Source-linked AI summary
A General End-to-end Diagnosis Framework for Manufacturing Systems
Ye Yuan, Guijun Ma, Cheng Cheng, Beitong Zhou, Huan Zhao, Hai-Tao Zhang, Han Ding
TL;DR
Manufacturing diagnosis and monitoring need a general framework that works across applications. This paper proposes an end-to-end CNN-based framework, which achieves the highest reported accuracy on both benchmark evaluations and supports applicability across manufacturing contexts.
Problem
Manufacturing diagnosis and monitoring lack a general framework that delivers satisfactory performance across applications.
Method
The proposed approach uses an end-to-end CNN to extract features for manufacturing-system monitoring.
Results
The framework achieves the highest reported accuracy for both benchmark evaluations.
Takeaways & Limitations
The results support applying the framework across benchmark manufacturing applications.
Takeaways & Limitations
The method requires tuning several hyperparameters to achieve its best performance.
Abstract
from arXiv · showhide
The manufacturing sector is envisioned to be heavily influenced by artificial intelligence-based technologies with the extraordinary increases in computational power and data volumes. A central challenge in manufacturing sector lies in the requirement of a general framework to ensure satisfied diagnosis and monitoring performances in different manufacturing applications. Here we propose a general data-driven, end-to-end framework for the monitoring of manufacturing systems. This framework, derived from deep learning techniques, evaluates fused sensory measurements to detect and even predict faults and wearing conditions. This work exploits the predictive power of deep learning to automatically extract hidden degradation features from noisy, time-course data. We have experimented the proposed framework on ten representative datasets drawn from a wide variety of manufacturing applications. Results reveal that the framework performs well in examined benchmark applications and can be applied in diverse contexts, indicating its potential use as a critical corner stone in smart manufacturing.
Introduction
The paper addresses the lack of generalizable manufacturing diagnosis methods by proposing a unified end-to-end CNN framework that automatically extracts features from measurement data. Across diverse manufacturing applications, the framework achieves high classification accuracy, small monitoring errors, and robustness to noise.
- Problem and contribution: Existing model-based and data-driven diagnosis approaches are highly application-specific and require substantial effort to generalize across manufacturing applications.The introduction identifies a need for convenient feature extraction combined with universality across diverse manufacturing contexts.
- Problem and contribution: The paper proposes a unified supervised-learning framework using a CNN to automatically extract features and solve diverse manufacturing monitoring problems.The framework is designed to reduce dependence on application-specific feature engineering and prior knowledge.
- Evaluation and results: Higher than 95% accuracies are achieved using the unified CNN framework for manufacturing diagnosis problems, while condition-monitoring tasks achieve small monitoring errors.The framework is evaluated on ten measurement datasets spanning different manufacturing problems.
- Evaluation and results: The framework achieves 100% accuracy for cooler and valve classifications, with 98.19% for pump leakage and 99.10% for hydraulic accumulator classifications.These results demonstrate performance across multiple hydraulic-system fault conditions.
- Robustness and interpretation: The proposed CNN framework remains robust under additive noise, retaining high classification accuracies within a certain noise range.The study also investigates how convolutional features capture useful signal information while ignoring redundant Gaussian noise.
Discussions
The proposed end-to-end framework was effective across diverse manufacturing diagnostic tasks, achieving the highest reported accuracy on benchmark and proprietary datasets. Its CNN interpretation capability and cross-industry applicability are promising, although hyperparameter tuning and data requirements remain limitations.
- The unified deep learning framework was tested on numerous critical diagnostic tasks across varied manufacturing applications.
- The end-to-end framework achieves the highest reported accuracy on both benchmark datasets and the authors’ own datasets.
- CNN fault-prediction interpretations provide information for understanding diagnosis decisions on manufacturing data with different frequencies, amplitudes, and phases.
- The framework could be applied to datasets from other industry applications.
- The method requires hyperparameter tuning and large training datasets, which may be infeasible for some applications.
Materials and Methods
The study uses diverse datasets spanning public, competition, laboratory, and industrial production sources. These datasets contain multiple sensor modalities and are processed for classification or regression tasks.
- Datasets: Datasets come from open-access, competition, laboratory experimental, and permission-based real production sources.The sources cover both publicly available and industrially provided data.
- Data processing: Measurements include current, force, vibration, and acoustic-emission signals, individually or in combination, for classification or regression.The signal combinations are processed according to the target task.
Main idea
The paper proposes an end-to-end deep-learning framework that converts manufacturing problems into supervised classification and regression tasks. A CNN automatically learns hidden features from raw sensory measurements for fault diagnosis and prediction, supporting closed-loop machine decision-making.
- Problem formulation: Practical manufacturing problems are formulated as supervised classification and regression tasks solved with deep learning.The framework applies deep-learning techniques to both classification and regression settings.
- Framework: The end-to-end algorithm automatically discovers hidden features needed for learning and prediction without prior knowledge.This removes the need to specify the relevant features in advance.
- Closed-loop operation: The fully automated closed-loop system feeds sensory measurements into the CNN and returns learned results to the machine for decision-making.A possible decision is whether to initiate a maintenance action.
Pre-processing
Measurements were normalized using dataset-specific procedures. Small time-course datasets were divided into constant-length samples without affecting periodicity, while prediction-task data underwent standardization.
- Pre-processing: Small time-course datasets, such as CWRU bearing data, were divided into constant-length samples without affecting periodicity.This preprocessing divided the total features within each sample.
- Pre-processing: Prediction-task datasets, such as Case 8, were transformed using standardization.The standardization procedure was specified in the Supplementary Information.
Parameter-tuning
The framework fine-tunes CNNs for classification and regression using task-specific output dimensions, activations, losses, and standard optimizers. Its convolutional design uses weight sharing to reduce parameters and improve computational efficiency.
- Convolutional neural networks: The CNN stacks input data, convolutional and pooling layers, flattening, fully connected layers, and an output layer.
- Parameter-tuning: Classification uses N output nodes, softmax activation, and cross-entropy loss, whereas regression uses one output, sigmoid activation, and least-squares loss.
- Convolutional neural networks: Weight sharing across convolutional filters greatly reduces training parameters and model complexity, improving computational efficiency.
- Parameter-tuning: Standard optimizers such as SGD or Adam update CNN weights during back-propagation until the predefined maximum iteration to reduce loss.
Interpretation of CNN model for manufacturing data
The CNN’s learned features were interpreted using synthetic sinusoidal time-series signals in binary classification experiments. The experiments achieved 100% accuracy, while robustness was assessed by adding noise at varying power levels and examining accuracy changes.
- Interpretation of CNN model for manufacturing data: The synthetic manufacturing signal was modeled as harmonically related sinusoids with Gaussian noise, parameterized by frequencies, amplitudes, phases, and noise.These coefficients were varied individually to interpret the convolutional layers’ extracted features.
- Interpretation of CNN model for manufacturing data: 100% accuracy was obtained in binary classification experiments that varied one signal coefficient while keeping the other three unchanged.The experiments used duplicated class A and class B signals with random training and validation samples.
- Interpretation of CNN model for manufacturing data: Extracted convolutional features were visualized in the frequency domain or polar coordinates for the different signal coefficients.Class A and class B signals were processed through the convolutional operation before visualization.
- Interpretation of CNN model for manufacturing data: Robustness was evaluated by adding noise whose power was expressed as a percentage of each original signal’s power and plotting accuracy against noise power.Similar cross-validation approaches were used for each case.