Previous research investigators have exploited machine-learning algorithms to diagnose the defects in rotating machinery. However, with increasing complexity in the design of rotating machinery, it is quite challenging to quantify the faults precisely. In this present study, an attempt has been made to predict the defect severity of the rotating machinery using Adaptive Neuro-Fuzzy Inference System (ANFIS). This ANFIS algorithm employs artificial neural networks to define the membership functions, rules and weights to construct the fuzzy inference system. Experiments are performed on a multi-stage spur gearbox model while it is subjected to fluctuating operating speeds. Two local defects on bearing race as well as on gear tooth with four different severity levels are seeded intentionally. Three condition monitoring (CM) strategies, namely, vibration, lubrication oil and acoustic signal analyses are executed, and the raw data is recorded synchronously. The raw vibration and acoustic waveforms are decomposed through discrete wavelet transform to extract the descriptive statistics from the wavelet coefficients. Among them, most discriminating features are selected and given as input to ANFIS classification tool to train the network for obtaining the Sugeno-type FIS, which in turn estimates the severity of the component. Later, the features from the individual CM strategies are combined to devise an integrated feature dataset which is further channelled as input to the ANFIS for predicting the defect severity levels. The investigation reveals that, the proposed integrated feature set in conjunction with ANFIS can discriminate between the defect severity conditions of the gears as well as bearings under fluctuating speeds.

以前的研究人员已经利用机器学习算法来诊断旋转机械中的缺陷。然而，随着旋转机械设计的日益复杂，精确量化故障是非常具有挑战性的。在本研究中，尝试使用自适应神经模糊推理系统 (ANFIS) 预测旋转机械的缺陷严重程度。该 ANFIS 算法采用人工神经网络来定义隶属函数、规则和权重，以构建模糊推理系统。实验是在多级直齿轮变速箱模型上进行的，同时它会受到波动的运行速度的影响。轴承座圈和齿轮齿上的两个局部缺陷具有四种不同的严重程度是有意播种的。三种状态监测 (CM) 策略，即 执行振动、润滑油和声学信号分析，并同步记录原始数据。通过离散小波变换分解原始振动和声波波形，从小波系数中提取描述性统计量。其中，选择最具辨别力的特征并将其作为 ANFIS 分类工具的输入，以训练网络以获得 Sugeno 型 FIS，进而估计组件的严重性。之后，将来自各个 CM 策略的特征结合起来设计一个集成的特征数据集，该数据集进一步作为输入到 ANFIS 以预测缺陷严重程度。调查显示，