A multi-fault diagnosis method for piston pump in construction machinery based on information fusion and PSO-SVM

2.1. Structure of piston pump in the construction machines

Piston pumps are not only rotating machines, but also reciprocating machines. They are designed to convert the rotary motion of the input shaft to the reciprocating motion of pistons. The typical structure diagram of piston pump in construction machines is shown in Fig. 1. The piston pump has 9 pistons installed in the piston holes which are evenly distributed along the cylinder block. The function of the compression spring is to keep the slipper contact with the swash plate and press the cylinder block against the valve plate. The tilt angle of the swash plate is controlled by the electromagnetic valve. The crescent bearing is used to adjust the tilt angle of the swash plate with the electromagnetic valve. The valve plate has two waist-shaped ports respectively connected with the suction and discharge ports of the piston pump. When the cylinder block is rotated by the drive shaft, the pistons reciprocate along the axis direction while rotating with the cylinder block. When the piston rotates from the upper side of the figure to the lower side, the slipper retracts with the swash plate and partial vacuum is formed in the piston hole. The hydraulic oil enters the piston hole from the oil suction port through the valve plate and the oil suction process has been completed. When the piston rotates to the upper side, the pressing force pushes the piston move to the right side. With the continuous reduction of volume of the piston hole, the hydraulic oil is discharged through the oil discharge port connected to the valve plate and the oil discharge process has been completed.

Fig. 1Structure diagram of the piston pump in construction machines: 1 – drive shaft, 2 – compression spring, 3 – slipper, 4 – piston, 5 – electromagnetic valve, 6 – valve plate, 7 – bearing, 8 – cylinder, 9 – crescent bearing, 10 – swash plate

2.2. Fault mechanism analysis of piston pump

As seen from section 2.1, there are four frictional pairs in the piston pump: valve plate/cylinder block, piston/piston bore, swash plate/slipper and piston head/slipper. Because of poor working condition, high load and long running time, wear in the pairs will become more and more serious. As a consequence, it will result in efficiency declining, pressure fluctuation, abnormal vibration, and shorten its service life. The common fault modes and mechanism analysis in the piston pumps are shown in Table 1 [25].

From Table 1, we can know majority of single fault can be identified by monitoring pressure signal or vibration signal. However, for the construction machines, which have to work in harsh environments over time. Multiple faults are most likely to happen simultaneously in the piston pump. For example, when solid contaminants the hydraulic oil, abrasion will occur not only in valve plate/cylinder block, but also in piston/piston bore. It increases the difficulty of fault diagnosis. Any single information is inaccurate and incomplete. Therefore, both pressure signal and vibration signal should be used for comprehensive evaluation.

4.1. Fault feature extraction from vibration signal

When the load pressure was 10 MPa, the time domain waveforms of the vibration signals of the test piston pump in 6 different states are shown in Fig. 4. As can be seen, all of the vibration signals exhibit obvious non-stationary characteristics. Different types of faults will cause different changes in the vibration signal. Most faults occur will result in increase of the amplitude of the vibration signal. Especially multi-fault, it will lead to more increase of the amplitude of the vibration signal. But it is difficult to judge the occurrence of different kinds of faults only by the vibration signals in time domain.

Fig. 4Time domain waveforms of vibration signals

Fig. 5Frequency spectra of vibration signals

Frequency spectra are calculated by using the Fast Fourier Transforms (FFT) and the corresponding results are shown in Fig. 5. From Fig. 5, the pulse frequency, 240 Hz, is clearly observed from all frequency spectra in spite of the health state of the piston pump. Additionally, there are obvious peaks at the frequencies of $m$×240 Hz $(m=$ 1, 2, 3, …). However, the fault characteristic frequency is not obvious due to the additional vibration caused by faults was very weak relative to which caused by the rotation and the oil supply. Therefore, it was difficult to diagnose different faults of piston pump by the frequency spectra of vibration signals. Furthermore, it also can be seen from Fig. 5 that the faults will bring periodic impacts accompanying the operation of the piston pump. Different types of faults will change the energy of different frequency bands of the vibration signal, so the energy in different frequency bands of the vibration signal contains rich fault characteristics information.

