Machine learning: The Innovation in Plastic Industry

In the past decade, it was observed that the companies that actively use big data grew 50% faster as compared to non-users. This has led to a trend of increased use of big data in the manufacturing sector as well. Converting big data into meaningful insights has become possible because of growing artificial intelligence (AI) capabilities, in particular, Machine Learning (ML). AI capabilities in industry sector only will add over $15 Trillion to the global economy by 2030.

ML is a specialized branch of AI in which the machines learn from its environment or the datasets given. Machine learning algorithms are employed in tasks where designing static instructions is difficult or even infeasible, such as finding patterns and anomalies, making predictions from large complex datasets.

AI capabilities in industry sector will add over $15 Trillion.

Although it has become a buzzword in the recent past, the concept of ML has been around for over 75 years. In last five years, big tech giants such as Google, Amazon, and Facebook have developed their own ML platforms. But the true beauty of ML is that its application is not limited to the tech industry. ML applications can benefit manufacturing, in particular, plastic industry, as well.

The need for injection molded plastic is increasing due to its use in a wide variety of industries such as automobile, manufacturing, electronics & consumer goods, building & construction, and healthcare. As the plastic industry expands, the need to utilize available resources in an optimal way is increasing. This results in the growing need of incorporating ML in plastic industry.

The main purpose of this article is to describe how ML allows improving quality control of plastic injection molding process. Next, we will also discuss how to avoid the equipment downtime using predictive maintenance.

The plastic injection molding process has a number of factors such as pressure, injection velocity, barrel temperature, etc. that dictate the final quality of the product. In order to have a high-quality product, it is crucial to keep the whole set of variables at an optimal level. However, lack of consistent relationships between the process parameters makes optimization through conventional mathematical modeling an extremely challenging problem. Thus, the traditional approach is to use manually preconfigured discrimination parameters based on rigid threshold levels. An innovative approach that is part of the Industry 4.0 transformation exploits machine-learning algorithms and provides a new regulation tool that allows optimizing the whole set of parameters in real-time.

Unlike conventional parameter adjusting based on operator’s experience, machine learning doesn’t rely on domain expertise. The main source of knowledge about process properties is the historical data, recorded by sensors from the particular molding equipment. Machine learning needs sets of intermediate readings from sensors along the process together with parameters of the final product. After processing thousands of such sets machine learning learns complex dependencies between intermediate parameters and quality of the product. In other words, the system builds required prediction mathematical model by itself. This is called training of machine learning model.

Quality prediction system: machine learning doesn’t rely on domain expertise.

After the machine learning model is trained, it can process live data from the sensors and predict the final quality of the plastic part. The accuracy of this prediction depends on a number of factors, such as quality and volumes of training data, level data preparation and cleansing, chosen machine learning algorithms, the experience of data scientists and so on. According to the latest research of data science companies for the plastic molding process prediction accuracy reaches 98%. In some cases, machine learning model is also capable to distinguish and predict types of fault, such as unfilled areas, streak or warped parts.

This new type of insight can be utilized in two ways: the first, if a machine-learning model predicts poor final quality, a manufacturer can stop further processing a particular batch at an early stage and save energy, material and time accordingly. The second, machine learning model provides an operator with additional precise insight into the process, which allows further optimization of the molding process parameters.

In order to implement machine learning solutions, an organization needs to establish a full data science project that includes a number of essential steps, starting from exploring opportunities and ending with building maintenance routine for the deployed machine-learning model.

Data acquisition and cleansing is one of the most important and also time-consuming phases of data science project. More than 20 independent parameters can be included into a dataset for plastic injection molding process model training. Here are the most important of them: Cycle time, Material blend, Injection time, Barrel temperature, Injection velocity, Pressure, Screw speed, Coolant temperature.

Data science project: The more diverse data you have in hands, the more accurate model you can build.

A general rule for initial data acquisition indicates that the more diverse data you have in hands, the more accurate model you can build. The manufacturer needs to keep a history of raw data from sensors and telemetry from tens of thousands of pressure cycles. And within single injection cycle signals from sensors needs to be recorded with sufficient sampling rate. For example, in order to pass a detailed representation of pressure dynamics to the machine learning model, the sampling rate needs to be from 100Hz to 1kHz.

In most cases, data obtained from real-world equipment is not usable as is. e.g. a broken sensor returns a constant value of 1500C temperature or doesn’t return anything for some time until it gets replaced. The task for a data scientist is to resolve issues such as missing values, incorrect values, constants, noise, and duplicates prior to building machine learning model.

Next, the mathematical model of the molding process is being built. To the date, there is a quite extensive list of algorithms for building a machine-learning model: logistic regression, decision tree, support vector machines, artificial neural networks. Data scientist chooses the most appropriate one for each particular case or uses even a combination of several algorithms. After the model is built and tested software engineers deploy it on a runtime platform. There are a number of ways of quick integration of machine learning system into existing processes without developing the complex application.

Another implementation of machine learning technology, which has gained a significant attention of the entire manufacturing industry is predictive maintenance systems. Downtime reduction – this is an outcome of predictive abilities of ML algorithms. Many manufacturing units run 24 hours for 365 days of the year. When the equipment breaks down, the resulting downtime leads to decreased production and a waste of raw materials. This can be prevented by conducting predictive maintenance, which raises an early warning for critical failures to avoid downtime.

The difference between classic rule-based approach and data-driven approach is that predictive maintenance solution based on machine learning recognizes complex patterns in readings from a wide number of diverse sensors and evaluate equipment’s current condition based on these patterns, not just comparing static values with reference numbers. ML models reliably estimate the times when the equipment would need maintenance servicing. When the production team takes preventative and corrective action at those times, a reduction in downtime occurs.

Data, required for building predictive maintenance system, can be divided into two main groups: first, parameters of machine health status that reflect degradation and second, failure history that allow machine mark a number of patterns as alarming.

Predictive maintenance: ML recognizes complex patterns in readings from sensors.

Plastic industry is adopting ML. Business leaders in the plastic industry have realized that ML offers a real opportunity to reduce the costs and also improve the overall quality of their production. As a next immediate step, the plastic companies are either starting to form their own data science department or start working with external data consulting agencies. Since ML algorithms for manufacturing industry is a highly sought-after skill, many companies find it difficult to retain talented employees and hence opt for consulting companies. In some instances, companies with their own ML department have collaborated with a consulting agency to shorten the timeline of the project.

It is estimated that by 2020, there will be 45 Zettabytes of data available in storage for various industries, including manufacturing. Furthermore, the global market for injection molded plastics would reach $162B and the compound annual growth rate (CAGR) would be about 4.9%. To summarize, the plastic industry is in an exciting growth phase. To be a part of these exciting times, the leadership in plastic businesses needs to take steps now and incorporate ML into their workflow. Tomorrow starts today.

Plastic industry is in an exciting growth phase.