Smart manufacturing has greatly benefited from the Internet of Things (IoT) revolution and advances in Deep Learning (DL) techniques as well. In this regard, several frameworks have been proposed that enabled the collections and analysis of diverse data from factory equipment and external sensors in order to make better decisions and optimize production towards zero-defect manufacturing strategies.
This paper focuses on a characteristic use case falling into this category that often arise in the electronics manufacturing industry. In particular, a system is proposed for the scanning of PCBs and the inspection of die attachment on an LCP substrate. The key quantity that needs to be monitored is the volume of glue that is deposited on the LCP since insufficient or excessive glue is the primary source of malfunctions. The current industry practice for fault diagnosis is to have human operators inspect the manufactured PCBs and identify potential defects, a process that is error prone, cumbersome and time consuming causing delays in the production process.
The proposed system uses a custom scanning device consisting of a high-resolution laser scanning sensor and a modular motion control framework in order to scan the areas of interest on a PCB and to extract geometric information in terms of 3D point clouds. Using these point clouds, glue volumes are estimated using a 3DCNN where the captured point clouds are quantized into 3D voxels. A graphical overview of our system is shown in Figure 1. In summary, the main technical contributions of our work can be pinpointed to,
deploying a modular scanning system for inspecting PCBs
introducing a process for developing a volumetric dataset of glue deposits
proposing a 3DCNN architecture for estimating the volume of glue
correlating glue volume measurements before and after die attachment, enabling quality inspection after die attachment where only a trace of glue is visible.
The proposed approach counters difficulties commonly occurring in manual inspection tools stemming from subjective aspects and provides fast and accurate assessments. It is worth noting that while deep learning has certainly improved state of the art in fault diagnosis for industrial applications, to the best of our knowledge there has not been any deep learning architectures used for quality inspection on 3D data. In this paper, the use of (R)egression-Net is yet another step closer to the exploitation of deep learning techniques in 3D industrial applications.
The remainder of the paper is organized as follows. Section II focuses on related work in deep learning and industrial applications, while Section IV provides insight on the examined microelectronics use case. In Section V the proposed system is described in detail including the deployed scanning system, the generated volumetric dataset and the RNet architecture. Next, Section VI is devoted to the experimental evaluation of our solution and finally Section VII draws the conclusions of our work.
Ii Related work
Deep Learning (DL) refers to a set of technologies, primarily Artificial Neural Networks (ANN), that has significantly improved state of the art in several tasks such as computer  and robotic  vision applications, image analysis [3, 4], audio signals [5, 6] and other complex signals . In industrial systems, DL applications focus on inspection and fault diagnosis in industrial shopfloors and manufacturing processes. CNNs are the most representative category of deep learning models and their architectures are able to deal with big multimodal datasets. For instance, in  a CNN framework is presented to convert sensor signals from manufacturing processes to 2-D images which are used to solve image classification tasks for fault diagnosis whereas in  a method is reported for analyzing tomography sensory data using a convolutional neural network pipeline. Moreover, in 
, a deep convolutional computation model is proposed to learn hierarchical features of big data in IoT, by using the tensor-based representation model to extend the CNN from the vector space to the tensor space. Also, a convolutional discriminative feature learning method for induction motor fault diagnosis is introduced in, an approach that utilizes a combination of back-propagation neural network and a feed-forward convolutional pooling architecture.
In another interesting application of DL in industry, the authors of  propose a semisupervised deep learning model for data-driven soft sensor in order to monitor quality variables that cannot be measured directly. In a related line of research, 
presents a soft sensor modeling method that combines denoising autoencoders with a neural network. Furthermore, a deep learning framework is introduced in for adaptively extracting features from raw data that are used for fault classification while  presents a framework that utilizes multiple sparse autoencoders for tool condition monitoring. In a relevant work, the authors of  propose a deep learning model for early fault detection of machine tools under time-varying conditions. The authors in , have proposed an optimized deep architecture, namely a large memory storage retrieval neural network, to diagnose bearing faults. In a related application, 
propose a convolutional deep belief network for bearing fault diagnosis applied in monitoring electric locomotive bearings while in the work of a deep residual network is used with dynamically weighted wavelet coefficients for planetary gearbox fault diagnosis. An unsupervised feature engineering framework is proposed in  that utilizes vibration imaging and deep learning for fault diagnosis in rotor systems. In another domain, 
reports a crack detection method for nuclear reactors monitoring using a CNN and a Naive Bayes data fusion approach. Also a deep neural network architecture for chemical fault diagnosis on imbalanced data streams is presented in.
