Optimising tool wear and workpiece condition monitoring via cyber-physical systems for smart manufacturing

Abstract

Smart manufacturing has been developed since the introduction of Industry 4.0. It consists of resource sharing and networking, predictive engineering, and material and data analytics (Kusiak 2018). This development is gradually reducing human operations and replacing them with computerized systems capable of varying a system’s response to different situations and requirements. In the manufacturing system, a major concern of manufacturers is utilizing tools and machines to the end of their useful life before they are replaced, while also avoiding scrapped workpieces, downtime, and poor product finish due to system failure based on avoidable conditions (Tayal et al. 2021). This has opened the way for several research works on Tool Condition Monitoring (TCM) systems to reduce production cost, lower production downtime, and improve product quality output. Available literature has referred to TCM systems that predict tool condition by indicating tool failure from generalized key tool condition features focused on an indirect TCM system. According to Kamarthi, Kumara and Cohen (2000), tool wear is the most common phenomenon that is considered in several manufacturing processes such as drilling, turning and milling operations. An extensive literature review indicates that many research works have considered TCM and product quality output, while others have attempted integrating both product surface quality (roughness parameter) and TCM. The challenge with the implementation of these approaches is that product quality (precision in workpiece dimension and surface finish requirements) output is dynamic and the tool condition for each product quality requirement may thus differ. The limitation to these approaches is that the feedback system is not dynamic, which may indicate that the method will fail to generalize under different operating conditions such as product quality requirements, workpiece material, and cutting tool type. Furthermore, another major drawback with the present Tool Condition Monitoring (TCM) system is that it focuses on both the pre-failure and post-failure (i.e., after the start of a catastrophic failure) phases. Therefore, it is imperative to develop an approach that can self-compare real-time cutting conditions with the consideration of tool condition and the workpiece surface quality requirements for determining the output of the process in terms of the cutting tool condition and the workpiece surface finish properties. The research has developed a smart precision machining system that captures cutting tool conditions through a non-obstructive approach that incorporates and integrates smart sensors, IoT controllers, cloud computing for data capturing, a machine learning algorithm for data, and signal analyses for decision-making. The approach has developed a tool and workpiece condition monitoring (TWCM) system that indicates the condition of the tool and workpiece in real time during the turning operation through the classification of the condition into developed knowledgebased classes of tool and workpiece parameters. The system captures, processes and analyses realtime process data and features the installed IoT sensor network using advanced signal processing techniques and machine learning techniques for indicating the condition of both the tool and workpiece during cutting processes. To develop the TWCM system, generalized features of the tool and workpiece are first correlated, and the offline threshold of the parameters is captured for condition mapping and analysis. This study has non-obstructively classified the real-time condition of the tool and workpiece into known knowledge-based classes indicating the tool condition and the corresponding surface profile parameters output of the workpiece to determine and monitor the deviation from target output requirement in real-time during machining. This step in the approach focuses on first measuring the surface parameters (of the tool flank face) of the new cutting tool (100% life), used good tool, rough tool, and worn tool classes, and also the surface of the workpiece before commencing the turning operation on the lathe machine. This is to establish the classes of tools using ISO 3685:1993 standards based on their flank wear parameters. This was done using a surface and edge wear measuring device that measures twenty-one (21) parameters. The parameters were filtered based on their sensitivity to tool wear and workpiece surface finish using MANOVA analysis, hence six (6) parameters, namely, arithmetic mean roughness, 𝑅𝑎, mean roughness depth, 𝑅𝑧 , max valley depth, 𝑅𝑦, root mean square deviation, 𝑅𝑞, total height of roughness profile, 𝑅𝑡 , and max roughness depth, 𝑅𝑚𝑎𝑥 were selected based on the acceptance of null hypothesis on the condition that 𝑝 − 𝑣𝑎𝑙𝑢𝑒 is less than the alpha value (α = 0.025) of the MANOVA analysis. The threshold of these parameters from both the cutting tool and workpiece were classified into four (4) classes, which are the new tool, good tool, rough tool and worn tool. The corresponding vibration signal of the tool and workpiece during the turning operation was progressively captured using an advanced industrial vibration sensor, IoT gateway, and cloud server in real-time, which together form the cyber-physical system. This was done to progressively establish the resultant effect of tool wear conditions on the surface condition of the workpiece during operation. Thereafter, the parameters were captured experimentally at the same heartbeat (30 seconds) as the vibration sensor in the time domain. The wear conditions (measured parameters) were grouped into classes using the knowledge base gathered from the experimental result and advanced signal techniques were applied for feature extraction from the vibration signal as a means of classifying the output. Digital filters were first applied to the vibration data to eliminate the varying low contribution due to the alignment of the accelerometer (vibration sensor) to the gravitational field. Since the signal from the condition monitoring is non-stationary and nonlinear, the empirical mode decomposition (EMD) method was applied to the signals to separate the signals into components for detailed insight into the inherent features rather than estimating the Power Spectra Density (PSD) of the signals after applying the digital filters that apply FFT using a uniform trigonometric function (sine, cosine) for its analysis. The