Elsevier

Computers in Industry

Volume 105, February 2019, Pages 164-176
Computers in Industry

Dynamic condition monitoring for 3D printers by using error fusion of multiple sparse auto-encoders

https://doi.org/10.1016/j.compind.2018.12.004Get rights and content

Highlights

  • Consider the healthy condition for 3D printers.

  • An error fusion of multiple sparse auto-encoders (EFMSAE) is proposed.

  • An attitude sensor which contained 9 channels is employed for collecting printer condition data.

  • The efficiency and effectiveness of EFMSAE is demonstrated by simulated and delta 3D printer experiments.

Abstract

With the development of 3D printers, the healthy condition is becoming more and more crucial for the printing quality. In this research, anerror fusion of multiple sparse auto-encoders (EFMSAE) is developed to monitor the condition of the 3D printer dynamically. To this end, an attitude sensor which contained 9 channels is employed for collectingprinter condition data of 3-axial angular velocity, 3-axial vibratory acceleration and 3-axial magnetic field intensity, simultaneously. To make use of the information of multiple sensorsmounted on the moving platform of the printer,multiple sparse auto-encoders (SAEs)are employed for the deep learning of these data. To integrate these information for extracting ingredients incondition monitoring, square prediction error (SPE)is applied as an error fusion tool to fuse multiple SAEs. The value of SPEis used as an indicator to describe the operation status of the printer. Both simulated and delta 3D printer experiments were carried for evaluating the performance of the addressed method. The results show that the present EFMSAE is capable of effectively monitoringthe dynamic healthycondition for 3D printers.

Introduction

3D printersarecriticalcomponentsin state-of-the-art manufacturing processes and have been widely deployed in modern industrial equipment. The 3D printing technique is based on the idea of material deposition. Its curing layer by layer can be implemented through different ways and materials [1]. However, such machines frequently operate in a harsh environment. As a result, a fault is likely to occur, which directly affects the precision of the 3D printing and even leads to catastrophic effects. Therefore, it is necessary to monitor the transmission condition of the 3D printer even if it has precision components [2]. Delta 3D printer is one of the most common and basic parallel mechanism, which is employed as the research object of the condition monitoring in this work. It should be pointed out that, delta 3D printers, even of well-designed, are more likely to faulty than serial mechanism ones due to more complicated transmission structures [3]. For this reason, condition monitoring and fault diagnosis of the weak fault are thus vital for minimizing the unscheduled process interruptions and avoiding economic loss for 3D printers. Nuchitprasitchai et al. [4,5] monitored the 3D printing platform by using a low-cost reliable real-time optical approach. Tlegenov et al. [6,7] proposed a physics-based dynamic modelfor monitoring nozzle clogging in fused deposition modelling machines and proposed a nozzle condition monitoring technique in fused filament fabrication 3D printing using a vibration sensor. Baumann et al. [8,9] designed a sensor array suitable for 3D printers and provided an approach to detect failures automatically. Wu et al. [10] used acoustic emission and support vector machines for in situ monitoring of FDM machine conditions. Holzmond et al. [11] presented a “certify-as-you-build” quality assurance system for in situ real time defect detection of 3D printed parts. Faes et al. [12] integrated a modular 2D laser triangulation scanner into an E3DP machine and analyzed feedback signals to monitor and control process state-variables online.Rao et al. [13,14] used advanced Bayesian non-parametric analysis of in situ heterogeneous sensor data to identify failure modes and detect the onset of process anomalies in additive manufacturing processes.He et al. [15] proposed an attitude monitoring approach to diagnose the fault of the delta 3D printer with support vector machines. Differing from the literature [15], in this work a cheaper attitude sensor (about 20$) is used for dynamic condition monitoring.

