Dr.-Ing. Hans-Peter Heim
The steady growth rates in the consumption of plastics over the past few years are evidence of the fact that plastic products are now an indispensable part of our everyday life and are set to gain even greater importance in future. In the year 2000, the plastics tonnage produced worldwide reached the 180-million-ton mark, with growth rates running at 4 to 6%.
The increasing requirements being placed on plastic products and the more technically refined raw materials that are available also require maximum precision from the process. This can only be translated into customer benefit, however, if the processor not only masters the process but is also familiar with the cause and effect relationships that prevail during the process chain and can document these relationships and use them to continually improve on the process.
One of the main routes to ensuring greater mastery of the process is process transparency, which can essentially be achieved through the acquisition and assessment of data. It is possible to acquire data from each step of a production line in order to characterize the starting parameters (raw material data, semi-finished product characteristics), the processing operation (process data) and the process results (intensity of quality attributes). Providing that all the parameters for the process and the process data remain constant, it can be assumed that, given otherwise identical boundary conditions, the product quality will also be identical (reproducibility condition). Plastics conversion processes are subject to natural fluctuations, however, with the actual values of the raw material data and process data deviating from their setpoint values. Both trends over time and cyclic fluctuations can emerge here, leading to corresponding quality deviations in the molded part.
There are different approaches that can be adopted in process monitoring and adaptive process control in order to master these fluctuations; these divide up into three chief directions:
1. Monitoring and control based on single parameters
2. Monitoring and control based on process characteristic values
3. Monitoring and control based on quality attributes
For some time now, the manufacturers of injection molding machines have been equipping their machines with more comprehensive and more advanced technology. The machines are becoming more accurate and the process sequence is being made more transparent by virtue of continuous data acquisition. The measurement and control techniques employed, however, are generally designed to ensure that specific system components are kept constant. These are individual parameters, such as temperatures, pressure and displacement.1 The state of the art for the machines currently being supplied is a measurement and control technology for the internal machine sequences that is based on a single-component control and open-loop control of this type.
A state-of-the-art machine control system will also generally allow simple documentation to be compiled from the machine data which can be used for quality-assurance purposes. The standard facilities generally include a tabular display of the actual process values recorded by the machine plus an assessment/graphic presentation of these actual process values on the basis of one-dimensional, statistical characteristic values (e.g. mean value and scatter). In most cases, this data can be exported from the machine control to quality assurance programs or to other systems. The individual process parameters are generally allocated upper and lower tolerance limits for monitoring purposes, so that parts can be singled out for elimination if the parameter goes above or falls below this limit. In other words, a signal is transferred to a handling system for purposes of eliminating the molded part produced and/or for putting the machine in alarm status.2
In order to enable additional information to be fed into this data monitoring system, it is possible to transfer measured data from external measuring stations or peripheral units to the machine control system and have this assessed in conjunction with the internal machine data. To ensure the clear-cut allocation of external and internal measured values to a specific molded part, however, it is necessary either to keep to a rigid process-chain sequence or to code molded parts for allocation purposes. Only very few systems on the market support this latter option. The 1:1 allocation of quality attributes of the molded part (as recorded in a measuring station, for example) to the corresponding actual values of the process parameters is, unfortunately, still far from being a matter of course. This approach is to be highly recommended, however, since it provides improved process transparency and permits the compilation of documentation.
Machine manufacturers and suppliers of automation systems have developed a range of special solutions for monitoring and control systems based on process characteristic values, as referred to above. The decisive difference compared with single-component monitoring and control is that it is not the actual values of the machine parameters that are taken as monitoring or control parameters in this case but so-called consequential parameters. These are derived from a combination of different parameters, or from the current process status, and can be recorded by measuring systems.
Examples include the monitoring of the cavity pressure curve, the monitoring of specific characteristic values for the cavity pressure 3, and the monitoring of melt viscosity by means of online rheometry. 4 In both cases it is assumed that, by complying with a specific reference value or reference curve that has been recorded during the production of high-quality molded parts, it will be possible to repeatedly produce high-quality molded parts, while any deviation from this value will inevitably lead to a deterioration in quality.
If, in addition to being aware of the maximum deviation from the reference curve that governs the good/bad decision, the processor also knows how this deviation can be influenced, i.e. is familiar with the quantitative cause/effect relationships that prevail, it is then possible to perform not only process monitoring but also process control by this means. With this type of process control, the actual values of given consequential parameters are recorded, compared with target values and then maintained within the specified limits by a controller that modifies a directly-adjustable parameter.
