Marine diesel engine degradation performance prediction method and device based on Bayesian reasoning and medium

Abstract

The invention discloses a marine diesel engine degradation performance prediction method and device based on Bayesian reasoning, and the method comprises the steps: collecting state information data, and making a training set and a test set; constructing a Bayesian neural network model and a Bayesian logistic regression model; jointly optimizing weight parameters of the Bayesian neural network model and parameters of the Bayesian logistic regression model; training a Bayesian neural network model and a Bayesian logistic regression model by adopting the training set; performing performance test on the trained model by adopting the test set; and obtaining state information data in the actual operation process of the marine diesel engine, and inputting the state information data into the trained model to respectively obtain a performance prediction result and a performance degradation curve of the marine diesel engine. The method has the advantages that the prediction precision is high, and the confidence interval can be generated to assist the adjustment of the operation and maintenance scheme of the diesel engine equipment.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, in particular to prediction of degradation performance of ship diesel engines, and in particular to a method, device and medium for predicting degradation performance of ship diesel engines based on Bayesian reasoning. Background Art

[0002] Remaining life prediction is a key part of mechanical equipment prediction and health management. Remaining life is the remaining time before the mechanical equipment fails due to the changes in its performance and status due to various factors during use. The main purpose is to analyze the current and historical operating conditions of equipment and parts, and predict the remaining life of mechanical equipment in advance, so that appropriate maintenance measures can be taken before the equipment fails to avoid equipment damage or accidents. At the same time, it can also reduce maintenance costs and improve the service life and efficiency of equipment.

[0003] Since marine diesel engines are a type of complex system with a long life span, if the end time of the equipment is taken as the scrapping time, the predicted remaining life range is too large and cannot accurately reflect the performance degradation of the current equipment. Therefore, the remaining life prediction of marine diesel engines is usually defined as the time from normal operation of the equipment to equipment failure, and the probability of failure is used as an evaluation index to predict the degradation performance of marine diesel engines. The performance prediction of marine diesel engines is conducive to the engineers to formulate maintenance plans and prevent failures.

[0004] Among the various thermal parameters of a marine diesel engine, the dynamometer diagram contains the most information. The dynamometer diagram can reflect the changes in parameters such as pressure, temperature, and power during the operation of the diesel engine, and can be used to evaluate the performance and working status of the diesel engine. Through the dynamometer diagram, you can understand the pressure and power output of the diesel engine, as well as other key information about the performance of the diesel engine. With the help of the dynamometer diagram information, the average indicated pressure of the diesel engine during operation can be calculated, thereby measuring the combustion efficiency and working quality of the diesel engine. However, the measurement of the dynamometer diagram is difficult, and the reliability of the pressure sensor is low due to the harsh installation environment.

[0005] In recent years, research has found that reliable diagnostic results can be obtained by monitoring the operating status of diesel engines through the fluctuation of crankshaft speed. The speed fluctuation caused by the crankshaft torsion fluctuation during diesel engine operation can reflect the ignition status of each cylinder of the diesel engine. Therefore, by studying the instantaneous speed fluctuation, the average indicated pressure in the cylinder can be estimated and the relevant diesel engine health status diagnosis can be performed accordingly.

[0006] An often overlooked problem in the prediction of ship equipment performance is that performance prediction methods are often interfered with by various sources of uncertainty. These sources of uncertainty can be divided into external condition uncertainty (caused by the randomness of future environment, load and operating conditions), modeling uncertainty (uncertainty of model parameters) and prediction method uncertainty (selection of various machine learning algorithms). For ships, due to the complexity of equipment, imprecise knowledge expression and limitations of testing methods, the mapping between fault symptoms and fault causes is random and uncertain. The prediction results of traditional data-driven models are point estimates, which cannot reflect the uncertainty of the current prediction results. Summary of the invention

[0007] In view of the deficiencies in the prior art, the purpose of the present invention is to provide a method, device and medium for predicting the degradation performance of a marine diesel engine based on Bayesian reasoning. By introducing Bayesian reasoning into the performance prediction of a marine diesel engine and considering the possibility of different parameters, Bayesian reasoning can generate a posterior distribution, thereby quantifying the uncertainty of the result in the form of a probability distribution. Since Bayesian reasoning integrates prior knowledge and various information of the observed object, accurate prediction of the performance of the target equipment can be achieved.

[0008] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0009] In the first aspect, the present invention proposes a method for predicting degradation performance of a marine diesel engine based on Bayesian reasoning, the key of which is that it comprises the following steps:

[0010] The state information data of the ship diesel engine from normal operation to fault state is collected, and after preprocessing, it is randomly divided into training set and test set;

[0011] Based on the operating status information of marine diesel engines, a Bayesian neural network model is constructed to predict the performance of marine diesel engines, and a Bayesian logistic regression model is constructed to quantify the uncertainty of the performance prediction results.

[0012] Variational inference algorithm is applied to jointly optimize the weight parameters of Bayesian neural network model and the parameters of Bayesian logistic regression model for diesel engine performance prediction and degradation trend analysis.

[0013] The Bayesian neural network model and the Bayesian logistic regression model are jointly optimized and trained using the training set, and the model parameters are jointly optimized and adjusted using the variational inference algorithm;

[0014] The test set is input into the trained Bayesian neural network model and Bayesian logistic regression model for performance testing. If the performance index reaches the preset standard, the training is completed; otherwise, the optimization training continues;

[0015] The status information data of the ship diesel engine in actual operation is obtained and input into the trained and optimized Bayesian neural network model to predict the real-time performance of the ship diesel engine. The performance prediction results are then input into the Bayesian logistic regression model to generate the degradation curve of the ship diesel engine for predicting future performance change trends.

