A contribution to online tool wear detection using deep learning methodology

Milling is a machining operation that is among the most common activities in the manufacturing industry. The term machining refers to removal of unwanted material from the workpiece, so that a designed product with specific size, shape and surface quality can be manufactured. Manufacturing businesses currently have to cope with growing demand for increased productivity and production quality (Abellan-Nebot & Subirón (2010)). Tool wear is one of the most common faults of machining process and an important threat to those demands which can reduce efficiency and deteriorate the parts quality.

During the machining process, temperature and forces on the cutting tool edge are high. This may cause the cutting tool to gradually wears out and lose part of its material due to friction, thermal effects, abrasion, plastic deformation, diffusion, chemical wear, and grain-pullout, etc. There are several wear mechanism which may occur during the cutting process (Altintas (1992)) which can be listed as abrasion, adhesion, fatigue, crater, flank and chemical wear.

This may cause damage to the tool, machine and workpiece and may lead to unscheduled downtime. Downtime refers to the time that a machine is not working which is a considerable factor in efficiency of a production line. The downtime due to tool faults and breakage is a major portion of total downtime in a machining environment. Moreover, to diminish the wear damage risks, cutting tools are usually replaced earlier than their maximum life which it will increase the production cost (Quintana & Ciurana (2011)). Instead, due to comparatively high prices of tools, industry is looking for approaches to maximize the tool usage while keeping low excessive wear and breakage risk.

Tool wear can also deteriorate surface finish (roughness) and is a limitation for production quality. Dimensional accuracy of the manufactured part, surface finish quality and tolerances are import requirements which can be endangered by tool wear.

Sensors and Data Acquisition

A sensor is a device which measures a physical quantity and converts its energy from one form to another. Selecting the most appropriate sensor is an important step in machining monitoring .

The ideal sensor for monitoring purpose is the one that is highly sensitive to the monitoring parameters and insensitive to the other process parameters. Reliability, sensor cost, ease of use in industrial environment also are among the other features must be taken into account (Abellan-Nebot & Subirón (2010)).

Various sensors are available for machining monitoring, which factors such as cost, accuracy and monitoring purpose can influence the sensor selection. Overall, four main types of sensors are employed in most of the monitoring machining and tool wear researches. Force and torque sensors, accelerometers, AE/sound sensors and motor power and current sensors. A multi-sensory system with sensor fusion of multiple sensors also is used in the literature for monitoring.

Data acquisition refers to the process of sampling signals and converting the resulting samples into digital numeric values that can be manipulated by a computer. The signals should be filtered after signal acquisition within the range of the frequency response of the sensor. Signal sampling means the reduction of a continuous signal to a discrete signal. The sampling frequency should be high enough to enable a perfect signal reconstruction after sampling. There is a necessary relationship between the highest frequency contained in a signal and the minimum required sampling rate which is called the Nyquist sampling theorem. Based on this rule the sampling frequency should be at least twice the highest frequency contained in the signal being sampled (Siddhpura & Paurobally (2012), Hurmuzlu & Nwokah (2016)). This step may followed by some quality improvement methods and noise reductions.

Cutting Force

Cutting force is considered as the signal which best describes the cutting process in terms of accuracy and fast responses to changes in cutting conditions. Cutting force signals have been widely used for monitoring purposes in all machining processes (Turning, Milling, Drilling, etc.). It is very sensible to tool faults and can reflect the fault very fast (Rehorn et al. (2005)). This signal can be used to monitor a wide range of machining variables such as tool wear, tool breakage/chipping, chatter vibration, and the quality and geometric profile of the cutting surface which makes it a perfect candidate to be used in a comprehensive machining monitoring system (Freyer et al. (2014), Zhang et al. (2012)). Torque signals are also used as another representation of force in machining monitoring application.

One of the drawbacks of using force signals is that they are highly dependent on cutting conditions, cutting material, workpiece material, etc. Therefore, an in increase in cutting force due to a fault is strongly dependent on other cutting variables. Furthermore, they are relatively expensive, difficult to install and their limited frequency response are among their limitations for industrial usage (Abellan-Nebot & Subirón (2010)), therefore they are not used in industry. As a research example, Wang et al. in a study employed force sensors and a distributed Gaussian ARTMAP (DGAM) network to solve non-uniformly distributed shapes and complex category boundaries problem for tool wear monitoring. They reported that, the classifier is insensitive to the noisy data and suitable for non-uniformly distributed data (Wang et al. (2013)).

Vibration

Interaction between workpiece and cutting tool produces vibration. Acceleration, velocity, or displacement sensors can measure vibration. Vibration is a significant factor affecting tool conditions, surface roughness, and dimensional accuracy in machining processes. As tool wears out or in case of tool breakage and other machining faults, amplitude and frequency of vibration signals changes which can be employed for monitoring purpose.

Many researches employed vibration signal for different applications of machining monitoring such as tool wear and breakage detection, surface quality control and chatter. The advantages of accelerometers are their simplicity and low cost. One of accelerometers practical problems is that it is sensitive to machining speed and it should remain within a specific range during the monitoring. Moreover, the placement of the sensor is a difficult task since it requires prior knowledge of the dynamic behavior of the machine tool(Delio et al. (1992)).

In an study by Gangadhar et al. condition of a single point cutting tool is monitored with help of the vibration signals acquired from an accelerometer (Gangadhar et al. (2014)). They obtained statistical features from vibration and the significant features were chosen using J48 algorithm. The accuracy of their monitoring system was 89.38% in distinguishing healthy and worn tools.

Lamraoui et al. (Lamraoui et al. (2015)) developed a chatter detection method in milling machine using vibration signals and cyclo-stationary analysis. They employed multi-band resonance filtering-envelope analysis. They claimed that chatter detection using dynamic cutting forces is not practical in industrial environments and used 3 accelerometers: 2 at spindle support in X and Y directions and 1 attached to work piece in cutting direction. Modal parameters are extracted using tap test and SLD is obtained in this research.