1 Introduction
Recently, new cutting processes such as hard turning, high speed cutting, and heavy duty cutting have been widely used, forcing cutting tools to undertake many challenges from high cutting temperatures and large cutting forces [1]. Tool wear and failure constitute about 20% of the total downtime, making cutting process monitoring quite substantial [2,3]. Signals such as cutting forces, acoustic emissions (AEs), vibrations, and images produced in cutting processes, together with multiple signal processing algorithms such as frequency domain analysis, wavelet transform, and empirical mode decomposition, are widely used. Based on that, machine learning methods such as support vector machine (SVM) and artificial neural network (ANN) are constantly introduced to construct the mappings between signals and tool conditions [4,5]. Özel and Nadgir [6] noticed that tool wear affects the proportion of every force component and used a three-layer fully connected neural network to conduct the wear fitting. Yen et al. [7] adopted a self-organizing feature map (SOM) to pretreat the AE acquired in micro milling. As a comparison, the SOM performed better than traditional methods when extracting features. Hsieh et al. [8] and Wang et al. [9] applied a multilayer perceptron and an SVM, respectively, for vibration distinguishing. Mikołajczyk et al. [10,11] used two neural networks. The first one conducted image binarizations for wear length determinations, based on which another network provided tool life predictions. Bergs et al. [12] adopted a convolutional neural network (CNN) for tool type identification and tool wear estimation.
The above methods have been widely discussed and applied, but selections and improvements are still necessary. For cutting force, the measurement (usually conducted by piezoelectric dynamometers coming from Kistler) is very expensive for industrial application. It requires many modifications in the tool turret, which also brings limits. For AE, the acquisition requires very high sampling rates due to its wide frequency range, which causes heavy burdens of transmission and storage. Signals such as vibrations and sounds are easily disturbed by noises. Previous studies noted that factors such as workpiece dimensions and system rigidities sometimes have more effects on vibrations than the tool itself [13]. The position of the sensor also largely affects the obtained signals, and Herwan et al. [14] even brought a discussion about the position to place the accelerometer. Though evidence such as surface roughness and wear images provide the most direct information about wear, they are inapplicable for a real-time system because they can only be acquired during the intervals of cutting processes. Finding a perfect method that suits all situations is difficult, and weighing pros and cons according to specific circumstances is needed.
Cutting temperatures, caused by cutting heats, are closely tied to extrusions, frictions, and deformations in cutting processes. High temperatures may soften cutting tools, and alternating temperature loads often leads to cracks in tools. Choudhury and Bartarya [15] applied a natural thermal couple to obtain cutting temperatures and used a neural network to conduct temperature-wear fittings. He et al. [16] also adopted thermal couples to acquire cutting temperatures and applied deep learning technics for wear prediction. However, factors that greatly affect temperatures such as process parameters and cutting lengths were strictly controlled, which indicated the difficulties in existing approaches to acquire and utilize temperatures. The temperature value at a moment is the result of heat balance and changes following the floating of the balance. The size of the workpiece and the coolant are reported to affect cutting temperatures much [17,18]. The relation between temperatures and tool wear seems to have great indeterminacy due to these irrelevant factors. However, if comparing the heat productions with their dissipations, another point of view is found. The heats completely come from the energy consumption when cutting and share 98%–99% of it. Thus, heat productions are directly affected by cutting fluctuations. When the processes are smooth, the fluctuation amplitudes are low, and the frequency distributions are even. When tool wear causes trembles, the amplitudes may grow, and abnormal peaks may appear at several frequencies. On the contrary, heat dissipations are affected by conduction and convection, which vary much slower and possess randomness. Owing to the differences between them, fluctuations are eventually presented by temperatures, that is, temperature undulations may have clear correlations with cutting conditions. For further discussions, an appropriate measuring method is needed.