[0016] Furthermore, the step of jointly optimizing and adjusting the model parameters by using the variational inference algorithm is as follows:

[0017] The variational inference algorithm dynamically adjusts the weight parameters in the Bayesian neural network model by calculating the approximate solution of the posterior distribution;

[0018] The Bayesian logistic regression model is based on the prediction results of the Bayesian neural network and uses the jointly optimized posterior distribution to adjust its parameters to quantify the uncertainty and generate the degradation curve of the diesel engine.

[0019] Furthermore, the preprocessing step includes data cleaning, feature engineering and normalization; the data in the training set and the test set are the instantaneous speed signal and the average indicated pressure of the ship diesel engine. Accordingly, the input data of the Bayesian neural network model is the instantaneous speed signal set, and its output data is the average indicated pressure set.

[0020] Furthermore, the construction process of the Bayesian neural network model is as follows:

[0021] Determine the performance indicators and performance evaluation criteria of the Bayesian neural network model;

[0022] Input the training set into the initial Bayesian neural network model, calculate its performance index according to the obtained prediction results, and optimize the number of hidden layers in the output layer of the initial Bayesian neural network model and the number of neurons in each hidden layer according to the performance evaluation criteria;

[0023] According to the parameters of the input layer and the output layer in the optimized Bayesian neural network model, a Bayesian neural network model for performance prediction of ship diesel engines is obtained.

[0024] Furthermore, the performance indicators of the Bayesian neural network model are the mean absolute error and root mean square error between the true value and the predicted value, specifically:

[0025]

[0026] Among them, y i , are the true value in the training set and the predicted value of the model output, respectively, and MAE is y i and The mean absolute error, RMSE, is yi and The root mean square error, n is the number of samples;

[0027] The performance evaluation standard of the Bayesian neural network model is the average value of the performance indicators MAE and RMSE after at least 30 tests.

[0028] Furthermore, the output layer parameters of the Bayesian neural network model are: the number of hidden layers is 2, and the number of neurons in each hidden layer is 30.

[0029] Furthermore, the specific steps of optimizing the Bayesian neural network model using the training set are:

[0030] The estimation problem of the posterior distribution is transformed into an optimization problem. The similarity between the two probability distributions q(z) and p(z|D) is calculated by KL divergence. The calculation formula is:

[0031]

[0032] Where KL[q(z)||p(z|D)] is the KL divergence between the approximate distribution q(z) and the posterior distribution p(z|D);

[0033] Use the Bayesian formula to transform the calculation formula of KL divergence:

[0034]

[0035] Let the evidence lower bound ELBO be L(q) = -∫q(z)logp(D,z) / q(z)dz. According to the mean field approximation theory, each parameter of L(q) is completely decomposed into two independent parts:

[0036] Part1=∫q(z)logp(D,z)dz,

[0037]

[0038] Use part1 and part2 to construct the loss function of the Bayesian neural network model:

[0039]

[0040] Among them, λ i for q i (z i ) weight, i = 1 ~ n;

[0041] Based on the obtained loss function, the gradient optimization algorithm is used to solve the maximized ELBO, update the weight parameters of the Bayesian neural network model, and train the Bayesian neural network model.

[0042] Furthermore, the specific steps of optimizing the Bayesian logistic regression model using the training set are:

[0043] Given a set of data: in The hypothesized functional form of the Bayesian logistic regression model is as follows:

[0044]

[0045] Among them, θ = (α, β), α and β are model parameters in the Bayesian logistic regression model, The prior distribution of β is N(0,I), and the prior distribution of α is N(0,1);

[0046] Get the posterior distribution of model parameters α and β:

[0047] p(β,α|y)=p(β)p(α)p(y|β,α),

[0048]

[0049] in, θ i =β T x i +α;

[0050] Based on the posterior distribution of model parameters α and β, the full conditional distribution of model parameters α and β is calculated:

[0051]

[0052] According to the full conditional distribution of the model parameters α and β, the MCMC algorithm is used to perform multiple sampling iterations to update the model parameters α and β, thereby completing the optimization training of the Bayesian logistic regression model.

[0053] In a second aspect, the present invention further proposes a device for predicting degradation performance of a marine diesel engine based on Bayesian reasoning, comprising:

[0054] The data acquisition module is used to collect the state information data of the ship diesel engine from normal operation to fault state, and randomly divide it into training set and test set after preprocessing;

[0055] A model building module for building a Bayesian neural network model to predict the performance of marine diesel engines and a Bayesian logistic regression model for quantifying the uncertainty of the performance prediction results;

[0056] Parameter optimization module, which is used to apply variational inference algorithm to jointly optimize the weight parameters of Bayesian neural network model and the parameters of Bayesian logistic regression model to achieve diesel engine performance prediction and degradation trend analysis;

[0057] The model training module is used to jointly optimize the Bayesian neural network model and the Bayesian logistic regression model using the training set, and jointly optimize and adjust the model parameters through the variational inference algorithm;

[0058] The performance testing module is used to input the test set into the trained Bayesian neural network model and Bayesian logistic regression model for performance testing. If the performance index reaches the preset standard, the training is completed, otherwise the optimization training continues;

[0059] The degradation performance prediction module is used to obtain the status information data of the ship diesel engine in actual operation, input it into the optimized Bayesian neural network model, predict the real-time performance of the ship diesel engine, and input the performance prediction results into the Bayesian logistic regression model to generate the degradation curve of the ship diesel engine for predicting its future performance change trend.

[0060] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method described in the first aspect.

[0061] The remarkable effects of the present invention are:

[0062] (1) According to the operating characteristics of ship diesel engines, the instantaneous speed signal of diesel engines is easy to measure. Therefore, the diesel engine average indicated pressure is predicted by the instantaneous speed signal of the diesel engine. The instantaneous speed is used as the status information data of the diesel engine, thereby indirectly realizing the performance prediction of the diesel engine.

