PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey of Predictive Maintenance: Systems, Purposes and Approaches

Tianwen Zhu, Yongyi Ran, Xin Zhou, Yonggang Wen

arXiv:1912.07383v2eess.SPeess.SY

TL;DR

Unplanned downtime and the limitations of reactive and preventive maintenance motivate the search for more effective maintenance strategies. This paper surveys PdM architectures, optimization purposes, and machine-learning approaches, finding that PdM offers a trade-off between repair and prevention costs while identifying deep-learning research needs.

  • Problem

    Reactive and preventive maintenance can incur high repair, prevention, or unnecessary-maintenance costs, while Industry 4.0 requires improved maintenance performance.

  • Method

    The paper conducts a comprehensive survey of PdM system architectures, maintenance purposes, and traditional machine-learning and deep-learning approaches.

  • Results

    PdM can achieve the best trade-off between repair cost and prevention cost among RM, PM, and PdM strategies.

  • Takeaways & Limitations

    The survey provides a high-level framework for understanding PdM technologies, protocols, purposes, and learning-based approaches.

Abstract

from arXiv · show

This paper highlights the importance of maintenance techniques in the coming industrial revolution, reviews the evolution of maintenance techniques, and presents a comprehensive literature review on the latest advancement of maintenance techniques, i.e., Predictive Maintenance (PdM), with emphasis on system architectures, optimization objectives, and optimization methods. In industry, any outages and unplanned downtime of machines or systems would degrade or interrupt a company's core business, potentially resulting in significant penalties and immeasurable reputation and economic loss. Existing traditional maintenance approaches, such as Reactive Maintenance (RM) and Preventive Maintenance (PM), suffer from high prevent and repair costs, inadequate or inaccurate mathematical degradation processes, and manual feature extraction. The incoming fourth industrial revolution is also demanding for a new maintenance paradigm to reduce the maintenance cost and downtime, and increase system availability and reliability. Predictive Maintenance (PdM) is envisioned the solution. In this survey, we first provide a high-level view of the PdM system architectures including PdM 4.0, Open System Architecture for Condition Based Monitoring (OSA-CBM), and cloud-enhanced PdM system. Then, we review the specific optimization objectives, which mainly comprise cost minimization, availability/reliability maximization, and multi-objective optimization. Furthermore, we present the optimization methods to achieve the aforementioned objectives, which include traditional Machine Learning (ML) based and Deep Learning (DL) based approaches. Finally, we highlight the future research directions that are critical to promote the application of DL techniques in the context of PdM.

I. INTRODUCTION

The paper frames Predictive Maintenance (PdM) as an Industry 4.0 response to costly downtime and limitations of reactive and preventive maintenance. It surveys PdM architectures, optimization purposes, and learning-based approaches, emphasizing deep learning and future research needs.

  • I. INTRODUCTION: Unplanned downtime can interrupt business and cause substantial penalties, lost sales, and economic or reputational losses.The paper cites Amazon’s 49-minute outage costing $4 million and average data-centre downtime losses of $138,000 per hour.
  • I. INTRODUCTION: Maintenance strategies evolved from Reactive Maintenance (RM) and Preventive Maintenance (PM) toward Predictive Maintenance (PdM) using sensing, IoT, and artificial intelligence.RM responds after failure, while PM follows schedules that may trigger unnecessary maintenance; PdM uses equipment condition information.
  • I. INTRODUCTION: PdM combines condition monitoring, fault diagnosis, fault prognosis, and maintenance planning to detect faults, predict their progression, and support maintenance decisions.The surveyed enabling technologies include IoT data acquisition, big-data preprocessing, deep learning, and deep reinforcement learning.
  • I. INTRODUCTION: The survey organizes PdM around system architectures, optimization objectives, and optimization methods, including cost, reliability, traditional machine learning, and deep learning.It emphasizes architectures that integrate industrial standards and emerging technologies while supporting data collection, diagnosis, and prognosis.
  • I. INTRODUCTION: Existing surveys mainly review equipment-specific fault diagnosis or prognosis, whereas this paper provides a holistic PdM perspective and covers cost, availability/reliability, and multi-objective models.The paper also reviews newer deep-learning approaches and identifies future research directions for PdM.

