Analyzing in-process tool condition from limited data is an important issue through which manufacturing process can be more precise and efficient. Tool wear is the most phenomenal condition people cared. Identifying the tool wear mechanism is intractable as it involves physical and chemical process, such as abrasion, adhesion, diffusion and other types of wear during cutting process. A few pioneered studies strived to understand the wear mechanism as the addition of brittle fracture, mechanical abrasion, physicochemical mechanism and others [
114]. A recent study [
115] developed a fundamental wear model by using a dedicated tribometer, which consists of cutting and thermal simulations. However, the formulation of all these factors is subject to certain simplifications and assumptions, and calibrating the pending coefficients of the model using limited testing data would introduce more statistical errors, which together make the physical model inaccurate and unstable towards real complicated machining process.
Instead of formulating complex and error-prone physical models for tool wear mechanism, most researchers intended to estimate tool wear in a more statistical manner, i.e., to estimate remaining useful life by fitting historical data into an empirical model. Endeavors to estimate the tool life can be traced back to the early 20th century when FW Taylor [
116] proposed the well-known Taylor equation, which is an empirical model with two unknowns. Ever since then, various empirical models [
117‐
119] and experimental studies [
120,
121] were presented targeting at different tool-workpiece combinations. A comprehensive list of variant tool wear empirical models for dry machining can be found in Ref. [
122]. Their procedures were in similar fashion: first a nonlinear formula describing the tool condition based on the observer’s domain expertise was established ahead of time, then factorial design of physical experiments were conducted to calibrate the unknowns of the formula, experimental validations were eventually conducted to prove the feasibility. Although the prediction accuracy reported in these works can reach as high as 95% in their experimental setups, it is perceived that any slight change of the actual cutting condition would devastate the accuracy. As the demand for accuracy and the complexity of manufacturing process keep growing rapidly, physical and empirical models have been widely deprecated. Zhao et al. [
123] argued that this was mainly due to the following reasons: First, the performance of these models was highly dependent on the domain expertise of the observer, whose robustness was unsecured due to the uncertainty and complexity of working conditions. Secondly, these models were unable to evolve along with the accumulation of data, and thus insensitive to the changing conditions, which lead to limited effectiveness and flexibility in real cases. These two deficiencies of model-based approach would introduce considerable amount of error, not to mention the error from the feature extractor, which together makes the physical model-based data analysis hardly compatible with wider applications.
The advance of volume and veracity of data makes it possible to adopt various machine learning algorithms to predict tool condition more accurately. Prevalent choices of machine learning techniques for tool condition analysis include support vector machine (SVM), artificial neural networks (ANN), Hidden Markov models (HMM) and decision tree. With sufficient training data, trained ANN model using back propagation can be comparatively accurate for tool wear estimation [
124]. Palanisamy [
125] compared ANN against classic regression model in terms of the capability in tool wear estimation, ANN was found to be more robust and accurate for its powerful fitting ability. As a statistical learning approach, SVM is superior for non-linear classification of data by mapping them into higher dimensional feature space, by which discretized state of tool wear can be classified. Support vector regression (SVR) is a variant of SVM for continuous regression of tool wear value. Tool breakage detection [
126] and tool wear estimation [
127] were successfully carried out via SVM/SVR with over 99% success rate when the design parameters of the SVM model was fine-tuned. It was also noticed in a more recent study that a hybrid estimator combining analytic fuzzy classifier (AFC) and SVM can reach higher accuracy in tool wear estimation [
128]. Other learning techniques, such as decision tree classifier [
129] and HMM [
130], were also applied in application of tool wear estimation and achieved plausible performance. It was however stated in Ref. [
129] that the performance of decision tree classifier combined with a PCA was case-sensitive. It is noticed that there has always been a hidden trade-off issue between the complexity of learning model and the training cost. To achieve high accuracy, a more complex learning model would thus require a larger training data set and heavier computational load, otherwise overfitting issue would lower the performance.
Most of the aforementioned tool condition analysis is majorly dependent on time series data such as cutting force and vibration. Shallow function approximators like ANN and SVM are technically incapable of dealing with such high-dimensional data and thus require a dedicated feature extractor beforehand [
131], as already elaborated in Section
4.1 and illustrated in Figure
6(b). Conceivably, the quality of the extracted features directly affect the accuracy of subsequent operations. Improper choice of feature extractor may fundamentally suppress the eventual performance. Therefore, it would be better if one can directly handle the raw data series and bypass the feature extraction stage. The development of deep neural networks such as convolutional neural network (CNN) [
132] and long short-term memory (LSTM) network [
133] can fully satisfy this requirement. Specifically in the application of tool condition analysis, Li et al. [
134] adopted CNN to detect tool breakage by spindle current signal, which achieved higher accuracy (93%) than that of the traditional BP neural network (around 80%). However, time-domain feature extraction was still adopted in this work, CNN was thus only regarded as a traditional machine learning methods with higher achievable accuracy. Another recent study [
135] was to monitor the tool wear level based on audio signal using CNN, which strived to eradicate the need of feature extraction by using the absolute values of Fourier transformation as input. As a result, the tool wear prediction accuracy reached to as high as 96.3%. A Convolutional Bi-directional Long Short-Term Memory networks (CBLSTM) was designed in Ref. [
123] to eliminate feature engineering in tool health monitoring. In this network, CNN was served as local feature extractor, while LSTM was to address sequences of varying length data and capture long-term dependencies, in that tool wear was a time-variant sequential progress.
