A study of micromanufacturing process fingerprints in micro-injection moulding for machine learning and Industry 4.0 applications

The manufacturing equipment used in this work is a state-of-the-art μIM machine (Wittmann-Battenfeld Micropower 15) which can be considered a scaled down version of regular injection moulding machines. However, the process itself is highly dynamic and dominated by various physical phenomena at the micro-scale due to the challenges in replicating extremely small mould cavities, providing components with only a few mg in masses for specific applications [19]. The details of the μIM process for thermoplastics are discussed elsewhere [14, 20].

The main consideration in controlling a μIM cavity filling process is to monitor the injection pressure (P_i) that is being recorded from a built-in strain gauge sensor (X-Sensors XB 120-1500) attached at the back of the injection piston which conveys the plasticised thermoplastic material. Another force sensor (Kistler 9210AA) is attached to the mould tool for recording cavity pressure (P_c) for each moulded part. The details of this sensor attachment is similar to our previous work in which the sensor was located at the back of an ejector pin towards the end of the moulding cavity [2]. P_c has been proved to be a very valuable data channel which can directly represent the quality of the micromoulded products as reported by several authors [8, 21]. The third data channel consists of the information coming from the encoder of the injection piston position which is readily available on the machine. The position data of the injection unit or the piston is also very important since this data is prone to the various physical processes such as friction and expansion. This makes this data channel capable of capturing cycle to cycle variations. These three data channels are connected to an ethernet network interface (National Instruments – cDAQ – 9185), and a custom-built LabVIEW code was created for capturing the multi-channelled data for each cycle. Data acquisition has been synchronised using a common trigger signal coming from the machine. Figure 2 shows the multi-channelled process data collected from a μIM cycle showing the main characteristics of the process.

The multi-channelled process data taken from a μIM cycle are characterised by two important stages. The first one is the injection phase which occurs in the vicinity of t = 6.5 s where the molten polymer is injected to the mould cavity by the piston movement. This phase is characterised by the peaks in P_i and P_c curves. The second feature of interest in the graph is the packing phase which can be traced with the constant piston position just after t = 6.5 s. In the following, a brief information regarding the moulded product will be given, and individual PFs that can be an indicator of moulding quality will be extracted from the multi-channelled data with different approaches.

The product used in the present study is based on a 0.5-mm-thick circular disc shape with 17 mm diameter which accommodates an array of microneedle cavities in 6×6 configuration in a μIM tool as discussed in our previous study [2, 22]. The quality proxy of the moulded products in this work has been defined as the microreplication efficiency of the microneedle features (M_μ) obtained from each part or manufacturing cycle. M_μ has been calculated by measuring the average microneedle height by using a telecentric optical machine vision system for one sample and dividing it by the average cavity depth of the needles on the mould insert which is 912 μm. This results in M_μ values between 0 and 1 where 1 would correspond to a fully replicated array of microneedle protrusions from the substrate. The conical structures of the needles provided necessary draft angles in the ejection stages of the moulding, and 3-dimensional warpage or deformations were found to be negligible from process variation perspective. Approximately 200 mg shot size has been used for the moulding of the parts. This shot size and high-cooling rates allowed the process to reach a steady state after few cycles. P20 grade tool steel also provided sufficient mechanical strength to the mould tool which makes the mould deformation insignificant during the process.

A commercial acrylonitrile butadiene styrene (ABS) polymer resin (CYCOLAC - HMG94MD, Sabic) was used. This particular thermoplastic material is suitable for thin-walled medical parts manufacture with superior flow capabilities and relatively low shrinkage (%0.5–0.8 [23]). The processing conditions given in Table 1 have been used for experimentation. Packing pressure is applied on the moulded parts just after the switch-over point, from velocity to pressure controlled regime for shrinkage compensation. One of the important considerations in process conditions is the mould temperature, where 3 different values were used, namely 50, 60, and 70°C. Twenty parts have been collected for each mould temperature and process data were recorded. Process fingerprints are studied for all 3 mould temperatures in a global approach for simplifying the assessments. At least 10 parts have been discarded for making sure that the temperature gradients and frictional affects are avoided while collecting the samples.