A. RM

Maintenance strategies progress from failure-driven RM and schedule-driven PM toward condition-based PdM, which uses operating data to balance maintenance frequency and cost. The survey frames PdM through architectures, modules, and enabling technologies for intelligent maintenance.

  • A. RM: RM repairs equipment only after breakdown, maximizing utilization but potentially increasing repair costs and causing additional equipment damage.Run-to-failure avoids maintenance spending before failure, but repairs or replacement can exceed the production value gained.
  • B. PM: PM schedules maintenance during normal operation to reduce unexpected breakdowns, relying largely on elapsed time and assumed bathtub-curve failure behavior.Its process includes statistically investigating failure characteristics and selecting policies that balance reliability, availability, safety, and maintenance cost.
  • B. PM: PM may perform unnecessary maintenance or miss equipment-specific wear-out timing, while reactive repairs normally cost three times more than scheduled repairs.These limitations can create planned downtime, inventory-management demands, or catastrophic damage before maintenance occurs.
  • C. PdM: PdM uses actual operating-condition data and predictive algorithms to identify fault trends, estimate likely failure timing, and select maintenance actions.Its aim is a trade-off between maintenance frequency and cost rather than running equipment to failure or replacing useful components prematurely.
  • D. Summary and Comparison: Compared with RM and PM, PdM is described as offering the best trade-off between repair cost and prevention cost while keeping maintenance frequency low enough to prevent unplanned RM.The paper also presents PdM 4.0 as continuous monitoring with predictive alerts, although two thirds of respondents remain below maturity level 3 and only around 11% reach level 4.
  • III. System Architectures of PdM: The survey introduces PdM architectures and modules spanning data acquisition, preprocessing, analysis, and emerging technologies such as CPS, IoT, big data, and cloud computing.The proposed high-level view is intended to clarify the modules and techniques needed to implement intelligent PdM.

B. OSA-CBM

OSA-CBM provides a uniform, layered framework for designing and implementing PdM systems, organizing condition-monitoring functions from data acquisition through advisory generation.

  • OSA-CBM, defined in ISO 13374, standardizes formats and methods for communicating, presenting, and displaying condition-monitoring information and data.
  • OSA-CBM functional blocks: Data Acquisition collects sensor data, while Data Manipulation transforms signals and extracts specialized features.
  • OSA-CBM functional blocks: State Detection compares features with expected values or operational limits to return condition indicators or alarms.
  • OSA-CBM functional blocks: Health Assessment evaluates degradation using health trends, operational status, and maintenance history.
  • OSA-CBM functional blocks: Prognostics Assessment projects current system health into the future using estimated future usage profiles.
  • OSA-CBM functional blocks: Advisory Generation recommends maintenance activities or configuration changes using histories, mission profiles, and resource constraints.

C. Cloud-enhanced PdM System

Cloud-enhanced PdM uses remote sensing, processing, diagnosis, prognosis, and planning to deliver maintenance capabilities as adaptable services. Cloud computing supports these services by addressing equipment resource, security, and multi-source data challenges.

  • Cloud-enhanced PdM builds on cloud computing and cloud manufacturing, which deliver IT and manufacturing resources as services over IoT.
  • Cloud-enabled PdM architecture: Its architecture remotely collects machine-condition data, processes it, performs diagnosis and prognosis, and executes PdM planning dynamically on the shop floor.
  • Supporting technologies: Supporting technologies include IoT, embedded systems, semantic web technologies, and machine-to-machine communication.
  • Cloud computing can support smart PdM services while addressing equipment memory capacity, processor power, data security, and multi-source data fusion.
  • Cloud-enhanced PdM characteristics: Cloud-based PdM functions are service-oriented, allowing users to avoid hosting and maintaining large numbers of computing servers or related software.

D. Digital Twin Driven PdM Framework

Digital twin-driven PdM links a physical manufacturing system with a virtual model through data and information connections. The section reviews cost components and optimization considerations for maintenance strategies.