[0063] (2) Two Bayesian models are proposed: a Bayesian neural network model and a Bayesian logistic regression model, which are used to realize the health status monitoring and degradation performance quantification of ship diesel engines, respectively. The parameters of the Bayesian neural network model and the Bayesian logistic regression model are inferred through variational inference and Markov chain Monte Carlo method to quantify the uncertainty of the parameters, and the posterior distribution of the model parameters can be calculated to finally obtain the confidence interval of the prediction result.

[0064] (3) A performance prediction framework based on the Bayesian neural network model and the Bayesian logistic regression model is proposed, which provides a new perspective and explanation for the quantification of uncertainty in the performance prediction of marine diesel engines. The prediction results in the form of probability distribution are helpful for adjusting the operation and maintenance plan of diesel engine equipment to reduce the failure rate of equipment, thereby ensuring the improvement of the efficiency of marine diesel engines under the premise of safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 is a general flow chart of the present invention;

[0066] Figure 2It is a diagram of the method steps of the present invention;

[0067] Figure 3 This is a comparison chart of the MAE and RMSE values ​​of the BNN model with different numbers of neurons;

[0068] Figure 4 Schematic diagram of the sampling process of each layer parameter of the Bayesian neural network model;

[0069] Figure 5 A comparison chart of the lower limit of evidence in each iteration;

[0070] Figure 6 It is a schematic diagram of the distribution of the posterior probability curve for predicting the average indicated pressure;

[0071] Figure 7 This is a comparison chart of the MAE and RMSE values ​​of each model in the verification experiment;

[0072] Figure 8 It is a comparison chart of the predicted value and the true value of the average indicated pressure by the Bayesian neural network model;

[0073] Fig. 9 The distribution diagram of normal samples and fault samples predicted by the Bayesian neural network model;

[0074] Fig.10 Schematic diagram of the parameter sampling process of the BLR model based on the MCMC method;

[0075] Fig.11 Schematic diagram of the degradation performance curve quantified for the Bayesian logistic regression model;

[0076] Fig.12 Schematic diagram for quantifying uncertainty for a Bayesian logistic regression model;

[0077] Fig.13 It is a principle block diagram of the degradation performance prediction device described in the present invention;

[0078] Fig.14 It is a schematic diagram of a computer device in the present invention. DETAILED DESCRIPTION

[0079] The specific implementation manner and working principle of the present invention are further described in detail below with reference to the accompanying drawings.

[0080] An often overlooked problem in the prediction of ship equipment performance is that the performance prediction method is often interfered by various sources of uncertainty. These sources of uncertainty can be divided into external condition uncertainty (caused by the randomness of future environment, load and operating conditions), modeling uncertainty (uncertainty of model parameters) and prediction method uncertainty (selection of various machine learning algorithms). For complex ship equipment, due to the complexity of the equipment, the imprecision of knowledge expression and the limitations of testing methods, the mapping between fault symptoms and fault causes is manifested as randomness and uncertainty. The prediction results of traditional data-driven models are point estimates, which cannot reflect the uncertainty of the current prediction results. Therefore, in this embodiment, Bayesian reasoning is introduced into the performance prediction of ship diesel engines. By considering the possibility of different parameters, Bayesian reasoning can generate a posterior distribution, thereby quantifying the uncertainty of the results in the form of a probability distribution.

[0081] Embodiment 1:

[0082] like Figure 1 and Figure 2 As shown, this embodiment is a method for predicting degradation performance of a marine diesel engine based on Bayesian reasoning, and the specific steps are:

[0083] Step 1: collect state information data of a ship diesel engine from a normal state to a fault state, perform data enhancement preprocessing on the data, and randomly divide the preprocessed state information data into a training set and a test set;

[0084] In this step, the preprocessing of data enhancement for the status information data includes data cleaning, feature engineering and normalization. Specifically:

[0085] Data cleaning processing refers to deleting drifting data in the data;

[0086] Feature engineering processing refers to extracting feature vectors from the instantaneous speed vector of a ship diesel engine;

[0087] Normalization processing refers to normalizing all the obtained eigenvectors, that is, dividing the obtained eigenvector by the sum of all elements in the eigenvector to obtain a normalized eigenvector;

[0088] Finally, the data set containing instantaneous speed characteristics and average indicated pressure is randomly divided into a training set and a test set, with 75% of the total number of samples being the training set and 25% of the total number of samples being the test set.

[0089] In this example, the data in the training set and the test set are the instantaneous speed signal and the average indicated pressure of the ship diesel engine. The reasons for selecting these two data are as follows:

[0090] The instantaneous speed signal of the diesel engine is easy to measure, so the diesel engine mean indicated pressure is predicted by the instantaneous speed signal of the diesel engine, and the instantaneous speed is used as the state information data of the diesel engine, thereby indirectly realizing the performance prediction of the diesel engine;

[0091] During the working cycle of the diesel engine, the gas pressure in the cylinder changes with the displacement of the piston to form a cycle curve. The area surrounded by the cycle curve can be expressed as the work done by the diesel engine, so it is called the indicator diagram. The indicator diagram comprehensively reflects the thermal conversion process of the diesel engine's output mechanical work and is an effective means to describe the performance of the diesel engine. The average indicated pressure can be calculated through the indicator diagram. The average indicated pressure is the indicated work done by the diesel engine's single cylinder working volume in one cycle. It is a very important indicator for measuring the actual cycle power performance of the diesel engine. The larger the average indicated pressure, the better the engine's working cycle is and the higher the utilization rate of the cylinder working volume. Therefore, this embodiment uses the average indicated pressure as the classification standard for the health status of the diesel engine. Table 1 shows the average indicated pressure of the test cylinder at 1000r / min and 75% load of the diesel engine. Under the fault conditions of different severity set in this embodiment, the greater the degree of fault, the smaller the average indicated pressure. Through the characteristic of the average indicated pressure, the average indicated pressure is set as the degradation performance index of the diesel engine. When the degree of diesel engine fault is greater, the average indicated pressure is smaller.