Deep learning has been also deployed in the field of semiconductor manufacturing which is the application of our work. In 
, authors introduce a method for wafer map defect pattern classification and image retrieval using CNN, so as to identify root causes of die failures during the manufacturing process. A deep structure machine learning approach is also proposed in
to identify and classify both single-defect and mixed-defect patterns by incorporating an information gain-based splitter as well as deep-structured machine learning. A fault detection and classification CNN model with automatic feature extraction and fault diagnosis functions has been proposed in to model the structure of the multivariate sensor signals from a semiconductor manufacturing process.
While deep learning has certainly fostered progress in fault diagnosis of 1D and 2D signals it has not found wide application in processing 3D sensory from industrial shopfloors. Nonetheless, outside industry, 3DCNN models have recently gained traction for traditional computer vision tasks. For instance, VoxNet integrates a volumetric occupancy grid representation with a supervised 3DCNN for object recognition. Similarly focussing on object recognition, a 3D volumetric object representation with a combination of multiple CNN models based on LeNet model is introduced in  while in , the authors proposed an algorithm for labeling complex 3D point cloud data and provide a solution for training and testing 3DCNN using voxels. A CNN has been also proposed in  that directly consumes 3D point clouds and has shown impressive results in 3D object recognition. In another interesting work  octree-based CNN is built upon the octree representation of 3D shapes and performs 3DCNN operations on the octants occupied by the 3D shape surface. Moreover, the authors of  use a subset of ModelNet  and propose a CNN combining a spatial transformation in an attempt to deal with the rotation invariance problem in 3D object classification. In the following section we will see that 3DCNN based deep models have potential for fault diagnosis on 3D data as is validated through a concrete microelectronics use case, namely die attachment on a LCP substrate.
Iii Industrial background
The proposed method is driven by a common problem faced in the electronics industry for the manufacturing of miniature electronic modules. The standard electronic equipment used in the industry for wire bonding, lacks the capability of automated optical inspection of the process. During manufacturing, silicon dies are attached to an LCP substrate using conductive glue that is placed by the glue dispenser machine.
The manufacturing process consists of two main stages performed sequentially. The first stage is the glue dispensing stage and the second is the die (IC) attachment stage. The most critical variable for the attachment stage is the volume of the dispensed glue that is controlled indirectly by parameterizing the pressure on the glue dispenser machine. The two fault conditions that can be faced in the production are the excessive and insufficient glue volume. The excess in the amount of the glue leads to internal short circuits and on the other hand, insufficiency in the glue volume leads to weak die bonding. The current practice for the detection of these two fault conditions is the manual inspection of the glue volume at the first stage and the glue fillet, defined as the visible overflown glue around the die border, at the second stage.
The manual inspection by a human operator who decides on the faulty conditions of the process, introduces an error that derives from the limitations of the human factor. This error corresponds to overhead costs on newly introduced parts in the production line that require adjustment period for the stabilization of the manual inspection process, resulting in an increase of the failure rate. The automation of the inspection process is considered an important step towards the automated quality control in the microelectronics industry and constitutes the motivation of the proposed system.
Iv Description of use case
For the development and testing of the proposed system we have used three PCBs depicted in Figure 2. As is illustrated in Figure 3, there are different types of dies annotated as A,B,C,D and E. The type of die defines the shape of the glue that needs to be placed on the LCP substrate for its attachment. This correspondence between the glue and die types is depicted in Figure 3. Each PCB comprises circuits over two rows and nine columns. Each of the circuits hosts four placeholders for each of the five glue types A,B,C,D and E, thus totally regions for the placement of the dies as illustrated in Figure 3.