  • D. Digital Twin Driven PdM Framework: A digital twin is a high-fidelity virtual copy that simulates, mirrors, predicts, and improves the life of a physical factory or machine.Its development is supported by sensor, IoT, and computation technologies.
  • D. Digital Twin Driven PdM Framework: The digital twin reference model contains a physical system in Real Space, a digital model in Virtual Space, and connecting data and information.Physical attributes are collected through smart sensing, while the virtual model combines physics-based models and data-driven analytics.
  • Cost Minimization: For RM, repair occurs only after equipment breakdown; PM schedules sequential maintenance actions using preventive replacement, inspection, downtime, and corrective replacement costs.The cited PM cost items are preventive replacement cost (Cp), inspection cost (Ci), unit downtime cost (Cd), and corrective replacement cost (Cc).
  • Cost Minimization: PdM maintenance actions follow failure predictions, so its cost model is associated with remaining useful life and system-specific conditions.One reviewed model includes corrective maintenance, PdM, production-capacity loss, indirect economic loss, and product-quality loss costs.
  • Cost Minimization: The reviewed comprehensive cost model sums five components: c = c1 + c2 + c3 + c4 + c5.These components represent corrective maintenance, PdM, production-capacity loss, indirect economic loss, and product-quality loss.

B. Availability/Reliability Maximization

Availability and reliability provide practical criteria for evaluating PdM policies because uptime and downtime are often easier to measure than cost-model parameters. Reviewed studies optimize inspection or maintenance decisions using these measures.

  • B. Availability/Reliability Maximization: Availability/reliability is a practical PdM metric because system uptime and downtime are often more accurately measured and easier to obtain than cost-model parameters.Cost models vary across maintenance strategies, systems, and application scenarios.
  • Reliability: Reliability denotes the probability that a system or equipment remains functional throughout a specified time interval.With equipment lifetime Tf, reliability is represented as R(t) = P(Tf > t).
  • Reliability: For degradation-based equipment, the first passage time Tf is when degradation signal X(t) first reaches a prespecified threshold D.This defines failure through a threshold-crossing event.
  • Availability: Availability represents the probability that a system is operational, with definitions depending on which uptime and downtime quantities are included.Reviewed models use mean time between maintenances and mean maintenance time, and cover series, parallel, and series-parallel systems.
  • Availability: Reviewed CBM studies seek inspection policies that maximize system availability, including settings with imperfect maintenance or excessive degradation treated as unavailability.One study optimizes the inspection time for a continuous degradation state.

C. Multi-Objective Optimization

Multi-objective PdM optimization addresses cases where a single criterion cannot represent operator preferences across competing objectives. The reviewed formulations seek feasible trade-offs among objectives such as cost, reliability, and risk.

  • C. Multi-Objective Optimization: Single-objective optimization may not adequately represent operator preferences when PdM criteria such as cost, reliability, downtime, risk, safety, and feasibility compete.The section motivates multi-objective formulations for selecting more representative solutions.
  • C. Multi-Objective Optimization: A general multi-objective problem optimizes a set of objective functions over decision variables x in a feasible space X.The formulation includes k objectives, with fi(x) denoting the i-th objective function.
  • C. Multi-Objective Optimization: The weight-sum formulation combines objectives as f(x) = Σ_i=1^k wi fi(x), with nonnegative weights summing to one.The weights encode the relative importance assigned to each objective.
  • C. Multi-Objective Optimization: Because objectives can conflict, optimization typically seeks a good trade-off rather than simultaneously achieving every objective’s optimum.One reviewed CBM study optimizes a risk threshold using maintenance cost and reliability models, with failure probability as a constraint.
  • C. Multi-Objective Optimization: A reviewed bi-objective CBM model minimizes maintenance costs and maximizes ship reliability using NSGA-II.The study applies nondominated sorting genetic algorithm II to solve the proposed model.

V. TRADITIONAL MACHINE LEANING BASED APPROACHES

Data-driven PdM uses machine learning to extract knowledge and support decisions from increasingly large sensor- and IoT-generated datasets. Before deep learning, shallow ML algorithms were widely developed for PdM.

  • V. TRADITIONAL MACHINE LEANING BASED APPROACHES: Growing sensor, IoT, and big-data resources have increased the attractiveness of data-driven PdM.Machine learning is presented as a powerful solution for extracting useful knowledge and making decisions from these data.
  • V. TRADITIONAL MACHINE LEANING BASED APPROACHES: Traditional PdM approaches include shallow machine-learning algorithms such as artificial neural networks and decision trees.The passage introduces these methods as alternatives developed before moving toward deep learning.

