Ship engine fault diagnosis method and system

Abstract

The invention provides a ship engine fault diagnosis method and system, and the method comprises the steps: firstly building a ship engine simulation model, and carrying out the simulation experiment of a ship engine in a normal operation state and a plurality of fault types through the ship engine simulation model; the method comprises the following steps: respectively acquiring key parameters of a ship engine in a normal operation state and under various fault types, further constructing a ship engine data set, and then performing interpolation completion on missing data in the ship engine data set by adopting an expectation maximization algorithm to obtain a completed ship engine data set, the method comprises the following steps: complementing a ship engine data set, performing dimensionality reduction on all data in the complemented ship engine data set by adopting a t-SNE algorithm to obtain a dimensionality-reduced new ship engine data set, dividing the dimensionality-reduced new ship engine data set into a training set and a test set, finally constructing an ELM network model based on a machine learning method, and completing fault diagnosis of the ship engine based on the ELM network model. The problem that the diagnosis accuracy is low under the condition of data missing during existing ship engine fault diagnosis is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of ship fault diagnosis, and particularly relates to a ship engine fault diagnosis method and system. Background Art

[0002] Water transportation plays a crucial role in the transportation system. It is an economical and efficient transportation mode that connects various regions through rivers, canals, lakes, and oceans, providing convenient transportation routes for people and goods. With the increasing demand for goods, the development of ports and coastal cities, the initiative of green transportation, and the growth of the tourism and cruise markets, water transportation will continue to play an important role in the development of sustainable transportation and has greater potential. Therefore, water transportation has broad development prospects. Ships are indispensable transportation tools in water transportation, ensuring the safety of personnel and property during navigation. As the power source, ship engines are widely used in modern society and play an important role in economic development. Currently, more than about 90% of ships use diesel engines as the main power plant. However, due to the complex structure of ship diesel engines and the harsh working environment, once the engine fails, it will seriously affect navigation, cause certain economic losses, and even endanger the personal safety of ship personnel. Therefore, it is crucial to diagnose these faults in a timely manner so that technicians can repair them and keep the ship engine in a normal working state.

[0003] Traditional fault detection methods are based on monitoring alarms or manual inspections, and diagnostic decisions are made only when the corresponding parameters exceed their boundaries. However, with the increasing intelligence of ships, more and more diagnostic information is obtained and constantly changing under different operating conditions. Intelligent fault diagnosis has been widely applied in ship engines. Common data-driven methods include support vector machines (SVM), BP neural networks, radial basis function (RBF) neural networks, multi-layer perceptrons (MLP), etc. Extreme learning machine (ELM) is a single-hidden-layer neural network that randomly selects the weights and biases of the hidden layer and completes learning by calculating the weights of the output layer. It has a fast training speed and high timeliness, and these advantages meet the requirements of ship engine fault diagnosis. However, intelligent fault diagnosis is still in its infancy and faces many challenging problems. For example, most current research results are based on the assumption of data integrity. In actual industrial processes, data missing (MD) phenomena often occur. There may be various reasons for the occurrence of data missing: values exceeding the instrument range, sensor failures, different sensor sampling rates, etc. Using a data set containing missing values for fault diagnosis usually results in low diagnostic accuracy. Since sensor missing data is completely randomly lost, the expectation-maximization (EM) algorithm is an iterative algorithm for calculating the maximum likelihood estimate or posterior distribution in the case of incomplete data, which can effectively solve the problem of data missing in ship engine fault diagnosis.

[0004] At the same time, industrial process data is usually high-dimensional, and multivariate data with highly correlated variables is likely to impede the detection of fault states. Such high-dimensional characteristics are prone to trap anomaly detection algorithms into the curse of dimensionality, resulting in the failure of the fault diagnosis model or achieving suboptimal effects. Therefore, feature selection is a key technology for extracting useful process information before performing data-driven fault diagnosis. t-Stochastic Neighbor Embedding (t-SNE) is an embedding model that can map data in a high-dimensional space to a low-dimensional space while preserving the local characteristics of the dataset. Summary of the Invention

[0005] To solve the problems in the current ship engine fault diagnosis process, such as the low diagnostic accuracy due to data missing and the failure or suboptimal effect of the fault diagnosis model caused by high-dimensional characteristics, the present invention provides a ship engine fault diagnosis method. By establishing a ship engine simulation model and simulating its normal operating state and multiple fault states to construct a ship engine dataset, and respectively using the designed Expectation-Maximization (EM) algorithm and t-SNE algorithm for data completion and dimensionality reduction, and inputting the training set into the constructed Extreme Learning Machine (ELM) network model for training to obtain a trained ELM network model, and then detecting the operating state of the ship engine, effectively solving the problem of low diagnostic accuracy in the case of data missing in the existing ship engine fault diagnosis. The present invention also relates to a ship engine fault diagnosis system.