Commonly used approaches such as thermocouples and infrared detections have several shortcomings. A natural thermal couple employs the tool and the workpiece as its two electrodes, which can only be applied under certain circumstances [19]. Thereafter, thin film thermocouples (TFTCs) on tool surfaces are proposed [20,21]. However, preventing TFTCs from damages in the cutting processes is difficult. Moreover, thermocouple-type measuring methods possess different degrees of thermal inertia, which limits the responsiveness facing alternating temperatures. Thus, temperature is regarded as a kind of slowly varying signal and is not fully utilized in many applications. Infrared measuring techniques with better responsiveness can meet the requirement of capturing the undulation temperatures and have been successfully applied [22,23]. The remaining problem is that the coolant and the chip may disturb the transmission of the radiations.
In this paper, a real-time tool condition monitoring method taking advantages of a novel in situ fiber optic temperature measuring system and an ANN is proposed. The fiber optic is embedded into the tool, ensuring transmission stability. A thermal simulation and an intermittent cutting experiment are carried out to prove that the measuring system has sufficient sensitivity facing rapid but slight cutting heat undulations. A fully connected neural network with straightforward structure is used for wear distinguishing. A learning rate adaption algorithm is used for optimizing the back propagation, reducing the dependence on the hyper parameters and the initial conditions. Thus, training and monitoring can be integrated in one program with graphical user interface, making the method more suitable for practical application. Other than giving alarms, the temperatures can provide more information when a human is on duty because they directly reflect cutting loads. The method has remarkable advantages such as low cost in measurement, decreased difficulty in operation, and fine stability, supplementing the industry with new ideas about tool condition monitoring.
2 Temperature measurement
2.1 System development
An in situ optical fiber measuring apparatus is built to meet the requirement of temperature acquisition. The apparatus consists of a set of customized tool and an optical path. Figure 1(a) shows that the tool insert and the tool holder are modified based on Sandvik TPGN-160308 cemented carbide insert and Shangin CTGPR-2525 tool holder series, respectively. Three micro holes with 0.6 mm diameters approaching three cutting edges are added to the insert. For each of them, the axis is 0.8 mm from the intersection line of the major flank and the minor flank while the end is 0.3 mm from the rake face. An optical fiber is embedded through it and fixed by a fixture. Experiments confirmed that the holes’ addition does not evidently decrease the tool rigidity and the service life. For experimental application, the holes are drilled by electrical discharge machining or laser grinding. If going into batch production, the holes can be reserved in powder metallurgy process. When machining, the tool insert is heated up to a high temperature and produces intense infrared radiations. The radiations are captured by the optical fiber probe, transmitted along the optical fiber jumper, and finally input to the optical platform. Figure 1(b) shows that the radiations’ transmissions in the platform are identified by the red arrows. They are shaped to a parallel beam by a plano-convex lens and separated into two orthogonal parts later by a beam splitter. When passing through the filters, the components whose central wavelengths are 2000 and 2500 nm remain while the remaining parts are removed. Then, the two beams reach 2 InGaAs amplified photodetectors (PDA10DT-EC, Thorlabs, Inc., USA) for photoelectric conversions. The voltages that represent radiation intensities are eventually converted into digital signals by a data acquisition card (Ni-9231, National Instruments Corp., USA) and input into a PC for temperature calculations. The temperatures are determined by comparing the radiation intensities at two different wavelengths.
For a kind of gray body, the radiation it produces can be described by the following formula:
where radiation intensity L is a function of wavelength λ and temperature T, ε is wavelength-dependent emissivity, and h, k, e, and c represent the Planck constant, the Boltzmann constant, the elementary charge, and the light velocity, respectively. When the wavelength is located in the infrared section and the temperature is lower than thousands of degrees Kelvin, the formula can be simplified to Wien’s form (Eq. (2)) because .
