One of the results of the maturation of technologies such as the Internet of Things (IoT), Augmented Reality (AR) and cloud computing is the rise of the smart factory. An increasingly familiar sight in smart factories are collaborative robots. Collaborative robots already play an important role in smart manufacturing and will take on more functions and provide greater value in the factory over time.
As smart factories increasingly rely on collaborative robots to fulfill their essential functions, it becomes even more important that they operate reliably and without unplanned downtime. This has prompted collaborative robot manufacturers to enable predictive maintenance in their products: early warning for users in the event of a failure that could ultimately jeopardize the operation of their collaborative robots. It offers the opportunity to fix faults within planned maintenance times without the disruption caused by unexpected machine failures.
In collaborative robots, predictive maintenance systems rely on sensors that detect small anomalies in the movement of limbs and joints, as well as in the motors that drive them:
- Sensors such as accelerometers and Inertial Measurement Units (IMUs) can detect vibrations caused by bearing wear, etc.
- Ultrasonic sensors pick up unique acoustic signatures to detect excessive friction.
Machine learning techniques, a branch of artificial intelligence (AI), are used to enable collaborative robots to detect differences in vibration and sound patterns from a reference point when the collaborative robot is brand new or in a known undamaged state. Analysis of abnormal patterns allows the system to diagnose early failures and trigger requests for planned repairs and maintenance to the plant management system.
In early implementations of machine learning, complex neural network algorithms for recognizing patterns in sensor signals typically ran remotely in powerful microprocessor-based embedded computing systems.
However, such centralized systems placed a heavy burden on processing equipment when dealing with inputs from a large number of collaborating robots, resulting in high power consumption and taking up a significant amount of bandwidth in the network connecting the collaborating robots to the central control system.
The advent of a new generation of sensors with embedded AI capabilities now offers collaborative robot manufacturers a new way to enable local machine learning. Using tools and software from STMicroelectronics, a pioneer in the development of machine learning sensors, collaborative robot design engineers can take advantage of a new, simpler way to build predictive maintenance capabilities into their products.
A wide range of MEMS sensors for vibration and ultrasonic measurements
ST offers one of the industry's largest portfolios of MEMS sensors, including accelerometers, IMUs, pressure sensors and microphones. The sensing elements are manufactured using specialized micromachining processes, while the IC interfaces are developed using specialized CMOS technology. This enables the design of specialized circuits that match the characteristics of the sensing element.
This technology underpins the high performance of the IIS3DWB, for example, the three-axis ultra-wide bandwidth MEMS accelerometer, which is ideal for detecting vibrations generated by faulty machines. ST also offers motion sensor modules based on its MEMS sensor ICs: the ISM330DHCX, for example, is a system-in-package product that includes high-performance 3D digital accelerometers and 3D digital gyroscopes tailored for Industry 4.0 applications.
Machine learning based on decision tree logic
The ISM330DHCX is one of ST's MEMS sensor offerings that includes embedded AI functionality in the form of a machine learning core (MLC). This machine learning capability enables system operators to transfer some predictive maintenance algorithms from the central application processor to the sensor, with the dedicated MLC consuming much less power.
So how can the sensor's small, low-power processing logic block provide the machine learning capabilities that would normally require a large, power-hungry application processor?
The answer lies in the decision tree logic that ST embeds in its smart sensors: ST-enabled decision tree algorithms are simpler than traditional neural network algorithms, and therefore consume far fewer instruction cycles and power.
A decision tree is a mathematical tool consisting of a series of configurable nodes. Each node represents an "if-then-else" condition that compares an input signal (i.e., a quantitative value calculated from raw sensor data) with a threshold value.
The ISM330DHCX can be configured to run up to eight decision trees simultaneously and independently. The decision trees are stored in the device and the results are generated in dedicated output registers. The results of the decision tree can be read at any time by the host microcontroller or application processor. The sensor can also generate interrupts for each change in the results generated by the decision tree.
How the Decision Tree Logic Works
The predictive model for the decision tree is constructed from a set of training data and stored in the ISM330DHCX. The training data is recorded in its desired state (i.e., in good condition, free of faults) during operation of the collaborative robot.
A decision tree is a method by which MLC analyzes common features in the raw sensor data. These common features will form the basis of a "model" that the sensors will use to compare the operation of the collaborative robot. If the sensor output highly matches the model, the collaborative robot is fault-free. If the sensor is unable to match its real-time measurements to the model, a potential malfunction is indicated and an alarm is sent to the machine operator.
Each node of the decision tree contains a condition under which the features are compared to a specific threshold. If the condition is true, the next node in the true path is evaluated. If the condition is false, the next node in the false path is evaluated, as shown in Figure 1. The state of the decision tree will evolve node by node until the result is found. The result of the decision tree defines a behavioral "category": in the case of a fitness wristband, such a category might be "walking" or "jogging". In predictive maintenance applications for collaborative robots, different workloads of collaborative robots correspond to different categories.
Decision tree consists of multiple nodes
The decision tree generates a new result for each time window, the length of which is set by the user to capture the characteristics of the relevant activity category. The result can also be modified by an additional optional filter called a "meta-classifier" that applies internal counters to the decision tree output.
The activity categories recognized by the MLC (in the form of filtered or non-filtered decision tree results) can be accessed through the registers of the ISM330DHCX module.