[0006] The technical solution of the present invention is as follows:

[0007] A ship engine fault diagnosis method, characterized by comprising the following steps:

[0008] Step of establishing a simulation model and constructing a dataset: Use engine performance simulation design software to establish a ship engine simulation model, and use the ship engine simulation model to conduct simulation experiments on the ship engine in the normal operating state and various fault types, respectively obtain the key parameters of the ship engine in the normal operating state and various fault types, and then construct a ship engine dataset;

[0009] Step of data completion and dimensionality reduction: Use the Expectation-Maximization (EM) algorithm to interpolate and complete the missing data in the ship engine dataset to obtain a completed ship engine dataset, then use the t-SNE algorithm to reduce the dimensionality of all data in the completed ship engine dataset to obtain a new ship engine dataset with reduced dimensionality, and divide the new ship engine dataset into a training set and a test set;

[0010] Steps for model establishment: Construct an ELM network model based on machine learning methods, input the training set into the ELM network model for training to obtain a trained ELM network model, and perform test analysis on the trained ELM network model through the test set to obtain the classification detection results of key parameters, and evaluate the current operating state of the marine diesel engine according to the classification detection results of the key parameters to complete the fault diagnosis of the marine engine.

[0011] Preferably, in the data completion and dimensionality reduction steps, using the Expectation-Maximization algorithm to impute the missing data in the marine engine dataset specifically includes the following steps:

[0012] S1: Establish a probability distribution model based on the marine engine dataset and set the initial model parameters of the probability distribution model;

[0013] S2: Expectation step, according to the initial model parameters and the non-missing data in the marine engine dataset, calculate the expected values of the missing data in the marine engine dataset;

[0014] S3: Maximization step, maximize the lower bound of its likelihood function according to the calculated expected values and then update the initial model parameters of the probability distribution model to obtain the updated model parameters;

[0015] S4: Repeat steps S2 and S3, calculate the difference between the initial model parameters and the updated model parameters, and then compare the difference with the preset threshold. If the difference is less than or equal to the preset threshold, stop the iteration, use the model parameters at the stop of the iteration as the optimal model parameters, estimate the missing data in the marine engine dataset using the optimal model parameters, and impute the missing data into the original marine engine dataset to obtain the completed marine engine dataset.

[0016] Preferably, in the data completion and dimensionality reduction steps, using the t-SNE algorithm to reduce the dimensionality of all data in the completed marine engine dataset specifically includes:

[0017] S1: Calculate the high-dimensional distance between any two data points in the completed marine engine dataset using the Euclidean distance, and convert the high-dimensional distance into a similarity probability using the Gaussian distribution function to obtain the similarity probability between any two data points in the high-dimensional space;

[0018] S2: Randomly initialize each data point in the completed marine engine dataset in the low-dimensional space using the normal distribution, then calculate the low-dimensional distance between any two initialized data points in the low-dimensional space using the Euclidean distance, and calculate the probability distribution between any two data points in the low-dimensional space according to the low-dimensional distance and using the t-distribution;

[0019] S3: Use the Kullback-Leibler divergence and construct a Kullback-Leibler divergence loss function based on the similarity probability and probability distribution.

[0020] S4: Use the gradient descent optimization algorithm to adjust the positions of data points in the low-dimensional space, so that the Kullback-Leibler divergence loss function is minimized, thereby completing the data dimensionality reduction process.

[0021] Preferably, the key parameters include effective power, mean effective pressure, maximum explosion pressure, supercharger outlet pressure, supercharger outlet temperature, intercooler outlet pressure, intercooler outlet temperature, exhaust pressure, and exhaust temperature.

[0022] Preferably, the probability distribution model includes a single normal distribution model and a mixture normal distribution model.

[0023] A ship engine fault diagnosis system, characterized in that it includes a parameter acquisition module, a data completion and dimensionality reduction module, and a model establishment module that are connected in sequence.

[0024] The parameter acquisition module uses engine performance simulation design software to establish a ship engine simulation model, and uses the ship engine simulation model to conduct simulation experiments on the ship engine in the normal operating state and various fault types, respectively obtain the key parameters of the ship engine in the normal operating state and various fault types, and then construct a ship engine data set.

[0025] The data completion and dimensionality reduction module includes a data completion module, a data dimensionality reduction module, and a data set division module that are connected in sequence. The data completion module uses the expectation-maximization algorithm to interpolate and complete the missing data in the ship engine data set to obtain a completed ship engine data set. Then, the data dimensionality reduction module uses the t-SNE algorithm to reduce the dimensionality of all data in the completed ship engine data set to obtain a new ship engine data set after dimensionality reduction. The data set division module divides the new ship engine data set into a training set and a test set.

[0026] The model establishment module constructs an ELM network model based on machine learning methods, inputs the training set into the ELM network model for training to obtain a trained ELM network model, and conducts test analysis on the trained ELM network model through the test set to obtain the classification detection results of the key parameters, and evaluates the current operating state of the ship diesel engine according to the classification detection results of the key parameters to complete the fault diagnosis of the ship engine.

[0027] Preferably, the data completion module in the data completion and dimensionality reduction module includes a probability distribution model parameter setting sub-module, an expectation sub-module, a maximization sub-module, and an interpolation processing sub-module that are connected in sequence.

