PNN: A Novel Progressive Neural Network for Fault Classification in Rotating Machinery under Small Dataset Constraint

TL;DR

This work tackles fault classification in rotating machinery under small, heterogeneous datasets by introducing a Progressive Neural Network (PNN) with FFT-based data standardization. The PNN progressively refines fixed-size features across layers, enabling efficient learning with few parameters and mitigating vanishing gradients, achieving high accuracy even with limited data. Across eight datasets, PNN6 (depth 6, Hd=100) attains state-of-the-art performance, with near-perfect AUROC at standard splits and robust results under few-shot scenarios, while requiring significantly fewer parameters than conventional deep nets. The framework supports cross-domain transfer learning, reducing data demands for new machines and layouts, and is well-suited for embedded deployment in Industry 4.0 settings, though further work is needed to extend to non-rotary and compound-fault cases.

Abstract

Fault detection in rotating machinery is a complex task, particularly in small and heterogeneous dataset scenarios. Variability in sensor placement, machinery configurations, and structural differences further increase the complexity of the problem. Conventional deep learning approaches often demand large, homogeneous datasets, limiting their applicability in data-scarce industrial environments. While transfer learning and few-shot learning have shown potential, however, they are often constrained by the need for extensive fault datasets. This research introduces a unified framework leveraging a novel progressive neural network (PNN) architecture designed to address these challenges. The PNN sequentially estimates the fixed-size refined features of the higher order with the help of all previously estimated features and appends them to the feature set. This fixed-size feature output at each layer controls the complexity of the PNN and makes it suitable for effective learning from small datasets. The framework's effectiveness is validated on eight datasets, including six open-source datasets, one in-house fault simulator, and one real-world industrial dataset. The PNN achieves state-of-the-art performance in fault detection across varying dataset sizes and machinery types, highlighting superior generalization and classification capabilities.

Figures (14)

• Figure 1: An overview of the proposed unified fault detection technique: Original datasets of various sizes are standardized to a single size (16384), and then PNN of $N$ layers and $C$ number of fault classes is trained on this data. The first layer $(l_1)$ of PNN has input $X$ of size $K$, and the output $z_h$ of size $H_d$, The intermediate layer $(l_n)$ receives the input from layer $l_{n-1}$ (size $H_d$) along with input from previous layers $l_{n-2}$ (size $K+(n-2)H_d$) and the final layer gives output $C_i$ ($i^{th}$ fault class). Here $n$ is the layer number & $N$ is depth of the PNN.
• Figure 2: Typical IC-Engine data set fault spectrum with processing by N-point FFT (N=16384) and max of bin operation.
• Figure 3: Detailed design of a PNN processing blocks
• Figure 4: t-SNE plot of fault classes for ICE, CWRU, and PB datasets. Each color represents a fault type in the plot.
• Figure 5: SpectraQuest machinery fault simulator with components.