To simulate the conditions that can be met during the production, the PCBs, in Figure 2, have been specifically manufactured to have a wide range of glue quantities, from insufficient glue to excessive glue. Concretely, the glue quantity progressively increases starting from the far left column of each of the three PCBs and moving to the right. To ensure the same quantity per column, aside appropriately parameterizing the pressure on the glue dispenser, glue quantity has been visually verified during manufacturing. Moreover, to simulate both the glue dispensing stage and the die attachment stage, the top PCB comprises regions at the die attachment stage, the middle PCB comprises regions at the glue dispensing stage and the bottom PCB comprises regions at both stages; top circuits comprises regions after die attachment while the bottom circuits comprises regions before die attachment.
V PCB inspection system
In this section, the components of the proposed system are described. Initially we present the custom scanning system that was developed for acquiring 3D measurements from a PCB. Next the processing of the collected data is described in order to generate an annotated data set of point cloud measurements and finally the proposed 3DCNN architecture is presented for glue volume estimation.
V-a Scanning system
For PCB inspection, a modular laser scanning system has been built. As depicted in Figure (a)a, it includes an Optimet Conopoint-10 sensor , a Newport XPS-RL2 motion controller , two linear stages and the support breadboards. In our system we use the FMS200CC and FMS300CC  linear stages with travel ranges of and respectively.
Since inspection time is important for any production line, effort has been put in order to significantly decrease scanning time without losing accuracy. In this regard, a hardware triggering mechanism has been developed. The mechanism uses the Pulse Acquisition Mode of the sensor with measurement rate that is driven by distance spaced trigger pulses generated by the Position Compare Output (PCO) of the controller. The PCO provides delay between the crossing of the positions defined by the sample step and sending out the signal. The maximum speed of of the stages produces uncertainty in position. The operational flow of the system is shown in (b)b.
In the bottom right image of Figure 7 we see the pulse diagram for a range for Y-stage. The PCO generates a series of pulses that trigger the data acquisition at positions that are defined by the sample step of . A total of 36 pulses are generated in Y axis. After the last pulse is generated, the X-stage moves by without being homed and a new series of pulses are generated in the reverse direction in Y-axis. Indicatively, by adapting this zig-zag scanning procedure as shown in the top right image of Figure 7 and implementing the hardware triggering mechanism, the scanning time is decreased approximately -fold, compared to taking single measurements for every step along the motion. For all of the following experiments, this scanning strategy is followed.
Scans on a specific type of glue region have a constant range along the axes. Some indicative scans over different regions of a PCB are shown in Figure 18. The 3D structure of the regions is captured quite well both before die attachment, where the glue deposit is clearly visible, and after die attachment. What is interesting for the more challenging second case, is that the glue fillet that exceeds the die border is captured as well as the 3D structure of the actual die. To verify this, in Figure 19 two same type regions are shown with the first one corresponding to a case of excessive glue on the left of the PCB and the second one to insufficient glue on the right of the PCB. Notice the much sharper gap between the die and the LCP substrate in the second case.
V-B Generating a dataset
To train and evaluate our monitoring framework an annotated point cloud dataset has been developed where the volumes of glue deposits have been analytically computed. To this end we used the circuits on the second PCB and second row of the bottom PCB in Figure 2 where no dies are attached and glue is clearly visible and assumed that the volume of deposited glue is equal for regions in the same column as explained in Section IV. For each regional scan, the plane equation for the LCP substrate has been calculated using the RANSAC algorithm  for plane fitting as implemented in PCL  and the 3D points corresponding to the glue mass have been manually cropped. In order to get a closed 3D surface, each 3D point of the cropped point cloud is projected on the estimated substrate plane as is illustrated on the example of Figure (a)a. Having a closed 3D point cloud, Poisson reconstruction  is applied to compute the surface mesh of glue as is depicted in Figure (b)b. Geometrically, having a surface , its volume can be computed using the formula,
In our case an analytical representation of is not available, therefore glue volume is approximated using its reconstructed mesh. More formally, let us denote the glue mesh as a set of 3D points’ triangles and the substrate plane as . Additionally, we denote with the projection of on and with the distance between the centers of and as is more clearly depicted in Figure (c)c. The glue volume is approximated as,
where is the area of the projected rectangle. Following this procedure, we end up with annotated point clouds for each of the five different types of glue deposits on the circuits with no dies attached. We extend these volume estimations to the rest of the circuits where dies have been attached in order to annotate the entire PCB dataset. Naturally, this extension process as well as the volume approximation is error prone due to several factors such as sensor noise, discrepancies between glue deposits due to the glue dispenser and the induced approximation error. Nonetheless, as shown in Section VI these estimations captures the manufacturer’s trend of progressively depositing more glue. This allows the correlation of the hidden glue after die attachment with actual measurements, thus enabling fault diagnosis even after dies are attached.