Mechanical device fault identification and classification methods can be divided into three major categories: model-based [16,17], signal processing-based [[18], [19], [20]] and knowledge processing-based ones [21,22]. Among them, the first 2 largely rely on experiences. In comparison, the knowledge processing-based method belongs to the concept of intelligent diagnosis technology that implements knowledge guidance from fault information detection to pattern extraction and from state identification to fault analysis and intervention decisions. This means that the diagnostic technology becomes a practical tool for normal users, not only for a few professionals. Conventional intelligent diagnosis and monitoring methods, such as principal component analysis and support vector machine, have been applied in mechanical device state identification and have achieved desirable results [[23], [24], [25]]. However, algorithm analysis shows that these intelligent processing technologies are “shallow mode” learning methods, which has the following major deficiencies: (1) the actual situation of the input information cannot be described completely to reveal the intrinsic rule of a complex input, and (2) these algorithms are designed for a specific data structure after feature calculation; and (3) the time domain, frequency domain or time-frequency domain features of a signal should be calculated in advance based on the experience to clearly distinguish different states.

These limitations restricted the application of these conventional algorithms and resulted in inferior generalization performance. To address this problem, in 2006, Hinton proposed a deep learning algorithm [26,27] that could effectively reveal an intrinsic pattern of complex inputs. This development attracted both academia and industry to deep learning algorithms. Bengio et al. [28,29] proposed a restricted Boltzmann machine and a greedy layer-wise deep network for machine vision and image processing.A review of the existing literature shows that some researchers have applied deep learning in mechanical fault diagnosis. For example, Tamilselvan and Tran et al. [30,31] integrated a deep belief network for the classification of faults in an aero-engine and a power transformer, piston compressor valve, respectively. He et al. [32] proposed large memory storage retrieval neural networks to diagnose the bearing faults. Qi et al. [33] built a stacked sparse auto-encoder for rotating machinery fault diagnosis. Ding et al. [34] employed Deep ConvNet in intelligent fault diagnosis of spindle bearings. Fu et al. [35] used a deep belief network in cutting equipment state monitoring. Hu et al. [36] utilized a noise reduction sparse auto-encoder to forecast fan rotation speed. Gan et al. [37] constructed a layered fault diagnosis model based on a deep belief network for bearing state identification. Wang et al. [38] applied deep belief network in fault detection for solar energy battery blades and noise of a vehicle suspension shock absorber. In order to prevent the overfitting of the deep neural network mode, Long et al. [39] applied the dropout algorithm for wind turbine gear-box failure identification. Sun et al. [40] presented a convolutional discriminative feature learning method for induction motor fault diagnosis. Wang et al. [41] employed deep belief networks to detect multiple faults in axial piston pumps.Lei et al. [42,43] employed a deep learning algorithm to monitor the health state of a planetary gear box. Shao et al. [44] proposed a continuous deep belief network with locally linear embedding method for rolling bearing fault detection. Ren [45] presented an integrated deep learning approach for multi-bearing remaining useful life collaborative prediction. Hu et al. [46] combined the deep Boltzmann machine and multi-grained scanning forest ensemble to deal with industrial fault diagnosis.Li et al. [47,48] presented a multimodal deep support vector classification (MDSVC) method and employed deep random forest for gearbox fault diagnosis. Liu et al. [49] proposed dislocated time series convolutional neural network for machine fault diagnosis. Li et al. [50] proposed the SAE-DBN method for bearing fault diagnosis using multi-sensor signals.Wang et al. [51] used deep learning for gearbox fault classification. Shao et al. [52] improved a deep belief networks (DBN) is method to learn features from sensor data of induction motor.Ma et al. [53] proposed a deep coupling autoencoder (DCAE) model for fault diagnosis with multimodal sensory data.

All these techniques have made great contribution to the fault diagnosis and condition monitoring fields, but they may not be completely applicable to 3D printers. Hence, this paper suggests an attitude monitoring incorporating machine leaning algorithm for the dynamic condition monitoring of the delta 3D printer.To this end, an error fusion of multiple sparse auto-encoders (EFMSAE) is developed. Amultiple SAEs model is proposed for a multi-sensor system.Squared prediction error (SPE) is adopted as an error fusion tool to evaluate the printing degradation performance.

