[Introduction]The concept of industrial predictive maintenance has been around for a long time, dating back to when people first said “machines will fail soon”. Predictive maintenance is everywhere, from lubricating the bearings inside a watch to maintaining and repairing large power-generating equipment, from simple appliances to complex space stations.
Whereas early predictive maintenance relies heavily on a technician’s expertise and intuition to solve problems or diagnose faults, today’s advanced diagnostic equipment and Industry 4.0 technologies have added Electronic and mechanical sensors to more accurately detect and diagnose problems. Sensors have become an important component of predictive maintenance applications.
Figure 1 Typical predictive maintenance application in Industry 4.0
As an important part of Industry 4.0, the local decision-making system collects sensor data in or near the equipment to make correct judgments, helping maintenance personnel to detect expensive and complex small problems that may be remote equipment in advance, and avoid causing big accident.This feature requires the sensor to have edge processingcapabilities and artificial intelligence (AI), as AI is a key technology for predictive maintenance applications. By implementing AI and edge processing directly on the sensor or host controller, e.g. STM32FP-AI-MONITOR1 indata analysis decisions can be performed locally.
Figure 1 shows a schematic diagram of a typical predictive maintenance application, where sensors detect information generated by the equipment and transmit the data to the main controller. In Industry 3.0, raw sensor data describing the condition of the machine is transmitted directly to the operator without any local processing or decision-making tasks involved. In Industry 4.0, the master controller processes sensor data locally and makes decisions locally. If the transmission conditions do not meet certain notification criteria, the host controller allows the wireless connection module to partially sleep. Operators only get involved after receiving a notification message from the cloud. This approach reduces the amount of data transmitted to the cloud and reduces the power consumption of local sensor nodes.
More deeply, there are four key steps to realize this perceptual decision module: important parameter identification; data analysis; sensor selection and decision tree location selection.
Identification of important parameters
Many parameters can indicate the health of the machine. Designers need to sift through important parameters based on their characteristics and ability to predict the state of the machine. In the application scenario of Figure 2, parameters such as acoustics, temperature, and physical vibration acceleration can all indicate wear on the machine’s heavy-duty bearings. Designers will study and analyze which parameters can be used to predict the 60% healthy state of the bearing. Ideally, just one parameter is enough to provide the most meaningful information and allow the decision tree to determine that the bearing health is 60% healthy.
In this example, the health of the machine is divided into four stages, as shown in Table 1:
Table 1 Stages of machine health status
Figure 2 Relationship between important parameters and machine health
Setting an early warning when the heavy bearing reaches 60% health, we captured and plotted acceleration, ultrasound and temperature versus time (weeks) to analyze and study important parameters, as shown in Figure 2, all three parameters were It can indicate the wear condition of the bearing. The research found the following:
● When the bearing enters the damage phase after t3, the acceleration data gives a strong signal. However, it does not track the health state well before t3, that is, it cannot effectively record the state of the machine before it reaches 50% health state, which means that we cannot accurately predict the health state of the machine before the bearing is damaged, so, only Relying on the indications from the accelerometer is not sufficient to predict the degree of early wear.
● The temperature data cannot accurately track the health of the bearing until the bearing enters the damage stage t4. Regardless of the cause of bearing damage, temperature parameters do not give a clear signal of bearing damage until friction increases dramatically.
● Ultrasonic parameters can effectively track the health of bearings and can signal as early as t1. As friction increases, it sends a clear signal when the bearing reaches 60% health. However, from the plotted data, when the bearing health drops below 50% around t3, the ultrasonic signal starts to lose track of the machine’s health, this is because the bearing is severely worn and cracked, greatly changing the bearing’s characteristics, And cause the vibration curve of the bearing to exceed the ultrasonic scanning range. The strong vibration at this stage can just be sensed by the accelerometer.
From this example, it is not difficult to see that ultrasonic detection is an important parameter for predictive maintenance to achieve 60% early warning of health.
