By Michael Tay, analytics platform lead at Rockwell Automation
In most enterprises, building machine-learning (ML) models is a long and complex process that requires subject matter experts to manually import data from different databases and control-system data archives. They then must align information sequentially so that comparisons are all relative to the same timeline, and then clean the dataset for model building, sweeping away any outlying information and unneeded input variables. This all requires custom programming code, manual conversions and querying using different tools.
It's a chore.
Existing analytics technologies within manufacturing organizations require deep expertise and knowledge in both data science and industrial processes. This means that many operational technology (OT) professionals who do not have the same skillset as expert data scientists can’t reap the benefits of basic analytics on the factory floor.
But the emergence of off-the-shelf, targeted machine-learning applications is changing the game—making ML accessible for OT workers without them having to learn new skills. Let’s explore how factory-floor workers can harness the power of ML to predict anomalies and estimate operational variables in production processes all without the help of a data scientist.
Anomaly detection
In manufacturing, identifying anomalies is key to ensuring products are up to standard. Anomaly detection is a machine-learning technique that identifies unusual or unexpected events in data. This is useful for preventing equipment or manufacturing issues, which can result in costly downtime and production losses.
Effectively using machine learning for anomaly detection requires manufacturers to collect and analyze large amounts of data from various sources—such as sensors, cameras and other monitoring devices on the plant floor—and use that data to train the algorithms to recognize normal patterns and identify anomalies. They can also use machine learning to analyze the relationships between different variables and identify correlations that may be indicative of normal or abnormal operations.
Machine learning-based anomaly-detection systems provide real-time monitoring and alert manufacturers when anomalies are detected. For instance, imagine a factory that produces car parts using a range of machinery. The machines generate data on their performance, including factors like temperature, vibration and speed. Using an anomaly-detection algorithm, workers can quickly identify when a machine is behaving unusually or outside of its normal parameters. This enables them to take corrective action before the machine breaks down, preventing costly repairs and downtime. Anomaly detection helps factory-floor workers quickly take action to prevent equipment failures, minimize downtime and avoid costly production losses.
Soft sensors
Soft sensors are another application of machine learning. Unlike anomaly detection, which focuses on identifying outliers in data, soft sensors are used to estimate process variables that are difficult or expensive to directly measure—think laboratory-determined quality. Machine learning-based soft sensors use available data from other sensors and process variables to build models that estimate the target variable.
The accuracy of soft sensors depends on the quality and quantity of data used for training. In many cases, soft sensors can provide estimates as accurate as direct measurement. This is because they can incorporate information from multiple sources and account for complex interactions between process variables that are not easily measurable. They can perform manual laboratory measurements that are difficult to measure and include variability with limited sampling, sampling variations and degradation.
Soft sensors can be deployed in real-time on the plant floor, providing operators with critical information about the process and enabling them to make informed decisions. They can also be used for process control and optimization, enabling manufacturers to improve efficiency, reduce waste and increase yield. For example, imagine a manufacturing plant that produces chemicals. It might be difficult or costly to measure certain process variables, such as the concentration of a particular compound in a reacting mixture. By using a soft sensor, workers can estimate these variables based on other measurable data, such as temperature, pressure and material balance. Manufacturers can better understand their processes and identify new opportunities for optimization by estimating process variables that are difficult to measure directly.
Machine learning for the masses
The emergence of off-the-shelf, targeted ML applications is transforming the manufacturing industry by making advanced analytics more accessible to factory-floor workers. It enables OT professionals to reap the benefits of basic analytics on the factory floor and empowers them to make data-driven decisions to improve efficiencies and reduce costs.
The acceleration of these enabling technologies, including no or extremely low code options for ready-made industrial solutions, should be an accelerator in producing more benefits and delivering solutions at higher success rates. It should be good news for digital initiatives across many industries. Then we. can start looking for more native edge-to-cloud answers as another avenue for scalability and management to support oversight and access for the hundreds of edge applications expecting to deliver value and to support plant-floor teams.
