This paper proposes an efficient learning based approach to detect the faults of an industrial oil pump. The proposed method uses the wavelet transform and genetic algorithm (GA) ensemble for an optimal feature extraction procedure. Optimal features, which are dominated through this method, can remarkably represent the mechanical faults in the damaged machine. For the aim of condition monitoring, we considered five common types of malfunctions such as casing distortion, cavitation, looseness, misalignment, and unbalanced mass that occur during the machine operation. The proposed technique can determine optimal wavelet parameters and suitable statistical functions to exploit excellent features via an appropriate distance criterion function. Moreover, our optimization algorithm chooses the most appropriate feature submatrix to improve the final accuracy in an iterative method. As a case study, the proposed algorithms are applied to experimental data gathered from an industrial heavy-duty oil pump installed in Arak Oil Refinery Company. The experimental results are very promising.

Maintenance and repair costs are often known to be the heaviest charges in industries. Various studies have been conducted on condition monitoring of equipment in industrial processes and have long been discussed by scientists in systems dynamic field. By choosing the appropriate way of maintenance, it may greatly reduce costs, machine downtimes, and spare parts consumption and improves the reliability of machines which consequently increases the safety of machine operators. Hereof, health evaluation and fault detection have been done by a series of measurements on the data carrier signals like vibration, acoustic emission, and temperature profile records by the properly mounted sensors. For accurate measurements and better expressing the structural dynamic behavior of a machine, vibrations are the most common indicator of failure occurrence in the machine operating time. Concerning this, signal processing techniques have always been efficient tools that are used regularly by researchers. Among all techniques, Time-Frequency transforms based on wavelets have been more popular due to their properties to represent the abnormalities and unpredictable components localized in the time domain [

To validate the effectiveness of the suggested procedure in this paper, the oil pump system installed in Imam Khomeini Oil Refinery Company located in Arak was utilized. The pump type is BB5 (High-Pressure Double Case Pump) made by Ebara company which is designed as radially split multi-stage and operates under the maximum flow rate of 1500 cubic meter per hour (m^{3}/h).

The most frequent faults which happen in a pump lifetime are analyzed in our practical study which includes casing distortion, cavitation, misalignment, looseness of interior components, and dynamically unbalanced mass. The examples of the measured signals in some machine conditions are shown in

To analyze vibration signals with variable frequency content, the well-known time-frequency method of wavelet transform (WT) is used to investigate the local and global content in analyzed signals.

Wavelet packet transform is the generalized form of the discrete wavelet transform. It breaks down the frequency domain to slighter intervals as it increases the frequency resolution. Wavelet packet coefficients of a finite energy function _{j,k,n} is given by the following equations [

where

The implemented feature extraction technique consists of five main steps as follows:

Wavelet signal denoising/smoothing

Segmentation

Wavelet packet decomposition (WPD)

Statistical analysis

Node selection

In parallel to the mentioned activities on the analyzing signal, the genetic algorithm was applied to the problem and corresponding search domain to find the optimum values for uncertain parameters due to the binary coding capability. Mentioned activities are investigated in detail as follows:

Step 1: Environmental noises are everlasting parts of the recorded signals in real plants. To attenuate the effects of these undesirable noises on predicted outcomes, we have obtained the wavelet-denoising stage to make the signal smoother and generate more stable outputs. Thus, the decomposition level in the wavelet packet structure is considered to be the first candidate for the optimization process which determines the decomposition and reconstruction level of the examined signal. In this stage, “biorthogonal3-1” was chosen to approximate the signals because of its suitable properties in signal reconstruction.

Step 2: By our visual investigations, it seems that due to the rotary nature of the equipment, vibration signals act approximately the same in cyclic time intervals. Thus, to analyze the signals temporarily, we segmented the signals into 40 sections to preserve the similar behavior in new intervals.

Step 3: In this step, the WPD algorithm at decomposition level of four was evolved on whole sections of all fault signals. The main point in this step is the selection of the proper mother wavelet function which can represent signals appropriately. So, GA will consider the mother wavelet function as the second candidate for optimization [

Step 4: To shorten the length of the generated coefficient matrix in the previous step and to overcome the computational complexity, all coefficients are subjected to statistical analysis. A statistical function is applied to coefficient vectors row by row. GA will seek an appropriate item among candidate functions including absolute mean, standard deviation, skewness, kurtosis, root mean square.

Step 5: The aim in this stage is to select the salient terminal nodes for two beneficial reasons: (1) To choose the most indicative features, (2) To reduce the needed CPU time for calculations; therefore, the last parameter for optimization is more effective wavelet nodes in the 4th level of WPD.

The inter-intra class criterion was employed as a score function to separate different faults in the optimization process which provides class discrimination information over the training set [

where

To verify the capability, the proposed method was implemented in Labview 2010 software. We used the “Inter/Intra class” criterion as an optimization fitness function [

Parameter | Value |
---|---|

Population size | 10 |

Crossover type | Single-point |

Crossover probability | 0.9 |

Mutation type | Bit inversion |

Mutation probability | 0.1 |

Natural selection rate | 50% |

Selection rule | Roulette wheel |

^{th} generation with the score level being constant for five consecutive generations. According to ^{th} generation. The parameters will be fixed to their optimal values and build the final feature set which is going to feed into the ANN classifier.

Candidate | Result |
---|---|

Denoising level | 3 |

Mother wavelet function | Db2 |

Statistical function | RMS |

Binary-coded node index | 1010001000010010 |

^{th} layer in WPD tree) can build distinct features from the fault signals. Thus, the features in wavelet terminal nodes numbered 1, 3, 7, 12, and 15 can make distinct class features set. We calculated RMS values in the selected nodes as input samples to drive the classifier. In this regard,

In the classification phase, the feature matrix with the size of 5 × 40 × 3 features was considered for each machine fault. The classifier network was verified by test data after the training stage.

Predicted output | ||||||
---|---|---|---|---|---|---|

Casing distortion | Cavitation | Looseness | Misalignment | Unbalanced mass | ||

Casing distortion | 24 | 0 | 0 | 0 | 1 | |

Cavitation | 0 | 17 | 0 | 0 | 0 | |

Looseness | 0 | 0 | 22 | 0 | 0 | |

Misalignment | 0 | 0 | 0 | 21 | 0 | |

Unbalanced Mass | 0 | 0 | 0 | 1 | 28 |

We outlined a learning based procedure to extract applicable features of an industrial pump (

The authors wish to thank H. Shafieepour, head of the instrumentation department of Arak Oil Refinery Company, for their anonymous supports in data acquisition, which led to this practical work.