|20 micrometer sampling step|
|50 micrometer sampling step|
Of course the number of annotated samples is very small for a 3DCNN network, therefore a data augmentation step has been implemented. To this end and in order to also take into account measurement noise we scan each PCB five times, thus artificially raising the number of PCB samples to fifteen. Moreover, the range along the is calculated for each regional scan. A cropping filter is applied on the plane using a bounding box that is shifted along the axes. For all our experiments the bounding box covers of the range on the axes and is shifted by a constant step ,
where is the respective axis range and is the minimum step of the linear stage which is either or for our experiments. This augmentation procedure, as is also shown in the example of Figures (a)a, simulates the potential mis-localization of the scanning device. The dataset is further augmented by adding Gaussian noise along the axis coarsely approximating noise during the scanning procedure. In our experiments, noise levels have been used by increasing the noise deviation by with respect to the range on . Following this procedure for different step of the linear stages, namely and , we get several thousands of point clouds for each of the different region types as is summarized in Table I. In the table we denote with att. and unatt. the cases with and without die attached and we adopt a to ratio for splitting between training and test sets.
V-C 3DCNN for glue volume estimation
To automate glue volume estimation we train instances of the proposed RNet architecture with each one corresponding to a particular type of glue before or after the respective die has been attached. Point clouds are parameterized using voxel grids similar to the works of [26, 39]. This simple representation works fine in our case since, in contrast to depth cameras, there are no occlusion issues caused by a single viewpoint while point cloud density is approximately constant. Throughout our experiments the size of voxel grids is . The dimensions of a single voxel are constant for each type of glue. Specifically, along the axes the dimensions are equal to the range of point cloud divided by the number of voxels whereas it is set to along the axis. For glue types where the dimension along the is not adequate for covering at least of the captured 3D points, the dimension of voxels increases in steps of until this criterion is fulfilled. An indicative voxel grid is shown in Figure (b)b corresponding to the point cloud of the upper left image in Figure 18. To facilitate visualization we have reduced the number of voxels along to .
The architecture of RNet is shown in Figure 27 along with the specific parameters of each layer. The network’s input is an occupancy grid and there are five network blocks consisting of sequences of 3D convolutional, leaky ReLU, batch normalization and max pooling layers that progressively drop the spatial dimensions of the input tensor while increasing the number of channels. Concretely the dimension of the input tensor for the first block is and the output tensor is which is then propagated to the following block. After these five network blocks, the last one generates a tensor that is subsequently flattened to a feature vector which is forwarded to a dense layer that produces the final volume estimation. Figure 27 includes the dimension of the output tensor under each network block along with the kernel size, stride and padding of the respective convolutional layer. The size of the max pooling window is with a stride of for all network blocks. It should be mentioned that we have also experimented with deeper architectures without noticing substantial accuracy improvement. The cost function that is minimized is the error between ground truth and estimation. Namely, if the glue volume of grid is denoted as and the volume estimation as the network loss is defined as,
The underlying optimization problem is solved using Adam optimization algorithm 
as is implemented in the PyTorch framework.
This section presents a series of experiments that were carried out in order to examine the performance of our framework in different aspects and compare it with alternative state of the art methods.
