Industrial Edge Analytics: Detect Issues Faster and Improve Maintenance

Edge computing processes data locally, transforming sensor data into decision-making inference close to the source. It delivers lower latency, higher accuracy, and faster decision-making for real-time industrial analytics at the edge. Embedded monitoring systems comprise sensors, DC/DC converters, and other signal-conditioning circuits, as well as microcontrollers (MCUs) or processors and communication interfaces.

Edge computing provides several advantages over a centralized control system:

• Scalability: Shifting decision-making into distributed nodes close to the machine or system enables easier scalability. Large plants with hundreds of machines face large communications overhead that can be harder to scale with a centralized system.
• Low latency: Fault detection and a corresponding response can be made more quickly, which is critical in maintaining system robustness and reliability. Continuous, real-time monitoring enables fast detection of abnormal behavior with deterministic control-loop timings. Machine monitoring can involve high-frequency data such as vibration signals, motor currents, and temperature, and continuous transmission of these signals on a centralized industrial network can overload its bandwidth and increase system latency.
• Greater immunity to network outages: Local decision-making allows the system to be less reliant on the network for fault processing and less dependent on internet connectivity. Offline operation can be maintained for a longer time.
• Lower-bandwidth networks: Because only high-level information is communicated on the central network, its bandwidth and storage requirements can be much lower.
• Enhanced privacy and security: By keeping sensitive information at the local device, edge computing can reduce the likelihood of data breaches and cyber threats.

Edge AI computing

The addition of AI model-based fault detection and inference can greatly enhance the system's robustness. Models can be trained for high-accuracy inference over a wide range of fault conditions. AI models can be run in the cloud or at the edge.

Historically, cloud-based AI computing is the more common choice. Larger models need high computing capabilities, restricting their operation to servers. However, edge AI is seeing strong success due to the development of smaller, more efficient models, as well as the increased performance of embedded MCUs and processors.

Unlike cloud servers, edge AI hardware runs continuously under harsh environments such as vibration and high temperatures. Hardware selection can involve a tradeoff between speed, low-power capability, size, memory capacity, and computational performance. AI-enabled MCUs have the advantage of deterministic real-time operation that ensures predictable timing for sampling and control loops.

Another approach is the use of smart sensors that incorporate AI chips. These devices can monitor sensor data and apply embedded AI algorithms to diagnose and generate fault alerts.

AI used in embedded applications is typically based on machine learning (ML). Models are trained to identify patterns and make inferences to accurately detect anomalies. Some common ML models are decision trees, support-vector machines, and random forests. A subset of machine learning is deep learning (DL), in which artificial neural networks are used. These can be used for complex patterns and require greater computational and memory resources.

Let's look at two examples of embedded edge AI analytics.

Microchip's dsPIC33A DSC for arc fault detection

Arc faults can occur from power spikes, insulation degradation in the power transmission path, degradation in semiconductor devices or passive components used for power conversion, or failures in the load. They create safety hazards for personnel and equipment. Arc faults in different applications have their own unique signature, e.g., arcs from an electric-vehicle battery disconnect versus those from a solar inverter.

DC power transmission has seen significant growth in recent years. Solar energy generation and EV charging stations are examples of DC power generation and load. Unlike AC systems, arc faults in DC systems do not have zero-crossing, which makes them self-sustaining. Series DC arcs in solar panels, created along a single conductor wire, can be hard to detect, as they cause a small drop in overall current and there are fluctuations in the load.

A challenge to creating useful ML models for DC arcs is the lack of widely available datasets. Customer arc datasets are proprietary and not typically shared by embedded controller manufacturers. One approach to test out the ML toolchain is the use of the publicly available IEEE low-voltage DC arc datasets, which Microchip Technology Inc. uses in its arc fault-detection reference design.

Microchip's dsPIC33 digital-signal controllers (DSCs) include a digital-signal processing (DSP) engine and integrated high-performance peripherals such as analog-to-digital controllers (ADCs). These controllers are suited for applications such as precision motor control, digital power conversion, and advanced sensing.