[0092] Table 1 Average indicated pressure of test cylinder under 75% load of diesel engine

[0093]

[0094] Therefore, the input of the Bayesian neural network model used in this embodiment is the instantaneous speed signal feature X = (x1, x2, ..., x n ), the output is the average indicated pressure Y=(y1,y2,…,y n ).

[0095] Step 2: Based on the ship diesel engine operating status information, a Bayesian neural network model is constructed to predict the ship diesel engine performance, and a Bayesian logistic regression model is constructed to quantify the uncertainty of the performance prediction results;

[0096] A Bayesian neural network is used to fit the relationship between the instantaneous speed signal and the indicated pressure. The BNN model introduces prior probabilities to the model parameters to achieve uncertainty in the estimation of model parameters. Ultimately, the BNN model can not only generate predictions, but also give the uncertainty of the predictions. Since the marginal probabilities of the BNN model parameters are difficult to calculate, a simple distribution q(z) is found through variational inference technology to approximate the posterior distribution p(z|D).

[0097] In this embodiment, the construction process of the Bayesian neural network model is as follows:

[0098] Step 2.1, determine the performance indicators and performance evaluation criteria of the Bayesian neural network model;

[0099] The hyperparameter setting of the Bayesian neural network has a significant impact on the final prediction performance. This embodiment studies the impact of different numbers of hidden layers and different numbers of neurons on the prediction performance of the model. The mean absolute error (MAE) and root mean square error (RMSE) are introduced as indicators of model performance for evaluation. The experiment was conducted 30 times, and the average value of the performance index was taken as the evaluation criterion of the model to reduce the impact of random factors on the model performance. The two performance indicators are defined as:

[0100]

[0101] Among them, y i , are the true value in the training set and the predicted value of the model output, respectively, and MAE is y i and The mean absolute error, RMSE, is y i and The root mean square error, n is the number of samples;

[0102] Step 2.2, input the training set into the initial Bayesian neural network model, and calculate its performance index according to the obtained prediction results, and optimize the number of hidden layers in the output layer of the initial Bayesian neural network model and the number of neurons in each hidden layer according to the performance evaluation criteria;

[0103] The number of neurons in the input layer and output layer of the neural network can be determined based on the number of features and predicted values ​​of the training set. However, in current research, there is no method to determine the number of neurons in the hidden layer. Therefore, this embodiment selects the optimal number of neurons through grid search. The number of hidden layers is set to 1, and the range of the grid search is [25, 35]. The optimal number of neurons in the hidden layer is determined by the model performance under MAE and RMSE evaluation. Figure 3 The MAE and RMSE values ​​of the BNN model with different numbers of neurons are shown. It can be seen that the MAE and RMSE values ​​are the lowest when the number of neurons is 30. Therefore, the number of neurons in the hidden layer in this embodiment can be set to 30.

[0104] In addition, this embodiment also studies the model performance of three different numbers of hidden layers, where the hidden layers are set to 1, 2, and 3 layers, respectively. The number of neurons in each layer is set to 30. Table 2 shows the model performance under different numbers of hidden layers. When the hidden layer is 2 layers, the MAE and RMSE values ​​of the Bayesian neural network are the lowest, which are 0.51 and 1.02, respectively.

[0105] Table 2 Performance of BNN models with different hidden layers

[0106]

[0107] The parameters of each layer in the Bayesian neural network are determined by the training set. The parameter training process is as follows: Figure 4 As shown, Figure 4 (a) shows the sampled values ​​of the input layer parameters. Figure 4 (b) shows the probability distribution of the input layer parameters. Figure 4 (c) shows the sampled values ​​of the first hidden layer parameters. Figure 4 (d) shows the probability distribution of the first hidden layer parameters, Figure 4 (e) shows the sampled values ​​of the second hidden layer parameters. Figure 4 (f) shows the probability distribution of the second hidden layer parameters. w_in_1 is the input hidden layer parameter of the Bayesian neural network model, w_1_2 is the first hidden layer parameter, and w_2_out is the second hidden layer parameter. The final parameters of the model are not single numbers, but probability distributions. The probability distribution of these parameters can ultimately generate the probability distribution of the predicted value. Figure 5 The iterative process of ELBO is shown.

[0108] Step 2.3: According to the parameters of the input layer and the output layer in the optimized Bayesian neural network model, a Bayesian neural network model for performance prediction of a marine diesel engine is obtained.

[0109] When the model training is completed, the mapping from input data to output data can be realized. When new samples are input into the model, the output value can be predicted.

[0110] The BLR model is used to model the ship diesel engine from normal state to fault state, and to quantify the uncertainty in this process. The logistic regression model is a nonlinear function with a target value between 0 and 1. In this patent study, the normal state of the ship diesel engine is set to 1 and the fault state is set to 0. Bayesian reasoning enhances the flexibility of the logistic regression function and quantifies the uncertainty of the prediction results. The uncertainty of the prediction results is quantified by the prior distribution of the unknown parameter θ and the confidence space of the prediction results. The uncertainty of the prediction results is quantified by the prior distribution of the unknown parameter θ and the confidence space of the prediction results. The sampling and estimation of the unknown parameter θ in the model are realized by the MCMC sampling method.

[0111] Step 3: Apply the variational inference algorithm to jointly optimize the weight parameters of the Bayesian neural network model and the parameters of the Bayesian logistic regression model for diesel engine performance prediction and degradation trend analysis, ensure the collaboration of the two in diesel engine performance prediction and degradation trend analysis, and improve the accuracy and robustness of the prediction, including:

[0112] (1) The variational inference algorithm dynamically adjusts the weight parameters in the Bayesian neural network model by calculating the approximate solution of the posterior distribution, so that it can provide more accurate performance prediction according to different operating conditions of the ship diesel engine (such as load changes and sea state changes);

[0113] (2) At the same time, the Bayesian logistic regression model uses the jointly optimized posterior distribution based on the prediction results of the Bayesian neural network to perform joint optimization adjustments to quantify the uncertainty and generate the degradation curve of the diesel engine;

[0114] During the operation of ship diesel engines, the variational inference algorithm can enable the Bayesian neural network model to maintain high-precision predictions under changing environmental conditions by jointly optimizing the parameters of the two models, while the Bayesian logistic regression model can accurately capture the degradation trend of the diesel engine.