Vi-a Experimental setup
The complexity of the PCB with its different types of glue regions and dies provide a challenging evaluation setup for the proposed framework. The conducted experiments have been divided in two categories based on the dataset described in Section V-B. In the first category dies have not been attached and glue deposits are clearly visible while in the second one dies have been placed and only glue on the sides of dies is partially visible. Another important experimental parameter that is examined is inspection time using different scanning step. In this regard, two different sets of experiments are presented where the sampling step for the scanning system of Section V-A is and micrometers.
For each of these sets of experiments, separate instances of RNet are trained corresponding to the five different glue types before and after die attachment. For each instance, point clouds are split in training and test set under a ratio. To avoid biasing the evaluation by the data augmentation described in Section V-B, the training and validation sets correspond to scans from different regions on the PCBs. Concretely, as is also explained in Figure 3, on each circuit module the top three glue deposits of each type are used for training and the last one for testing.
Each RNet instance has been trained for epochs using a batch size of , a learning rate of and having trainable parameters. Training and validation mean square error remained approximately constant after epochs thus we assume that the training procedure converges to a stable minimum in all cases. Training time was up to
hours for the cases with the largest training set and was performed on two Tesla K40m GPUs. For parameter initialization all convolutional layers are initialized using a Gaussian distribution of mean 0 and standard deviation of 0.02 while batch normalization layers are initialized using a mean of 1 and standard deviation of 0.02.
Vi-B Data augmentation analysis
|20 micrometer sampling step|
|50 micrometer sampling step|
In this paragraph we examine the contribution of the data augmentation step. In this regard, Table II contains the achieved Mean Square Error (MSE) on the validation set when training with and without data augmentation. The rows denoted as no IC and IC correspond to before and after die attachment. In all cases it is apparent that the data augmentation step on Section V-B clearly improves the generalization of the trained model, achieving lower MSE by at least an order of magnitude.
Vi-C Volume estimation evaluation
|20 micrometer sampling step|
|50 micrometer sampling step|
|20 micrometer sampling step|
Although the volume estimation task that we focus on can have a wider industrial applicability, to the best of our knowledge it has not been studied before. Thus there are no obvious alternative deep architectures that we can compare with. To this end, we have adjusted two well known methods, namely VoxNet  and PointNet , to our volume estimation task. For VoxNet we have replaced its final classification layer with a fully connected layer in order to produce a single volume estimation whereas for PointNet we followed the same procedure for its classification sub-network. Also since PointNet takes as input 3D points we have uniformly subsampled each of our 3D scans to sample this number of points.
). Although PointNet has shown excellent results for 3D object classification, it performs poorly in the current task and its architecture needs to be further altered. Probably in its current form it cannot encode accurately the subtle differences between different quantities of glue from direct point cloud input. RNet outperforms VoxNet as well in almost all experiments.
|Scanning Step||Scanning Time||Prediction Time||Total Time|
As is anticipated the error is significantly larger after dies have been attached since glue mass can not be directly scanned and inevitably there are discrepancies in the ground truth values as explained in Section V-B. Nonetheless, test error is not substantially affected by having a larger sampling step, showcasing the ability of RNet and to a lesser degree VoxNet to infer the volume of glue even from a sparser point cloud. Having a larger sampling step is important for accelerating the scanning process and not delaying the PCB manufacturing process. The required scanning and prediction time for a single PCB with a step of or is shown in Table V. During prediction point clouds are examined one by one and not in a batch allowing an earlier notification in case of a defect. In both cases the total time is below the critical time threshold of one hour after which glue viscosity drops and any repair activity becomes much more difficult.
