Fault analysis and diagnosis system for induction motors

Highlights

• This study analyzes the vibration signals of the 10 most common motor fault types.

• Extension theory and NN are combined to build the motor fault diagnosis system.

• The proposed motor fault diagnosis system is fast, requires less training data.

• The proposed fault diagnosis system exhibits first-rate identification capability.

Abstract

This study analyzes the vibration signals of fault induction motors for establishing an intelligent motor fault diagnosis system by using an extension neural network (ENN). Extension theory and a neural network (NN) are combined to construct the motor fault diagnosis system, which identifies the most likely fault types in motors. First, the vibration signal spectra of the 10 most common fault types are measured and organized into individual motor fault models. Subsequently, according to the motor fault data, representative characteristic frequency spectra are identified, and the correlation between the motor fault types and their corresponding characteristic frequency spectra are established to develop the motor fault diagnosis system. Finally, the test results confirm that the proposed motor fault diagnosis system is fast, requires less training data, and demonstrates first-rate identification capability.

Introduction

Currently, motor fault assessment involves detecting signals from motors by using instruments. The signals are then analyzed by professionals, who identify the causes of excessive motor vibrations. However, because the knowledge, experience, and analytical approaches of professionals vary, motor condition assessment results can also vary. In addition, the signals must then be analyzed by on-site workers, who can use their knowledge and experience to identify the fault types, maintain drives, and replace components. This conventional troubleshooting method may cause errors in assessing fault points, engendering unnecessary human resource use and time costs in maintaining and replacing components [1], [2].

In response to consumer demands, electric equipment has recently become large-scale, automated, and complex, rendering conventional troubleshooting methods inapplicable to current equipment. Therefore, a new troubleshooting technique for identifying system fault points must be formulated. Scholars worldwide are committed to creating new system troubleshooting techniques [3], [4], [5]. An artificial intelligence algorithm [3] was applied to improve the efficiency of motor drive fault diagnosis and save human resources and time. In addition, approaches involving neural networks (NNs) [4], [5] have been employed to establish motor drive fault diagnosis systems. These approaches require calculation of connection weights among neurons for identifying fault types. Although these approaches have improved troubleshooting speeds, they require voluminous system data [6], [7], [8]. Other types of algorithms, such as neural-fuzzy theory [1] and approximate reasoning [2], have also been applied for motor drive fault diagnoses. Neural-fuzzy theory does not require structuring motor drive systems; hence, establishing a fault diagnosis system on the basis of this theory involves only input–output relationships. However, when algorithms that are based on this theory are used, more time is required for identifying optimal solutions, and data identification must be performed individually. By contrast, approximate reasoning entails incorporating a simpler calculation approach for troubleshooting. Nevertheless, if differences in motor drive characteristics are not substantial, errors may occur when approximate reasoning is employed for troubleshooting. The researchers in [9] reported three motor fault diagnosis types, namely abnormal air-gap eccentricity, bearing failure, and broken rotor bar, on the basis of motor stator current power spectra, and they monitored the diagnoses through smart wireless sensor networks. However, these motor fault diagnoses cannot clearly identify the exact fault position. Hilbert and wavelet transforms and knowledge extraction by using data mining [10], [11], [12] have been applied for converting time–frequency plots into frequency–amplitude plots. In addition, the genetic algorithm selects the fault characteristics and diagnoses motor fault types. However, it can diagnose only six motor fault types: bowed rotor, broken rotor bar, fault bearing, unbalanced rotor, stator fault, and voltage unbalance. This implies that all of the 10 common motor fault types have yet to be diagnosed.

A new approach to intelligent fault diagnosis of three-phase induction motors using a signal-based method was proposed in [13]. In the classifier design, the artificial ant clustering technique, inspired by the behavior of real ants, was used as an unsupervised classification method for optimizing fault diagnosis. This technique was tested in different situations to demonstrate its effectiveness in detecting failures, even when the information about operating modes is limited or difficult to obtain. Nevertheless, the obtained results proved the efficiency of the proposed approach only in diagnosing broken bars and bearing failure at various load levels. In addition, to accurately diagnose motor faults, such as broken bars, mixed eccentricity, and shaft bearing fault, the characteristic spectra of motor vibration signals and stator current have been obtained in [14], [15], [16], [17], [18], [19]. However, other motor fault types demonstrating the same fault characteristic spectra still exist, but with different amplitudes, thus resulting in the false detection of a motor fault type during motor fault diagnosis.

Developing a new fault diagnosis technique to shorten the time required by on-site workers for determining fault points and maintaining systems is highly imperative. In addition, on-site professional diagnosis staff can determine the required duration of system maintenance according to the signals acquired using such a new technique, thus increasing economic benefits.

The rest of this paper is organized as follows. Section 2 described the motor types and frequency spectrum analysis. And then, Section 3 expounded the proposed motor fault diagnosis system. An analysis of the measured vibration signal spectra for 10 fault types was made in Section 4. In Section 5, the proposed ENN fault diagnosis system was used for fault diagnosis of the induction motors. Finally, some conclusions and discussions are given in Section 6.