The rest of the paper is structured as follows. The EFMSAE is developed in Section 2. In Section 3 and Section 4, a simulation experiment and a delta 3D printer experiment are carried out, respectively. Peer methods are also compared in Section 4. Finally, Section 5 concludes the whole works.

Section snippets

Condition monitoring by attitude sensor data

As shown in Fig. 1(a), the attitude sensor employs three-axis gyros assisted by three-axis accelerometers and three-axis magnetometers. Hence, with an attitude sensor based on MEMS, one can obtain the raw data of three-axial gyros, three-axial accelerometers and three-axial magnetometers [54]. This means that there are nine channels in the attitude monitoring to be selected in this paper.

It should be noticed that the attitude sensor is fixed on top of the moving platform as is shown in Fig. 1

Evaluation using simulated signals

To evaluate the addressed EFMSAE method, simulated signals are generated for the simulation experiment. To describe the waveform generated by a printer bearing under the constant radial load with a single localized defect, the vibration signature can be expressed as [56]y(t)=yd(t)yq(t)yr(t)ye(t)+yn(t)where yd is a series of impulses at bearing fault frequency, yq is the bearing radial load distribution, yr represents the bearing induced resonant frequency, ye reflects the exponential decay due

Real printer experiments and discussion

The effectiveness of the proposed method for dynamic condition monitoring is demonstrated using real data acquired from a delta 3D printer (SLD-BL600-6), as shown in Fig. 8. The experimental test rigcontains a delta 3D printer, an attitude sensor (BWT901) and a laptop (DELL Inspiron N4110). A cylindrical shell model (75 mmradius×0.3mmheight)was first established in a 3D modeling software. The G-codes were then generated by importing the 3D model into the slicing software and the trail was set

Conclusions

The healthy condition of the 3D printer is becoming more and more important for the printing quality. In this paper, a dynamic condition monitoring method based on EFMSAE has been proposed to accurately assess the delta 3D printer states. To reduce the sensor cost, a cheap attitude sensor, which is just about 20$ and contains 9 channels, was used in the experiments. By collecting the monitored data in different condition patterns, the attitude samples were applied for EFMSAE modelling in

Acknowledgements

This work is supported in part by the National Natural Science Foundation of China (51605406, 51775112, 71801044), the Research Program of Higher Education of Guangdong (2016KZDXM054), the Natural Science Foundation of Guangdong Province (2018A030310029), and the DGUT Grant (GC300501-08, GC300501-12).

References (57)

  • V.T. Tran et al.

    An approach to fault diagnosis of reciprocating compressor valves using teager-kaiser energy operator and deep belief networks

    Expert Syst. Appl.

    (2014)
  • Y. Fu et al.

    Analysis of feature extracting ability for cutting State monitoring using deep belief networks

    15th CIRP Conference on Modelling of Machining Operations

    (2015)
  • Q. Hu et al.

    Transfer learning for short-term wind speed prediction with deep neural networks

    Renew. Energy

    (2016)
  • M. Gan et al.

    Construction of hierarchical diagnosis network based on deep learning and its application in the fault pattern recognition of rolling element bearings

    Mech. Syst. Sig. Process.

    (2016)
  • S. Wang et al.

    A data indicator-based deep belief networks to detect multiple faults in axial piston pumps

    Mech. Syst. Sig. Process.

    (2018)
  • F. Jia et al.

    Deep neural networks: a promising tool for fault characteristic mining and intelligent diagnosis of rotating machinery with massive data

    Mech. Syst. Sig. Process.

    (2016)
  • H. Shao et al.

    Rolling bearing fault detection using continuous deep belief network with locally linear embedding

    Comput. Ind.

    (2018)
  • L. Ren et al.

    Multi-bearing remaining useful life collaborative prediction: a deep learning approach

    J. Manuf. Syst.

    (2017)
  • G. Hu et al.