captured vibration signals were decomposed using EMD into a finite number of Intrinsic Mode Functions (IMFs) and residuals. Hilbert Huang Transform (HHT) was used to determine the instantaneous properties of the signal that were used for discriminating signals under different conditions. HHT was applied to the IMFs to evaluate instantaneous properties such as instantaneous frequencies, amplitude, and energy of the signal. A total of twelve (12) features were computed from the IMFs after applying HHT to the decomposed signals. The aim of the features was to precisely capture the tool and workpiece conditions by classifying the class of the extracted features from the analyzed vibration signal from both the tool and workpiece during the cutting operation. The process of classifying the features indicating the condition of the tool and workpiece during operation was done using a machine learning approach. To optimize the computational time and cost of the classification algorithm, the genetic algorithm (GA), using the Roulette Wheel (RW) method was used for feature selection. The convergence curve after 100 iterations showed that the model converged at the twenty-second (22nd) iteration even though the iteration still proceeded to 100 iterations as shown in Figure 5:5. Four features were selected from the twelve (12) computed features and these features were used for the ML classification algorithm. The classification was first performed using Neural Network Feed Forward Backprop with SCG algorithm. The input layers were four (4) while the output layers were four (4), with one hidden layer with eighteen (18) nodes in the network. Precision in manufacturing considering the tool and workpiece condition is reflected in the minutest range of several classes of parameters indicated through on-condition real-time signal analysis, which predicts the conditions accurately with a confidence level of about 89.8% and an error of 0.102 after 44 iterations. K-nearest neighbour (KNN) and Support Vector Machine (SVM) ML algorithms are also applied for classifying the tool and workpiece condition to evaluate the best classification algorithm in terms of performance. The K-fold cross-validation technique was applied and the error loss of each classification model was determined and plotted with K being five (5) and ten (10). With the 5-fold cross-validation, the overall error loss for the five SVM models was 0.5031, while for the KNN model it was 0.0318. This indicated that KNN models performed better under 5-fold cross-validation than SVM models and Feed Forward neural network with the SCG model for tool condition classification during the machining process. The overall error loss for the SVM models with 10-fold classification was 0.5009 while the error loss for KNN models was 0.0343, which also showed that KNN is a better model in terms of performance accuracy. Toledo-Pérez et al. (2019) reviewed the SVM-based model of EMG signal classification and reported that many sounds, vibration signals and images have been classified using the SVM classification algorithm, achieving more accuracy without feature selection and 5% less with feature selection. Therefore, to determine if the SVM model would perform better without feature selection, the models were evaluated with the 12 features, and the loss function was determined. The performance of both models for 5-fold cross-validation with all the feature vectors showed that for SVM models, the performance improved much more compared to when feature selection was implemented. The overall average error loss when 5-fold cross-validation was performed on all the features was 0.1668 compared to 0.5031 when feature selection was performed. However, for KNN models using 5-fold cross-validation with feature selection, the overall average error loss increased from 0.0318 to 0.2202. These results showed that while feature selection improved the performance of KNN models in classifying the conditions of the tool, it was not the case with the accuracy and performance of the SVM model. The 10-fold cross-validation error loss for SVM and KNN classification models developed without applying feature selection also indicated that the performance of SVM models was more accurate without feature selection. The error loss for 10-fold cross-validation for SVM models when feature selection was applied was 0.5009, while it was reduced to 0.1578 without feature selection. On the other hand, the performance of KNN models when feature selection was adopted was 0.0343, while it increased to 0.2172 without feature selection. Therefore, the KNN algorithm performed better overall in classifying the condition of the cutting tool during the machining operation with the KNN8 model being the best-performed model with an error loss of 0.0106. The fitted KNN8 was then optimized by applying hyperparameter optimization with the objective function being the error loss of the model and the constraints being the distance metrics. The optimal distance metrics using the kdtree neighbour searcher method was then determined. The ‘crossbar’ distance metrics were observed to minimize the 10-fold KNN8 model, while the estimated objective function value was 0.01416 and the observed objective function value was 0.014. The condition of the tool and the workpiece during operation can be classified at an interval of the sensor’s heartbeat which is 30 seconds but can be set to one (1) second. Since the system makes use of IoT-enabled industrial sensors, it implies that the condition of the tool and workpiece can both be remotely monitored and re-configured. The knowledge-based classification differentiates the condition of the tool and workpiece during operation into classes that detail the range of correlated surface profile parameter values with the corresponding tool parameters. Varying conditions of the tool can be matched with the product requirement and the classes and the tool can invariably be put to optimal use which makes this novel approach a useful method in precision manufacturing. Therefore, the model was again tested with a test set, and the error loss in the classification was evaluated as 0.0106. This research directly impacts the local manufacturing industries through product quality improvement by matching manufacturing operations with product quality requirements through real-time condition classification and avoiding a lower bound approach (damage to product), and also optimizing their system through optimal tool useful life usage and tool and workpiece wear condition classification.