Soft sensor for in-line quality control of turning processes based on non-destructive testing techniques and advanced data fusion

When processing quenched and tempered steels a variety of surface changes can occur, with a particular focus on the formation of White Layers. These White Layers exhibit a microstructure characterized by an ultra-fine-grained appearance on the surface. In order to effectively control surface conditions throughout the machining process, a thorough understanding of the underlying mechanisms is essential. The importance of using quantitative assessment techniques and models for in-process monitoring of surface integrity is highlighted. These findings are supported by various studies. Gauder et al. [1] investigated turning processes involving AISI 4140 QT steel using multivariate regression. Böttger et al. [2] introduced a soft sensor concept for predicting surface layer modifications during longitudinal turning of AISI 4140. Softs sensor is defined as Their approach included tool wear and component models based on process forces, acoustic emission, and micromagnetic material parameters. Combining one or more physical measurement principles linked with expert knowledge and mathematical models is called a soft sensor [3]. Meurer et al. [4] worked extensively on orthogonal cutting of AISI 4140 QT steel, concentrating on dynamic recrystallization and the development of White Layer. Their analytical model, based on process forces and temperature fields, facilitated the prediction of White Layer formation. Sada [5] used neural networks to predict surface conditions of mild steel after turning. Based on process parameters, their approach provided accurate forecasts for material removal rate and surface roughness. Zemzemi et al. [6] tackled the prediction of White Layer during orthogonal turning of AISI 52100 hardened steel. Their research involved the integration of different models to anticipate cutting forces while analyzing their correlation with temperature distribution and mechanical stresses. Kuntoğlu et al. [7] studied the impact of cutting parameters on vibration and roughness when turning AISI 5140 steel. Using the response surface methodology and regression models, they optimized parameters and achieved precise predictions. Uhlmann et al. [8] studied the interaction between tool wear and machining parameters when milling AISI 4140 steel. Their work demonstrated the utility of power measurements for real-time tool wear detection and used machine learning to predict magnetic properties based on machining parameters and tool wear. Ankener and Böttger [9] performed a comprehensive micromagnetic and microstructural characterization of ferromagnetic steels under different heat treatment conditions. Their work showed the relationship between properties such as hardness, micromagnetic properties (BR, HCm, Bmax), and retained austenite content (RA). Stampfer [10, 11] discovered that lower base hardness, indicating reduced dislocation densities, is associated with higher susceptibility to predominantly mechanically induced loads compared to higher base hardness. Stampfer’s analysis of AISI 4140 QT samples revealed that an increase in hardness is not always a reliable measure of surface layer modifications. This phenomenon is attributed to tempering effects caused by exceeding the tempering temperature during machining. Key findings include: X-ray analysis showed a reduction in the average FWHM (full width at half maximum) below the base material level, reflecting tempering-induced decreases in dislocation density and, consequently, hardness. Near the surface the effect is amplified by mechanically induced hardening, leading to an increase in FWHM and hardness. Significant thermal and mechanical loads can induce surface layer modifications, and their impact on hardness can be partially or fully compensated, which explains the limited sensitivity of hardness increase to process parameters compared to other tempering levels. In summary, the studies mentioned above improve our understanding of the effects of machining on steel surfaces and propose models to predict surface changes, tool wear and relevant parameters. However, there is currently a gap in the development of a model for in-situ prediction of UFG formation in AISI 4140 steel using micromagnetic.

The starting material is low-alloy, lower-eutectoid quenched and tempered steel AISI 4140 (German standard 42CrMo4 Q+T). The chemical composition is listed in Table 1. The sample was austenitized in a vacuum furnace and quenched. The cooling rate achieved in this way was comparable to the cooling rate of conventional oil quenching. During hardening, the austenitizing temperature was 850 \(^{\circ }\)C. After quenching, the specimens were annealed by holding them at different tempering temperatures of 300 \(^{\circ }\)C, 450 \(^{\circ }\)C and 650 \(^{\circ }\)C for 60 min. As a result, high strengths are achieved with high toughness at the same time, which ensures suitability for stresses acting on the machined components. Due to the variation of the different tempering temperatures, the specimen show a different microstructure.

As previously mentioned, the domains of ferromagnetic materials lead to the formation of nonlinear hysteresis, which forms the basis of our output parameters. Experiments were conducted on ten different samples, each of which had six different surfaces created with different cutting velocities and feed rates.