[0028] The probability distribution model parameter setting sub-module establishes a probability distribution model based on the ship engine dataset and sets the initial model parameters of the probability distribution model;

[0029] The expectation sub-module calculates the expected values of the missing data in the ship engine dataset according to the initial model parameters and the non-missing data in the ship engine dataset;

[0030] The maximization sub-module maximizes the lower bound of its likelihood function according to the calculated expected values, and then updates the initial model parameters of the probability distribution model to obtain the updated model parameters;

[0031] The imputation processing sub-module controls the expectation sub-module and the maximization sub-module to work repeatedly, calculates the difference between the initial model parameters and the updated model parameters, and then compares the difference with a preset threshold. If the difference is less than or equal to the preset threshold, the iteration is stopped, and the model parameters at the end of the iteration are used as the optimal model parameters. The optimal model parameters are used to estimate the missing data in the ship engine dataset, and the missing data is imputed and added to the original ship engine dataset to obtain the completed ship engine dataset.

[0032] Preferably, the data dimensionality reduction module in the data completion and dimensionality reduction module includes a similarity probability calculation sub-module, a t-distribution low-dimensional calculation sub-module, a KL divergence loss function construction sub-module, and a gradient descent optimization sub-module connected in sequence.

[0033] The similarity probability calculation sub-module calculates the high-dimensional distance between any two data points in the completed ship engine dataset using the Euclidean distance, and converts the high-dimensional distance into a similarity probability using the Gaussian distribution function to obtain the similarity probability between any two data points in the high-dimensional space.

[0034] The t-distribution low-dimensional calculation sub-module randomly initializes each data point in the completed ship engine dataset in the low-dimensional space using the normal distribution, then calculates the low-dimensional distance between any two initialized data points in the low-dimensional space using the Euclidean distance, and calculates the probability distribution between any two data points in the low-dimensional space according to the low-dimensional distance using the t-distribution.

[0035] The KL divergence loss function construction sub-module constructs a KL divergence loss function using the KL divergence and according to the similarity probability and the probability distribution.

[0036] The gradient descent optimization sub-module adjusts the positions of the data points in the low-dimensional space using the gradient descent optimization algorithm to minimize the KL divergence loss function, thereby completing the data dimensionality reduction process.

[0037] Preferably, the key parameters include effective power, mean effective pressure, maximum explosion pressure, supercharger outlet pressure, supercharger outlet temperature, intercooler outlet pressure, intercooler outlet temperature, exhaust pressure, and exhaust temperature.

[0038] Preferably, the probability distribution model includes a single normal distribution model and a mixture normal distribution model.

[0039] The beneficial effects of the present invention are as follows:

[0040] A ship engine fault diagnosis method provided by the present invention first uses engine performance simulation design software to establish a ship engine simulation model, and uses the ship engine simulation model to conduct simulation experiments on the ship engine in the normal operating state and various fault types, respectively obtaining the key parameters of the ship engine in the normal operating state and various fault types, constructing a ship engine data set, and dividing the ship engine data set into a training set and a test set; then uses the Expectation-Maximization algorithm (EM) to complete the missing data in the ship engine data set, obtaining a new ship engine data set after completion, effectively solving the problem of data loss in ship engine fault diagnosis; then uses the t-SNE algorithm to reduce the dimension of all data in the new ship engine data set after completion, obtaining a new ship engine data set after dimension reduction. The t-SNE algorithm can map the data in the high-dimensional space to the low-dimensional space and retain the local characteristics of the data set, effectively improving the accuracy of ship engine fault diagnosis; finally, an ELM network model is constructed based on machine learning, and the training set is input into the ELM network model for training, obtaining a trained ELM network model, and the trained ELM network model is tested and analyzed through the test set to obtain the classification detection results of the key parameters, and the current operating state of the ship diesel engine is evaluated according to the classification detection results of the key parameters to complete the fault diagnosis of the ship engine. The present invention can be understood as a ship engine fault diagnosis method based on EM-TELM. By establishing a certain two-stroke ship diesel engine model, simulating its normal operating state and fault state, collecting simulation data to construct a benchmark data set for ship fault detection, aiming at the data loss phenomenon in the actual industrial process, the Expectation-Maximization method (EM) can effectively solve the problem of data loss in ship engine fault diagnosis; using the completed data to input into the ELM model to detect the operating state of the ship engine, solving the problem of low diagnostic accuracy in the case of data loss in the existing ship engine fault diagnosis, and at the same time using t-SNE for data dimension reduction, improving the accuracy of ship engine fault diagnosis. The ship engine fault diagnosis method based on EM-TELM proposed by the present invention reduces the influence of missing data and improves the accuracy of ship engine fault diagnosis.

[0041] The present invention also relates to a ship engine fault diagnosis system, which corresponds to the above-mentioned ship engine fault diagnosis method and can be understood as a system for implementing the above-mentioned ship engine fault diagnosis method. The system includes a parameter acquisition module, a data completion and dimensionality reduction module, and a model establishment module that are connected in sequence. Each module works in coordination with each other. By establishing a ship engine simulation model and simulating its normal operating state and multiple fault states, a ship engine data set is constructed. The maximum expectation algorithm and the t-SNE algorithm are respectively used for data completion and dimensionality reduction, and the training set is input into the constructed ELM network model for training to obtain a trained ELM network model, thereby detecting the operating state of the ship engine and solving the problem of low diagnostic accuracy in the case of data loss during the existing ship engine fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 is a flowchart of the ship engine fault diagnosis method of the present invention.

