Physics-Driven Deep Feature Fusion: A Lightweight CSAKansformer Architecture for Tool Wear Diagnosis in P25 Turning

Shuqiang Wang, Tianyue Zhang, Ximin Liu, Wei Liu, Huanqi Zhang, Feng Chang · 2026

Read Physics-Driven Deep Feature Fusion: A Lightweight CSAKansformer Architecture for Tool Wear Diagnosis in P25 Turning by Shuqiang Wang, Tianyue Zhang, Ximin Liu, Wei Liu, Huanqi Zhang, Feng Chang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/pmc-13211143.txt.

Introduction

Shuqiang Wang and colleagues spot cutting tool wear from fused sensor images, in work from 2026. Tool wear breaks machining flow and harms part quality if it goes unseen for long. Multi sensor fusion often runs heavy and slow, which blocks live use on the shop floor.

Sound and vibration streams become colour images through a physics guided angle method. A slim custom network then reads those images to call the tool state. The network holds few layers and answers in a few milliseconds per step. Accuracy passed ninety percent on turning tests in the study runs. Scores beat two leading light image models by clear margins.

Catalog record

CREATORShuqiang Wang, Tianyue Zhang, Ximin Liu, Wei Liu, Huanqi Zhang, Feng Chang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE77 KB
VERSION1.0
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC13211143

