1. Introduction
A diesel engine is a kind of internal combustion engine that converts thermal energy into mechanical energy. It plays an important role in the field of national defense, in the chemical industry, in the marine industry, for nuclear power, and so on. Once a diesel engine fails, it not only causes economic losses directly or indirectly in terms of the shutdown of equipment, but it may also threaten the personal safety of users [
1,
2]. To enhance the availability of the diesel engine, it is imperative to monitor the engine condition and detect early faults. However, the detection of faults and the diagnosis of diesel engines [
3] are not simple tasks due to the complex structure and fickle working conditions. If the operating conditions are not considered in detection and diagnosis activities, it is likely to lead to false alarms or missed detection [
4,
5]. With the information of operating conditions, the engineering applicability of a fault detection and diagnosis method [
6,
7,
8] can be improved to avoid fatal performance degradation and huge economic losses at an early stage. Unfortunately, most fault detection methods are carried out under stable operating condition to avoiding variable operating conditions. Therefore, condition recognition is an important and urgent task in practical engineering applications.
In a diesel engine, the flywheel is attached to the crankshaft, and they rotate together. They convert the reciprocating motion of the piston into the rotational motion of the crankshaft, which outputs torque for the driving of the car and other power-driven mechanisms. Therefore, the operating conditions of a diesel engine can be determined by two parameters: load and the rotation speed of the crankshaft. The load is the output torque of the engine through the flywheel. However, the multi-factor operating conditions of a diesel engine are difficult to measure in many situations, such as for the power systems of vehicles, propulsion devices of ships, and other dynamic equipment. Therefore, the demand for automatic recognition of multi-factor operating conditions is urgent.
During the operation of a diesel engine, the corresponding status information can be obtained by using vibration analysis [
9], oil analysis [
10], thermal performance analysis [
11], and visual inspection. Vibration is an intrinsic mechanical phenomenon, and the vibration signals contain rich information about the diesel engine’s status; thus, vibration monitoring is a powerful tool for condition recognition, as well as fault detection and diagnosis. In this paper, we aim at recognizing the multi-factor operating conditions of a diesel engine based on vibration signals.
Thanks to the development of computing calculation power and powerful signal processing techniques, the recognition tasks based on vibration signals made great progress. At present, some recognition algorithms based on vibration signals exist, and most of them focus on designing various handcrafted features, fusing multiple features and training different classifiers. In Reference [
12], the Hilbert spectrum entropy, which combines the Hilbert spectrum and information entropy, was proposed for the pattern recognition of diesel engine working conditions. In Reference [
13], the frequency domain features of vibration signals were extracted for back propagation (BP) and radial basis function (RBF) neural network training to recognize the cylinder pressure. In Reference [
14], based on the cylinder head vibration signals measured under stable operating conditions, an engine cylinder pressure identification method using a genetic algorithm with BP neural network was proposed. In Reference [
15], combustion evaluation parameters were extracted using time–frequency coherence analysis and the cylinder pressure could be estimated based on the parameters and an RBF neural network. In Reference [
16], the measured signal was converted into a crank angle degree signal using the rotational speed monitored by magnetic pickup sensors. Then, a real-time engine load classification algorithm was proposed based on an artificial neural network.
Most pattern recognition studies focused mainly on single-factor conditions or recognition under stable operating conditions. For single-factor conditions, the number of categories is generally no greater than five. In practical engineering applications, a single factor cannot describe complex operating conditions, and this drawback results in ambiguous boundaries among different operating conditions. As for multi-factor operating conditions, as the number of operating conditions increases, so does the complexity of condition recognition. Simultaneously, as the vibration signals are random, transient, and cyclostationary, and as the corresponding feature extraction requires rich domain knowledge, it is difficult to extract sensitive characteristics of significant importance for multi-factor operating condition recognition.
Over the last few years, with the development of deep learning, many researchers exploited deep neural networks (DNNs) as the feature extractor and classifier [
17,
18]. Benefiting from the powerful feature extraction ability of neural network, especially convolution neural networks (CNNs) [
19], these approaches and their variations exhibit good performance in the related tasks. In Reference [
20], time domain and frequency domain feature representations were selected to form a vector to act as the input parameters of a CNN. The trained CNN classifier could diagnose the fault patterns of a gearbox with outstanding performance. In Reference [
21], the vibration signals of rolling bearings were analyzed using continuous wavelet transform to get time–frequency representations in grayscale. Then, all compressed time–frequency representations were taken as the input for CNN training, and the trained CNN classifier could identify the faults of rolling bearings with strong generalization ability. In Reference [
22], a deep convolutional neural network of up to 38 layers, which could provide high classification accuracy, was proposed for gas classification. For CNN applications with vibration signals, there are different approaches to network input. In other words, the CNN is taken as a classifier, and the input of the CNN is mainly based on other feature extraction methods. At the same time, state-of-the-art CNN models have several parameters, which leads to problems related to storage, computation, and energy cost. In addition, recurrent neural networks (RNNs) and long short-term networks (LSTMs) [
23,
24] were validated in terms of their performance on one-dimensional (1D) signals. In Reference [
25], a CNN and a fully connected neural network were both incorporated into a deep neural network framework to improve LSTM. The framework outperformed the original LSTM for the early diagnosis and prediction of sepsis shock. In Reference [
26], an end-to-end model combining a CNN and RNN was proposed for the automatic detection of atrial fibrillation. Compared to the state-of-the-art models evaluated on standard benchmark electrocardiogram datasets, the proposed model produced better performance in detecting atrial fibrillation. The ideas in References [
25,
26] are very good references for multi-factor operating condition recognition based on vibration signals.
Therefore, a multi-factor operating condition recognition algorithm is proposed herein based on a 1D CNN and LSTM. In the proposed neural network framework, the 1D CNN was designed to extract local features of vibration signals through 1D convolution, and the LSTM was designed to describe the temporal relationship between local features. The contributions of this paper are summarized as follows:
A multi-factor operating condition recognition method is proposed using a 1D convolutional long short-term network (1D-CLSTM). As far as we know, this is the first study to combine a 1D CNN and LSTM to recognize operating conditions based on a time series of vibration signals;
Considering the particularity of engine vibration signals, batch normalization (BN) is introduced to regulate the input of some layers by fixing the mean value and variance of input signals in each convolutional layer;
Adaptive dropout is proposed for improving the model sparsity and preventing overfitting;
The designed 1D convolutional long short-term network (1D-CLSTM) classifier can achieve high generalization accuracy for recognizing multi-factor operating conditions.
The rest of this paper is organized as follows:
Section 2 presents the test bench of a diesel engine and the experimental data acquisition.
Section 3 introduces the technical background for the 1D CNN and LSTM.
Section 4 describes the designed 1D-CLSTM and the flowchart of the multi-factor operating condition recognition algorithm.
Section 5 shows the training performance of the designed 1D-CLSTM classifier, with generalizability verification, a performance comparison with different methods, and a training performance comparison with different dropout ratios. Finally, conclusions and future prospects are presented in
Section 6.