This article was written and contributed by Sathish Ravichandran, Machine Learning Engineer at Bsquare.
Sathish Ravichandran helps companies implement IoT initiatives by building scalable software platforms to deploy various stages of the data science process – such as pre-processing and feature engineering – in production. He also develops machine learning models in serverless architecture to generate real-time predictions and enhance data processing efficiency. Sathish runs the School of AI, Seattle and holds an MS from the University of Washington.
Most manufacturers today realize the critical importance of collecting equipment data to improve their operations. From better understanding machine performance to finding inconsistencies in the production line, data is often the start to solving many manufacturing woes. In addition, the availability of data in an industrial setting opens the door for use of emerging technologies such as IoT and machine learning which can deliver next-level asset optimization.
However, not every business has pools of data to draw from right out of the gate. Luckily for them, there are ways to introduce new technologies into their operations regardless of the amount of data available - even if the answer is zero. In fact, there are varying levels of machine learning maturity that can help manufacturers achieve their asset optimization goals, starting with rules engines (minimal, or no data) and ending with deep learning (complex, built-out data sets). Let’s take a look at these stages in a little more detail:
Stage #1: Just Beginning the Data Journey
Manufacturers just starting out rarely have the data they need to analyze their operations and leverage machine learning in a meaningful way - or so they think. While they can’t quite jump ahead to more complex machine learning systems, such as deep learning, they can apply rules engines to translate analytical insights into actionable logic.
Rules engines pair machine learning with the institutional knowledge of subject matter experts (SMEs) to create logic and automate actions based on first-hand insights, making them a great fit for inventory optimization. Once an SME has determined the desired outcome for the engine and written specific rules, the engine uses an ‘if this, then that’ function to execute the appropriate actions when specific conditions arise.
With IoT-enabled assets, businesses can integrate data from their supply chain to forecast demand and ensure they right-size the quantities of raw materials on hand. Then, through rules-based actions, they can automatically pre-order the precise amount of material needed for future production. For example, a window manufacturer could use the IoT and supply chain data to forecast the exact amount of glass needed for the window units it will produce each month. This level of accuracy helps them avoid ordering too much inventory, which can increase the cost of storage, and ordering too little inventory, which can result in production holdups.
Another example might be an LED lamp manufacturer: they need to have LED chips, epoxy resin, wire and more. With too little wire, they won’t be able to produce the number of lamps they need to meet customer demand. Yet, with too much wire, they will need to warehouse the surplus amount until they need it.
Stage #2: Big Data and Analysis
After several months, manufacturers will have collected enough data to progress beyond rules engines to something much more scalable: machine learning. This method of data analysis allows software applications to become more accurate in predicting outcomes without being explicitly programmed. Through machine learning, manufacturers can more easily achieve their yield optimization goals. But first, they must establish a well-defined problem to solve and ensure they have relevant historical data sets (which is why this stage occurs after they’ve collected more data), so the machine can compare new data to the old data and arrive at a solution.
Businesses use machine learning to analyze a combination of on-site sensor data from equipment and external and historical data to forecast the maximum number of products they can produce while still ensuring output quality remains high. For example, a manufacturer of electric bicycles might leverage IoT data to determine the productivity of machines, while also turning to external data to determine the popularity of models, and as a result, the demand of each bicycle. With this information, the manufacturer could produce the number and type of bicycles required to meet customer demands without over producing any of the models across its lines.
Stage #3: Years of Collected Data Provides Greatest Insight
Finally, manufacturers that have been collecting data for years and have a mature enough data set can layer deep learning into their operations for more insight and achievements like optimized quality. Deep learning is a machine learning technique that businesses use to teach artificial neural networks to learn by example. In deep learning, a computer model learns to perform tasks directly from images, text, or sound, with the aim of exceeding human-level accuracy.
Businesses can use deep learning to detect imperfections in their operations. By training neural networks to recognize the difference between unblemished and defective images, they can become aware of even the most minor imperfections that would otherwise be missed by the human eye. For example, a wine bottling operator might teach the system what a perfect wine bottle looks like, so that it can detect anomalies that don’t match the spec, such as the label falling ever so slightly outside the approved label area. Once flagged, production managers can decide if it the imperfect bottle is okay to process and ship to customers.
In another example, a car manufacturer may require holes placed two and a half centimeters from the edge of a car door. If these holes are misplaced, additional door components will be affected down the assembly line, such as the screw that will be drilled in. If the machine is trained to distinguish what a perfect car door looks like, then it can detect anomalies in production. Over time, these businesses can begin to identify and analyze error patterns to make improvements upstream and create a more consistent product.
There is no more avoiding it: data has become essential to optimizing manufacturing processes and staying competitive. While the argument “I don’t have enough data to utilize these new technologies” may have been suitable in the past, it is no longer a valid response with the various levels of adoption available. Instead of accepting the costs of underperforming assets, manufacturers can leverage the data they do have (no matter how little or how much), along with IoT and machine learning, to identify the characteristics of machines operating at peak efficiency. They can then apply those traits to subpar equipment for lower total cost of ownership and better return on assets.