[0043] Figure 2 is a schematic diagram of the ship engine simulation model of the present invention.

[0044] Figure 3 is a preferred structural block diagram of the ship engine fault diagnosis system of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0045] The present invention will be described below with reference to the accompanying drawings.

[0046] The present invention relates to a ship engine fault diagnosis method, which can also be referred to as a ship engine fault diagnosis method based on EM-TELM. It can use AVL BOOST software to simulate the fault data of a certain two-stroke ship diesel engine, organize it into a ship engine fault data set, which includes several main fault types, can simulate and test nine main parameters, and randomly miss some data, and divide the obtained data set into a training set and a test set. Perform EM data interpolation on the training set to complete the missing data, and use t-SNE for data dimensionality reduction, establish an ELM network model, initialize the network parameters, train the model, and update the parameters of the network model; input the test set data into the trained model for fault diagnosis. The flowchart of this method is as Figure 1 shown and successively includes the following steps:

[0047] I. The steps of establishing a simulation model and constructing a data set, which can also be called the parameter acquisition step: Use engine performance simulation design software to establish a ship engine simulation model, and use the ship engine simulation model to conduct simulation experiments on the ship engine in the normal operating state and various fault types, respectively obtain the key parameters of the ship engine in the normal operating state and various fault types, and then construct a ship engine data set.

[0048] This step establishes a diagnostic object. First, an engine performance simulation design software (preferably AVL BOOST software, which is a software for simulating and analyzing the thermodynamics of internal combustion engines and can simulate and analyze the steady-state and transient performance of internal combustion engines) is used for simulation to establish a ship engine simulation model. Specifically, the engine involved in the simulation is a certain two-stroke ship diesel engine, and its main components include an exhaust gas turbocharger, an intercooler, cylinders, etc. As Figure 2 shown, a simulation model of a certain two-stroke ship engine is established in AVL BOOST software. Components such as the cylinder block and supercharger are selected from the model tree, then connected according to the system diagram, and finally data is input into each module. Among them, TC1 is the turbocharger, CO1 is the intercooler, C1-C6 are the cylinders, PL is the pipeline, SB is the system boundary, 1-21 are the pipelines, and MP is the measuring point.

[0049] After the ship engine simulation model is established, fault simulation experiments are carried out on several common faults, and relevant parameters are collected to obtain a ship engine fault data set. Since there are many types of ship engine (preferably diesel engine) faults, the common faults of ship diesel engines are studied, and typical fault types such as reduced turbocharger efficiency, reduced intercooler efficiency, single-cylinder fault, injection angle advance and injection angle delay are selected. Obtain the fault data set. According to the selected fault type and deviation degree, modify the component parameters to simulate several fault types of ship diesel engines and obtain a certain scale of fault data set. Specifically, according to a selected fault type, analyze the cause of the fault. For example, 1) the pipes and blades of the turbocharger are dirty and blocked with carbon deposits, and the sealing performance decreases, resulting in a reduction in turbocharger efficiency, and the fault can be simulated by modifying the turbocharger efficiency value; 2) seawater is the preferred cooling medium for the intercooler of the diesel engine, and the salts and other impurities contained in seawater accumulate on the water side of the intercooler, which will cause fouling on the water side, and the fouling on the water side deteriorates the heat transfer of the intercooler, and the fault can be simulated by modifying the intercooler efficiency value; 3) injection angle advance and injection angle delay will lead to too long ignition delay period, and the combustion mixture is discharged before being well cooled, resulting in a decrease in the power of the diesel engine, and the fault can be simulated by modifying the size of the injection advance angle, etc. By modifying the parameters of the main components, various different fault types of ship diesel engines can be simulated; in addition, set the deviation degree, and take 30% of the deviation from the normal value as the fault condition parameter and input it into the main components, so as to simulate and measure nine different key parameters including effective power, mean effective pressure, maximum explosion pressure, turbocharger outlet pressure, turbocharger outlet temperature, intercooler outlet pressure, intercooler outlet temperature, exhaust pressure and exhaust temperature. Take the nine different key parameters as a group of key parameters, and respectively obtain the key parameters of the ship engine in the normal operating state and the key parameters under various fault types. 400 groups of key parameters are obtained for each state, and then a ship engine data set is constructed.

[0050] II. Data Completion and Dimensionality Reduction Steps: The Expectation-Maximization algorithm is used to impute the missing data in the ship engine dataset to obtain a completed ship engine dataset. Then, the t-SNE algorithm is used to reduce the dimensionality of all the data in the completed ship engine dataset to obtain a new ship engine dataset with reduced dimensionality, and the new ship engine dataset is divided into a training set and a test set.