where and are the constants that contain values such as the Planck constant and the Boltzmann constant. If building relationships between temperatures and radiation intensities directly, emissivity calibration for every tool is needed, but things become easier when comparing the radiation intensities at two different wavelengths (λ1 and λ2). The ratio of the two intensities (R) can be expressed by Eq. (3):
Considering that the two wave bands are close to each other, the corresponding emissivities possess approximately equal values. Thus, a relation between the ratios and the temperatures that does not depend on emissivities can be obtained, as shown by Eqs. (4) and (5):
Not only emissivity but also those factors such as transmission attenuation are almost counteracted when calculating the ratio, which decreases the dependence on calibration and improves the measuring repeatability. A K-type thermocouple is used for calibrating the outputs of two photodetectors, which indicates that the complex expression can be replaced by a linear relation for engineering application [24]. Limited by the sensitivity of the photodetectors and the cut-off frequencies of the filters, the range of the measuring temperature is from 220 to 1000 °C. The photos of the customized tool, the mounting of the tool on the turret and the optical platform are shown in Figs. 1(c)‒1(e), respectively. Subsequent experiments are developed in a turning center (QT200N series, LGMazak, Japan). More details about the apparatus are provided in Ref. [24].
2.2 Transient thermal simulation
A commercial finite element analysis software (Ansys) is used for a transient thermal simulation to verify how the cutting heat affects the location a small distance away from the cutting interface. The finite element model and its boundary conditions are depicted in Fig. 2(a), where Parts I, II, and III represent the tool holder, the tool insert, and the insert fixture, respectively. They are reconstructed according to their real configurations. Part IV is the tool turret with several simplifications. In actual turning processes, the tool turret and the lathe bed absorb a substantial part of heat while remaining cool because of their large heat capacities. Thus, Part IV is introduced as a heat storage for simulating this situation, instead of imitating the real shape. Three surfaces at its end are regarded as isothermal surfaces whose temperatures are equal to those of the environment. Most of the other surfaces are exposed to air and treated as convection boundaries. The convective heat transfer coefficient is set to 50 W/(m2·K) [25]. A very small area (marked with slashes) is behind the cutting edge, representing a tool–chip interface. The shape and the size of the area are determined according to the rake face photos obtained in the cutting experiments. To simplify the simulation, the heat flows can be assumed evenly distributed on the interface, and the specific thermal loads will be discussed later. The contact thermal conductivity between every two parts is set to 2550 W/(m2·K).
Two kinds of materials are involved, namely, cemented carbide for the tool insert and AISI 1045 steel for the other parts. AISI 1045 steel possesses relatively fixed thermo–physical properties, and its thermal conductivity, heat capacity, and density are 45 W/(m·K), 473 J/(kg·K), and 7860 kg/m3, respectively. Carbide insert’s density is 14830 kg/m3, and several other temperature-dependent thermal physical properties are provided in Table 1 [26].
Tab.1 Temperature-dependent thermo–physical properties of carbide insert |
| Temperature/°C | Thermal conductivity/(W·m−1·K−1) | Heat capacity/(J·kg−1·K−1) |
|---|---|---|
| 20 | 110 | 220 |
| 100 | 105 | 244 |
| 300 | 98 | 290 |
| 500 | 90 | 320 |
| 700 | 82 | 328 |
| 900 | 75 | 337 |
Figure 2(b) shows the meshed finite element model, where hexahedron elements are used. The sizes of the elements range from 0.1 to 4 mm. Details are provided in Table 2. Parts with larger temperature gradients or more complex configurations are meshed more densely.
Tab.2 Sizes of meshes for different parts of the tool system |
| Part | Size/mm |
|---|---|
| Tool turret | 4.0 |
| Tool holder | 2.0 |
| Insert fixture | 0.5 |
| Tool insert | 0.5 |
| Tool–chip interface | 0.1 |
Considering that the system needs a while to heat up, the simulation time is set to 0.5 s. According to the experiences obtained when machining, the heat flows loaded to the model should fluctuate at a frequency of 50 Hz and in a range from 40 to 42 W. Figure 2(c) shows that the heat flows change 25 times during the whole process. For transient simulation, one iteration can be regarded as one sampling. To meet the requirement of sampling rate in engineering (i.e., five times the signal bandwidth), 250 iterations are used in the process.
The temperatures at the location that the fiber optic probe can reach are investigated. Figure 2(d) shows that the simulated temperatures keep rising as a whole but fluctuate at the same frequency as the heat flows. The temperatures at the measuring position positively follow the rapidly changing heat flows, indicating that the floating cutting heats can be reflected by temperatures whether equilibrium between heat productions and dissipation is achieved or not.
