Thermal fault diagnosis of marine diesel engine based on LSTM neural network algorithm

2.1. Setting of thermal parameters

Thermal fault is a very common type of diesel engine fault, which is generally divided into three types: blocking, leakage and phase error. Thermal parameters of diesel engine are an important basis for judging the type of fault. During the practical navigation, the thermal parameters of diesel engine will vary widely with the change of sea state. When the thermal parameters change beyond the preset range, it will be considered as a failure.

To sum up, based on thermal faults and other research experience, this paper sets 13 thermal faults and seven fault parameters, and sets the normal range and fault range values of each parameter, which detailed in Table 1.

2.2. Diesel engine simulation model

This paper uses Matlab and Simulink software platform to build the diesel engine simulation model, according to the thermodynamic formula and combined with the actual work of the diesel engine to build the simulation interface, which is shown in the Fig. 1.

Fig. 1Dynamic simulation model diagram of diesel engine

Based on the simulation model of diesel engine, by changing the input parameters and physical structure of the simulation model, the operation condition of diesel engine when thermal failure occurs is simulated, and finally the simulation model of diesel engine thermal failure is established, which makes a good foundation for building the diesel engine fault simulation system in the next step. The thermal parameters corresponding to different thermal faults can be obtained by running the model.

3.1. LSTM neural network

The traditional recursive neural network model focuses on the time dimension, which can model the sequence sample. However, the recursive neural network usually contains a large number of parameters, which is not only difficult to train but also easy to dissipate the gradient in the time dimension. Long Short Term Memory Network (LSTM) has made breakthroughs in language translation, image classification and sequence prediction in medicine, which can solve the problem of long-term dependence. In industrial control, LSTM is mainly used in fault prediction and analysis [6-11].

LSTM is essentially a special Recurrent neural Network (RNN), LSTM networks have more complex hidden layers. The hidden layer of RNN has only one state function H, which is sensitive short-term input. LSTM adds a new state function C in the hidden layer, which is suitable for long-term output. The function of state function c is realized through the forgetting gate, the input gate and the output gate. These three structures realize the long-term preservation of information by controlling the input and output of the hidden layer.

The forgetting gate controls whether to forget the upper long-term state $C_{t-1}$ with a certain probability. The input gate and the output gate are responsible for processing the input and output of the current sample. The input gate inputs the immediate state of the current sample into the long-term state $c$, and the output gate determines whether to output the long-term state c as the current state variable, so as to realize the long-term memory of information.

The LSTM network algorithm is divided into two parts: information forward propagation and error back propagation. The forward propagation algorithm of a single LSTM module is derived as follows:

1) The forgetting gate reads the output information $h_{t-1}$ of the previous hidden layer and the current input $x_t$. After reading, the forgetting gate outputs a specific value itself as part of the $C_{t-1}$. Then the value of $f_t$ determines whether the information is retained or not. The value of 1 will completely retained the information, and the value of 0 will forget the next layer of information:

2) The input gate consists of two parts. The first part uses sigmod activation function to output $i_t$; the second part uses tanh activation function, output $at$. The input gate mainly depends on these two activation functions to realize the function of sample information input:

3) After the treatment of forgetting gate and input, the cell state function $C_t$ has changed:

4) The output information consist of two parts, the first part is $o_t$, the second part is $h_t$:

In summary, LSTM network changes the information transmission form of RNN network, adds the gate structure control unit, generates the path of long-term continuous flow of gradient, and introduces the linear self-circulation operation, which well solves the long-term dependence of RNN.

4.1. Result display

From the data obtained by fault simulation, the sample data of standard working conditions and each faults state are, extracted. 10000 groups of data are selected from each case. The supervision value is marked as 1-14. According to the code, the training set samples are divided into 90 % of the total sequence data, and the test set samples account for 10 %.

The network completes the training and learning after 1000 iterations. After the completion of the training, only the numerical values of the thermal parameters are input into the network. The predicted supervision value can be obtained by the algorithm model, and then the specific thermal fault types of diesel engines can be judged. Fig. 2 shows the performance diagram of LSTM neural network training results provided by Matlab.

Fig. 2Performance diagram of LSTM neural network training results

In addition, the training sequence is drawn by predictive value and compared with test data. Update the network using observations to predict each time step. For each prediction, the observation values of the previous time step are used to predict the next time step, and Fig. 3 is obtained.

Through the figures, we can clearly see that LSTM neural network can make accurate prediction for diesel engine thermal fault diagnosis, and the final training results are highly fitted. The root mean squared error of LSTM model is small, indicating that the variable has a strong ability to explain the predicted value.

Fig. 3LSTM neural network prediction comparison chart

4.2. Comparison of diagnostic results between LSTM model and BP model

Taking the same sample data and selecting the neural network toolbox for network training, different fault diagnosis models can be obtained. By testing two different fault diagnosis models with test data, the comparison results of fault diagnosis accuracy can be obtained. It can be seen from the following table that the diagnostic accuracy of LSTM fault diagnosis model is higher.