[0051] Specifically, using the Expectation-Maximization (EM) algorithm to impute the missing data in the ship engine dataset specifically includes the following steps:

[0052] S1: Establish a probability distribution model based on the ship engine dataset and set the initial model parameters of the probability distribution model as θ, where the initial model parameters θ can be set as the mean and variance;

[0053] S2: Expectation step, i.e., E step, make the best guess for the missing data according to the known information, that is, calculate the expected value of the missing data in the ship engine dataset according to the initial model parameters θ and the non-missing data in the ship engine dataset, and calculate according to the following formula:

[0054] Q i (z i )=p(z i |x i ;θ) (1)

[0055] In the above formula, z i represents the latent variable, that is, the missing data, x i represents the non-missing data, and θ represents the initial model parameters. That is to say, calculate the probability that each sample belongs to z i according to the initial value of the parameter θ, and obtain the Q function.

[0056] S3: Maximization step, i.e., M step, this step is to make a maximum likelihood estimate, maximize the lower bound of its likelihood function according to the calculated expected value, and then update the initial model parameters θ of the probability distribution model, that is, perform maximum likelihood estimation on the initial model parameters θ, optimize the initial model parameters θ, and obtain the updated model parameters θ i+1 , and perform according to the following formula:

[0057]

[0058] S4: Repeat S2 step and S3 step, and calculate the initial model parameters θ of the current iteration and the updated model parameters θ of the next iteration of the current iteration i+1The difference is then compared with a preset threshold. If the difference is less than or equal to the preset threshold, the iteration stops, that is, the Q function or the iteration value no longer changes significantly. The model parameters at the end of the iteration are taken as the optimal model parameters. The missing data in the ship engine dataset is estimated using the optimal model parameters, and the missing data is imputed and added to the original ship engine dataset to obtain a completed ship engine dataset.

[0059] It should be noted that for each missing data point, its possible potential values are calculated. That is, for each missing data point, the probability that it belongs to each possible state (normal state or fault state) under the current model parameters is calculated. Here, the state can be discrete (such as normal, fault 1, fault 2, etc.) or continuous parameter values.

[0060] Then, to improve the diagnostic accuracy, t-SNE is used to reduce the dimensionality of the data. The main dimensionality reduction process is to first calculate the high-dimensional distances between points, normalize the distances, convert them to probabilities, adjust the distance weights using the parameter perplexity Perp, fit the distribution function, then randomly initialize the low-dimensional distribution of points, calculate the low-dimensional distances, convert the distances to probabilities using the t-distribution, and finally calculate the information loss of dimensionality reduction. The loss function is the KL divergence, and the low-dimensional distribution is optimized to retain the high-dimensional features until a preset condition is reached. The t-SNE algorithm is used to reduce the dimensionality of all the data in the completed ship engine dataset, which specifically includes the following steps:

[0061] S1: For the input sample X, set the parameters Perp, the number of iterations, the learning rate, and the momentum of the objective function, and calculate the conditional probability p under the given Perp j|i and p ij , that is, convert the Euclidean distance to the probability p ij denotes the similarity between x i and x j in the completed ship engine dataset, or in other words, calculate the high-dimensional distance (i.e., the distance in the high-dimensional space) between any two data points in the completed ship engine dataset using the Euclidean distance, and convert the high-dimensional distance to the similarity probability using the Gaussian distribution function to obtain the similarity probability between any two data points in the high-dimensional space, as follows:

[0062]

[0063] where σ i is the variance, and k represents an index used to sum over all other data points in the new ship engine dataset when calculating the denominator. k iterates over all data points different from i to calculate the weighted sum of each data point in the denominator.

[0064] S2: Randomly initialize each data point in the completed ship engine dataset using a normal distribution in a low-dimensional space (two-dimensional or three-dimensional). For example, randomly generate a two-dimensional coordinate (y i, y j ) for each data point; then calculate the low-dimensional distance between each pair of data points in the low-dimensional space using the Euclidean distance, and calculate the probability distribution between any two data points (i.e., each pair of data points) in the low-dimensional space using the t-distribution based on the low-dimensional distance between each pair of data points.

[0065] S3: Use the KL divergence to measure the difference in similarity between the high-dimensional space and the low-dimensional space, that is, use the KL divergence and based on the similarity probability P ij between any two data points in the high-dimensional space, and the probability distribution Q ij between any two data points in the low-dimensional space to construct the KL divergence loss function C, as shown in the following formula:

[0066]

[0067] S4: Use the gradient descent optimization algorithm to optimize (i.e., minimize) the KL divergence loss function C, that is, calculate the gradient of the KL divergence loss function with respect to the position of each data point in the low-dimensional space, and update the position of the low-dimensional data points according to the calculated gradient (which can also be understood as adjusting the position of the data points in the low-dimensional space) to minimize the KL divergence loss function, thereby completing the data dimensionality reduction process, and then obtaining the new ship engine dataset after dimensionality reduction, and dividing the new ship engine dataset into a training set and a test set.

[0068] III. Model establishment steps: Construct an ELM network model based on machine learning, input the training set into the ELM network model for training to obtain a trained ELM network model, and perform test analysis on the trained ELM network model through the test set to obtain the classification detection results of key parameters, and evaluate the current operating state of the ship diesel engine according to the classification detection results of key parameters to complete the fault diagnosis of the ship engine.