In order to obtain the energy characteristics in different frequency bands of the vibration signal, empirical mode decomposition (EMD) was used to decompose the vibration signals and the results were shown in Fig. 6.

Fig. 6EMD of vibration signals

a) N

b) SF1

c) SF2

d) SF3

e) MF1

f) MF2

As we know, EMD can decompose the signal into a group of Intrinsic Mode Functions (IMFs), which contain components of original signal in different frequency bands from high frequency to low frequency [26]. It can be seen from Fig. 6 that vibration signals in different modes were decomposed into 8 to 9 IMFs. Each IMF was a stationary signal with a single component and the IMFs in same frequency band were obviously different under 6 different modes. Therefore, the energy variation of each frequency band implied the healthy state of the piston pump.

The steps for extracting the energy characteristics after EMD of vibration signals were as follows:

1) Calculate the total energy $E_i$ of each IMF:

where $C_i$ is IMF obtained after EMD.

2) Construct feature vectors $T$:

3) Normalize the feature vector:

where ${\sum }_{i=1}^n{\left({\left|E_i\right|}^2\right)}^{1/2}.$

Using the above calculation method, it was found that the sum of the energy of the first six IMF components has reached more than 95 % of the total energy for each state. It is mean that the energy of vibration signal is mainly concentrated in the high-frequency band and the first 6 IMF components contained the main healthy state information of the piston pump. Therefore, the energy of the first 6 IMF components extracted from the vibration signal constitutes the feature vector for the fault diagnosis of piston pump.

4.2. Fault feature extraction from pressure signal

When the load pressure was 10 MPa, the time domain waveforms of the pressure signals of the test piston pump in 6 different states are shown in Fig. 7. It can be seen that pressure signals fluctuate around 10 MPa and there is a significant pressure pulsation every 0.04 s, which is consistent with the pulse frequency 240 Hz. When the piston pump has different types of faults, the time domain waveform of the pressure signal will change. Therefore, the 9 time domain characteristics parameters of the pressure signal constitute the feature vector for the fault diagnosis of piston pump. The selected 9 time domain characteristic parameters are defined as Table 4. In which $x_i$ is a signal series for $i=$ 1, 2, …, $n$, $n$ is the number of data points.

Fig. 7Pressure signals

5.1. PSO-SVM classifier

In this study, Support Vector Machine (SVM) was selected as the classifier for fault diagnosis. SVM is a machine learning technique which based on statistical theory and structural risk minimization. It is widely used in the pattern recognition and fault diagnosis [28, 29].

However, before training the SVM model, penalty factor and kernel parameter should be predetermined, which has great impact on the classification accuracy. Using PSO algorithm particle swarm optimization (PSO) to optimize the two parameters of SVM, the steps are listed as follows.

Step 1. Initialize the penalty factor $C$ and the kernel parameter $\gamma$ to build a SVM model.

Step 2. A two-dimensional particle swarm ($P_1$, $P_2$) is established, in which the parameter $C$ is optimized by particle swarm $P_1$ and the parameter $\gamma$ is optimized by particle swarm $P_2$ respectively. Setting the initial values of the particles (including the inertia weight $\omega$, positive acceleration constants $c_1$ and $c_2$, the constraint factor $\beta$ and the number of iterations) and generating the initial position and velocity of the particle swarm are randomly.

Step 3. The initial velocity and position are substituted into the SVM to obtain the fitness values of each particle, which is:

where $f_i$ denotes the fitness value of the $i$th particle, $l$ is the number of samples. $y_{ij}$ is the SVM output of the $j$th sample of the $i$th particle, and ${\widehat{y}}_j$ is the average output of all samples.