The dsPIC33A is Microchip's most advanced series within the family. This DSC features a 32-bit CPU capable of 200-MHz operation, a high-speed DSP engine that can perform single-cycle multiply-accumulate instructions for numerical processing, and a double-precision floating-point unit for math operations.

High-speed peripherals include a 12-bit ADC capable of up to 40 mega samples per second (MS/s), 100-MHz gain-bandwidth op amps, 5-ns comparators, and 1-MS/s digital-to-analog converters. While a general-purpose MCU uses multistep interrupt service routines (ISRs), the dsPIC33 uses a single-step ISR with context register switching to save critical context. This allows a typical 20-instruction ISR to be compressed down to four to six instructions, enabling low interrupt latency.

The combination of the fast ISR response and 200-MHz operation reduces interrupt latency to less than 0.18 µs, compared with 0.4 µs with a general-purpose MCU running at 120 MHz. This enables a faster response to time-critical tasks such as real-time arc fault detection and system shutdown.

Software development support is critical to the successful deployment of MCU hardware in a real-time application. Microchip's MPLAB ML development suite converts raw sensor data in both time and frequency domains into statistically meaningful features that are used to train and evaluate multiple algorithms to identify the best-performing model. It supports model training, validation, and optimization and can generate a C API that can run on the MCU.

Using the IEEE DC arc datasets, the MPLAB ML was able to train a model in just 13 minutes. The decision-tree-based model generated for the dsPIC33A DSC was highly efficient, with a size of only 1,142 bytes, utilizing 1,336 bytes of SRAM and 4,522 bytes of flash. A latency of 0.6584 ms was achieved with 100% accuracy on the DC arc dataset. This compares with only 57% accuracy achieved on a rules-based series fault-inference approach. A reference design (see Figure 1) based on the dsPIC33AK128MC102 DSC is available.

TI's MSPM0G5187 MCU for gearbox fault detection

A neural processing unit (NPU) is a specialized hardware accelerator optimized for neural-network computations. While not as flexible as a DSP engine, NPUs can provide improved speed/performance/power tradeoffs for DL models that require extensive matrix-multiplication operations, as is required in convolutional neural network (CNN)-based AI inference. The integration of an NPU into an MCU allows computationally intensive tasks to be offloaded from the main CPU to the NPU, which improves inference speed while reducing power consumption.

Texas Instruments Inc.'s (TI's) MSPM0G5187 MCU features a TinyEngine NPU integrated into the 32-bit Arm Cortex M0+-based CPU. This NPU is capable of 640-2,560 mega operations per second on pretrained CNN models. It has been shown to use 120× less energy per inference and have 90× lower latency than software-based AI.

TI's CCStudio Edge AI Studio provides a graphical environment for model development, training, and compilation. In addition to a variety of pretrained models in the TI Model Zoo, developers can use their own models or train models with custom data. The neural-network compiler translates the trained ML models into optimized C code for leveraging the specific hardware and enables the seamless deployment to either the NPU or CPU.

TI has demonstrated the use of edge AI analytics with its MSPM0G5187 MCU for gearbox fault detection. Monitoring vibration patterns through sensors and applying ML algorithms to detect anomalies can allow for predictive rather than reactive maintenance of industrial systems such as gearboxes and motors. 1D CNNs are highly effective for time-domain analysis of sensor data such as vibration patterns.

The vibration dataset used for the model was generated using SpectraQuest's Gearbox Fault Diagnostics Simulator. This dataset includes recordings from four sensors placed in four different directions, with the load varying from 0% to 90%. This open-source dataset covers both healthy and broken-tooth gearboxes. Eight trained models were created to run on the MCU using the vibration dataset. Figure 2 shows the improved accuracy of 99.9% achieved on the optimized AI-based model versus traditional rules-based detection.

The post Industrial Edge Analytics: Detect Issues Faster and Improve Maintenance appeared first on Embedded.