2.3 Intermittent cutting experiment
An intermittent cutting experiment is carried out to investigate if the optical fiber measuring system can capture the fast but slight fluctuations. A steel rod with 12 grooves around (Fig. 3(a)) is used for producing pulsing heat flows. The diameter is 90 mm, and the available length is 120 mm. The heat flows suddenly decrease when the tool goes through a groove. To generate fluctuations whose frequency is no less than that in simulation and ensure an appropriate cutting speed, the spindle rate is set to 300 r/min. At this speed, the tool goes through the grooves 60 times per second and generates 60 heat flow declines at the same time. The feed rate and the cutting depth are set to 0.1 mm/r and 1 mm, respectively. Whether the obtained temperatures reflect the fluctuations is noteworthy.
A sequence of temperatures is obtained and plotted in Fig. 3(b). (Limited by the sensitivity of the photodetectors, the temperatures lower than 220 °C are not exactly displayed.) Its frequency domain form is shown in Fig. 3(c). Peak I at 60 Hz is distinct, indicating that the apparatus successfully captures the fast but slight undulations. However, several peaks at other frequencies also appear and cannot be interpreted as harmonics of the first one. Considering that they are around 180, 360, 190, and 380 Hz, Peaks III and V can be inferred as harmonics of Peaks II and IV. To explain this phenomenon, the raw voltages given by two photodetectors before cutting are both checked. Figures 3(c) and 3(d) show that voltages produced by the 2000 nm photodetector possess strong 180 and 360 Hz components, which have evident correlations with Peaks II and III detected previously. Similarly, voltages given by another photodetector have correlations with Peaks IV and V. Those unexpected peaks appear before the tool comes into contact with the workpiece, which do not profit measuring and monitoring. To avoid frequency aliasing, the sampling rate for temperature acquisition must be at least twice the noise frequencies (i.e., approximately 760 Hz).
A straightforward approach to evaluate the undulations is fast Fourier transform (FFT), which reveals regular quivers of the signals at several frequencies. More complicatedly, several time–frequency analysis methods such as short time Fourier transform and discrete wavelet transform can be deployed, which consider time resolution and frequency resolution while bringing larger computation consumptions. Whether a time–frequency analysis method is needed depends on the stationary behavior of the signals. A sequence lasting for 30 s is intercepted from the temperatures when the cutting is relatively stable, and its statistical features are discussed. Figure 4(a) shows that the standard deviations, the kurtoses, and the skewness of the sequence are randomly distributed, indicating that the statistical features do not vary with time or change very slowly. Thus, the cutting temperature can be regarded as a kind of stationary signal in most instances (except the moment the tool contacts with or retracts from the workpiece). Time resolution does not need to be considered, and FFT is sufficient for feature extractions. Pre-experiments [24] indicate that tool life can be divided into four stages according to its flank wear. In the three previous stages, the flank wear rapidly rises and keeps around 0.3 mm for a long time. In the last stage, the flank wear starts to grow again quickly, and the cutting starts to deteriorate evidently. The last stage can be considered to represent the condition of failure. Therefore, a series of cutting is conducted by a new tool and a tool whose flank wear is larger than 0.3 mm to explore the possible differences in corresponding temperature data. Five different speeds are applied to involve varying frequencies. The temperatures are converted into spectrograms whose resolution is 1 Hz, where the parts below 5 Hz are discarded to remove slow-varying, low-frequency components. Figure 4(b) shows that yellower regions that mean higher amplitudes are clearly visible, illustrating that the resolution is sufficient to reveal differences between two kinds of tools. The temperatures obtained by a worn tool show two peaks in the frequency domain while the others are smooth. When the cutting parameters change, the peaks appear at different positions but remain in the selected range. In this range, the measuring apparatus has good responsiveness and avoids the intrinsic disturbances from the photodetectors. Comprehensively considered, 45 amplitudes at integer frequencies between 6 and 50 Hz are selected as a group of features for tool condition analyses.