[0115] Step 4: Use the training set to jointly optimize the Bayesian neural network model and the Bayesian logistic regression model, and optimize the parameters of both models through the variational inference algorithm to ensure efficient processing of diesel engine performance prediction and degradation trends under different environmental conditions and working conditions;

[0116] The optimization training process of the Bayesian neural network model is as follows:

[0117] Step A1: The estimation problem of the posterior distribution is transformed into an optimization problem. The similarity between the two probability distributions q(z) and p(z|D) is calculated by KL divergence. The specific process is as follows:

[0118] Assume that the training set is D = {x (i) ,y (i)}, the likelihood probability is p(D|z)=∏ i p(y (i) |x (i) ,z). According to Bayes’ theorem, the product of the likelihood probability p(D|z) and the prior distribution p(z) is proportional to the posterior distribution p(z|D), so the posterior distribution can be expressed as p(z|D)∝p(D|z)p(z).

[0119] The process of maximizing p(D|z)p(z) is to perform the maximum a posteriori estimation of the parameter z. In the variational inference algorithm, the calculation of the posterior distribution p(z|D) is avoided, and a traceable model q of the unobserved variable z is introduced to indirectly approximate the posterior distribution p(z|D):

[0120] q(z)≈p(z|D)

[0121] In variational inference, the estimation problem of the posterior distribution is transformed into an optimization problem, and the KL divergence is used to calculate the similarity between two probability distributions q(z) and p(z|D). The KL divergence of q(z) and p(z|D) can be expressed as:

[0122]

[0123] Where KL[q(z)||p(z|D)] is the KL divergence between the approximate distribution q(z) and the posterior distribution p(z|D);

[0124] Step A2: The basic idea of ​​variational inference is to find a distribution q(z) similar to distribution p(z|D) by minimizing KL divergence. Bring Bayes' theorem into KL divergence and use Bayes' formula to transform the calculation formula of KL divergence:

[0125]

[0126] Step A3: Let the lower bound of evidence ELBO be L(q) = -∫q(z)logp(D,z) / q(z)dz. The problem of minimizing the KL divergence can be transformed into maximizing the ELBO, making the calculation easier. In the process of maximizing the ELBO, it is not necessary to calculate the probability distribution p(D) in the KL divergence.

[0127] However, in Bayesian neural networks, when q(z) contains multiple variables, the calculation is more complicated. Therefore, the mean field approximation is introduced to simplify the assumptions so that each parameter is completely decomposed into independent parts, as shown below:

[0128] q(z)=q1(z1)·q2(z2)·…·q n (z n )=∏ n q i (z i )

[0129] L(q)=∫q(z)logp(D,z)dz-∫q(z)logq(z)dz

[0130] Part1=∫q(z)logp(D,z)dz

[0131]

[0132] Step A4, the two parts Part 1 and Part 2 work together to maximize ELBO. Part 1 drives the variational distribution and places weights where the posterior distribution is high. Part 2 attempts to diffuse the variational distribution by placing weights around its parameter space. The loss function of the Bayesian neural network model is constructed using Part 1 and Part 2:

[0133]

[0134] Among them, λ i for q i (z i ) weight, i = 1 ~ n;

[0135] Step A5: Based on the obtained loss function, the gradient optimization algorithm is used to solve the maximized ELBO, the weight parameters of the Bayesian neural network model are updated, and the Bayesian neural network model is trained. The updating method is as follows:

[0136]

[0137] Where λ is a constant.

[0138] In this way, the weight parameters of the Bayesian neural network model are updated and the Bayesian neural network model is trained.

[0139] This embodiment uses variational reasoning to derive the joint probability distribution of all uncertain parameters in the Bayesian neural network model, so that the joint probability distribution gradually approaches the posterior distribution of the parameters, and removes the parameter mean from the obtained posterior distribution of the parameters as the model parameter.

[0140] The optimization training process of the Bayesian logistic regression model is as follows:

[0141] Step B1: Given a set of data: in The hypothesized functional form of the Bayesian logistic regression model is as follows:

[0142]

[0143] Among them, θ = (α, β), α and β are the model parameters of logistic regression, The model parameters β and α in the Bayesian logistic regression model are considered as random variables. Assume that the prior distribution of β is N(0,I) and the prior distribution of α is N(0,1).

[0144] Step B2, α and β are independent of each other, and the sampled samples are independent and identically distributed, then the posterior distribution of the parameters is:

[0145]

[0146] in:

[0147]

[0148] Likelihood function:

[0149]

[0150] θ i =β T x i +α

[0151] The likelihood function can be written in this form because y i ∈{0,1}, when y i =1, We can get:

[0152]

[0153] When i =0, we can get:

[0154]

[0155] Step B3: Based on the posterior distribution of the model parameters α and β, obtain the full conditional distribution of the model parameters α and β:

[0156]

[0157] Step B4: According to the full conditional distribution of the model parameters α and β, the MCMC algorithm is used to perform multiple sampling iterations to update the model parameters α and β, thereby completing the optimization training of the Bayesian logistic regression model.

[0158] Step 5: Input the test set into the trained Bayesian neural network model and Bayesian logistic regression model for performance testing. If the performance index reaches the preset standard, the training is completed; otherwise, return to step 4 to continue optimization training;

[0159] Step 6: Obtain the status information data of the ship diesel engine in actual operation, input it into the trained and optimized Bayesian neural network model, predict the real-time performance of the ship diesel engine, and then input the performance prediction results into the Bayesian logistic regression model to generate the degradation curve of the ship diesel engine for predicting its future performance change trend.