While Tables III, IV and V are useful for comparing different experimental setups, from the end-user’s point of view it is more important to examine the ability of the trained models to discern the fluctuations of glue quantity. In this regard, Figure 48 shows the performance of VoxNet and RNet for each test set in comparison with the ground truth. The vertical axis refers to the volume of glue, while the lateral one to the index of the test set sample. On each plot the red curve shows the ground truth glue volume while the green and blue curves the predictions of RNet and VoxNet respectively. We have excluded PointNet from this comparison since its high MSE would hinder visualization. Each test set has been sorted according to the annotated quantity of glue in the respective samples, thus the red curves are monotonically decreasing. The diagrams of Figure 48 verify Table III since there is no substantial difference when increasing the scanning step while accuracy drops after die attachment. Also RNet obviously follows more tightly ground truth in several instances (e.g. type E unattached) whereas in other cases is covered by VoxNet due to VoxNet’s wider fluctuations. However in general, the trained models can indeed capture the different levels of glue since their predictions follow the ground truth curves. Of course when dies are attached there are stronger deviations from the ground truth, especially for type B and D glue depositions where the scanned area is smaller and deposited quantity of glue is less. Nonetheless the general downward trend is efficiently captured in all cases. Even after die attachment RNet can infer glue volume from the exceeding glue around the die and from the elevation of the die above the LCP substrate due to the intermediate glue layer, therefore providing meaningful predictions for fault diagnosis.
Vi-D Fault diagnosis evaluation
Volume estimations can be used for automated fault diagnosis based on predefined threshold of normal quantities. To this end, using expert knowledge we have defined such threshold and evaluated the performance of RNet, VoxNet and PointNet for all experiment scenarios. In Figures (a)a and (b)b we see the classification accuracy for RNet using a or scanning step. Again RNet outperforms VoxNet and PointNet with an average classification accuracy of % compared to % and % respectively.
In this paper a fault diagnosis system is proposed for inspecting glue dispensation and die attachment on PCBs. To realize this, a custom scanning module has been built using a commercially available laser scanner and two high-accuracy linear stages while a hardware triggering mechanism has been implemented in order to accelerate the scanning process. This paper also introduces RNet architecture in order to monitor a geometric quantity, namely the volume of glue from the scanned point clouds corresponding to different regions of interest on a PCB. To train and deploy the 3DCNN data parameterization and augmentation steps have been introduced.
The efficiency of the system is demonstrated through several experiments on different types of glue regions. A main finding of our work is that glue volume can be estimated even after dies have been attached and most of the glue is occluded. Although the error increases in this case, the predictions of RNet are still useful for quality control. This enables the automatic inspection at a later stage, after dies are attached thus postponing a time consuming scanning process that might affect the viscosity of glue especially in even more complex PCBs. Another interesting outcome of our research is that prediction accuracy is not significantly affected by a larger scanning step as similar performance is noticed for sparser point clouds at a significantly reduced scanning time.
As a next step it would be interesting to deploy and evaluate the proposed framework in other industrial use cases where a geometric parameter needs to be monitored and further research challenges would arise that have not been addressed in the current work. Such challenges could be related to the misalignment of the sensor and specimen that could result in partial scans and the detection of multi-scale defects that would require adjustments in both data parameterization and network architecture.
-  A. Aakerberg, K. Nasrollahi, and T. Heder, “Improving a deep learning based rgb-d object recognition model by ensemble learning,” in International Conference on Image Processing Theory, Tools and Applications (IPTA), 2017.
-  M. R. Loghmani, S. Rovetta, and G. Venture, “Emotional intelligence in robots: Recognizing human emotions from daily-life gestures,” in IEEE International Conference on Robotics and Automation (ICRA), 2017.
-  T. H. Chan, K. Jia, S. Gao, J. Lu, Z. Zeng, and Y. Ma, “Pcanet: A simple deep learning baseline for image classification?” IEEE Transactions on Image Processing, vol. 24, no. 12, 2015.
-  J. Lee, T. Kim, J. Park, and J. Nam, “Raw waveform-based audio classification using sample-level CNN architectures,” CoRR, vol. abs/1712.00866, 2017. [Online]. Available: http://arxiv.org/abs/1712.00866
A. Vafeiadis, D. Kalatzis, K. Votis, D. Giakoumis, D. Tzovaras, L. Chen, and R. Hamzaoui, “Acoustic scene classification: from a hybrid classifier to deep learning,” inWorkshop on Detection and Classification of Acoustic Scenes and Events (DCASE2017), 2017.
