Source-linked AI summary
Deep Learning and Its Applications to Machine Health Monitoring: A Survey
Rui Zhao, Ruqiang Yan, Zhenghua Chen, Kezhi Mao, Peng Wang, Robert X. Gao
TL;DR
Modern connected machinery produces large datasets for machine health monitoring, while conventional approaches depend heavily on manually designed features. This survey reviews deep-learning architectures and their applications to MHMS, concluding that DL-based methods support representation learning and prediction across varied monitoring tasks while remaining dependent on suitable data and domain knowledge.
Problem
Machine health monitoring must analyze increasingly large machinery datasets, but conventional data-driven systems rely on difficult-to-design handcrafted features and feature-selection procedures.
Method
The paper systematically surveys DL-based MHMS and organizes applications around auto-encoders, RBM variants, CNNs, and RNNs.
Results
The reviewed studies apply DL-based MHMS to varied monitoring tasks, with reported capabilities including representation learning for complex machinery data and target prediction.
Takeaways & Limitations
Deep learning serves as a bridge between big machinery data and data-driven machine health monitoring across multiple architecture families.
Takeaways & Limitations
DL-based MHMS performance depends heavily on dataset scale and quality, while model depth is constrained by available machinery data.
Abstract
from arXiv · showhide
Since 2006, deep learning (DL) has become a rapidly growing research direction, redefining state-of-the-art performances in a wide range of areas such as object recognition, image segmentation, speech recognition and machine translation. In modern manufacturing systems, data-driven machine health monitoring is gaining in popularity due to the widespread deployment of low-cost sensors and their connection to the Internet. Meanwhile, deep learning provides useful tools for processing and analyzing these big machinery data. The main purpose of this paper is to review and summarize the emerging research work of deep learning on machine health monitoring. After the brief introduction of deep learning techniques, the applications of deep learning in machine health monitoring systems are reviewed mainly from the following aspects: Auto-encoder (AE) and its variants, Restricted Boltzmann Machines and its variants including Deep Belief Network (DBN) and Deep Boltzmann Machines (DBM), Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN). Finally, some new trends of DL-based machine health monitoring methods are discussed.
I. INTRODUCTION
Industrial IoT and connected sensors are producing large machinery datasets for data-driven machine health monitoring. This survey examines how deep learning can process these data, reviews major DL architectures and MHMS applications, and discusses emerging trends.
- Industrial IoT enables connected machines to generate large datasets that can support fault diagnosis, prognosis, and remaining-useful-life prediction.
- Deep learning can bridge big machinery data and intelligent machine health monitoring by learning high-level representations.
- GPU hardware, improved software, faster networks, and larger datasets have reduced computational barriers and supported deep-learning adoption.A four-layer DBN with 100 million free parameters reportedly went from several weeks of training to around one day.
- Conventional data-driven MHMS relies on handcrafted features, feature extraction or selection, and shallow model training.Typical shallow models include SVM, Naive Bayes, and logistic regression.
- The survey reviews DL-based MHMS using auto-encoders, RBM variants, CNNs, and RNNs, aiming to provide researchers and engineers with an overview and practical entry point.
II. DEEP LEARNING
Deep learning is a fast-growing branch of machine learning that uses multiple nonlinear layers to learn hierarchical data representations. The section briefly introduces four architectures applied to machine health monitoring.
- Deep learning uses multiple nonlinear processing layers to learn hierarchical representations of data.
- Because the field is rapidly developing, the paper limits its introduction to major architectures used in machine health monitoring.
- The reviewed deep architectures are auto-encoders, restricted Boltzmann machines, convolutional neural networks, and recurrent neural networks.
A. Auto-encoders (AE) and its variants
Auto-encoders learn hidden representations by encoding inputs and decoding them toward the original data, with variants adding reconstruction, sparsity, or denoising objectives. Their unsupervised training also supports deep-network pretraining.
- An auto-encoder encodes input x into hidden representation h and decodes h back toward the original representation.
- Auto-encoder parameters are optimized to minimize reconstruction error over training samples.
- Sparse auto-encoders add a KL-divergence penalty to encourage sparse hidden representations and avoid identity transformations.The sparsity target is controlled by a predefined mean activation target p.
- Denoising auto-encoders corrupt their inputs and learn to reconstruct the clean data, commonly using dropout or binary masking noise.
- Stacked denoising auto-encoders provide unsupervised layer-wise pretraining before supervised fine-tuning of a deep neural network.A softmax or regression layer can map the final representation to prediction targets.
B. RBM and its variants
RBMs model visible and hidden units as a bipartite network, while DBNs and DBMs extend this structure through multiple hidden layers with different connectivity and training procedures.
- RBM: An RBM is a two-layer bipartite neural network with visible and hidden units, symmetric cross-group connections, and no within-group connections.