Once the important parameters have been identified, the next step is to study the data profiles. Designers must evaluate different data processing capabilities and artificial intelligence algorithms to reliably predict machine health.
There are many data processing methods available for predictive maintenance applications, which can be grouped into two broad categories: time domain and frequency domain. Each method has various advantages and disadvantages.
● The time domain method is simple and easy to understand and requires low computing power. The output of the sensor is always in the time domain. Root mean square (RMS), average or peak detection of time domain signals are typical tracking values. Comparing thresholds or magnitudes of raw or processed data can yield decision flags. The disadvantage of this method is that it is only suitable for simple waveform analysis. In practical industrial applications, some data analysis is complex because they may contain vibrations of different mechanical components and environmental vibrations of other machines. Figure 3 shows an example of data analysis in the time domain.
Figure 3 Time-domain acceleration waveform example
In this example, the magnitude of the vibration produced by the motor unbalance is much greater than that produced by the output shaft. If RMS or averaging or other time domain signal processing methods are used, the sensor cannot effectively identify the vibration level of the output shaft.
Figure 4 Complex waveform composed of multiple waveforms
● However, there is a powerful signal processing method to manage complex signals. This type of complex waveform consists of multiple simple waveforms, as shown in Figure 4. The Fast Fourier Transform (FFT) is an effective waveform analysis tool that converts time-domain data into frequency-domain data, placing vibrations from different components in different frequency spectra, as shown in Figure 5.
Figure 5 Spectrum
The Fourier transform method divides the vibration amplitudes of different sources into different frequency spectra. In addition to Fourier transform, data processing can also use other technical methods, such as average, RMS, peak, neural network, etc., to perform accurate data filtering, provide more reliable data for decision trees, and achieve more intelligent decision making.
Parameter identification and data analysis requires some tools. Here are some commonly used tools:
✦ Professional measurement tools
Accurate and detailed measurement data can be obtained using off-the-shelf professional measurement equipment, which is highly recommended for demanding high-precision applications.
✦ Evaluation Demo Kit
Sensor manufacturers such as STMicroelectronics offer evaluation kits that write-free software (Figure 6). These small boards, for example, the STEVAL-MKI109V3, have sockets for the sensor boards. Designers can choose to plug their favorite sensor board into the motherboard. Some manufacturers also provide graphical user interface (GUI) software for controlling the sensor. These GUI software can access all registers of the sensor, configure and retrieve data without writing code, and provide practical data processing arithmetic functions, such as Fourier transform FFT is one of them (Figure 7).
Figure 6 The connection between the STEVAL-MKI109V3 evaluation board and the sensor board
Figure 7 STEVAL-MKI109V3 GUI screenshot
To evaluate the sensor’s features and applicability, it is recommended to use a code-free evaluation board. These boards can also perform initial data acquisition, start engineering algorithms and data analysis processes. When it comes to later prototyping or proof-of-concept stages, sensor manufacturers may offer another powerful development tool to greatly simplify development tasks and shorten development cycles. Take the STWIN development kit as an example:
✦ STWIN Wireless Industrial Node (STEVAL-STWINKT1B)is a development kit and reference design that simplifies prototyping and testing of advanced industrial IoT applications such as condition monitoring and predictive maintenance.
Figure 8 STEVAL-STWINKT1B
Figure 9 SensorTile Box interacts with mobile phone
STWIN development kit is based on STM32 ultra-low-power microcontroller and integrates various industrial-grade sensors, including inertial sensors (vibration sensor, accelerometer, 6-axis IMU, magnetic sensor), environmental sensors (high-precision temperature sensor, pressure sensor, humidity sensors) and high-performance microphones (digital and analog, with ultrasonic sensing capabilities), support all types of condition monitoring, especially those related to vibration analysis. The development kit also comes with rich software packages and optimized firmware libraries, as well as a cloud-based dashboard application to speed up the design cycle of end-to-end total solutions.