A. Artificial Neural Network (ANN)

Traditional machine-learning methods support fault diagnosis and prognosis in PdM, but their characteristics and applicability vary, and ANN-based approaches often require engineered features.

  • ANNs have been applied to fault diagnosis and prognosis across machinery, components, and power systems.
  • ANN performance depends strongly on network architecture, including hidden neurons, connections, and activation functions.
  • ANN models have been used for bearing RUL prediction, including a model with three hidden layers containing (7-3-7) neurons.
  • ANNs usually require hand-crafted feature extraction or selection because shallow learning does not effectively extract informative features from raw sensor data.
  • Decision trees, SVMs, and k-NN are also used for PdM classification or prognosis, with differing learning strategies, advantages, limitations, and applications.

VI. DEEP LEARNING BASED APPROACHES

Deep learning approaches are widely applied in PdM because they learn representations through multilayer transformations, with autoencoders supporting feature learning, health assessment, fusion, diagnosis, and prognosis.

  • Autoencoders, CNNs, DBNs, and other deep models are widely applied to feature learning, fault classification, and fault prediction in PdM.
  • An autoencoder encodes input data into a compressed hidden representation and decodes it into a reconstruction close to the original input.
  • Autoencoder training minimizes reconstruction error, commonly measured by mean squared error over the dataset.
  • Autoencoder-based models can estimate degradation and distinguish fault severity when failure histories or labeled data are scarce.
  • Autoencoders support health-indicator construction, multi-sensor feature fusion, classifier integration, and RUL prediction through combinations with regression models.
  • One integrated deep denoising autoencoder method reported an error of about 20% between predicted and true RUL values.

B. Convolutional Neural Network (CNN)

CNNs extract local features hierarchically for PdM diagnosis and prognosis, including RUL prediction, while hybrid architectures share representations across fault diagnosis and RUL tasks.

  • A typical CNN combines input, convolution, pooling, and fully connected layers to transform local input features into representations for classification or regression.
  • Pooling reduces model parameters while retaining effective information, potentially limiting overfitting and improving training speed.
  • CNNs have been applied to raw vibration signals, time-frequency representations, infrared images, and multi-scale vibration components for fault diagnosis.
  • About 99% accuracy was reported for an end-to-end 1D-CNN fault classifier after hyperparameter tuning.
  • CNN-based models automatically learn features and construct health indicators or degradation representations for fault prognosis and RUL prediction.
  • The JL-CNN shares a CNN representation between fault diagnosis and RUL prediction while using independent fully connected networks for task-specific outputs.
  • 82.7% and 24.9% MSE decreases were reported relative to SVR and traditional CNN, respectively, for the proposed JL-CNN.

C. Recurrent Neural Network (RNN)

RNNs model sequential data by retaining prior inputs in hidden states, while LSTM and GRU address recurrent gradient problems and better capture long-term dependencies.

  • C. Recurrent Neural Network (RNN): RNNs update hidden outputs from current input and the previous hidden output, preserving sequential information in the network state.The final hidden output represents the whole input sequence.
  • C. Recurrent Neural Network (RNN): RNNs suffer from gradient vanishing or exploding during Back Propagation Through Time, motivating LSTM and GRU architectures.LSTM adds forget gates to improve long-term dependency modeling.
  • C. Recurrent Neural Network (RNN): RNN-based models have been applied to rotating-machinery fault diagnosis, including deep RNN and GRU-based multistage approaches.The reviewed Tennessee Eastman evaluation reports that LSTM better separates different faults.
  • C. Recurrent Neural Network (RNN): RNN variants, particularly LSTM and GRU, are also used for remaining useful life prediction from nonlinear degradation and selected health indicators.One bearing approach selects sensitive features using monotonicity and correlation metrics before constructing an RNN-based health indicator.

D. Deep Belief Networks (DBN)

DBNs learn hierarchical representations through stacked RBMs and support diagnosis, anomaly detection, and RUL estimation; related GAN and transfer-learning methods extend PdM under data constraints.