[0069] Specifically, first construct an ELM network model based on machine learning methods, and input the key parameter data in the training set into the ELM network model for model training to obtain a trained ELM network model. Among them, the extreme learning machine (ELM) is a feedforward single-hidden layer network, the weights and biases of its hidden layer nodes are randomly generated, and the output weights are calculated using the least squares method. That is, train the network, minimize the cost function to reach the maximum number of epochs, and complete the training.

[0070] Training set {(x i ,y i ) N i＝1}, yi For the corresponding label vector using One - Hot encoding, N is the number of training samples, and the output of its hidden layer is:

[0071]

[0072] Among them, g is the activation function, which is a non - linear piece - wise continuous function satisfying the ELM universal approximation ability theorem. The sigmoid function is used. w is the input weight of the hidden layer, b is the bias of the hidden layer, and L is the number of hidden layer nodes. The output of the entire network is:

[0073] f L (x) = H(x)β

[0074] Among them, β is the output weight between the hidden layer and the output layer, and its objective function is:

[0075]

[0076] Since H is invertible, according to the least - squares method, its optimal solution is

[0077]

[0078] Among them, H + is the generalized inverse matrix of H. To prevent overfitting problems, a regularization coefficient can be added to the objective function, and the objective function becomes:

[0079]

[0080] Its optimal solution is:

[0081]

[0082] Then, set the number of hidden - layer neurons and the regularization parameter, and perform network training. That is, the input weight w and the hidden - layer bias b of the hidden - layer parameters are randomly initialized. Then, the sigmoid function is used as the activation function to map the input data to a new feature space, calculate the hidden - layer output H, and calculate the output - layer weight β using the solution formula. After training, a trained ELM network model is obtained, and the trained ELM network model is tested and analyzed through the test set to obtain the classification detection results of the key parameters. That is, the test set or test samples are input into the trained ELM network model, the diagnostic results are output, the classification accuracy is calculated, and the current operating state of the ship diesel engine is evaluated according to the classification detection results of the key parameters to complete the fault diagnosis of the ship engine.

[0083] The present invention also relates to a ship engine fault diagnosis system, which corresponds to the above-mentioned ship engine fault diagnosis method and can be understood as a system for implementing the above method. The system includes a parameter acquisition module, a data completion and dimensionality reduction module, and a model establishment module that are connected in sequence, as Figure 3 shown. Specifically,

[0084] The parameter acquisition module uses engine performance simulation design software to establish a ship engine simulation model, and uses the ship engine simulation model to conduct simulation experiments on the ship engine in the normal operating state and various fault types, respectively obtaining the key parameters of the ship engine in the normal operating state and various fault types, and then constructing a ship engine data set;

[0085] The data completion and dimensionality reduction module includes a data completion module, a data dimensionality reduction module, and a data set division module that are connected in sequence. The data completion module uses the expectation-maximization algorithm to interpolate and complete the missing data in the ship engine data set, obtaining a completed ship engine data set. Then, the data dimensionality reduction module uses the t-SNE algorithm to reduce the dimensionality of all the data in the completed ship engine data set, obtaining a new ship engine data set after dimensionality reduction, and the data set division module divides the new ship engine data set into a training set and a test set;

[0086] The model establishment module constructs an ELM network model based on machine learning methods, inputs the training set into the ELM network model for training, obtaining a trained ELM network model, and conducts test analysis on the trained ELM network model through the test set, obtaining the classification detection results of the key parameters, and evaluating the current operating state of the ship diesel engine according to the classification detection results of the key parameters to complete the fault diagnosis of the ship engine.

[0087] Preferably, as Figure 3 shown, the data completion module in the data completion and dimensionality reduction module includes a probability distribution model parameter setting sub-module, an expectation sub-module, a maximization sub-module, and an interpolation processing sub-module that are connected in sequence,

[0088] The probability distribution model parameter setting sub-module establishes a probability distribution model based on the ship engine data set and sets the initial model parameters of the probability distribution model;

[0089] The expectation sub-module calculates the expected values of the missing data in the ship engine data set according to the initial model parameters and the non-missing data in the ship engine data set;

[0090] The maximization sub-module maximizes the lower bound of its likelihood function according to the calculated expected values and then updates the initial model parameters of the probability distribution model, obtaining updated model parameters;

[0091] The control expectation sub-module and the maximization sub-module work repeatedly, calculate the difference between the initial model parameters and the updated model parameters, and then compare the difference with a preset threshold. If the difference is less than or equal to the preset threshold, stop the iteration. Take the model parameters at the stop of iteration as the optimal model parameters, use the optimal model parameters to estimate the missing data in the ship engine dataset, and impute and add the missing data to the original ship engine dataset to obtain the completed ship engine dataset.

