenging box pressure, air cooler outlet temperature, fuel consumption rate, and exhaust manifold temperature.

8. A fault diagnosis system for marine diesel engines based on an improved Kolmogorov-Arnold network, characterized in that, include: The data acquisition and processing module is used to collect thermodynamic parameter data during the operation of the marine diesel engine, preprocess the thermodynamic parameter data, and obtain a standardized fault feature sample set. The model building module is used to construct a Kolmogorov-Arnold network diagnostic model. The Kolmogorov-Arnold network diagnostic model includes an input layer, at least one hidden layer, and an output layer. Learnable nonlinear activation functions are set on the connection edges between adjacent layers. The nonlinear activation functions are parameterized using B-spline curves. The nodes of the input layer are used to receive feature vectors from the standardized fault feature sample set. The nodes of the hidden layer are used to sum and aggregate the function values ​​output by each connection edge of the previous layer. The output of the output layer is the fault classification result. The training module is used to construct the total loss function to train the Kolmogorov-Arnold network diagnostic model; The fault diagnosis module is used to input the preprocessed thermodynamic parameter data of the marine diesel engine to be diagnosed into the trained Kolmogorov-Arnold network diagnostic model to obtain the fault diagnosis results.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is executed by a processor to implement the method as described in any one of claims 1 to 7.

10. A fault diagnosis device for marine diesel engines based on an improved Kolmogorov-Arnold network, characterized in that, include: Storage unit, used to store executable instructions; as well as A processing unit is configured to be connected to a memory to execute executable instructions to perform the method as described in any one of claims 1 to 7.

Images