Step 4. For each particle, compare its fitness value to the best position it has experienced and use the best value as the current best position.

Step 5. For each particle, compare its fitness value to the best position experienced by the group and set the best value to the current position of the group.

Step 6. Adjust the velocity $v_{ij}$ and position $x_{ij}$ of the particles according to Eq. (7) and Eq. (8):

where, $i$ is a particle in the group and $j$ is the dimension of the particle. $v_{ij}$ and $x_{ij}$ represent the velocity and the position of particle $i$ in the $j$-dimensional space respectively. $p_{ij}$ denotes the individual optimal solution of particle $i$ and $p_g$ denotes the group optimal solution of all particles. $\beta$ is a constraint factor used to control the velocity weight, which is usually set to 1. $\omega$ is the inertia weight, which is used to balance the capabilities of global exploration and local exploitation. $c_1$ and $c_2$ are two positive acceleration constants regulating the relative velocities, and they are usually set to 2. $r_{1j}$ and $r_{2j}$ are random variables in the range [0,1].

Step 7. Check whether the end condition is met (the minimum error requirement or the maximum number of iterations is reached). And if it is yes, stop the iteration, or return to step3.

Step 8. The optimized parameters $C$ and $\gamma$ are brought into the Equation to establish the SVM classification model.

Furthermore, the original SVM is a typical two-class classifier. However, the fault diagnosis of piston pump in this paper needs a multi-classification to separate 6 modes. For a case of $k$ classes, $k$-1 two-class SVM classifiers are needed. So, the diagnosis model which comprises five two-class SVM classifiers was build. With all training samples of the six modes, PSO-RVM1 was trained to separate the normal state from fault states. When the input fault feature was a sample representing normal state, the output was set to 1, otherwise the output was set to 0. PSO-SVM2, PSO-SVM3 and PSO-SVM4 were designed to separate 3 single fault states. PSO-SVM5 was designed to separate 2 multi-fault states. The diagnosis process of PSO-SVM classifiers was shown in Fig. 8.

Fig. 8Diagnosis process of PSO-SVM classifier

5.2. Fault diagnosis results

The diagnosis results of multi-source information fusion were compared with the results of vibration signal and pressure signal diagnosis alone.

The fault diagnosis process based on EMD energy feature of vibration signal and PSO-SVM was shown in Fig. 9 The fault diagnosis results are shown in Table 5. It can be seen that the recognition accuracy of the normal state and three single failure modes are 96.3 % and 92.4 % respectively. This shows that the energy characteristics extracted after EMD decomposition reflect the health status of the piston pump well. At the same time, PSO-SVM can also distinguish between the normal state and several single fault modes of piston pump. However, we also can see the recognition rate of the multi-fault is relatively low, which is 87.1 %. Therefore, relying on vibration signals to detect multi-fault does not achieve satisfactory result.

Fig. 9Fault diagnosis process based on vibration signal

The fault diagnosis process based on time domain feature of pressure signal and PSO-SVM was shown in Fig. 10. The fault diagnosis results are shown in Table 6. The recognition accuracy using pressure signal is lower than which based on vibration signal. Especially, the recognition accuracy for multi-fault is only 78.9 %. Therefore, it is necessary to carry out subsequent research on fault feature extraction method and fault pattern recognition method based on pressure signal.

The fault diagnosis process based on multi-source information fusion and PSO-SVM was shown in Fig. 11. The fault diagnosis results are shown in Table 7. It can be seen that the recognition accuracy of the normal state, three single failure modes are 98.3 % and 97.6 % respectively. These recognition accuracies are higher than which using a single signal. Moreover, two multi-fault modes can be well recognized and the recognition accuracy reaches 94 %. Therefore, using the method proposed in this paper can not only identify the single fault, but also effectively identify multi-fault of the piston pump.

Fig. 10Fault diagnosis process based on pressure signal

Fig. 11Fault diagnosis process based on multi-source information fusion