The sPFs are quality indicators that do not require any alterations to the data channel and can be extracted directly for each μIM cycle. One sPF has been selected from each data channel mentioned in Fig. 2. Maximum injection pressure value (P_i-max) has been taken as a PF representing the peak value of the injection pressure data (Fig. 3a). This particular FP gives insight regarding the deceleration of the injection piston after hitting a certain pressure threshold that can be affected by several phenomena during the process such as viscosity of the material and friction. Maximum cavity pressure value (P_c-max) has been extracted from cavity pressure data in a similar fashion to P_i-max. Intuitively, it can be stated that the cavity pressure is a more sensitive measurement than the P_i because of the measurement’s proximity to micro-features due to its location in the cavity.

The 3^rd PF has been extracted from the encoder data of the injection piston position where the amount of packing (∆x) has been calculated for each cycle after the injection unit decelerates. ∆x is a measure of shrinkage compensation during the packing phase and can provide useful insights regarding the packing dynamics. Figure 3 is a zoomed-in version of Fig. 2 where the details for extracting sPFs are shown. The differences in the scales for Fig. 3 a and b show the important features of data channels. For instance, the packing behaviour and the resolution limit of the encoder can be seen, which is approximately 60 μm (see Fig. 3b).

The first derivatives of the pressure measurements and how they are sustained in the data channels shown in Fig. 2 might also be of interest for detecting cycle to cycle variations. Derivatives and integrals within certain intervals were considered PFs for capturing these aspects of the data. For instance, the first derivative of cavity pressure data has been calculated (Fig. 4). Data taken from 3 random cycles in Fig. 4 show that the rate of increase during the injection varies considerably during the cycles. The rate of change in cavity pressure data can provide insights regarding the flow and hence the replication capability of the molten polymer. Thus, the maximum rate of cavity pressure increase ([dP_c/dt]_max) has been selected as a dPF.

Pressure sustained during at any stage of μIM is of importance and can be a reliable process quality indicator [8, 22, 24]. Viscosity of a polymer can be considered as resistance to flow, and higher viscosities result in bigger pressure drops in polymer processing. The viscosity of the polymers and hence the pressure sustained in the cavity during μIM can differ significantly due to the extreme process conditions such as shear heating effects [25]. Lower viscosities during the moulding process will result in better replication of micro-features. Hence, pressure curves taken from μIM can be integrated within certain intervals, and this property can be quantified for quality indication. For this purpose, P_i and P_c data that has been integrated in the vicinity of the pressure peaks and values have been taken as dPFs, namely ∫P_i and ∫P_c. The integrations have been carried out within a 100 ms interval from the peak value towards the pressure decay using a numerical integration based on the functions present in the LabVIEW package. This calculation has been schematically depicted in Fig. 5 for P_i. The calculation for P_c was done in a similar way and omitted for preventing redundancy.

Another way of extracting process quality proxies is to analyse the samples comparatively for being able to quantify the deviations. Initial analysis showed that only cavity pressure measurements are suitable for doing such a comparative analysis. The following steps were used for extracting comparative process fingerprints (cPFs), and the method is also depicted in a graphical format in Fig. 6:

Previous work in the literature covered how the process quality indicators could change according to the process parameters such as melt temperature, mould temperature, injection speed, and holding pressure [19, 20]. It was demonstrated that both process and quality proxies were dependent on the processing conditions. However, a cycle by cycle analysis has been missing in the literature with respect to characterising the dimensional features of the products. Moreover, the process quality indicators or fingerprints can be changed significantly when a smaller or bigger mould cavity used. This changes the cooling rates, temperatures, and pressure measurements completely. For filling this gap in the literature, an approach based on cycle by cycle monitoring of the process fingerprints has been adopted.

These 9 process quality indicators or PFs are summarised in Table 2. They are extracted for each 60 cycles resulting in 9 PF series for individual products. PFs are analysed with corresponding M_μ values of the parts in the following section, and a ranking system has been made for evaluation of the process quality indicators. Finally, the resulting linear relationships are used for forming a MLR model for predicting the microreplication efficiency of the process.