The kit has an onboard Bluetooth® low energy wireless connectivity module and can plug into a Wi-Fi wireless connectivity daughter board (STEVAL-STWINWFV1). Wired connection can be achieved through the onboard RS485 transceiver.
With the data analysis tool at hand, the next step is to choose the right sensor:
a) Select the sensor type according to the important parameters found in one
STMicroelectronics provides various sensors such as accelerometers, gyroscopes, magnetometers, vibration sensors, microphones, pressure sensors, humidity sensors, temperature sensors, laser sensors, infrared sensors, and more. Industrial-grade sensors typically offer higher performance and accuracy, better temperature and time stability, and even product life-cycle guarantees.
b) Select the sensor range according to the maximum measurement range and sensitivity or important frequency range (bandwidth) found in 2;
Each sensor has its own maximum range and frequency response bandwidth. Designers must carefully study these two parameters to select the most suitable sensor. Figure 9 shows a range of models we recommend for predictive maintenance application scenarios.
Figure 10 Selecting sensors according to application scenarios
Decision tree location selection
As an industry-recognized pioneer in MEMS technology, STMicroelectronics is the first to embed edge processing in sensor products. Designers can partition edge processing in sensors or embed decision trees within the main controller. The best choice depends on the data processing and the complexity of the decision tree. Decision-making functions in ST’s sensors fall into three categories:
● Embedded simple logic
STMicroelectronics MEMS sensors have simple embedded threshold comparison logic. The interrupt flag is triggered once the amplitude and time window thresholds reach the preset values.
● Finite State Machine (FSM)
A state machine is a mathematical abstraction for designing logical connections (Figure 10). An FSM is a behavioral model consisting of a predetermined number of states and transitions between states, similar to a flowchart. Sensors can be set to generate decision flags as soon as a user-defined mode is met. To facilitate decision-making, some of ST’s sensors have embedded 16-state machines.
Figure 11. Embedded finite state machine of the sensor
● Machine Learning Core (MLC)
The MLC machine learning core is not designed to handle complex data, so it cannot do the work of a finite state machine. MLC can indeed offload some low-density algorithms that would otherwise run on the application processor to the MEMS sensor, thereby significantly reducing system power consumption. MLC recognizes data patterns when they match a user-defined set of classes. The sensor filters the input data using a configurable dedicated computation module containing filters and features computed over a fixed time window set by the user. The basic principle of machine learning processing is a logical process of comparing preset thresholds and “feature” values on an “if-then-else” condition through a series of configurable nodes (Figure 11).
Figure 12 Decision-making process within the sensor’s MLC
In conclusion, as a fundamental part of Industry 4.0 applications, sensors are essential components in predictive maintenance and, with built-in intelligence, sensors can reduce the load on the main controller, thereby increasing the energy efficiency of the entire system. As a leader in the MEMS sensor industry, STMicroelectronics offers a full range of sensors (accelerometers, gyroscopes, magnetometers, vibration sensors, microphones, pressure sensors, humidity sensors, temperature sensors, laser sensors and infrared sensors, etc.). In application areas such as predictive maintenance, this wide range of products forms an important bridge between innovative concepts and practical applications.
 Industrial Evolution
 0.061mg/LSB=0.061×9.8milm meter/s2/bit
 Sensors with Machine Learning
 Finite State Machine in MEMS Sensor
 FP-AI-Monitor1: STM32Cube function pack for ultra-low power STM32 with artificial intelligence (AI) monitoring application based on a wide range of sensors
 STM32: 32-bit Arm Cortex MCUs provided by STMicroelectronics
 Capacitive MEMS accelerometer for condition monitoring
 STWIN SensorTile Wireless Industrial Node development kit and reference design for industrial IoT applications
 How to use the STEVAL-STWINKT1B SensorTile Wireless Industrial Node for condition monitoring and predictive maintenance applications
 IIS3DWB Sensor Adaptor Board
 Edge Processing (Edge Computing)