[0092] Preferably, as Figure 3 shown, the data dimensionality reduction module in the data completion and dimensionality reduction module includes a similarity probability calculation sub-module, a t-distribution low-dimensional calculation sub-module, a KL divergence loss function construction sub-module, and a gradient descent optimization sub-module connected in sequence. Therefore, the data dimensionality reduction module uses the t-SNE algorithm to perform dimensionality reduction on all data in the completed ship engine dataset, specifically including:

[0093] S1: The similarity probability calculation sub-module calculates the high-dimensional distance between any two data points in the completed ship engine dataset using the Euclidean distance, and converts the high-dimensional distance into a similarity probability using the Gaussian distribution function to obtain the similarity probability between any two data points in the high-dimensional space.

[0094] S2: The t-distribution low-dimensional calculation sub-module randomly initializes each data point in the completed ship engine dataset in the low-dimensional space using the normal distribution, then calculates the low-dimensional distance between any two initialized data points in the low-dimensional space using the Euclidean distance, and calculates the probability distribution between any two data points in the low-dimensional space according to the low-dimensional distance using the t-distribution.

[0095] S3: The KL divergence loss function construction sub-module constructs a KL divergence loss function using the KL divergence based on the similarity probability and the probability distribution.

[0096] S4: The gradient descent optimization sub-module uses the gradient descent optimization algorithm to adjust the positions of the data points in the low-dimensional space to minimize the KL divergence loss function, thereby completing the data dimensionality reduction process.

[0097] Preferably, the key parameters include effective power, mean effective pressure, maximum explosion pressure, supercharger outlet pressure, supercharger outlet temperature, intercooler outlet pressure, intercooler outlet temperature, exhaust pressure, and exhaust temperature.

[0098] Preferably, the probability distribution models include a single normal distribution model and a mixture normal distribution model.

[0099] The present invention provides an objective and scientific method and system for diagnosing faults of a ship engine. A simulation model of the ship engine is established and its normal operating state and multiple fault states are simulated to construct a ship engine data set. The expectation-maximization algorithm and the t-SNE algorithm are respectively used for data completion and dimensionality reduction, and the training set is input into the constructed ELM network model for training to obtain a trained ELM network model, thereby detecting the operating state of the ship engine, and solving the problem of low diagnostic accuracy in the case of data loss in the existing diagnosis of ship engine faults.

[0100] It should be noted that the above specific embodiments can enable those skilled in the art to understand the present invention more comprehensively, but do not limit the present invention in any way. Therefore, although this specification has described the present invention in detail with reference to the drawings and embodiments, those skilled in the art should understand that the present invention can still be modified or equivalently replaced. In short, all technical solutions and their improvements that do not depart from the spirit and scope of the present invention should be covered by the protection scope of the patent of the present invention.

Claims

1. A method for diagnosing a ship engine fault, characterized in that: The following steps are involved: Simulation model establishment and data set construction steps: Use engine performance simulation design software to establish a ship engine simulation model, and use the ship engine simulation model to simulate the ship engine in normal operation and various fault types, and obtain the key parameters of the ship engine in normal operation and various fault types, and then construct the ship engine data set; Data completion and dimension reduction steps: The maximum expectation algorithm is used to interpolate and complete the missing data in the ship engine dataset to obtain the completed ship engine dataset, and then the t-SNE algorithm is used to reduce the dimension of all data in the completed ship engine dataset to obtain a new ship engine dataset after dimension reduction, and the new ship engine dataset is divided into a training set and a test set; Model building steps: construct an ELM network model based on the machine learning method, and input the training set into the ELM network model for training to obtain the trained ELM network model, and test and analyze the trained ELM network model through the test set to obtain the classification detection results of key parameters, and evaluate the current operating status of the ship diesel engine according to the classification detection results of key parameters to complete the fault diagnosis of the ship engine.

2. The ship engine fault diagnosis method according to claim 1, characterized in that: In the data completion and dimensionality reduction step, the maximum expectation algorithm is used to interpolate and complete the missing data in the ship engine data set, which specifically includes the following steps: S1: Establish a probability distribution model based on the ship engine data set and set the initial model parameters of the probability distribution model; S2: Expectation step, based on the initial model parameters and the non-missing data in the ship engine data set, the expected value of the missing data in the ship engine data set is calculated; S3: maximization step, maximizing the lower bound of the likelihood function according to the calculated expected value and then updating the initial model parameters of the probability distribution model to obtain the updated model parameters; S4: Repeat steps S2 and S3, and calculate the difference between the initial model parameters and the updated model parameters, and then compare the difference with the preset threshold. If the difference is less than or equal to the preset threshold, stop the iteration, and use the model parameters when the iteration is stopped as the optimal model parameters. Use the optimal model parameters to estimate the missing data in the ship engine dataset, and interpolate the missing data into the original ship engine dataset to obtain the completed ship engine dataset.