Two of the sPFs (P_i-max and P_c-max) have been known to be linked to the dimensional features of the micromoulded products in the literature; however, no reports have been found citing their direct link to the micro-feature replication efficiency [8, 20, 21]. Data given in Fig. 7 a and b show that an obvious positive linear relationship is present between peak pressure values and M_μ. P_c-max resulted in a slightly higher R² (0.7819) than P_i-max as expected. Moreover, injection pressure (P_i) measurements are taken at the back of the injection piston, and these also include inertial effects due to mass of the whole injection unit resulting in slightly less relevant data than P_c data. Both PFs suggest that pressure measurements can provide high-quality data regarding M_μ. For instance, P_i-max and P_c-max change significantly for 950 to 1150 and 150 to 450 bar range, respectively. Any impact that can alter the flow or injection behaviour of the molten material will be manifested on these pressure values and be detected for fault prevention.

∆x data presented in Fig. 7c resulted in an R² value of 0.498 for predicting M_μ. This can be attributed to the relatively low resolution in the measurement of the position as shown in Fig. 3b. The encoders inability to measure displacements below 60 μm is evidenced by the exact same values detected for ∆x. Certain accumulations in this particular scatter plot are due to the different mould temperatures employed and a linear relationship between the amount of packing and M_μ has been found, however with much lower correlation. ∆x is a measure of the movement of the injection piston after the switch-over from speed controlled regime to the pressure controlled packing regime, where lower ∆x values might suggest a relatively late switch-over or more material conveyance to the cavity due to dosage.

[dPc/ dt]_max data shown in Fig. 8a represent a positive trend with respect to M_μ, however with low sensitivity and an R² of 0.4336. The similar behaviour is also valid for ∫P_i. This behaviour might suggest that not enough data points were acquired while capturing the process data and shows that this particular PF types might require higher data capturing rates than usual as the increase and decreases happen very rapidly. ∫P_c data performed better than [dPc/ dt]_max and ∫P_i, resulting in R² of 0.763. This shows that cavity pressure decay profiles occurring after a cavity filling event are the most reliable of the dPF containing information that can be linked to microreplication directly. The inferior performance of ∫P_i can also be explained by the fact that the injection pressure overshoots near the piston (see Fig. 5) can vary significantly from cycle-to-cycle resulting in more scattered values especially for higher pressure integral values (614–618 bar.s) as illustrated in Fig. 8b. The high amount of momentum of the injection unit makes the control of the deceleration difficult leading to variable pressure overshoots and hence integral values.

The quality indicators extracted from comparative analysis and their scatter plots are given in Fig. 9. The analysis shows that the main trends are mostly the same for the whole integration range, 1.5 s and 3 s. However, differences can be quantified by analysing the R² values. For instance, R² values of 0.750 and 0.789 corresponding to Δ_all and Δ_1.5s, respectively, clearly show that the quantification of the comparative deviation near the peak (see Fig. 6b) is of more relevance to the microreplication quality detection. Once again, this can be attributed to the importance of cavity pressure decays as it is the case in ∫P_c. A comment can be made that Δ_1.5s and ∫P_c are too similar and present no significant difference, albeit a 3.5% improvement in R² values is present while using the comparative analysis approach. The difference between Δ_1.5s and Δ_3s is only marginal at 0.89% which shows that the prior one is preferable for quality prediction and machine learning purposes. In overall, cPF analysis of the cavity pressure data proved to be logical and resulted in best R² values amongst all PFs for representing the microreplication efficiency.

A Pearson correlation coefficient (PCC) matrix depicting the interactions between 9 process fingerprints and M_μ is given in Fig. 10. PCC between two variables takes values between −1 and 1 and is a statistical measure indicating the linear correlations between two variables. Positive values indicate a positive linear trend between two variables, whereas negatives represent a negative linear relationship. The intuitive approach could be the selection of best performing PFs from each class of process fingerprints in a multiple linear regression model involving PFs and quality prediction or monitoring of micro-products. However, lower correlations are favoured between the selected process quality indicators for building a multiple linear regression model for making it more realistic.