-  P. Ghamisi, B. Höfle, and X. X. Zhu, “Hyperspectral and lidar data fusion using extinction profiles and deep convolutional neural network,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 10, DOI 10.1109/JSTARS.2016.2634863, no. 6, 2017.
-  L. Wen, X. Li, L. Gao, and Y. Zhang, “A new convolutional neural network-based data-driven fault diagnosis method,” IEEE Transactions on Industrial Electronics, vol. 65, DOI 10.1109/TIE.2017.2774777, no. 7, 2018.
-  O. Costilla-Reyes, P. Scully, and K. B. Ozanyan, “Deep neural networks for learning spatio-temporal features from tomography sensors,” IEEE Transactions on Industrial Electronics, vol. 65, DOI 10.1109/TIE.2017.2716907, no. 1, 2018.
-  P. Li, Z. Chen, L. T. Yang, Q. Zhang, and M. J. Deen, “Deep convolutional computation model for feature learning on big data in internet of things,” IEEE Transactions on Industrial Informatics, vol. 14, no. 2, 2018.
-  W. Sun, R. Zhao, R. Yan, S. Shao, and X. Chen, “Convolutional discriminative feature learning for induction motor fault diagnosis,” IEEE Transactions on Industrial Informatics, vol. 13, no. 3, 2017.
-  L. Yao and Z. Ge, “Deep learning of semisupervised process data with hierarchical extreme learning machine and soft sensor application,” IEEE Transactions on Industrial Electronics, vol. 65, DOI 10.1109/TIE.2017.2733448, no. 2, 2018.
-  W. Yan, D. Tang, and Y. Lin, “A data-driven soft sensor modeling method based on deep learning and its application,” IEEE Transactions on Industrial Electronics, vol. 64, DOI 10.1109/TIE.2016.2622668, no. 5, 2017.
-  J. Pan, Y. Zi, J. Chen, Z. Zhou, and B. Wang, “Liftingnet: A novel deep learning network with layerwise feature learning from noisy mechanical data for fault classification,” IEEE Transactions on Industrial Electronics, vol. 65, DOI 10.1109/TIE.2017.2767540, no. 6, 2018.
-  C. Shi, G. Panoutsos, B. Luo, H. Liu, B. Li, and X. Lin, “Using multiple-feature-spaces-based deep learning for tool condition monitoring in ultraprecision manufacturing,” IEEE Transactions on Industrial Electronics, vol. 66, DOI 10.1109/TIE.2018.2856193, no. 5, 2019.
-  B. Luo, H. Wang, H. Liu, B. Li, and F. Peng, “Early fault detection of machine tools based on deep learning and dynamic identification,” IEEE Transactions on Industrial Electronics, vol. 66, DOI 10.1109/TIE.2018.2807414, no. 1, 2019.
-  M. He and D. He, “Deep learning based approach for bearing fault diagnosis,” IEEE Transactions on Industry Applications, vol. 53, no. 3, 2017.
-  H. Shao, H. Jiang, H. Zhang, and T. Liang, “Electric locomotive bearing fault diagnosis using a novel convolutional deep belief network,” IEEE Transactions on Industrial Electronics, vol. 65, DOI 10.1109/TIE.2017.2745473, no. 3, 2018.
-  M. Zhao, M. Kang, B. Tang, and M. Pecht, “Deep residual networks with dynamically weighted wavelet coefficients for fault diagnosis of planetary gearboxes,” IEEE Transactions on Industrial Electronics, vol. 65, DOI 10.1109/TIE.2017.2762639, no. 5, 2018.
-  H. Oh, J. H. Jung, B. C. Jeon, and B. D. Youn, “Scalable and unsupervised feature engineering using vibration-imaging and deep learning for rotor system diagnosis,” IEEE Transactions on Industrial Electronics, vol. 65, DOI 10.1109/TIE.2017.2752151, no. 4, 2018.
-  F. Chen and M. R. Jahanshahi, “Nb-cnn: Deep learning-based crack detection using convolutional neural network and naive bayes data fusion,” IEEE Transactions on Industrial Electronics, vol. 65, DOI 10.1109/TIE.2017.2764844, no. 5, 2018.
