Manufacturers, to keep up with the latest changes in technology, need to explore one of the most critical elements driving factories forward into the future: machine learning. Let’s talk about the most important applications and innovations that ML technology is providing in 2022.
Machine Learning vs AI: What’s the Difference?
Machine learning is a subfield of artificial intelligence, but not all AI technologies count as machine learning. There are various other types of AI that play a role in many industries, such as robotics, natural language processing, and computer vision. If you’re curious about how these technologies affect the manufacturing industry, check out our review below.
Basically, machine learning algorithms utilize training data to power an algorithm that allows the software to solve a problem. This data may come from real-time IoT sensors on a factory floor, or it may come from other methods. Machine learning has a variety of methods such as neural networks and deep learning. Neural networks imitate biological neurons to discover patterns in a dataset to solve problems. Deep learning utilizes various layers of neural networks, where the first layer utilizes raw data input and passes processed information from one layer to the next.
Factory in a Box
Let’s start by imagining a box with assembly robots, IoT sensors, and other automated machinery. At one end you supply the materials necessary to complete the product; at the other end, the product rolls off the assembly line. The only intervention needed for this device is routine maintenance of the equipment inside. This is the ideal future of manufacturing, and machine learning can help us understand the full picture of how to achieve this.
Aside from the advanced robotics necessary for automated assembly to work, machine learning can help ensure: quality assurance, NDT analysis, and localizing the causes of defects, among other things.
You can think of this factory in a box example as a way of simplifying a larger factory, but in some cases it’s quite literal. Nokia is utilizing portable manufacturing sites in the form of retrofitted shipping containers with advanced automated assembly equipment. You can use these portable containers in any location necessary, allowing manufacturers to assemble products on site instead of needing to transport the products longer distances.
Using neural networks, high optical resolution cameras, and powerful GPUs, real-time video processing combined with machine learning and computer vision can complete visual inspection tasks better than humans can. This technology ensures that the factory in a box is working correctly and that unusable products are eliminated from the system.
In the past, machine learning’s use in video analysis has been criticized for the quality of video used. This is because images can be blurry from frame to frame, and the inspection algorithm may be subject to more errors. With high-quality cameras and greater graphical processing power, however, neural networks can more efficiently search for defects in real-time without human intervention.
Using various IoT sensors, machine learning can help test the created products without damaging them. An algorithm can search for patterns in the real-time data that correlate with a defective version of the unit, enabling the system to flag potentially unwanted products.
Another way that we can detect defects in materials is through non-destructive testing. This involves measuring a material’s stability and integrity without causing damage. For example, you can use an ultrasound machine to detect anomalies like cracks in a material. The machine can measure data that humans can analyze to look for these outliers by hand.
However, outlier detection algorithms, object detection algorithms, and segmentation algorithms can automate this process by analyzing the data for recognizable patterns that humans may not be able to see with much greater efficiency. Machine learning is also not subject to the same number of errors that humans are prone to make.
One of the core tenants of machine learning’s role in manufacturing is predictive maintenance. PwC reported that predictive maintenance will be one of the largest growing machine learning technologies in manufacturing, having an increase of 38 percent in market value from 2020 to 2025.
With unscheduled maintenance having the potential to deeply cut into a business’s bottom line, predictive maintenance can enable factories to make appropriate adjustments and corrections before machinery can experience more costly failures. We want to make sure that our factory in a box will have as much uptime with the fewest delays possible, and predictive maintenance can make that happen.
Extensive IoT sensors that record vital information about the operating conditions and status of a machine make predictive maintenance possible. This may include humidity, temperature, and more.
ML Models Used for Predictive Maintenance
A machine learning algorithm can analyze patterns in data collected over time and reasonably predict when the machine may need maintenance. There are several approaches to achieve this goal:
- Regression Models: these predict the Remaining Useful Life (RUL) of the equipment. This uses historical and static data and manufacturers can see how many days are left until the machine experiences a failure.
- Classification Models: these models predict failures within a predefined time span.
- Anomaly Detection Models: These flag devices upon detecting abnormal system behavior.
Thanks to the IoT sensors powering predictive maintenance, machine learning can analyze the patterns in the data to see what parts of the machine need to be maintained to prevent a failure. If certain patterns lead to a trend of defects, it’s possible that hardware or software behaviors can be identified as causes of those defects. From here, engineers can come up with solutions to correct the system to avoid those defects in the future. This enables us to reduce the margin of error of our factory in a box scenario.
Digital twins are a virtual recreation of the production process based on data from IoT sensors and real-time data. They can be created as an original hypothetical representation of a system that doesn’t yet exist, or they could be a recreation of an existing system.
The digital twin is a sandbox for experimentation in which machine learning can be used to analyze patterns in a simulation to optimize the environment. This helps support quality assurance and predictive maintenance efforts as well. We can also use machine learning alongside digital twins for layout optimization. This works when planning the layout of a factory or for optimizing the existing layout.
ML Models for Energy Consumption Forecasting
If we want to optimize every part of the factory, we also need to pay attention to the energy that it requires. The most common way to do this is to use sequential data measurements, which can be analyzed by data scientists with machine learning algorithms powered by autoregressive models and deep neural networks.
- Autoregressive models: Great for defining trends, cyclicity, irregularity, and seasonality of power consumption. To improve accuracy, data scientists can transform raw data into features that can help specify the task for prediction algorithms.
- Deep neural networks: Data scientists use these to process large datasets to find patterns of data consumption quickly. These can be trained to automatically extract features from input data without feature engineering like autoregressive models.
- Neural networks for sequential data: RNN (Recurrent neural networks), LSTM (Long short-term memory)/GRU (Gated recurrent unit), Attention-based neural networks to store information of previously inputted energy usage data using internal memory.
We’ve used machine learning to optimize the factory’s production processes, but what about the product itself? BMW introduced the BMW iX Flow at CES 2022 with a special e-ink wrap that can allow it to change the color (or more accurately, the shade) of the car between black and white. BMW explained that “Generative design processes are implemented to ensure the segments reflect the characteristic contours of the vehicle and the resulting variations in light and shadow.”
Generative design is where machine learning is used to optimize the design of a product, whether it be an automobile, electronic device, toy, or other items. With data and a desired goal, machine learning can cycle through all possible arrangements to find the best design.
ML algorithms can be trained to optimize a design for weight, shape, durability, cost, strength, and even aesthetic parameters.
Generative design process can be based on these algorithms:
- Reinforcement learning
- Deep learning
- Genetic algorithms
Improved Supply Chain Management: Cognitive Supply Chains
Let’s step away from the factory in a box example for a bit and look at a broader picture of needs in manufacturing. Production is only one element. The supply chain roles from a manufacturing center are also being improved with machine learning technologies, such as logistics route optimization and warehouse inventory control. These make up a cognitive supply chain that continues to evolve in the manufacturing industry.
Warehouse Inventory Control
AI-powered logistics solutions use object detection models instead of barcode detection, thus replacing manual scanning. Computer vision systems can detect shortages and overstock. By identifying these patterns, managers can be made aware of actionable situations. Computers can even be left to take action automatically to optimize inventory storage.
At MobiDev, we have researched a use case of creating a system capable of detecting objects for logistics. Read more about object detection using small datasets for automated items counting in logistics.
How much should a factory produce and ship out? This is a question that can be difficult to answer. However, with access to appropriate data, machine learning algorithms can help factories understand how much they should be making without overproducing. The future of machine learning in manufacturing depends on innovative decisions.