The extrusive challenge for the adoption of deep learning to make accurate analysis is the demand of large volume of labeled data, the acquisition of which is extremely costly and time-consuming for many manufacturing applications. For example, the identification of tool tip dynamics for a newly inserted tool needs hundreds of impact tests at different tool postures. In this situation, the utilization of historical data to facilitate the training of a new case becomes a potential and appealing solution. Chen et al. [
136] proposed a transfer learning-based prediction for pose-dependent tool tip dynamics in five-axis machine, by which the number of required impact tests is highly reduced. Sun et al. [
137] utilized deep transfer learning to predict tool life, by taking advantage of the learnt similar characteristic across different objects. A recent study on tool wear prediction based on meta-learning was proposed by Li et al. [
138]. Meta-learning has the ability of learning the hidden rules behind a variety of different but similar tasks/models. The adoption of meta-learning in this study successfully predicts the tool wear status in changing cutting conditions with enhanced accuracy, while only a few training samples are needed upon a new learning task. This meta-learning approach provides a new perspective to solve manufacturing problems where the acquisition of data samples are expensive and time-consuming. Table
6 lists the evolution of tool condition analysis.
Table 6
Evolution of tool condition analysis
4.2.2 Process Condition Analysis
Process condition analysis is a typical classification task. In machining process, the condition can be categorized into idling, stable and chatter state. Timely and precise identification of process condition is always desired to make adaptive adjustment of process control. Previous studies made some important progress in identifying the mechanism of cutting process. Budak and Altintas [
139] explored the mechanism of chatter during milling process and came up with a physical model to identify the chatter stability induced by the dynamic milling forces. According to this study, the cutter is simplified as a two degree-of-freedom system subject to a dynamic radial force, based on which the theoretical chatter stability lobe was derived. On the other hand, the calculation of dynamic cutting force is also simplified using numerical method. This plausible offline solution may not practically satisfy real machining cases [
45], as it requires a complete analysis of machine dynamics including the spindle, tool holder, tool and the workpiece, which is not only intractable to precisely identify but also requires tedious calibration works for different process conditions. The simplifications and unpredictable systematic bias further reduced the accuracy in offline analysis. Although researchers carried forward this theory to adapt to more complex situations, e.g., five-axis machining [
140], they were still of limited usage since the fundamental gap was not completely filled.
When it comes to online identification of process condition, the preferred option is to make diagnosis as early as possible, in order to prevent workpiece damage ahead of time. Traditional estimation algorithm, such as maximum likelihood [
141] though achieved great successful rate, but lacked the ability for early prediction. The main reason is that subtle features are prone to be overlooked before they become phenomenal. Machine learning methods have been employed in this task for the superiority in classification, especially in those hard-to-recognize scenario. In particular, acceleration signals were analyzed based on wavelet transform and SVM, this combination was able to detect transition state between stable and chatter state, showing excellent performance with over 95% accuracy rate [
142]. In this way, chatter could be firmly suppressed in its infancy stage. Later on, neural network approaches were also developed for process condition classification using vibratory signal [
143]. In addition to the feature generation which is mandatory for traditional machine learning approaches, this work introduced a feature selection strategy based on envelope analysis to rank the features according to their entropy, and only those high-ranking features were selected for classification. This operation essentially reduced the error from irrelevant features and hence increased the final accuracy.
To further reduce the error induced by feature extraction, deep learning methods were also utilized for machining process condition analysis. Among existing deep learning algorithms, CNN is known for its powerful image (second order tensor) processing and classification capability. However, most captured data from machining process is in the form of first order tensor (time sequence), which is not practical to be processed via CNN. Fu et al. [
144] innovatively transformed measured signals into plotted image and employed convolutional neural network to achieve real-time identification of cutting vibration state. This work realized directly use of the original signal sequence for cutting state monitoring with significant performance of over 99.5% accuracy in most testing cases. Deep Belief Network (DBN) has been majorly dealing with voice and speech recognition [
145]. The in-process vibration signal is similar to the voice. Fu et al. [
146] got inspired by this and came up with a DBN approach for cutting state monitoring. It turned out that DBN can steadily achieve high performance on the raw vibration signal without much data preparation.
Since data is relatively convenient to acquire during the manufacturing process, most deep learning approaches can already achieve very promising accuracy in their case studies. Still, conditions can be quite different in real machining situation where various materials, tools and parameters are combined in each individual task. Transfer learning has been attracting more attention to deal with varying conditions [
147] and proved to be effective for chatter detection with accuracy up to 95%. This new learning technology will not only reduce the data needed for training a deep model, but also increase model versatility to adapt to complex manufacturing process scenario. Table
7 lists the evolution of process condition analysis.
