AI Hardware for Condition Monitoring of Machines: Key Components and Novel Trends

By Eugene Zetserov, POLYN Technology

The use of artificial intelligence is rapidly expanding in areas such as robotics, machine learning, and computer vision. However, comprehensive hardware must be integrated right from the beginning to fully utilize AI’s capabilities.

In the past, there was a widespread assumption that AI would rely on immense supercomputers housed in data centers. While this notion continues to fuel innovation, AI is no longer restricted to the cloud.

Advancements in semiconductor technology are enabling companies like Google and Facebook to create specialized chips. These chips excel in executing the highly parallel operations essential for AI tasks and enabling concurrent computations alongside other accelerators across disparate servers within the same data center. 2024 could be called the Year of AI at the Edge, and thus power efficiency and low latency become the most essential criteria for AI components.

Until now, AI solutions in the Industrial Internet of Things (IIoT) market have centered on cloud analytics and big data models. This is primarily because effective solutions for optimizing sensor nodes are still lacking. Consequently, existing options require complex and costly products regarding capital and operational expenditures or non-continuous monitoring alternatives. The main elements of the current solution to collect data from machines and mechanisms are MEMS or Piezo sensors (sometimes used by other methods) and MCUs with radio or wire connections.

The increasing number of non-expensive and mass diploid IIoT sensors pose a challenge regarding connectivity, whether to the CPU host or toward the cloud. More IIoT sensors are now connected to the host device via wireless radio channels to tackle this challenge. Transmitting all the raw data directly to a gateway or cloud requires substantial communication channels and considerable energy consumption.

As for the computing method, while algorithms can handle raw data from IIoT sensors, using neural networks for signal processing offers a more compelling approach. Vibration signals can be intricate, especially in complex machinery operating at various speeds and loads. Additionally, background noise and measurement noise further complicate signal analysis.

Traditional methods such as spectrum analysis and adaptive filtering often struggle with such complexity, prompting the need for machine learning techniques. Algorithm-based data compression isn’t effective for noisy signals due to its linear data processing nature.

In contrast, neural networks can extract valuable information from highly noisy data thanks to their nonlinear processing capabilities. They excel at extracting useful features from raw signals, clustering and classifying the data, and addressing vibration monitoring challenges. Specific deep neural network architectures are particularly adept at this.

Neuromorphic Reduction of Sensor Data

A key point in vibration data reduction is the neuromorphic extraction of useful information from the raw data stream by an autoencoder neural network consisting of an encoder and a decoder. The encoder extracts data features from the input signal, forming an embedding (a descriptor, a latent space). From a mathematical point of view, embedding is a projection of the processed signal into a space with a significantly lower dimension.  The decoder reconstructs the signal data from the encoded representation.

In this case, variational autoencoders that impose additional structure to the latent space offer a more flexible approach than simple ones by learning the parameters of a latent space distribution that help generate new data.

Once the autoencoder is trained, its encoder section is used to preprocess raw signals. This reduces the amount of data transmitted wirelessly from the encoder’s output to a classifier placed at the MCU, on the server, or in the cloud by more than 1,000 times.

Fig.1 Converting the input vibration signal into a latent representation

The decoder section is replaced by a classifier that uses the data classes learned during training.

For classifiers, on-device learning or uptraining can be applied when the classifier is retrained on labelled data from the sensor node and the classes are updated or newly formed. On-device learning is a powerful tool that significantly increases the precision of classification.

Cost-Effective Implementation as Neuromorphic Front End

Virtually all vibration sources suitable for machinery condition monitoring involve rotating parts. As a result, the vibration spectrum typically consists of sinusoidal signals at different frequencies alongside non-deterministic broadband vibrations linked to defects, structural issues, and background noise. Decomposing vibrations enhances the utilization of distinct diagnostic methods for each signal component. An autoencoder neural network can identify non-linear components within such spectra, even when these distortions are minor, generating the character of each signal as embeddings.

Embedding generation is a universal process, so a trained encoder can extract embeddings from any vibration raw data. It will extract features even from a new type of input signal that differs from those it was trained on. This allows the clustering of all kinds of data after the neuromorphic embeddings are generated and their classification afterward.

After an efficient autoencoder is found for data volume reduction, the challenge of running this neural network efficiently at the sensor level is still present.  The Neuromorphic Front End (NFE) is a novel hardware component of analog neuromorphic signal processing directly at the sensor node via a tiny ultra-low power silicon chip. It is responsible for extracting only useful data from a raw data stream, reducing the amount of transmitted data by more than 1,000 times, from megabits to kilobits of data; it reduces it to an extent suitable for narrowband protocols for wireless communication with reduced power consumption. With its true parallel data processing, NFE power consumption typically is only about 100 microwatts, enabling long-life battery operation or energy harvesting-based designs.

Various methods can be used for the next step, classification, from simple ones to complex neural network classifiers. They generally fall into several categories: probabilistic methods, nearest neighbors, decision trees, and neural networks. Classifiers implement linear or non-linear transformations on input data.

In vibration analysis, the classes represent different stable states of the machine in question. Examples include different rotation speeds of bearings, a broken sawtooth noise, oil shortage, an abrasive in the bearing’s lubricant, bearing clip deformation, a wind turbine blade imbalance, and cracking and erosion.  If a data point does not fall into the known classes, it probably indicates a malfunction or a new machine operation mode.

Fig.2 From the raw signal on the left, the vibration signals are processed by MLP (multilayer perceptron) until the classification on the right.

In the NFE solution, a classifier receives the encoded representations of various vibrations and assigns each data point to one of the classes defined during the autoencoder’s training.

This universal method allows the extraction of embeddings from data of various mechanisms, such as bearings, chain and gear drives, gearboxes and electric motors, compressors and generators, wind turbines, and car tires. The classifier must be trained on data sets of a specific mechanism collected in different operating modes.

Classifier selection depends on the specific application and dataset characteristics. Linear classifiers suffice for linear embedding spaces, while more complex neural network models, such as MLP, excel in handling nonlinear regions in the embedding space of vibrations.

Eugene Zetserov is Vice President of Marketing and Business Development for POLYN Technology (polyn.ai), a fabless semiconductor company offering ultra-low-power, high-performance Neuromorphic Analog Signal Processing (NASP) technology, IP, and Tiny AI chips based on NASP.