3 Tool condition analysis
3.1 Neural network
Though novel machine learning models such as CNN and recurrent neural network and their superiorities have been widely discussed, the fully connected feedforward neural network is still a proper approach under certain circumstances. The strong nonlinear mapping ability is sufficient for handling 1D signals, and low calculation consumption ensures fine practicability. Decreasing the uncertainties brought by hyper parameters instead of pursuing the complexities of the network is urgent. In this paper, a four-layer feedforward neural network depicted in Fig. 5(a) is introduced to undertake tool condition classification. The input layer consists of 45 nodes that accept 45 amplitudes. Two hidden layers possessing adjustable dimensions follow. Hyperbolic tangent activation function is applied in the three previous layers. The output layer has two nodes, each of which produces a component to form an output vector. SoftMax activation function is used in the last layer to ensure the sum of all the outputs is 1. Multi classification cross entropy loss function given by Eq. (6) is applied to evaluate output deviation L:
Fig.5 (a) Architecture of neural network, (b) changes in losses when adopting different learning rates (using common back propagation algorithm), (c) changes in losses when adopting different learning rates (using learning rate adaption algorithm), and (d) contrast of losses when applying different hidden layer sizes (using learning rate adaption algorithm). |
where N is the number of categories, which is 2 in this paper, indicating the two states of a tool, and is an output vector, which is a fitting of the target vector y. Only two cases are allowed for target vectors: (0, 1) represents healthy samples, whereas (1, 0) represents worn samples. The first component means the probability of being worn, and the second one is the contrary.
An inappropriate learning rate often leads to failure in convergence, which is unacceptable in a practical system. Five hundred samples are fed into the neural network, and 3 different learning rates are adopted to carry out a test. The initial loss is 0.7225, and the losses at ends of every other 50 iterations are shown in Fig. 5(b). The losses increase to quite large values and begin to vibrate when the learning rates (abbreviated to lr in the figure, the same below) are 0.1 and 0.01. Though the network can converge when the learning rate is small enough, the relatively flat curve indicates a relatively low training efficiency. Determining an appropriate learning rate is even more difficult if several changes happen to the scale of the network or the input samples. Here, a learning rate adaption algorithm is introduced to decrease the dependence on hyper parameters. The method can be considered as a kind of linear search, which enables a neural network to determine a step size automatically during descent. The neural network discards the current iteration, is restored to the previous stage, and reduces the step size if deviation increases. By contrast, the neural network cautiously adopts greater step sizes to accelerate descent if the loss keeps decreasing. Figure 5(c) shows that the losses fast converge to 10−6 magnitude, which is irrelevant to initial learning rates. The requirements on other initial conditions also become less strict. Different hidden layer scales are applied, and corresponding training curves are shown in Fig. 5(d). The 45-10-10-2 structure means every hidden layer has 10 neurons, and so on. The discrepancy in losses tend to become smaller after several hundreds of iterations, indicating that the reduced difficulty in determining an appropriate structure. A small scale can be selected as far as possible to avoid overfitting. For the following tool condition classification, a 45-18-18-2 structure is applied.
3.2 Classification test
Titanium alloy has high strength and low thermal conductivity, bringing difficulties in machining. Monitoring the machining of titanium alloy appears to be substantial. A titanium alloy (Ti-6Al-4V) bar is used for a series of tests. A new tool and a worn tool are used for carrying out five periods of cutting separately. (The new tool means a tool whose flank wear is much less than 0.3 mm, and the worn tool means a tool whose flank wear is larger than 0.3 mm.) Every period lasts for 40 s, and corresponding temperatures are obtained as classification experiences. The cutting speed and the cutting depth are 80 m/min and 1 mm, respectively. Five different feed rates evenly ranging from 0.1 to 0.14 mm/r are adopted to simulate a variable feed rate process. The temperatures keep rising in a whole period of cutting because reaching an equilibrium between heat production and heat dissipation takes a long time, that is, the temperatures are simultaneously affected by the wear condition, the feed rate, and the time of cutting, which can be described by a box diagram in Fig. 6(a). The effects of the feed rates on the temperatures are complex, and the temporal distribution degrees are uncertain. As the cutting time increases, the temperatures probably distribute in a wider range, making distinguishing a worn tool using raw temperature values difficult. The temperature sequences are then divided into segments and converted into a frequency domain to form a training set. To produce sufficient segments using a sequence with a limited length, overlaps between every two segments are allowed. In actual operation, the head and the rear of each sequence are discarded to reduce the effects brought by the tool’s feeding in and retracting while the middle part lasting for 37 s is retained for sampling. One hundred segments are intercepted from one temperature sequence, and 1000 segments are eventually formed from the whole 10 sequences. Figure 6(b) shows that every segment is converted into its frequency domain, where stable differences irrelevant to feed rates and sampling time appear between two kinds of tools. They are all sent to the neural network for training.
