An end-to-end machine learning approach with explanation for time series with varying lengths

Aug 5, 2026

Upload by Aurelia

An end-to-end learning approach addresses accurate prediction of complex product quality parameters from industrial process time series, where batch operations produce variable-length sequences. A 1D CNN with a masking layer is proposed to handle varying lengths, alongside a 1D CNN plus class activation mapping (CAM) strategy for result interpretation and highlighting regions of interest. The approach is compared with an unsupervised 1NN method using dynamic time warping (DTW), including FastDTW and DTAIDistance, under balanced/unbalanced classes and scaled/unscaled inputs. Sensor data from plastic-part production is used to predict hard-to-measure quality parameters, achieving 83.7% accuracy and reducing CNN training time versus DTW.

LanguageEnglish

CategoryTechnology

FormatPDF

Pages18

Reading Time45 min

AccessFree to download

Abstract

Problem and motivation

Proposed method: 1D CNN with masking and CAM

Comparative baselines: 1NN + DTW

Dataset and application in plastic-part production

Results and accuracy improvements

Neural Computing and Applications (2024) 36:7491–7508 [https://doi.org/10.1007/s00521-024-09473-9](https://doi.org/10.1007/s00521-024-09473-9) An end-to-end machine learning approach with explanation for time series with varying lengths Manuel Schneider1 • Norbert Greifzu1,2 • Lei Wang3 • Christian Walther4 • Andreas Wenzel1,2 • Pu Li5 Received: 24 June 2023 /Accepted: 14 January 2024/Published online: 19 February 2024 􀀂 The Author(s) 2024 Abstract An accurate prediction of complex product quality parameters from process time series by an end-to-end learning approach remains a signiﬁcant challenge in machine learning. A special difﬁculty is the application of industrial batch process data because many batch processes generate variable length time series. In the industrial application of such methods, explainability is often desired. In this study, a 1D convolutional neural network (CNN) algorithm with a masking layer is proposed to solve the problem for time series of variable length. In addition, a novel combination of 1D CNN and class activation mapping (CAM) technique is part of this study to better understand the model results and highlight some regions of interest in the time series. As a comparative state-of-the-art unsupervised machine learning method, the One-Nearest Neighbours (1NN) algorithm combined with dynamic time warping (DTW) was used. Both methods are investigated as end-to-end learning methods with balanced and unbalanced class distributions and with scaled and unscaled input data, respectively. The FastDTW and DTAIDistance algorithms were investigated for the DTW calculation. The data set is made up of sensor signals that was collected during the production of plastic parts. The objective was to predict a quality parameter of plastic parts during production. For this research, the quality parameter will be a difﬁcult or only destructively measurable parameter and both methods will be investigated for their applicability to this prediction task. The application of the proposed approach to an industrial facility for producing plastic products shows a prediction accuracy of 83.7% . It can improve the reverence method by approximately 1.4% . In addition to the slight increase in accuracy, the CNN training time was signiﬁcantly reduced compared to the DTW calculation. Keywords Time series classiﬁcation (TSC) varying length 􀀂 CNN with masking-layer 􀀂 CAM for TSC 􀀂 Quality prediction 􀀂 Injection moulding & Manuel Schneider m.schneider@hs-sm.de & Norbert Greifzu [norbert.greifzu@iosb-ast.fraunhofer.de](norbert.greifzu@iosb-ast.fraunhofer.de) 1 Faculty Electrical Engineering, Embedded Diagnostics Systems, Schmalkalden University of applied sciences, Blechhammer 6, 98574 Schmalkalden, Thuringia, Germany 2 Fraunhofer Institute of Optronics, System Technologies and Image Exploitation, IOSB-AST Ilmenau, Fraunhofer IOSB, Am Vogelherd 90, 98693 Ilmenau, Germany 3 Department Intelligent Energy Systems, Fraunhofer Institute for Integrated Systems and Device Technology IISB, Fraunhofer IISB, Schottkystraße 10, 91058 Erlangen, Bavaria, Germany 4 Bauhaus-Universit¨at Weimar, Institute for Structural Mechanics, Marienstraße 15, 99423 Weimar, Thuringia, Germany 5 Department of Computer Science and Automation, Technische Universit¨at Ilmenau, Ehrenbergstraße 29, 98684 Ilmenau, Thuringia, Germany 1 Introduction Artiﬁcial intelligence (AI) has been widely applied in the industry over the last two decades [1, 2] . Machine learning (ML), including deep learning (DL), as a subﬁeld of AI, is becoming increasingly important for research in image and signal processing, as well as for industrial applications [3, 4] . In [5], the focus was on the detection of defects in plastic parts with a convolutional neural network (CNN) through image analysis in combination with edge computing and Internet of Things (IoT) systems. Some challenges of quality prediction in the context of big data in the ﬁeld of Indu

FAQs

Why is variable-length time series a challenge in industrial batch processes?

Many batch processes generate time series with different lengths across operations, which complicates modeling and prediction when input lengths vary.

What does the proposed method use to handle varying time-series lengths and explain predictions?

It uses a 1D CNN with a masking layer for variable length handling and combines 1D CNN with class activation mapping (CAM) to interpret model results and focus on regions of interest.

How does the study evaluate the proposed approach against other methods?

It compares the end-to-end CNN-based method with an unsupervised 1NN approach using DTW, testing both balanced and unbalanced class distributions and scaled and unscaled inputs, including FastDTW and DTAIDistance for DTW computation.

Aug 5, 2026

Upload by Aurelia