Table 7
Evolution of process condition analysis
4.2.3 Part Condition Analysis
The well-being of in-process part directly affects the quality of final product. Surface roughness [
77] and part dimensional error [
148] are the two most concerned aspects, since they respectively reflect the manufacturing quality in microscopic and macroscopic view. For the formal one, physical models and experimental data based regression are the two mainstream solutions people utilized to understand the surface roughness mechanism. Lin and Chang [
149] established a surface topography simulation model incorporating the effects of tool geometry, cutting parameters and tool motions to simulate the surface finish profile during turning operation. Kim and Chu [
150] determined the surface roughness by proposing a geometrical model to calculate the maximum height of the effective scallop. This model was particularly complicated as it considered the cutter runout effect and cutter marks. Others conducted experimental studies to unveil the relationship between tool life, surface roughness and vibration [
101]. Regression analysis was adopted to handle the experimental data.
The dimensional error of in-process part can be categorized into plastic deformation caused by residual stress and elastic deformation caused by large cutting load. Finite element method (FEM) was a primary choice for the evaluation of these two types of deformation, due to the large uncertainty of part shape and stress distribution during the process. The distortion of thin-walled workpiece induced by machining residual force was predicted using a modified finite element model [
151]. The combination of experimental results with FEM was proposed to predict the shape deviation of complex geometry [
152]. Elastic deflection also induces machining error, especially for thin-walled part. Wan et al. [
153] estimated the cutter deflection using a simple cantilever beam model, and the workpiece deflection using FEM simulation. The induced error was compensated accordingly [
154].
Both analytical model and FEM have to make a great deal of simplifications since accurate prediction of surface roughness and part deformation require tedious trial-and-error process and excessive computing power. Targeting at online analysis, trade-off between model complexity and its performance has always been a puzzling task. In light of this issue, machine learning algorithms started to take over online quality analysis with higher performance. As for the surface roughness prediction, although people spent great effort investigating its mechanism, it however varied with different processes and conditions. Any sophisticated physical models will only take effect in a limited range of applications. In this case, ANN has been widely adopted [
155,
156] in both turning and milling process. Using a small number of training samples, ANN is capable of generating accurate prediction values but would essentially require a good design of network structure. As compared to linear and exponential regression model [
155], neural networks were found to be capable of better predictions for surface roughness. Support vector regression (SVR) method was also utilized for the prediction of roughness. A comparison of three types of SVRs and ANN was conducted in Ref. [
157], results showed that SVR can achieve prediction accuracy as high as 95%, while for ANN it was slightly lower (91.4%) and required more computational time at the same time.
When it comes to dimensional error prediction, online prediction and real time compensation has always been a preferable choice. Li et al. [
158] developed a soft-touch sensor which provides proximity information when the tool is approaching the workpiece, and a neuro-fuzzy network for predicting machining errors. This hybrid learning system succeeded in precise prediction of the aggregate sum of thermal error, force-induced deflection error and other source error in turning process. Another dimensional error prediction in milling process was achieved using ANN [
159]. In this work, data set of process parameters that can affect dimensional errors was yielded via experiments. The large number of influencing parameters led to the choice of ANN, which generated more accurate models than the previous empirical models after training process.
Conventional machine learning approaches suffice the demand for real-time prediction of surface roughness and part deformation. A foreseeable trend in this section would be more precise identification of part conditions, such as the types of defect and crack, by further exploiting the advanced vision-based sensors. Towards this goal, traditional shallow learning approaches require artificially defined feature descriptors from the captured raw pixels, while deep networks are able to directly process raw data. In particular, CNN serves as a primary choice for surface inspection task. A max-pooling CNN was developed in Ref. [
160] to identify steel defect with an error rate of 7%, which outperformed the best trained classifier using artificial feature descriptors (15%). Part et al. [
161] showed that using CNN can achieve 250 times faster inspecting speed compared to manpower inspection, without sacrificing the accuracy. Ren et al. [
162] proposed a CNN based feature extractor for pixel-wise surface inspection, which did not require large-scale training data using pretrained model. The heat map showing distribution of defects was then generated for the identification of seven types of defects using image processing algorithms. This work showed improved accuracies in both classification and segmentation tasks for all seven defect types. Crack identification was also realized using a deep RBM from consumer-grade camera images [
163], which provided an alternative option in addition to CNN. In terms of part deformation prediction and control, the utilization of responsive fixture made it possible to measure and accumulate online deformation data for different parts in different machining stages. Such data potentiates the training of a mixed deep learning model, as proposed by Zhao et al. [
164], to predict the part deformation and make process adjustments in an early stage. As can be concluded from previous studies, most deep learning based part condition analysis takes image as raw input. It is conceivable that when the amount of training data is limited, deep neural network, such as CNN, can be easily over-fitted to jeopardize the accuracy. In order to reduce data dependency, Ferguson et al. [
165] trained a CNN using openly-available image datasets and leveraged transfer learning to adapt the pre-trained CNN model to the detection of defects, by using small X-ray dataset. Cheng et al. [
166] applied a parameter-based transfer learning in modeling shape deviations during additive manufacturing, as one particular example to represent the future trend. Table
8 lists the evolution of part condition analysis.
Table 8
Evolution of part condition analysis