Artificial intelligence (AI) in industry can help to detect wear and tear and impending damage to expensive machines in good time. Companies thus avoid expensive downtimes. T-Systems expert Roberto Rösler used a very clear example to explain to his customers (and me) how anomalies can be detected and how predictive maintenance works.
It doesn't have to be the most expensive solution to detect anomalies. Initially, a Raspberry Pi is enough. The mini-computer, which is popular with hobby computer scientists and professionals alike, is no bigger than a palm of your hand. Its main memory has 1024 MB. Watson's, IBM's super-brain, is more than 16,000 times more powerful. But for the purposes of Rösler, an artificial intelligence expert at T-Systems, the computing power of the Raspberry Pi is completely sufficient.
Industrial machines often produce vast amounts of data via sensors and log files. If this data is analyzed continuously and automatically, patterns can be identified. If these deviate from the standard, a failure is likely to occur. Then the production manager can take countermeasures in good time. "As experts in predictive maintenance, we always think from end to end of the process," emphasizes Rösler. "But my demo is first about explaining the principle of anomaly detection."
Learn the model, then apply the model
This is where AI comes in. The art is to identify critical data patterns. Rösler trained the Raspberry Pi in a day and a half. Experts speak of "model creation" or model learning. In simple terms, this means that on the basis of previously recorded data, an abstract model is created that describes the normal state of a machine. If the trained software detects deviations from the learned routines, it sounds the alarm. In other words, it applies what it has learned (Model Application).
Rösler's demo object is only at first sight far away from a highly modern and expensive machine park of a world market leader. He is the complex "machine" himself. The sensors of his smartphone register his movements. And the AI monitors whether he is moving within a defined frame. Walking forward and standing up are the normal states, as Rösler taught the AI on the basis of acceleration data. If the Raspberry Pi registers other movements, such as walking backwards or jumping, with the help of the smartphone, it reports an anomaly and causes an LED alarm to flash.
If an industrial robot is out of sync ...
Now you don't have to worry when an expert in artificial intelligence bounces. If, however, an industrial robot on a production line is only minimally out of step with the strict rhythm, the shift supervisor should take a closer look. Perhaps the technicians will have to wait for the machine before the planned inspection date. In this way, they prevent major damage.
Smart anomaly detection is part of an overall predictive maintenance solution from T Systems. "We can automatically examine a wide variety of data streams for anomalies," explains Rösler. It is also important to know exactly what the application scenario looks like. A model in the cloud does not always make economic sense. "AI applications for anomaly detection usually require many resources for learning, but very few for application," says Rösler. "In addition, the speed of response is sometimes the most important thing. Then there is no point in transferring the entire amount of data to the cloud first. It would then be better to evaluate the data directly on the machine. Latency times and transmission costs are then eliminated. "So we're moving intelligence to the edge of the cloud, to the edge cloud. This is where networked mini-computers work, such as the Raspberry Pi.
Whether AI from the cloud or on the edge device: Industrial companies have three applications for anomaly detection:
- The early warning system.
- The cause analysis, in order to identify and eliminate the causes more quickly in the event of damage.
- Risk assessment to obtain a detailed picture of the condition of the entire machinery.
T-Systems will be demonstrating at the Hannover Messe how this works in industrial practice. An oil company, for example, offers an application scenario. If a lot of data is generated, communication channels are limited and fast decisions are required, data processing on the edge of the cloud can help. Edge computing enables an energy company to control several hundred pumps and line controllers in real time.