[0160] When the model training is completed, the mapping from input data to output data is realized. When new samples are input into the model, the output value can be predicted. Figure 6 (a) shows the probability distribution of the average indicative force value generated by new samples in the model. Figure 6 The dashed line in (a) corresponds to the most likely value. At the same time, this embodiment also shows the probability distribution of multiple groups of average indicated pressures predicted by the model when multiple samples are input, such as Figure 6 (b) as shown.

[0161] The corresponding maximum probability value in the probability distribution is used as the predicted value. The error between the true value and the predicted value is shown in Table 3. As can be seen from Table 3, the average indicated pressure value predicted by the BNN model is very close to the actual average indicated pressure value, with a maximum error of only -0.79.

[0162] Table 3 Error between the predicted value and the true value of the BNN model

[0163]

[0164] In the present invention, support vector regression (SVR), linear regression (LR), decision tree (DT), random forest (RF) and Bayesian linear regression (BLR) are introduced as comparison models to compare and verify the performance of the Bayesian neural network model used in this embodiment. At the same time, R2 is introduced as the third model prediction performance index to evaluate the prediction performance of different models under the same data set. Figure 7 Table 4 shows the MAE, RMSE and R2 of different models respectively. Figure 7 (a) shows the MAE value of each model. Figure 7 (b) shows the RMSE value of each model.

[0165]

[0166] In the formula, y i and is the average indicated pressure true value and the model predicted value. is the average of the predicted values, and n is the number of samples.

[0167] Table 4 R of each model 2 value

[0168]

[0169] MAE and RMSE are used to measure the deviation between the model prediction value and the true value. The lower the value, the better the model prediction performance. 2 is a dimensionless parameter that is the percentage of variation in the outcome variable that the model predictor variables can explain, ranging from 0 to 1. 2The higher the value, the better the prediction performance of the model. Among the models used in this embodiment, the Bayesian neural network model achieved the lowest values ​​of MAE and RMSE, and R 2 The highest value. Compared with the linear regression model, the Bayesian linear regression model achieves better prediction results. Because Bayesian reasoning can constrain the regression model through the specified prior distribution and obtain the uncertainty of the predicted value in the form of the posterior distribution, it means that better prediction results can be obtained through the entire estimation range. The MAE, RMSE and R of the Bayesian neural network model used in this embodiment 2 The values ​​are 0.505, 1.02, and 97.34%, respectively, which are better than other models. Figure 8 The predicted values ​​and true values ​​of the mean indicated pressure by the Bayesian neural network are shown. It can be seen that the predicted values ​​are relatively consistent with the true values.

[0170] The predicted mean indicated pressure is denoted as the Bernoulli variable {I = (P k ),k=1,2,…,n}, I=1 indicates that the device has not reached the fault state, and I=0 indicates that the device has reached the fault state. Fig. 9 As shown in the figure, the horizontal axis represents the average indicated pressure value of the diesel engine, and the vertical axis represents whether the equipment is currently in normal working condition or faulty working condition. It can be found that when the diesel engine is in normal working condition, most of the average indicated pressure values ​​are in the interval [15, 20], while when it is in faulty working condition, most of the average indicated pressure values ​​are in the interval [5, 15]. The difference in average indicated pressure under different conditions provides the possibility for quantitative evaluation of diesel engine performance.

[0171] Logistic regression is a nonlinear function in which multiple independent variables jointly determine a dependent variable, which is represented by a Bernoulli variable. The Bernoulli variable of the diesel engine operating condition depends on the mean indicated pressure value, which can be described by the logistic function as:

[0172]

[0173] Where C is the current state of the diesel engine and can be 0 or 1. α and β obey the normal distribution. i is the predicted average indicated pressure value.

[0174] In the present invention, the Bayesian logistic regression model is used to fit the nonlinear process of the diesel engine from normal operation to fault operation. The Markov Chain Monte Carlo (MCMC) algorithm is used to estimate the parameters α and β. The posterior distribution estimation process of the parameters is as follows: Fig.10 For this example, the sampling rate of the MCMC algorithm is set to 5000. Fig.10 (a) contains all α and β sample values, Fig.10 (b) shows the posterior distribution of parameters α and β. After MCMC sampling, the maximum likelihood values ​​of parameters α and β are 19.2 and -1.31 respectively. It should be noted that if β is 0, it means that the average indicated pressure cannot well reflect the working condition of the diesel engine.

[0175] The purpose of the Bayesian logistic regression model proposed in this embodiment is that when a new average indicated pressure value is input, the model can output the performance degradation probability of the current device. Fig.11 The figure shows the performance curve of diesel engine degradation based on mean indicated pressure. When the mean indicated pressure is greater than 15 and close to 20, the diesel engine is likely to be in normal working condition. When the mean indicated pressure is less than 10, the current diesel engine is likely to be in fault condition. When the mean indicated pressure is equal to 15, Fig.11 It can be seen that the diesel engine degradation performance is at 60%, indicating that the current equipment has not reached a fault state. As the mean indicated pressure decreases, the performance of the ship diesel engine decreases from 100% to 0.

[0176] This example quantifies uncertainty by using a 95% confidence interval for the prediction results, such as Fig.12 The middle curve area represents the uncertainty of the prediction results, which can be interpreted as the confidence of the model in the current prediction results. As more data is collected, the uncertainty of the prediction results will be further reduced. Fig.12 It can be seen that the uncertainty interval of the prediction result is relatively small, which means that there are enough training samples to train the model, and the Bayesian logistic regression model is confident in the current prediction result. It should be noted that when the confidence interval of the prediction result is large, the operator needs to pay more attention to the current decision-making. The probability of diesel engine degradation performance obtained by the BLR model can provide a reference for the predictive maintenance of ship diesel engines and improve the engine manager's understanding of the current operating status of the diesel engine.