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Physics-Driven Deep Feature Fusion: A Lightweight CSAKansformer Architecture for Tool Wear Diagnosis in P25 Turning Authors: Shuqiang Wang, Tianyue Zhang, Ximin Liu, Wei Liu, Huanqi Zhang, Feng Chang Published: 2026 ABSTRACT Accurate tool wear identification is essential for ensuring the continuity of intelligent machining and workpiece quality. To address the challenges of multi-source fusion inefficiency and inadequate feature extraction, this study proposes a novel identification architecture combining physics-guided multi-channel Gramian angular field (PG-MGAF) with a minimalist 14-layer CSA-Kansformer network. Multi-source signals are preprocessed via PG-MGAF to convert 1D time-series into 2D RGB images, effectively characterizing spatial coupling and interactive energy across three channels. Subsequently, the minimalist network maps these composite features to tool states, significantly reducing computational overhead. Experimental results demonstrate that the proposed model achieves an average accuracy of 93.6% with a single-step inference latency of only 5.90 ms, significantly outperforming mainstream methods such as MobileNet-V2 and ConvNeXt. This architecture provides a high-efficiency, low-latency solution for real-time tool condition monitoring under complex industrial conditions. 1. Introduction With the deep advancement of Industry 4.0 and intelligent manufacturing, high-speed, high-precision, and unmanned machining impose higher requirements on the modern equipment manufacturing industry. The inevitable wear or sudden failure of cutting tools during continuous operations directly leads to workpiece rejection, shortened equipment lifespan, and severe unplanned production interruptions. Statistics show that tool-related costs account for 25% to 30% of total machining costs, and downtime caused by tool failures constitutes 1/3 of total automated line downtime [1]. Therefore, achieving accurate, real-time tool condition monitoring (TCM) is essential. Existing TCM technologies are divided into offline inspection and online monitoring. Offline inspection is inefficient and difficult to adapt to continuous automated production due to required machine shutdowns [2]. While online monitoring is the mainstream, direct visual measurement is highly susceptible to harsh working conditions like cutting fluid splashing and chip occlusion. In contrast, indirect monitoring methods based on physical signals—such as cutting force, vibration, and acoustic emission—effectively overcome these limitations. Offering real-time, non-destructive, and uninterrupted tracking, they have become the most widely applied technical route in TCM [3]. Current indirect monitoring-based TCM research mainly branches into two paradigms. The first is end-to-end deep learning based on one-dimensional (1D) time-series signals. Employing architectures like 1D-CNN, Temporal Convolutional Networks (TCN), and Long Short-Term Memory (LSTM) networks, this branch directly extracts features from raw signals, eliminating reliance on manual feature engineering [4]. However, the restricted local receptive field of 1D convolutions struggles to simultaneously capture global temporal dependencies and local high-frequency mutations. Consequently, weak early-stage wear signals are easily submerged by intense operational noise [5,6]. Furthermore, the fusion of multi-source heterogeneous signals is often limited to shallow channel concatenation, failing to uncover coupled evolution mechanisms among different physical signals [7,8]. The second branch is the visual recognition method based on the two-dimensional (2D) imaging of 1D signals. Through specific encoding rules, 1D signals are converted into 2D images, transforming hidden wear characteristics into visual features such as texture, grayscale, and spatial structure for end-to-end recognition via mature computer vision models [9]. This fully exploits the advantages of multi-scale feature extraction and global receptive fields, simultaneously capturing global evolution patterns and local mutation features [10]. It provides a high-dimensional representation space for deep multi-source fusion, offering a novel pathway to break through the inherent limitations of the 1D route [11]. Recent studies have extensively explored 2D imaging and feature extraction for single-sensor time-series signals, with technological evolution predominantly centering on imaging method optimization and network architecture upgrades. For instance, Martínez-Arellano et al. pioneered the use of the Gramian angular summation field (GASF) to losslessly encode 1D cutting force signals for CNN-based classification, laying the foundation for 1D-to-2D signal conversion. To capture high-frequency transients indicative of early wear, the Continuous Wavelet Transform (CWT) has been widely adopted [12]. Subsequent studies integrated CWT with multi-scale feature pyramids [13], improved ResNets [14], and MobileViT [15] to balance recognition accuracy and low-latency online monitoring. Despite these advances, single-sensor imaging exhibits inherent limitations. A solitary imaging mechanism captures only one facet of the signal (e.g., GASF emphasizes temporal dependencies, while CWT focuses on time-frequency distributions), failing to comprehensively depict the non-stationary wear evolution. Furthermore, single-sensor data cannot span the full degradation spectrum—from macroscopic load shifts to microscopic material fractures—making it highly vulnerable to working condition fluctuations and industrial noise. To overcome single-sensor representation bottlenecks, multi-source fusion imaging has emerged as a crucial research frontier. Researchers have attempted to expand feature dimensions by combining multiple imaging mechanisms and sensor modalities. For example, methods integrating GASF and Markov transition fields (MTF) [16], time-frequency MTF [17], and deep domain adaptation networks [18] have demonstrated improved accuracy and robustness under variable conditions. Kou et al. further pushed this boundary by concatenating GAF-encoded time-series features with infrared thermal images [19]. Nevertheless, existing multi-source fusion strategies have yet to resolve fundamental pain points. First, current fusion paradigms predominantly rely on horizontal concatenation or shallow channel stacking. By ignoring the inherent physical topological correlations and coupled evolution mechanisms of heterogeneous signals (e.g., force, vibration, and acoustic emission), these methods fail to achieve deep, channel-level energy interactions. This often amplifies wear-irrelevant noise and degrades overall robustness. Second, the absence of a unified, full-life-cycle health baseline normalization framework makes it difficult to isolate wear-specific degradation increments from amplitude disparities across varying sensors, thereby leaving early weak wear features susceptible to noise submergence. Regarding feature extraction networks, while the introduction of CNNs, Transformers, and attention mechanisms has elevated performance, two critical bottlenecks persist. First, feature mapping in mainstream architectures heavily relies on multi-layer perceptrons (MLPs). Given that tool wear is a strongly coupled, highly non-linear dynamic process, the static “linear weights + fixed activation” mechanism of MLPs struggles to accurately approximate complex degradation boundaries. Second, conventional attention mechanisms are tailored for general visual images and are fundamentally misaligned with physically encoded multi-channel fused images. They fail to adaptively focus on wear-critical feature flows, resulting in the severe underutilization of physical degradation information. To address the bottlenecks of multi-source feature fusion and non-linear feature extraction under complex working conditions, and to achieve high-accuracy identification of tool wear states, the main contributions of this paper are as follows:A WGAN-DIV-based minority-class augmentation and high-dimensional physical map reconstruction mechanism oriented toward complex turning conditions is proposed. Addressing the issues of sample imbalance across wear stages and the limited representation capability of single-channel imaging, this paper first introduces a divergence-regularized Wasserstein generative adversarial network (WGAN-DIV) to generate minority-class samples for the slight-wear and severe-wear stages. The generated samples are further evaluated through MMD, SSIM, and statistical consistency tests to verify their distributional fidelity. Subsequently, the augmented multi-source features are physically rearranged and adaptively normalized according to the “force–vibration–acoustic emission” hierarchy, and then mapped into three RGB channels, namely R: static spatial coupling based on GASF, G: dynamic degradation divergence based on GADF, and B: energy compensation based on cross OPM. This mechanism enhances the physical coherence of the input representation and improves the robustness of subsequent wear-state recognition.A minimalist recognition model (CSA-Kansformer) integrating physical prior attention and the KAN architecture is constructed. Aiming at the extremely non-linear degradation process of tool wear, this paper introduces the cutting-edge KAN architecture to replace the traditional fixed MLP module, relying on its

Read the complete text