Table 2 and Fig. 1 show the cutting velocities \({v_{c}}\) and feed rates f of the differerent section of the workpieces. Three of the ten samples were subjected to tempering temperatures of 300 \(^{\circ }\)C, four to 450 \(^{\circ }\)C, and three to 600 \(^{\circ }\)C as shown in Table 3. The total length of a sample was 180 mm, but measurements started at 9 mm and continued until the sensor head reached 155 mm. The experiments did not start from the zero-point due to the initial need for an area for the sensor head. The total length of the measured area was 146 mm, and the diameter of the sample was 48 mm. Abrasive tool wear on the flank face was measured after each experiment.

As a result of the turning operation, properties of the material such as surface quality, residual stress, and hardness exhibit micro-scale changes. The purpose of non-destructive material evaluation is to model these changes using sensor output values. The samples were exposed to turning operations using cutting tools with abrasive flank wear ranging from 0 to 400 \(\upmu \)m. Later, the samples were clamped in the turning machine and measured as shown in Fig. 2. All the samples were subjected to measurements using the 3MA sensor at room temperature (24 \(^{\circ }\)C) to confirm that the heating did not result in any microstructural changes.

During the experiments, variable magnetizations were utilized. The internal structure of the sensor head is shown in Fig. 3. The magnetic yoke (3) driven by the magnetization coil (2) leads to a low-frequent, high-amplitude magnetization in the sample. A Hall probe (6) and transmitter / receiver coils (7, 8) are attached in the middle between the pole shoes of the magnetic yoke. The coils are responsible for generating the high-frequent, low-amplitude magnetization for the Eddy-Current (EC) and the Incremental Permeability (IP) methods as well as for receiving the signals from EC, IP and the Barkhausen Noise (BN) method. The Hall probe is used for detecting the tangential magnetic field needed for the Harmonic Analysis (HA) method and for the control of the magnetization amplitude for all methods. Table 4 shows the cutting parameters of the components for modelling White Layer formation based on 3MA data.

A key finding of the study presented here is the need to address lift-off compensation and in-process detection of fine-grained microstructures known as White Layers on AISI4140 through newly developed training, testing and validation methods based on 3MA data. This approach enables the prediction of product quality and the identification of undesirable surface conditions during turning processes. The development of the \({\text {UFG}_{\text {3MA}}}\) parameter plays a central role in the creation of a soft sensor. In summary, the control of the process is initiated when the value surpasses the unfavorable range (NOK) leading to adjustments in process parameters aimed at maintaining the \({\text {UFG}_{\text {3MA}}}\) parameter within or near the critical range, thereby facilitating high production efficiency. Different control models suited for establishing a robust closed-loop control system aimed at managing adverse surface conditions during the turning production process in real-time in the next section.

Starting from the control system where the machine parameters are readjusted between the individual cuts, an online control system for the turning process is to be implemented in the future. Building on the study of the 3MA variables and their influence on the characterisation of machining processes, this section provides a comprehensive discourse on quality control strategies. The primary goal is to proactively control and prevent the formation of UFG while archiving the maximum material removal rate and therefore maintaining the lowest possible production times. In order to do so a soft sensor will be developed with input of 3MA and output process control variables. Process control implementation necessitates both hardware and software to manipulate process parameters. In the case of CNC turning machines, the goal is to link the feed rate and cutting velocity overrides with NC code parameters and externally regulate them through an industrial PC. Converting discrepancies in target values to control variables involves the application of control design principles. For this, the three control theories, PID controller, Model Predictive Controller (MPC) and Reinforcement Learning (RL) were compared. The widely used PID controller is effective for simple linear systems but falls short in handling complex tasks and non-linear systems [13]. Despite its adaptability, limitations arise in addressing uncertainties and time delays [14]. MPC, formulated as an optimization problem, proves promising for complex tasks by considering multiple objectives and constraints [15]. While requiring an accurate model and careful modeling, MPC’s theoretical applicability to various systems makes it a viable choice for cutting process control [16]. RL, with adaptive learning properties, excels in handling complex dynamics but faces challenges in setting hyperparameters and requiring substantial data for optimization [17‐19]. MPC emerges as the most suitable option for precise control of cutting processes due to its applicability to demanding and complex tasks [20]. It allows simultaneous optimization of multiple quality measures and simplifies problem formulation through mathematical objectives and constraints [21].