The aim can be to precisely observe a reference curve for the cavity pressure, for example. This is either done during individual phases or by taking the phases of injection, compression and holding pressure together. 5 In the meantime, manufacturers are supplying machines which will also permit the plasticizing process to be regulated on the basis of a reference curve for the screw-drive torque. 6 To sum up, all the different phases of the injection molding process can thus be controlled as a function of a characteristic value or reference curve with any unevenness being smoothed out through appropriate modification of the adjustment parameter. This makes it possible to control the process within an individual cycle in order to reduce the level of fluctuations.
Adaptive control concepts to operate across all the different cycles, by contrast, pursue the aim of reducing longer-term trends, such as those due to fluctuations brought about by batch fluctuations in the raw material. Concepts have so far been presented for adaptive control of the viscosity that is measured online, or for optimizing the cavity pressure curve (attaining a reference curve).7
The underlying assumption behind all monitoring and control concepts that observe just one-dimensional monitoring variables, i.e. that consider each parameter in isolation or observe just a single characteristic value, is that a high quality will be produced providing that the individual parameters/characteristic values are kept sufficiently constant. When tests are conducted to verify this assumption, they regularly produce a result that confirms it. A deliberate modification in the moisture content of the raw material, with the corresponding change in viscosity, will naturally lead to changed flow behavior in the mold and, with otherwise identical conditions, will probably result in deviations in the quality attributes of the molded part too. Deviations from the reference pressure curve are observed, which will similarly be detected by an online rheometer. This test will therefore serve to confirm the above assumption.
The situation is completely different, however, if, at the same time, changes occur in the boundary conditions that are not taken into account for the parameter/characteristic value being monitored. This can be illustrated with a simple example. Assuming that the moisture influence already referred to gives rise to an increase in viscosity for a reduction in the moisture content, then a control system for the injection and holding pressure phase based on cavity pressure would now attempt to precisely maintain the reference curve. It is clear to see that precise reproduction of the reference curve will not produce a successful result in every case, and that a higher cavity pressure is required to achieve complete filling of the cavity. Corresponding cases can be drawn up for the other variants of single-component and characteristic-value monitoring and control systems.
The basic problem is the so-called interactive effects – parameters which, if varied simultaneously, have a combined influence on the quality of the molded part. Monitoring methods or controllers that monitor or keep constant individual components or process parameters that do not correlate with the quality attributes will thus generally not suffice to monitor the quality of the molded part or keep it constant. The target ought to be the observation of quality attributes rather than the observation of specific process parameters.
What is important here is not to observe individual parameters in isolation but to model the combined effect of the parameters that exert an influence. Different methods are available for this modeling, including pattern recognition, statistical regression methods, fuzzy logic and neural networks, as well as combinations of these processes. 8 Each of these methods has its strengths and weaknesses for different fields of application. From the angle of increased process transparency, which was the requirement voiced at the outset, it is particularly important to establish how far a method is able to provide users with information on the cause-and-effect correlations of their process. One particularly appropriate method for this task is process modeling with the aid of the multiple linear regression calculation, which will be briefly explained below. Further-reaching information and a detailed presentation of examples of process optimization and process monitoring using statistical models can be found at www.KTP.cc together with information on the MPC software (Multistage Process Control).
In the light of the above-mentioned reproducibility condition, it is possible to compile a mathematical description of the correlations that exist between the input and output variables for an individual processing stage and hence for the process chain as a whole, by taking all the different process stages together. This correlation is formulated in statistical terms via multiple regression analysis. The machine parameters are allocated to molded-part quality data for the individual cycles, enabling a process model to be compiled in the manner shown in Fig. 1 for each of the quality attributes to be described. The quality attributes are taken as the regressands (Yi) and the process data as the regressors (A, B, C, ...). 9
The process model that is established in the course of the experiments and process analyses can be used for an estimate of the intensity of the individual quality attribute Yi with different machine-parameter combinations.
Continuous process monitoring with process models shifts quality control from the final inspection of the finished part to the correlated process parameters. It does this by having molded-part attributes predicted on the basis of the prevailing production conditions by means of a multiple regression calculation. The attributes are then assessed on the basis of their agreement with the specifications.10
To monitor production, it is necessary to have a model which describes a small process window around the set operating point. Actual values are used here so as to record the true fluctuations that occur in the process. The data is established either at the fixed point of operation on the basis of the prevailing scatter or through an experiment design with a low variation width, while all the available actual values are recorded. The requirements on the accuracy of the measuring instruments and measuring methods are thus correspondingly stringent when it comes to the compilation of a monitoring model.