  • D. Deep Belief Networks (DBN): DBNs stack restricted Boltzmann machines and can be pretrained greedily without supervision before fine-tuning for likelihood or labeled classification.A softmax layer can be added for supervised fault classification.
  • D. Deep Belief Networks (DBN): DBNs extract high-level monitoring-signal features for fault classification, with one reviewed multisensor system achieving 98.8% accuracy under load changes.The approach fuses classifier outputs using DS evidence theory.
  • D. Deep Belief Networks (DBN): DBNs can replace manually extracted features and operate with or without an additional softmax classifier for fault identification.They have been applied to aircraft-engine, transformer, and other benchmark health-classification tasks.
  • D. Deep Belief Networks (DBN): DBN-based pipelines support RUL prediction and early anomaly detection by combining learned features with FNNs, RVMs, clustering, optimization, or Mahalanobis distance.Applications include bearings, wind turbines, and lithium-ion batteries.
  • D. Deep Belief Networks (DBN): GANs generate synthetic labeled or unlabeled sensor data for augmentation and class-imbalance mitigation, while GANomaly detects abnormal samples using anomaly scores.GAN-based prognosis can generate future bearing health-indicator trajectories for RUL estimation.
  • D. Deep Belief Networks (DBN): Transfer learning addresses scarce target-system failure data by adapting representations or parameters learned from another domain.A reviewed method achieves near-100% test accuracy across three mechanical datasets, including 99.64% on a gearbox dataset.

G. Deep Reinforcement Learning (DRL)

Deep reinforcement learning combines reinforcement learning with deep networks to support complex PdM decisions, diagnosis, prognosis, and signal or health-indicator optimization.

  • G. Deep Reinforcement Learning (DRL): DRL combines reinforcement learning and deep learning to solve complex decision-making tasks by evaluating action values in system states.Action values can represent the cost or profit of taking an action at a state.
  • G. Deep Reinforcement Learning (DRL): DRL has been applied to operation and maintenance management for power grids equipped with prognostics and health-management capabilities.The framework uses information gathered from prognostic health management.
  • G. Deep Reinforcement Learning (DRL): A DRL fault-diagnosis architecture maps raw fault data to fault modes by integrating a stacked autoencoder with a DQN agent.The stacked autoencoder senses fault information before the agent maps it to Q values.
  • G. Deep Reinforcement Learning (DRL): DRL can optimize signal processing by controlling a bandpass filter to select frequency bands with high signal-to-noise ratios before fault diagnosis.The reviewed method uses the reciprocal of a smoothness index as its control signal.
  • G. Deep Reinforcement Learning (DRL): Hybrid deep-learning architectures combine complementary capabilities, such as autoencoder feature extraction, LSTM sequence processing, and DRL control-policy learning.The survey notes that hybrid architectures can achieve better performance.

1) Auto-encoder & LSTM:

The survey reviews hybrid autoencoder–LSTM and related deep-learning approaches for PdM, then identifies guidance and research needs involving standards, data, interpretability, imbalance, and multicomponent systems.

  • 1) Auto-encoder & LSTM:: Autoencoders and LSTMs can be combined so autoencoders learn representations while LSTMs process temporal information for anomaly detection or RUL prediction.The section presents hybrid architectures as a way to combine complementary deep-learning capabilities.
  • 1) Auto-encoder & LSTM:: Hybrid deep models have been applied to denoise vibration signals and diagnose rotating-machinery faults, reaching 96.65% and 97.25% accuracy in bearing and gearbox experiments.The denoising autoencoder supplies cleaned signals to the convolutional diagnosis network.
  • 1) Auto-encoder & LSTM:: The survey provides a table comparing advantages, limitations, and typical applications to guide selection of deep-learning methods for specific PdM applications.This guidance follows the review of commonly used architectures and approaches.
  • VII. FUTURE RESEARCH DIRECTIONS: Future PdM research needs standards for emerging technologies, large shared datasets, visualization for interpretability, and solutions to class imbalance.The survey links deep-learning performance to dataset scale and quality.
  • VII. FUTURE RESEARCH DIRECTIONS: Existing deep-learning approaches often target individual components, leaving multicomponent systems and component dependencies as a stated research challenge.Increasing component count raises the complexity and difficulty of deep-learning PdM.
  • VIII. CONCLUSION: The paper concludes by synthesizing PdM architectures, maintenance purposes, traditional ML, DL approaches, and future directions for applying DL in PdM.Its stated purposes include cost minimization, availability or reliability maximization, and multiple objectives.
Loading 1912.07383v2…