[0177] Embodiment 2:

[0178] This embodiment proposes a marine diesel engine degradation performance prediction device based on Bayesian reasoning, such as Fig.13 As shown, it includes:

[0179] The data acquisition module is used to collect the state information data of the ship diesel engine from normal operation to fault state, and randomly divide it into training set and test set after preprocessing;

[0180] A model building module for building a Bayesian neural network model to predict the performance of marine diesel engines and a Bayesian logistic regression model for quantifying the uncertainty of the performance prediction results;

[0181] Parameter optimization module, which is used to apply variational inference algorithm to jointly optimize the weight parameters of Bayesian neural network model and the parameters of Bayesian logistic regression model to achieve diesel engine performance prediction and degradation trend analysis;

[0182] The model training module is used to jointly optimize the Bayesian neural network model and the Bayesian logistic regression model using the training set, and jointly optimize and adjust the model parameters through the variational inference algorithm;

[0183] The performance testing module is used to input the test set into the trained Bayesian neural network model and Bayesian logistic regression model for performance testing. If the performance index reaches the preset standard, the training is completed, otherwise the optimization training continues;

[0184] The degradation performance prediction module is used to obtain the status information data of the ship diesel engine in actual operation, input it into the optimized Bayesian neural network model, predict the real-time performance of the ship diesel engine, and input the performance prediction results into the Bayesian logistic regression model to generate the degradation curve of the ship diesel engine for predicting its future performance change trend.

[0185] Embodiment 3:

[0186] like Fig.14 As shown, this embodiment proposes a computer device, including a memory, a network interface and a processor, wherein the processor generally controls the overall operation of the computer device; the network interface enables the electronic device to communicate with other terminals or servers through a network; the memory stores a computer program that can be run on the processor, and the processor implements the method described in Example 1 when executing the computer program.

[0187] The memory is configured to store instructions and applications executable by the processor, and can also cache data to be processed or processed by various modules of the processor (for example, image data, audio data, voice communication data, and video communication data), which can be implemented through flash memory (FLASH) or random access memory (Random Access Memory, RAM). Data can be transmitted between the processor, network interface and memory through a bus.

[0188] It should be noted here that the description of the above storage medium and device embodiments is similar to the description of the above method embodiments, and has similar beneficial effects as the method embodiments. For technical details not disclosed in the storage medium and device embodiments of this application, please refer to the description of the method embodiments of this application for understanding.

[0189] A person skilled in the art can understand that all or part of the steps of implementing the above method embodiment can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiment; and the aforementioned storage medium includes: mobile storage devices, read-only memories (ROM), magnetic disks or optical disks, etc., various media that can store program codes.

[0190] Alternatively, if the above-mentioned integrated unit of the present application is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present application can be essentially or partly embodied in the form of a software product that contributes to the relevant technology. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROMs, magnetic disks, or optical disks.

[0191] Embodiment 4:

[0192] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the method described in Embodiment 1 is implemented.

[0193] In summary, this embodiment proposes a method, device, equipment and storage medium for predicting degradation performance of a marine diesel engine based on Bayesian reasoning, wherein the method introduces Bayesian reasoning into the performance prediction of a marine diesel engine. By considering the possibility of different parameters, Bayesian reasoning can generate a posterior distribution, thereby quantifying the uncertainty of the result in the form of a probability distribution. Bayesian reasoning combines prior knowledge and various information of the observed object to achieve accurate prediction of the performance of the target equipment;

[0194] Two Bayesian models are proposed for the performance prediction of marine diesel engines: the Bayesian neural network model and the Bayesian logistic regression model. The Bayesian neural network model is used to realize the health monitoring of marine diesel engines, while the Bayesian logistic regression model is used to quantify the process of the equipment from normal operation to failure state. The instantaneous speed of the diesel engine is selected as the state monitoring data to indirectly predict the average indicated pressure of the diesel engine, and further evaluate and quantify the current performance of the diesel engine;

[0195] Finally, in order to verify the superiority of the proposed Bayesian reasoning method, data sets were obtained under normal conditions and fault conditions of different severity of the ship diesel engine. The experimental results show that the diesel engine performance prediction method based on Bayesian reasoning proposed in this embodiment not only has higher prediction accuracy than other algorithms, but also can generate confidence intervals for the prediction results to assist in the adjustment of diesel engine equipment operation and maintenance plans, reduce unexpected interruptions of diesel engines, and improve the efficiency of ship diesel engines while ensuring safety.

[0196] The technical solution provided by the present invention is described in detail above. In this embodiment, specific examples are used to illustrate the principle and implementation method of the present invention. The description of the above embodiment is only used to help understand the method of the present invention and its core idea. It should be pointed out that for ordinary technicians in this technical field, without departing from the principle of the present invention, the present invention can also be improved and modified, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. A method for predicting degradation performance of marine diesel engines based on Bayesian reasoning, characterized in that: The steps include: The state information data of the ship diesel engine from normal operation to fault state is collected, and after preprocessing, it is randomly divided into training set and test set; Based on the operating status information of marine diesel engines, a Bayesian neural network model is constructed to predict the performance of marine diesel engines, and a Bayesian logistic regression model is constructed to quantify the uncertainty of the performance prediction results. Variational inference algorithm is applied to jointly optimize the weight parameters of Bayesian neural network model and the parameters of Bayesian logistic regression model for diesel engine performance prediction and degradation trend analysis. The Bayesian neural network model and the Bayesian logistic regression model are jointly optimized and trained using the training set, and the model parameters are jointly optimized and adjusted using the variational inference algorithm; The test set is input into the trained Bayesian neural network model and Bayesian logistic regression model for performance testing. If the performance index reaches the preset standard, the training is completed; otherwise, the optimization training continues; The status information data of the ship diesel engine in actual operation is obtained and input into the trained and optimized Bayesian neural network model to predict the real-time performance of the ship diesel engine. The performance prediction results are then input into the Bayesian logistic regression model to generate the degradation curve of the ship diesel engine for predicting future performance change trends.