Fig.6 (a) Temporal distributed temperatures under different feed rates (for training), (b) frequency domain samples produced by two kinds of tools and under different feed rates, (c) temporal distributed temperatures under different feed rates (for test), and (d) contrasts of ANN outputs and expected values. |
Four other periods of cutting are carried out by two tools for a test, and every period lasts for 25 s. The feed rates are set to 0.115 and 0.125 mm/r, which are not adopted before. The cutting speed and the cutting depth remain unchanged. Corresponding temperatures are presented in Fig. 6(c). The head and the rear of every sequence are discarded while 50 segments are generated from the 22-second middle part. Two hundred samples are collected and input into the neural network. For a healthy tool, the first component of the output vector is expected to be 0. For a worn one, the first component is expected to be 1. Expected values and actual outputs are plotted in Fig. 6(d), where the scatter points represent the outputs provided by the ANN, and the lines represent the expected values. None of the outputs evidently deviates from the expected values, indicating a successful tool condition classification.
Things become more complicated as variable speed processes because those peaks move following spindle speeds, bringing more confusion information. However, considering that the workpiece diameters also influence the spindle speeds, a variable speed process better simulates practical machining. Thus, a normal tool and a worn tool are used for carrying out five periods of cutting separately, carrying out another experiment. The feed rate and the cutting depth are 0.1 mm/r and 1 mm while five different cutting speeds evenly range from 60 to 100 m/min. The cutting time remains 40 s. Corresponding temperatures are expressed by a box diagram in Fig. 7(a). As expected, the temperatures not only increase following cutting speeds but also distribute over cutting time. Overlaps are observed between the temperatures produced by two kinds of tools, which is probably because the differences of wear degrees are not excessively large. The same as the previous experiment, 1000 segments are extracted from the whole 10 sequences and converted into spectrums to make up a training set. Figure 7(b) shows that the yellower regions distribute obliquely as expected, indicating that the spindle speeds affect the feature peaks.
Fig.7 (a) Temporal distributed temperatures under different cutting speeds (for training), (b) frequency domain samples produced by two kinds of tools and under different cutting speeds, (c) temporal distributed temperatures under different cutting speeds (for test), and (d) contrasts of ANN outputs and expected values. |
Four other periods of cutting, each lasting for 25 s, are carried out for test. The feed rate and the cutting depth remain unchanged while the cutting speeds are set to 75 and 85 m/min. Figure 7(c) shows the temperatures. Two hundred samples are then collected and input into the neural network. The classification results are shown in Fig. 7(d), where the scatter points represent the outputs provided by ANN, and the lines represent the expected values. If regarding 0.5 as the threshold, a 90% accuracy is obtained for the healthy tool, and a 96% accuracy is obtained for the worn tool. Though a mild reduction in accuracy is observed compared with the previous experiment, it makes correct decisions most of the time, which can provide valuable assistance in practical machining.
