Abstract:The industrial fault prediction based on time series data inference is a very meaningful work. Generally, this task has two key difficulties including the limited amount of fault data and learning of timeline-based knowledge. In this paper, we propose a novel dual enhancement method to conduct the time series fault prediction task. It consists of a time series sample augmentation (TSSA) strategy and a timing integration attention enhancement model, named time series enhanced transformer (TSET). Specifically, we decompose the original signal by variational mode decomposition and then add the noise signal for recombination, so as to augment the time series samples. Furthermore, to solve the problem that the traditional transformer is difficult to adequately extract the features from time series data, we propose a new timing association attention module to enhance the time series feature learning. It obtains the comprehensive temporal related features formed by a time integration function and then inputs them to the different channels of the attention mechanism in the transformer architecture. A TSET is constructed by the dual enhancement. We validate the proposed method on both public and industrial datasets. The experimental results show that our method can outperform the mainstream baseline methods. Note to practitioners—The model in this paper is applied to fault prediction in industrial automation processes. The log and fault record data generated in industrial processes are natural time series data. Time series prediction models generated in the field of Natural language processing need a lot of corpus data to generate reasonable inference results. However, the fault data of industrial processes usually cannot reach an order of magnitude sufficient for these models to operate well. On the other hand, these corpus based data models are not friendly to industrial fault data. The model in this paper is established from two aspects: data enhancement and temporal correlation enhancement, so that the industrial fault sample data can meet the needs of the prediction model as much as possible. By enhancing temporal correlation features, the prediction model becomes friendly to industrial temporal fault data. The method is validated through experiments on actual industrial process data and public datasets. In practical applications, inputting data into the model proposed in the paper generates prediction results based on time series, and distinguishes whether the predicted data is a fault, thereby providing fault warning, which is helpful for the automation of the entire industrial life cycle.

Time Series Prediction Method of Industrial Process With Limited Data Based on Transfer Learning

Multi Time Scale Inception-Time Network for Soft Sensor of Blast Furnace Ironmaking Process

Short-term Traffic Prediction with Deep Neural Networks and Adaptive Transfer Learning

Transfer Learning Based on Incorporating Source Knowledge Using Gaussian Process Models for Quick Modeling of Dynamic Target Processes

Energy Predictive Models with Limited Data using Transfer Learning

A Review of Transfer Learning Approaches for Load, Solar and Wind Power Predictions

Subspace Metric-Based Transfer Learning for Spindle Thermal Error Prediction under Time-Varying Conditions

A novel transfer learning strategy for wind power prediction based on TimesNet-GRU architecture

Mat-Transformer-Based State Prediction Method for Information Equipment

Transfer Learning in Deep Learning Models for Building Load Forecasting: Case of Limited Data

A general transfer learning-based framework for thermal load prediction in regional energy system

A Novel Time Diffusion Model for Industrial Time Series Data Generation

A Comprehensive Survey of Deep Transfer Learning for Anomaly Detection in Industrial Time Series: Methods, Applications, and Directions

Industrial Time Series Prediction Based on Incremental DBSCAN-KNN with Self-learning Scheme

Prediction of photovoltaic power generation based on LSTM and transfer learning digital twin

Pre-training enhanced unsupervised contrastive domain adaptation for industrial equipment remaining useful life prediction

Time Series Fault Prediction Via Dual Enhancement

An Integrated Transfer Learning Method for Power Generation Prediction of Run-off Small Hydropower in Data-scarce Areas

MTTLA‐DLW: Multi‐task TCN‐Bi‐LSTM transfer learning approach with dynamic loss weights based on feature correlations of the training samples for short‐term wind power prediction

Time Prediction in Ship Block Manufacturing Based on Transfer Learning

Transfer Learning in the Transformer Model for Thermal Comfort Prediction: A Case of Limited Data