- RBM: The RBM energy function uses visible-hidden weights and visible and hidden bias terms to define the model.
- DBN: DBNs stack multiple RBMs, using each layer’s hidden output as the next layer’s input and training layers greedily without supervision.
- DBM: DBMs organize hidden units into multiple layers with full connectivity only between neighboring layers and no within-layer or non-neighboring connections.
- DBM: Unlike DBNs, DBMs are fully undirected and trained jointly, making DBM training more computationally expensive.
C. Convolutioanl Neural Network
CNNs learn abstract representations from sequential or spatially arranged data by alternating convolution and pooling operations, followed by prediction layers.
- Architecture: CNNs use spatially shared weights and spatial pooling, properties originally developed for image processing and later applied to sequential data.
- Architecture: Convolutional layers apply local filters to raw inputs, while pooling layers retain significant features over sliding windows and produce fixed-length representations.
- 1D-CNN: A 1D-CNN processes sequential inputs by applying filter vectors to concatenated windows and generating feature-map activations.
- 1D-CNN: Max-pooling reduces feature-map length and can further minimize the number of model parameters by taking maxima over consecutive values.
- Architecture: A one-layer CNN combines one convolutional layer, one pooling layer, one fully connected layer, and one softmax layer for prediction.
D. Recurrent Neural Network
RNNs represent sequential data by retaining information from previous inputs, while stacked and bidirectional structures increase capacity and temporal coverage.
- RNN fundamentals: RNNs connect units through directed cycles and can map the history of input sequences to target vectors using internal memory.
- RNN fundamentals: At each time step, an RNN updates its hidden state from the current input and previous hidden output, then maps the final representation to targets.
- Advanced RNNs: Vanilla RNNs may fail to capture long-term dependencies because backpropagated gradients can vanish, motivating LSTM and GRU architectures with gates.
- RNN variants: Stacked RNNs propagate one recurrent layer’s hidden output into the next layer, increasing model capacity and flexibility.
- RNN variants: Bidirectional RNNs process sequences forward and backward with separate hidden layers and concatenate the resulting representations.
III. APPLICATIONS OF DEEP LEARNING IN MACHINE HEALTH MONITORING
Deep learning methods are being applied to machine health monitoring to improve representation learning and reduce reliance on manually designed features and expert knowledge.
- Applications: Layer-by-layer pretraining with auto-encoders or RBMs can facilitate DNN training and improve its discriminative characterization of machinery data.
- Applications: CNNs and RNNs provide more advanced composition mechanisms for learning representations from machinery data.
- Target modeling: In DL-based monitoring systems, softmax layers usually predict discrete diagnostic targets, whereas linear regression layers predict continuous prognostic targets.
- Applications: End-to-end DL-based monitoring structures can be constructed with less human labor and expert knowledge.
A. AE and its variants for machine health monitoring
AE-based models learn representations from machinery data, using raw or engineered features and several variants to address dimensionality, overfitting, multisensory inputs, and classification. A two-layer SAE-based DNN also supports visualization of learned representations, whose discriminative power improves with additional layers.
- AE-based representation learning: AE models learn high-level representations from machinery data automatically, while stacked auto-encoders are widely used in deeper architectures.One-layer AE models have also been used when training data are limited, with dropout applied to reduce overfitting.
- Input representations and dimensionality: Raw sensory time-series inputs can exceed one thousand dimensions, creating computation-cost and overfitting concerns from large parameter counts.Some studies therefore construct AE models on features extracted from raw inputs.
- Input representations and dimensionality: SAE studies use frequency spectra, multisensory synchronized windows, and handcrafted FFT/WPT features as inputs for rotating-machinery diagnosis, anomaly detection, and fault classification.Frequency spectra may expose discriminative distributions of constitutive components across discrete frequencies.
- AE variants: AE variants include a two-phase SAE framework with separate representation learning and classification, continuous sparse auto-encoders, and stochastic-unit or extreme-learning-machine-based encoders.The stochastic unit is described as changing gradient direction to prevent overfitting, while the extreme-learning-machine variant learns its decoder through a single least-squares step.
- AE variants: Visualization of a two-layer SAE-based DNN provides a way to evaluate learned representations, whose discriminative power improves as the number of layers increases.The visualization offers an additional perspective for assessing DL-based machine health monitoring.
B. RBM and its variants for machine health monitoring
RBM-based methods and their deeper variants learn representations for machinery prognosis, degradation assessment, and fault recognition. Reviewed systems combine unsupervised pretraining with regression, softmax classification, feature fusion, or multimodal inputs.
- Deep variants: Most reviewed RBM-related applications use deep belief networks to pretrain deep neural networks for machine health monitoring.The section also includes work developing RBM and DBM representations from machinery data.