To apply the classification ability to tool condition monitoring, another examination is conducted to track the outputs of the neural network in a whole cutting process. The cutting speed, the feed rate, and the cutting depth are 100 m/min, 0.16 mm/r, and 1.0 mm, respectively. A normal tool and a worn tool are used for carrying out four periods of cutting, each lasting for 30 s, to produce 800 samples for training. As the monitored object, the third tool with a good initial state works continuously until the cutting condition evidently changes. Owing to the limitation of the workpiece length, the tool is required to retract from the workpiece to begin another period every 80 s. Eventually, five periods of cutting are conducted, and the relatively stable middle part lasting for 70 s of every period is intercepted to splice into a continuous sequence lasting for 350 s, as shown in Fig. 8(a). As a control, vibration signals are recorded at the same time by an accelerometer (356A24, PCB Piezotronics, USA). The time frequency diagram of the vibrations is shown in Fig. 8(b), where several changes happen to the spectral distribution at the 212th second, indicating that the cutting stability is disrupted.
As a verification, 350 samples are generated in chronological order from the spliced sequence and sent to the trained classifier. The outputs are plotted by scatters in Fig. 8(c). In the range of 0–220 s, all the outputs are 0 except five isolated points, indicating that the tool works smoothly. From the 220th second, the classifier intensively outputs 1, reflecting the change of the cutting condition. Though the exports of the classifier slightly lag behind the vibrations, an approximate agreement between two measures is obtained, confirming the reliability of our method. The result does not mean that the vibration performs better in tool condition monitoring. In actual application, vibration acquisition is easily affected by the position of the sensor. Figure 8(d) shows that repeated mounting and dismounting of the accelerometer remarkably affect the obtained vibration. The method based on cutting temperature successfully monitors the third tool utilizing the previous experiences, instead of being affected by mounting and dismounting.
3.3 Application program
A software with graphical user interface is constructed specially for the optical fiber measuring system, realizing the integration of measuring, training, and monitoring. It can be used to train a customized neural network for specific tools and workpieces. To collect samples for training, the first step is to conduct cutting experiments separately using two kinds of tools and record the corresponding files. Figure 9(a) shows that 10 files representing two kinds of tools are loaded, and their labels are presented on the far left of the interface. With these files prepared, a control panel shown in Fig. 9(b) provides guidance in training a neural network. The operator is demanded to click the “Add” button, informing the software of what kind of samples will be added (healthy or worn) and then drag the cursor on the displayed curve, appointing a range for sampling. A limit of iteration times can be assigned. The training ends when an acceptable loss is obtained or the limit is reached. With the training finished, the software can analyze the tool conditions when acquiring the temperatures. A queue that can exactly accommodate one sample is filled when acquiring the temperatures. An independent thread is launched for monitoring, which keeps querying the length of the queue until it is fully filled. To ensure a sufficient frequency resolution, gathering data takes at least 1 s, which is much longer than the neural network classifier’s consumption. Thus, the method can be regarded real time, and the software updates the tool condition every other one second in actual application. Figure 9(c) shows that a caution is given if the tool is worn. When cutting using a worn tool, a conspicuous caution appears a few seconds after the tool comes into contact with the workpiece. As a comparison, when using a normal tool, the software keeps silent.
Fig.9 (a) Interface displaying measuring files. ① Currently displayed file, ② file tab, ③ currently displayed channel, ④ channel tab, ⑤ region displaying the measuring curve. (b) Control panel for training neural networks. ⑥ Buttons enabling the addition of samples, ⑦ edit box for assigning iteration times, ⑧ region displaying the training process, ⑨ “reset” button and the “start” button, ⑩ edit box accepting a file name, ⑪ “save” button. (c) Interface displaying real-time measuring curve and cautions. ⑫ Caution updated every other second. |
4 Conclusions
In summary, a real-time tool condition monitoring method based on cutting temperatures and an ANN is proposed, taking advantages of an in situ fiber optic temperature measuring system. During a cutting process, temperatures are acquired, divided into segments, converted into spectrums, and eventually input into the neural network classifier. The neural network immediately produces a result. A learning rate adaption algorithm is adopted to accelerate the back propagation, which can effectively reduce the dependence on initial conditions and hyper parameters when training. The algorithm shows apparent superiority in convergence, allowing training a customized model conveniently for specific circumstances. The classification accuracies stay 90% and higher when cutting parameters change, illustrating that the method can well suit actual machining. This method can offer industry alternative and reliable solutions for real-time tool monitoring and considerable conveniences in future application.