    A deep boltzmann machine and multi-grained scanning forest ensemble collaborative method and its application to industrial fault diagnosis

    Comput. Ind.

    (2018)
  • C. Li et al.

    Multimodal deep support vector classification with homologous features and its application to gearbox fault diagnosis

    Neurocomputing

    (2015)
  • C. Li et al.

    Gearbox fault diagnosis based on deep random forest fusion of acoustic and vibratory signals

    Mech. Syst. Sig. Process.

    (2016)
  • P. Wang et al.

    Virtualization and deep recognition for system fault classification

    J. Manuf. Syst.

    (2017)
  • Y. Liu et al.

    Attitude angle estimation for agricultural robot navigation based on sensor fusion with a low-cost IMU

    IFAC Proceedings

    (2013)
  • Y. Wang et al.

    Modeling of low shaft speed bearing faults for condition monitoring

    Mech. Syst. Sig. Process.

    (1998)
  • L.M. Galantucci et al.

    Analysis of dimensional performance for a 3D open-source printer based on fused deposition modeling technique

    3Rd CIRP Global Web Conference on Production Engineering Research - Advancement Beyond State of the Art (CIRPe)

    (2014)
  • X. Song et al.

    Development of a Low-cost parallel kinematic machine for multidirectional additive manufacturing

    J. Manuf. Sci. Eng.

    (2015)
  • S. Nuchitprasitchai et al.

    Factors effecting real-time optical monitoring of fused filament 3d printing

    Progr. Addit. Manuf.

    (2017)
  • S. Nuchitprasitchai et al.

    Three hundred and sixty degree real-time monitoring of 3-d printing using computer analysis of two camera views

    J. Manufact. Mater. Process

    (2017)
  • Cited by (48)

    • Machinery condition monitoring in the era of industry 4.0: A relative degree of contribution feature selection and deep residual network combined approach

      2022, Computers and Industrial Engineering
      Citation Excerpt :

      Deep learning models are an ideal choice in the field of CM, considering their advantages in handling large-scale data and automatic learning functions in terms of multi-scale and multi-level (Hassan & Mahmood, 2017; Souza et al., 2020). In recent years, numerous deep learning architectures and their variants such as Auto-Encoder (AE) (Zhang et al., 2019), Deep Belief Network (DBN) (N. Lu et al., 2020), Convolutional Neural Network (CNN) (Hsueh et al., 2019; Y. Wang et al., 2021) and Recurrent Neural Network (RNN) (Y. Wu et al., 2018) have been proven successful in CM applications by various researchers. As a branch of machine learning, deep learning is acting as a bridge between a large amount of data and CM of complex equipment and is a very important application method in the era of large machinery data.

    • A perspective survey on deep transfer learning for fault diagnosis in industrial scenarios: Theories, applications and challenges

      2022, Mechanical Systems and Signal Processing
      Citation Excerpt :

      Besides the survey articles mentioned above, there are some other related articles which focus on the deep learning-based [28], transfer learning-based [29], convolutional neural network-based [30,31], AI-Enabled-based [32] or special components-based [33–35] methods for machine fault diagnosis, which will not be enumerated here in detail. Admittedly, as the valuable and systematic scholarly sources on the IFD, these literature reviews have contributed to the development of fault diagnosis from many aspects and guided the researchers towards a clearer future direction [36,37]. Nevertheless, there are still some aspects that have not yet been comprehensively summarized by the previous literature articles:

    • Mechanical fault time series prediction by using EFMSAE-LSTM neural network

      2021, Measurement: Journal of the International Measurement Confederation
      Citation Excerpt :

      However, it needs a good data input mode to achieve better prediction results. Zhang et al. [35] put forward an error fusion of multiple SAEs (EFMSAE) for 3D printers, which has a good effect on feature extraction and health status detection. However, the EFMSAE cannot predict the healthy condition.

    View all citing articles on Scopus
    View full text