3. The ship engine fault diagnosis method according to claim 1, characterized in that: In the data completion and dimensionality reduction step, the t-SNE algorithm is used to reduce the dimensionality of all data in the completed ship engine data set, specifically including: S1: The high-dimensional distance between any two data points in the completed ship engine data set is calculated using the Euclidean distance, and the Gaussian distribution function is used to convert the high-dimensional distance into a similarity probability to obtain the similarity probability between any two data points in the high-dimensional space; S2: Each data point in the completed ship engine data set is randomly initialized in the low-dimensional space using normal distribution, and then the low-dimensional distance between any two initialized data points is calculated in the low-dimensional space using Euclidean distance. The probability distribution between any two data points in the low-dimensional space is calculated based on the low-dimensional distance and t distribution; S3: Use KL divergence and construct the KL divergence loss function based on similarity probability and probability distribution; S4: Use the gradient descent optimization algorithm to adjust the position of the data points in the low-dimensional space so that the KL divergence loss function is minimized, thereby completing the data dimensionality reduction process.

4. The ship engine fault diagnosis method according to claim 1, characterized in that: The key parameters include effective power, mean effective pressure, maximum burst pressure, supercharger outlet pressure, supercharger outlet temperature, intercooler outlet pressure, intercooler outlet temperature, exhaust pressure and exhaust temperature.

5. The ship engine fault diagnosis method according to claim 2, characterized in that: The probability distribution model includes a single normal distribution model and a mixed normal distribution model.

6. A ship engine fault diagnosis system, characterized in that: It includes a parameter acquisition module, a data completion and dimension reduction module, and a model building module connected in sequence. The parameter acquisition module uses the engine performance simulation design software to establish a ship engine simulation model, and uses the ship engine simulation model to simulate the ship engine in normal operation and under various fault types, respectively obtaining the key parameters of the ship engine in normal operation and under various fault types, and then constructing a ship engine data set; The data completion and dimension reduction module comprises a data completion module, a data dimension reduction module and a data set division module which are connected in sequence. The data completion module uses a maximum expectation algorithm to interpolate and complete the missing data in the ship engine data set to obtain a completed ship engine data set. The data dimension reduction module then uses a t-SNE algorithm to reduce the dimension of all data in the completed ship engine data set to obtain a new ship engine data set after dimension reduction. The data set division module divides the new ship engine data set into a training set and a test set. The model building module constructs an ELM network model based on a machine learning method, and inputs a training set into the ELM network model for training to obtain a trained ELM network model, and tests and analyzes the trained ELM network model through a test set to obtain classification detection results of key parameters, and evaluates the current operating status of the ship diesel engine according to the classification detection results of the key parameters to complete the fault diagnosis of the ship engine.

7. The ship engine fault diagnosis system according to claim 6, characterized in that: The data completion module in the data completion and dimension reduction module includes a probability distribution model parameter setting submodule, an expectation submodule, a maximization submodule and an interpolation processing submodule connected in sequence. The probability distribution model parameter setting submodule establishes a probability distribution model based on the ship engine data set and sets initial model parameters of the probability distribution model; The expectation submodule calculates the expected value of the missing data in the ship engine data set according to the initial model parameters and the non-missing data in the ship engine data set; The maximization submodule maximizes the lower bound of its likelihood function according to the calculated expected value and updates the initial model parameters of the probability distribution model to obtain updated model parameters; The interpolation processing submodule controls the expectation submodule and the maximization submodule to work repeatedly, and calculates the difference between the initial model parameters and the updated model parameters, and then compares the difference with a preset threshold. If the difference is less than or equal to the preset threshold, the iteration is stopped, and the model parameters when the iteration is stopped are used as the optimal model parameters. The optimal model parameters are used to estimate the missing data in the ship engine data set, and the missing data are interpolated and added to the original ship engine data set to obtain the completed ship engine data set.

8. The ship engine fault diagnosis system according to claim 6, characterized in that: The data dimension reduction module in the data completion and dimension reduction module includes a similarity probability calculation submodule, a t-distribution low-dimensional calculation submodule, a KL divergence loss function construction submodule and a gradient descent optimization submodule connected in sequence. The similarity probability calculation submodule uses Euclidean distance to calculate the high-dimensional distance between any two data points in the completed ship engine data set, and uses Gaussian distribution function to convert the high-dimensional distance into similarity probability to obtain the similarity probability between any two data points in the high-dimensional space; The t-distribution low-dimensional calculation submodule uses normal distribution to randomly initialize each data point in the completed ship engine data set in the low-dimensional space, and then uses Euclidean distance to calculate the low-dimensional distance between any two initialized data points in the low-dimensional space, and calculates the probability distribution between any two data points in the low-dimensional space based on the low-dimensional distance and using t-distribution; The KL divergence loss function construction submodule adopts KL divergence and constructs a KL divergence loss function according to similarity probability and probability distribution; The gradient descent optimization submodule uses a gradient descent optimization algorithm to adjust the position of data points in the low-dimensional space so that the KL divergence loss function is minimized, thereby completing the dimensionality reduction process of the data.

9. The ship engine fault diagnosis system according to claim 8, characterized in that: The key parameters include effective power, mean effective pressure, maximum burst pressure, supercharger outlet pressure, supercharger outlet temperature, intercooler outlet pressure, intercooler outlet temperature, exhaust pressure and exhaust temperature.

10. The ship engine fault diagnosis system according to claim 6, characterized in that: The probability distribution model includes a single normal distribution model and a mixed normal distribution model.

Images