The best performing PFs from each class in terms of R² values are P_c-max, ∫P_c, and Δ_1.5 for predicting/monitoring microreplication during the μIM process. These PFs are somewhat derived from the cavity pressure data and show strong correlations between each other. Three PCC values of P_c-max vs ∫P_c, P_c-max vs Δ_1.5, and ∫P_c vs Δ_1.5 are 0.96, −0.99, and 0.98, respectively. For making the multiple regression model for machine learning more realistic, and rather than flooding the model with redundant data, P_i-max, [dPc/ dt]_max, and Δ_1.5 have been selected. These PFs not only provide useful insights from different channels but also represent a selection of various PF extraction methods such as analysis of peak values (P_i-max), pressure increases rates ([dPc/ dt]_max), and comparative analysis (Δ_1.5). Corresponding PCC values for P_i-max vs [dPc/ dt]_max, P_i-max vs Δ_1.5, and [dPc/ dt]_max vs Δ_1.5 are 0.71, −0.89, and −0.72, respectively. Table 3 summarise the best PFs according to R² values and selected PFs for the multiple linear regression model.

Linear regression is a statistical method that has been widely used for modelling process behaviours, product quality, and predictions based on dependant and independent variables [2, 10, 26]. More variables can be added to the linear regression models for training the predictive model towards a more accurate machine learning approach. In this case, the method becomes a multiple linear regression (MLR) combining the behaviours of different independent variables for predicting a dependant variable which is usually a quality criterion of the product. This can be generalised as: where y is the dependant, x is independent variable with individual slopes (m_i), and c is the intercept of the curve. In our model, microreplication efficiency (M_μ) is the dependant variable that is being affected by the dynamic phenomena. The independent variables are the process fingerprints (PFs) which are the proxies of the physical processes from μIM.

The PFs in Table 2 and M_μ are organised in data columns in a spreadsheet file for representing 60 moulding cycles and are processed using code written in Python using Jupyter Notebook IDE (integrated development environment) for creating the MLR model. The code can be found in the Appendix, and it assigns the columns of data to independent (PFs) and dependant (M_μ) variables followed by creating a test and training data groups. After the usage of the linear regression function in the code, all coefficients are calculated, and the MLR equation for the model can be stated as:

To check the validity of the MLR model, a test size of 0.2 is set where 20% of the manufacturing cycles are randomly chosen by the code to test the model. After generating the code, the remaining 80% is used for training the model. The test dataset is then used as inputs in Eq. (2), and corresponding predicted M_μ values (M_μ_pred) are generated. Test values (M_μ_test) which were randomly selected in the beginning were then compared with predicted y values (M_μ_pred) which is the outcome of the training set. The scatter plot in Fig. 11 shows the relationship between M_μ_test and M_μ_pred. The test and predicted values are also provided. The success of the model is then quantified by the R² value calculated by the Origin Pro software.

The R² value of 0.835 has been found from the linear fit calculated from M_μ_test and M_μ_pred scatter plot. The result suggests that M_μ_test can be explained at about %84 accuracy by using predicted values generated by the MLR model. It can be concluded from the analysis that by training the MLR model with different PFs, the reliability and accuracy of the models can be increased as compared with single quality indicators vs M_μ as given in the previous sections. Moreover, since the MLR method also includes testing datasets for checking the validity of the models, it is a fairer way of quantifying the behaviours of process fingerprints. A comment can be made that ~84% accuracy might not be appropriate for such a quality assessment procedure; however, the main purpose is to present that the quality proxies or PFs work in the first place and there is a lot of room for improvement such as using additional layers of instrumentations on the machine and using more quality indicators in the MLR models. Furthermore, relatively low accuracy of data-based in-line quality assurance models can still be very cost-effective since the PFs are directly physical metrics and they can change significantly in the case of mechanical or electrical failures on the machines. For instance, frictional effects in injection units, heater failures, or changes in the material viscosities can be thought as examples that can lead significant differences in the values of the PFs that can easily be detected and faulty products can be sorted. The cost-effectiveness of the models usually overshadows the low accuracies when it comes to manufacture of high added values micro-nano featured components due to their high margins.

Another aspect of the method presented can be its applicability in the shop floor and automation scenarios. The total time spent for DAQ system setup and process fingerprint extractions can be measured with a few days provided that some coding and data analysis skills are present. Moreover, the time investment for coding and data analysis is a onetime effort and provides the necessary flexibility in the shop floor for the manufacture of different products.

The following suggestions can be stated for making the model even more accurate: