Machine Learning for Real-Time Acoustic Monitoring
Signals are all around us and we use them to make sense of the world. Xi has long experience in signal analysis and processing, and evermore, machine learning techniques are being used to supplement classical methods. One example of growing interest is in maintenance planning, especially in the case of gearbox driven wind turbines. Here we discuss some concepts related to machine learning, and how they may be applied to noise analysis.
Noise Characterization for Science and Engineering
Understanding the causes and properties of noise sources is important in a wide variety of fields, ranging from noise mitigation in renewable energy systems to the effects of underwater noise on marine life, to the detection of earthquakes and even nuclear tests. “Noise” may physically just be a series of oscillating waves moving through the Earth, air or sea (or perhaps through the wall when your neighbour’s party is a bit too loud), but correctly sampling, analyzing, and understanding what it means is no simple task. A variety of classical signal processing techniques for analysis of acoustic processes is rapidly being supplemented by machine learning methods that can provide even deeper insights.
One area of great interest is the real-time monitoring of public infrastructure. Xi has worked for many years in renewable energy, and we know that over time all power systems age and wear. Sometimes the wear occurs suddenly and is obvious – a sudden failure due to improper installation, or damage from a once-in-a-century storm. More often, however, wear occurs gradually over long periods of time. One particularly vulnerable type of system we’ve long worked with is wind turbine gearboxes. Gearboxes allow for greater power generation than their direct-drive counterparts, but are more mechanically complex, and thus more prone to wear. This article will look at what neural networks are and how they can be used in machine learning to anticipate and manage the performance of gearboxes.
Using Neural Networks to Improve Maintenance Planning
Early identification of gearbox wear and tear can allow operators to perform repairs sooner and at much lower cost. SCADA (Supervisory Control and Data Acquisition) systems are standard components of any wind farm, and they oversee wind turbine performance in terms of power output, blade and generator spin rates, component temperatures, etc… Continuous noise monitoring, though, is not widely used.
Machine learning for the analysis of gearbox acoustic signatures is an exciting new area gaining interest for real-time monitoring and maintenance planning. People can recognize the sound of their own car when a family member returns home or can identify the voice actor in a new cartoon or can tell the difference between cheap and expensive headphones when listening to their favourite song. This is just people being a machine learning system!
The addition of acoustic machine learning techniques to modern SCADA systems may provide a significant enhancement in maintenance planning. Fundamentally, neural networks are designed in a similar way to how the human brain works: a neural net consists of multiple “layers”. The first layer is an input signal (the sound of a turning gearbox), and the output layer gives an answer to a question: “tell me about the age and wear of this gearbox?”.
This an example of classification: the gearbox acoustic profile corresponds to a “brand new” system, or one with “minor wear”, “moderate wear”, or “major wear”. Neural networks may also, however, provide regressions, i.e. predictions of future behaviour based on current information.
Between the input and output layers of the neural network are the hidden layers, consisting of “neurons”, which simply represent many, many different mathematical operations on the data as it travels through the network between the input (sound) and the output (answer to question). The data is transformed at each neuron, and a decision is made whether to forward it to another neuron or multiple neurons in a process called activation. There, another calculation is performed, and so on… The trick lies in how all the neurons are connected, meaning that a nearly endless number of possible connectivity configurations are possible. Enormous numbers of possible configurations mean an enormous number of possible applications: speech analysis, image processing, identifying cats and dogs in images, text recognition, and even the completely autonomous generation of paintings, poetry or music! (People can debate the quality of this art… Even impressionist cats are possible).
One neural network structure that is widely used for time series data, such as in audio data, is the recurrent neural network. This structure allows, for example, the ability to hold a coherent conversation with other humans (some of us are better at this than others!). We remember what was said 10 minutes ago as well as 10 seconds ago, lending simple speech (words, or more simply, audio waves) context and meaning.
Using Big Data for Supervised Learning
Neural networks don’t work this magic completely on their own, though. Just like people, they need to be trained to do their job well, and just like people, they may not be suited for all jobs! Training involves feeding the neural network raw data – lots of raw data! – and seeing how well the answer (output) matches what you already know about the data (label or ground truth). The network keeps telling you that a training sample recorded from an operating gearbox is anomalous, but you know it’s fine and that it corresponds with a brand-new generator! The mathematical operations the neurons perform in the hidden layers need to be adjusted. One of the most common ways to do this is called stochastic gradient descent, which is a way to reduce the error between the answers the net spits out and the ground truth about the samples you already know.
Training requires large data sets with ground truth labels because:
- There are a lot of neurons and a lot of mathematical operations being performed in order to get from input signal to output answer. The net is, in the end, an EXTREMELY complex nonlinear mathematical function.
- There are a lot of features of the acoustic profile that that may indicate the wear state of the gearbox, such as high-frequency roll-off, the quality of peaks in the spectral distribution (i.e. width, magnitude), energy centres.
It’s complicated! One of the biggest parts of a data analyst’s job is to decide what features in the signal are useful for determining the wear of the gearbox and what features are not. A lot of experimentation is involved, and this has to be done with a knowledge of the network’s structure – i.e. it’s strengths and weaknesses. But many networks already exist which allow for transfer learning to new specific applications, and once a net has been shown to do a particular job well, it can save enormous amounts of time spent on human analysis and quality control.
How Xi Can Help
Xi Engineering are experts in acoustic engineering, but we also have long experience in a wide range of engineering fields where handling, understanding, and presenting results from Big Data is essential. If you have a project that involves signals and want to explore new ways to interpret and understand them, we may be able to help.