- RBM-based methods: RBM-based methods support bearing RUL prediction by adding a linear regression layer after pretraining and using predicted RMS trajectories to calculate RUL.A separate RBM formulation adds a trendability regularization term to the hidden-node training objective for machine RUL prediction.
- Deep variants: Some systems combine multiple modalities or multiple DBN outputs, including time, frequency, and time-frequency features and fused probabilistic softmax outputs.One approach fuses outputs from three GDBMs through a support vector classification framework.
- Deep variants: A DBN-based framework assesses bearing degradation using RMS features fitted by a Weibull distribution together with frequency-domain features.The framework is presented as an application under a bearing accelerated life test.
- Deep variants: DBN systems use direct vibration signals or multiple feature sets, including raw vibration, MFCC, and wavelet features, for induction-motor diagnosis and cutting-state monitoring.The reviewed cutting-state work reports comparative performance across the corresponding inputs.
C. CNN for machine health monitoring
CNNs are applied to machinery data represented as 2D time-frequency images, sensor-indexed matrices, or 1D time series. Reviewed architectures combine convolutional feature extraction with pooling, fully connected, softmax, regression, or multiscale components for diagnosis and prognosis.
- CNN representations: CNN filters extract local patterns from 1D or 2D machinery data, while stacked convolutional layers build more complex representations.Machinery inputs may be time-frequency spectra or time-series data arranged in one-dimensional form.
- 2D CNN applications: A 2D CNN can use the number of sensors as the input height, with convolutional and fully connected layers followed by a softmax classifier.A related RUL model uses normalized-variate sensor time series, average pooling, and a linear regression output.
- 2D CNN applications: Other 2D CNN systems reshape statistic-feature vectors into matrices or transform signal vectors into 32 × 32 matrices before convolutional processing.These designs adapt non-image machinery features to conventional CNN input formats.
- 1D CNN applications: A 1D CNN integrates feature extraction and classification directly on raw time-series data for real-time motor fault detection.The reviewed section also describes 1D CNN use with normalized vibration signals for damage detection and localization.
- Architecture summary: CNN architectures in the reviewed MHMS studies are summarized by configurations specifying input dimensions, convolutional layers, pooling layers, and fully connected layers.The configuration notation records filter counts, filter sizes, and layer repetitions.
D. RNN for machine health monitoring
RNNs, including LSTM and GRU architectures, address the temporal structure of machinery sensor data. Reviewed methods use recurrent encoders and decoders for sequence reconstruction and health-index-based RUL estimation, while CNN figures and table material provide adjacent MHMS context.
- Sequential modeling: RNNs, LSTMs, and GRUs are used for machinery sensor sequences because they encode temporal information.LSTM and GRU architectures are presented as alternatives that help address training difficulties in vanilla RNNs.
- Sequential modeling: An LSTM encoder-decoder transforms a multivariate input sequence into a fixed-length vector and then produces a target sequence.The encoder and decoder together form the reviewed sequence-modeling structure.
- RUL estimation: For RUL estimation, one reviewed approach trains on raw signals corresponding to normal behavior, uses reconstruction error to compute a health index, and estimates RUL from that index.The stated assumption is that larger reconstruction error corresponds to a more unhealthy machine condition.
IV. SUMMARY AND FUTURE DIRECTIONS
The survey synthesizes DL-based machine health monitoring across four architecture categories and identifies future directions involving datasets, domain knowledge, visualization, and difficult class distributions.
- Systematic overview: The survey organizes DL-based machine health monitoring research into Auto-encoders, Restricted Boltzmann Machines, Convolutional Neural Networks, and Recurrent Neural Networks.It presents these categories as a systematic overview of state-of-the-art work.
- Systematic overview: DL-based machine health monitoring can use end-to-end structures to map raw machinery data to targets with less extensive human labor and expert knowledge.The survey states that this general approach is not restricted to specific kinds of machines.
- Future directions: The performance of DL-based machine health monitoring depends heavily on dataset scale and quality, motivating large-scale open machinery datasets.The survey contrasts relatively shallow machinery models with deeper models supported by large datasets such as ImageNet.
- Future directions: Domain knowledge can improve DL-based machine health monitoring by reducing model size through discriminative features and boosting performance through task-specific regularization.The survey cautions that deep learning is not a universal solution for every machine health monitoring problem.
- Future directions: Visualization of learned representations and model behavior can provide insight into otherwise unexplainable deep neural network computations.The survey mentions t-SNE and layer-activation visualization as examples.
- Future directions: Research has also explored imbalanced-class problems using class-resampling or cost-sensitive CNN training and combinations of bootstrapping with CNNs.These methods are presented as recent developments in DL-based machine health monitoring.