-  Z. Hu and P. Jiang, “An imbalance modified deep neural network with dynamical incremental learning for chemical fault diagnosis,” IEEE Transactions on Industrial Electronics, vol. 66, DOI 10.1109/TIE.2018.2798633, no. 1, pp. 540–550, 2019.
-  T. Nakazawa and D. V. Kulkarni, “Wafer map defect pattern classification and image retrieval using convolutional neural network,” IEEE Transactions on Semiconductor Manufacturing, vol. 31, no. 2, 2018.
-  G. Tello, O. Y. Al-Jarrah, P. D. Yoo, Y. Al-Hammadi, S. Muhaidat, and U. Lee, “Deep-structured machine learning model for the recognition of mixed-defect patterns in semiconductor fabrication processes,” IEEE Transactions on Semiconductor Manufacturing, vol. 31, no. 2, 2018.
-  K. B. Lee, S. Cheon, and C. O. Kim, “A convolutional neural network for fault classification and diagnosis in semiconductor manufacturing processes,” IEEE Transactions on Semiconductor Manufacturing, vol. 30, no. 2, 2017.
-  D. Maturana and S. Scherer, “Voxnet: A 3d convolutional neural network for real-time object recognition,” in IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2015.
-  X. Xu, A. Dehghani, D. Corrigan, S. Caulfield, and D. Moloney, “Convolutional neural network for 3d object recognition using volumetric representation,” in 2016 First International Workshop on Sensing, Processing and Learning for Intelligent Machines (SPLINE), 2016.
J. Huang and S. You, “Point cloud labeling using 3d convolutional neural
International Conference on Pattern Recognition (ICPR), 2016.
-  R. Q. Charles, H. Su, M. Kaichun, and L. J. Guibas, “Pointnet: Deep learning on point sets for 3d classification and segmentation,” in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), DOI 10.1109/CVPR.2017.16, 2017.
-  P. Wang, Y. Liu, Y. Guo, C. Sun, and X. Tong, “O-cnn: Octree-based convolutional neural networks for 3d shape analysis,” ACM Transactions on Graphics, vol. 36, no. 4, 2017.
-  L. Shao and P. Xu, “Neural network for 3d object recognition,” Stanford University, Computer Science Department, Tech. Rep., 2016.
-  Z. Wu, S. Song, A. Khosla, F. Yu, L. Zhang, X. Tang, and J. Xiao, “3d shapenets: A deep representation for volumetric shapes,” in 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), DOI 10.1109/CVPR.2015.7298801, 2015.
-  “Conopoint-10, 3d non contact measurement,” http://www.optimet.com/conopoint-10.php, accessed: 2018-07-04.
-  “Xps-rl universal high-performance motion controller/driver,” https://www.newport.com/f/xps-rl-universal-high-performance-motion-controller, accessed: 2018-07-04.
-  “Linear metrology stages,” https://www.newport.com/f/fms-metrology-stages, accessed: 2018-07-04.
-  M. A. Fischler and R. C. Bolles, “Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography,” Communications of the ACM, vol. 24, 1981.
-  R. B. Rusu and S. Cousins, “3d is here: Point cloud library (pcl),” in IEEE International Conference on Robotics and Automation (ICRA), 2011.
-  M. B. M. Kazhdan and H. Hoppe, “Poisson surface reconstruction,” in Eurographics Symposium on Geometry Processing, 2006.
-  Z. Wu, S. Song, A. Khosla, F. Yu, L. Zhang, X. Tang, and J. Xiao, “3d shapenets: A deep representation for volumetric shapes,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015.
-  D. Kingma and J. Ba, “Adam: A method for stochastic optimization,” in International Conference on Learning Representations (ICLR), 2015.
-  A. Paszke, S. Gross, S. Chintala, G. Chanan, E. Yang, Z. DeVito, Z. Lin, A. Desmaison, L. Antiga, and A. Lerer, “Automatic differentiation in pytorch,” in Conference and Workshop on Neural Information Processing Systems, 2017.