Each quality attribute that is to be assessed requires a model which is supplied on a cycle-by-cycle basis, with all the requisite actual values from the production process, via a link between the machine control system and the model computer. The data transferred from the control system is introduced into the regression functions, as shown in Fig. 2, and the intensity of the individual quality attribute is calculated.
A decision regarding the use of the product can be taken on the basis of the statements from the forecast model. To do this, the upper and lower specification limits of the drawing dimensions are allocated to the regression models established for the corresponding attribute and compared with the calculated values for each cycle. In addition to this, the accuracy of the estimation has to be taken into account.
In the regression models, all the parameters of relevance are taken into account simultaneously, with allowance being made for their interaction where technically feasible. This permits a highly accurate calculation of the quality attributes for each individual cycle and thus allows both a very short reaction time to non-complying parts and a pronounced reduction in the number of quality inspections, with the exception of the uncertainty zones. This monitoring will be equivalent to 100% control.11
It is generally necessary to run through a number of process steps in order to produce a plastic part. If the part has to be welded, for example, then both the process control during the injection molding and the process control during joining will play a decisive role in respect of the properties of the final product. The difficulties associated with the documentation and monitoring of multi-stage process chains lie first and foremost in synchronizing the recorded process parameters for purposes of the model calculation so as to ensure that each molded part can be allocated to the process data that go with it from all the different process stages.12
This synchronization can only be achieved with the appropriate software support and with appropriate marking of the parts (such as with barcodes), on account of the potentially complex production situation that prevails in industry.
If these preconditions have been fulfilled, it is possible to model the process correlations of all the steps involved in achieving the overall result using a multiple regression calculation. The dependency structures become transparent and the deviations introduced by a preceding process step are suitably taken into account in the process that follows.
The benefits of process analysis and process optimization for the value-added chain of product creation can be worked out in cost terms by employing the appropriate management-economic acquisition and assessment systems. Each reduction in the number of reject parts that is achieved through optimization serves to avoid errors and waste. In cases where there is an extensive vertical range of production, the early recognition of faults through continuous process monitoring can also prevent further mistakes from being made. Depending on the subsequent use to which reject parts are put, the company may also incur costs in the form of disposal and recycling charges if a more holistic view is taken of the process environment. Hence, when observing production processes in conjunction with the costs closely associated with the process, it makes sense to take into account all the consequential activities that result from faults, and also the costs that are incurred through these. It is clear just what potential model-aided quality monitoring holds when it comes to effects that can be expressed directly in monetary terms, such as reduced reject numbers and a lesser use of resources.
Model-based statistical quality monitoring additionally has a number of different positive impacts that it is impossible to put a precise figure on. The benefit to the company in respect of what are primarily tactical and strategic considerations should not, however, be underestimated The advantages basically divide up into three categories:
These benefits of the CPC concept that cannot be expressed in direct monetary terms are set out in a publication by Potente et al, together with a quality/cost analysis.13
The process model compiled for monitoring the process can also be employed for adaptive process control. In order to permit the machine settings to be corrected when the monitoring model establishes a deviation between the actual quality and the target quality, it is necessary to have three models for quality-based adaptive control. These are a time-independent quality model that formulates the correlation between the process variables and the quality attributes, a time-independent machine model for describing the correlation between the machine settings and the quality attributes, and a time-dependent machine model for formulating the relationship between the machine parameters and the process variables.14 While the machine producers have engaged in greater efforts to harmonize the interfaces for communication between the computer and the machine control over the past few years, and are coming to regard open-loop/closed-loop control as less of a company-constrained solution, the implementation of the model-based adaptive control system set out above would still involve a high outlay for the adaptation/communication of the machine control system and the model computer. No standardized solutions are available at present.
As already mentioned, a model-based adaptive process control system needs to be built up on the basis of a monitoring system with a high information value, i.e. a methodology that will supply sufficient information on the process, as is shown in the example above. Against this background, it is possible to ask the justified question as to whether a fully-automated adaptive control is necessary at all, or whether the cause/effect knowledge obtained by modeling the process could not be put to better use in providing the basis for a continuous improvement process and for accumulating the corresponding know-how, and dispensing with an adaptive control altogether. The process itself will doubtless provide the answer to this – and the implementation of a model-based process monitoring system certainly constitutes a good first step