2. The method for predicting degradation performance of a marine diesel engine based on Bayesian reasoning according to claim 1 is characterized in that: The steps of jointly optimizing and adjusting the model parameters by using the variational inference algorithm are as follows: The variational inference algorithm dynamically adjusts the weight parameters in the Bayesian neural network model by calculating the approximate solution of the posterior distribution; The Bayesian logistic regression model is based on the prediction results of the Bayesian neural network and uses the jointly optimized posterior distribution to adjust its parameters to quantify the uncertainty and generate the degradation curve of the diesel engine.

3. The method for predicting degradation performance of a marine diesel engine based on Bayesian reasoning according to claim 1 is characterized in that: The preprocessing step includes data cleaning, feature engineering and normalization; the data in the training set and the test set are the instantaneous speed signal and the average indicated pressure of the ship diesel engine. Correspondingly, the input data of the Bayesian neural network model is the instantaneous speed signal set, and its output data is the average indicated pressure set.

4. The method for predicting degradation performance of a marine diesel engine based on Bayesian reasoning according to claim 1 is characterized in that: The construction process of the Bayesian neural network model is as follows: Determine the performance indicators and performance evaluation criteria of the Bayesian neural network model; Input the training set into the initial Bayesian neural network model, calculate its performance index according to the obtained prediction results, and optimize the number of hidden layers in the output layer of the initial Bayesian neural network model and the number of neurons in each hidden layer according to the performance evaluation criteria; According to the parameters of the input layer and the output layer in the optimized Bayesian neural network model, a Bayesian neural network model for performance prediction of ship diesel engines is obtained.

5. The method for predicting degradation performance of a marine diesel engine based on Bayesian reasoning according to claim 4 is characterized in that: The performance indicators of the Bayesian neural network model are the mean absolute error and root mean square error between the true value and the predicted value, specifically: Among them, y i , are the true value in the training set and the predicted value of the model output, respectively, and MAE is y i and The mean absolute error, RMSE, is y i and The root mean square error, n is the number of samples; The performance evaluation standard of the Bayesian neural network model is the average value of the performance indicators MAE and RMSE after at least 30 tests.

6. The method for predicting degradation performance of a marine diesel engine based on Bayesian reasoning according to claim 4 is characterized in that: The output layer parameters of the Bayesian neural network model are: the number of hidden layers is 2, and the number of neurons in each hidden layer is 30.

7. The method for predicting degradation performance of a marine diesel engine based on Bayesian reasoning according to claim 1 is characterized in that: The specific steps of optimizing the Bayesian neural network model using the training set are: The estimation problem of the posterior distribution is transformed into an optimization problem. The similarity between the two probability distributions q(z) and p(z|D) is calculated by KL divergence. The calculation formula is: Where KL[q(z)||p(z|D)] is the KL divergence between the approximate distribution q(z) and the posterior distribution p(z|D); Use the Bayesian formula to transform the calculation formula of KL divergence: Let the evidence lower bound ELBO be L(q) = -∫q(z)logp(D,z) / q(z)dz. According to the mean field approximation theory, each parameter of L(q) is completely decomposed into two independent parts: Part1=∫q(z)logp(D,z)dz, Use part1 and part2 to construct the loss function of the Bayesian neural network model: Among them, λ i Q i (z i ) weight, i = 1 ~ n; Based on the obtained loss function, the gradient optimization algorithm is used to solve the maximized ELBO, update the weight parameters of the Bayesian neural network model, and train the Bayesian neural network model.

8. The method for predicting degradation performance of a marine diesel engine based on Bayesian reasoning according to claim 1 is characterized in that: The specific steps of optimizing the Bayesian logistic regression model using the training set are: Given a set of data: in The hypothesized functional form of the Bayesian logistic regression model is as follows: Among them, θ = (α, β), α and β are model parameters in the Bayesian logistic regression model, The prior distribution of β is N(0,I), and the prior distribution of α is N(0,1); Get the posterior distribution of model parameters α and β: p(β,α|y)=p(β)p(α)p(y|β,α), Among them, i i =b T x i +a; Based on the posterior distribution of model parameters α and β, the full conditional distribution of model parameters α and β is calculated: According to the full conditional distribution of the model parameters α and β, the MCMC algorithm is used to perform multiple sampling iterations to update the model parameters α and β, thereby completing the optimization training of the Bayesian logistic regression model.

9. A device for predicting degradation performance of marine diesel engines based on Bayesian reasoning, characterized in that: include: The data acquisition module is used to collect the state information data of the ship diesel engine from normal operation to fault state, and randomly divide it into training set and test set after preprocessing; A model building module for building a Bayesian neural network model to predict the performance of marine diesel engines and a Bayesian logistic regression model for quantifying the uncertainty of the performance prediction results; Parameter optimization module, which is used to apply variational inference algorithm to jointly optimize the weight parameters of Bayesian neural network model and the parameters of Bayesian logistic regression model to achieve diesel engine performance prediction and degradation trend analysis; The model training module is used to jointly optimize the Bayesian neural network model and the Bayesian logistic regression model using the training set, and jointly optimize and adjust the model parameters through the variational inference algorithm; The performance testing module is used to input the test set into the trained Bayesian neural network model and Bayesian logistic regression model for performance testing. If the performance index reaches the preset standard, the training is completed, otherwise the optimization training continues; The degradation performance prediction module is used to obtain the status information data of the ship diesel engine in actual operation, input it into the optimized Bayesian neural network model, predict the real-time performance of the ship diesel engine, and input the performance prediction results into the Bayesian logistic regression model to generate the degradation curve of the ship diesel engine for predicting its future performance change trend.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images