Neuronal networks and AI to reshape maintenance operations in the aerospace sector

The aerospace industry, as well as any industry heavily invested in mechanical and electronic machinery, works with the assumption that any system will eventually fail. Aerospace product manufacturers use their knowledge of the reliability of their products to build maintenance programs that increase the life span of their systems and reduce undesired events. But traditional maintenance programs based on Reactive or Preventive Maintenance methods don’t manage to fully optimize the risks and costs, forcing replacement of valid parts by rigid schedules and conservative life limits for critical components.

Predictive Maintenance aims to apply Artificial Intelligence methods to predict the failure at the optimum time, minimizing downtime and maximizing equipment availability and life span. It is a bold approach that would allow for a flexibilization of the standard maintenance schedules applied nowadays in the aerospace industry.

With the coming of a new era of aerospace technology, where innovation is coming in the shape of more efficient and less CO2 intensive energy systems (electrical systems, hydrogen, etc.) the proliferation of sensors in aerospace products will bring a cascade of new data to work with. The best methodology to treat and exploit this data in benefit of aerospace operations are deep learning algorithms such as neural networks.

Machine learning allows our system to learn critical information about our data to perform a desired task. Deep Learning is a machine learning subset that requires less human interaction and gives our system the capability to analyze even bigger amounts of data. In this article we are going to review the basic concepts around deep learning, neural networks and the best kind of network for our problem.

Neural networks are the quintessence for Deep Learning. They try to approximate an f* function. For example, a classifier, y = f*(x) maps an input x to a category y. To summarize, a mapping is defined y = f*(x;α) where the model tries to learn the α parameters that made the best approximation to the real function.  They are called networks since they represent a composition of functions where each one is attributed to a layer. In fact, they are also known as Deep Learning due to the greater number of layers, the greater the depth of the network.

They are also called neural since they are highly inspired on neuroscience. Each layer is represented for a vector and each element inside the vector represents a neuron. Model’s width is estimated using vector’s dimensionality.

We are going to short review the basic different types of neural networks:

Convolutional Neural Networks (CNN) are specialized processing grid-based data. Images can be seen as a 2D grid of pixels.

The first set of layers identifies features of the image whilst the tail end classifies the features. Feature identification is done by applying three operations. First, convolution and ReLu (rectified linear unit) are applied and then pooling is performed.

The output is then flattened to a smaller dimensional vector and the fully connected layer is applied to include all of the learning that was done. Each value is then passed through a softmax function that outputs a set of probability values. The maximum probability value for the class is then used to classify the image.

Recurrent Neural Networks (RNN) are specialized processing sequential data. They take profit of the idea behind sharing parameters between different parts of the model. Such concept makes possible to extent model for temporal data.

To summarize, we refer to RNN as an operation over a sequence of vectors x(t) where t is the temporal index. These networks operate by sequence batches with a different length for each member of the batch. Network has backward connections making feedback possible.

Therefore, a RNN has two inputs: the present and the recent past. This is important because the sequence of data contains crucial information about what is coming next, which is why a RNN can successfully perform tasks which other models cannot.

LSTM Networks are RNN that include LSTM units. LSTM units are composed for a cell, an input gate, an output gate and a forget gate. The cell remembers values for a certain amount of time and the other gates manage data flux inside and outside the cell. They improve traditional RNN fixing some optimization instabilities.

RNNs have been used for the following purposes:

• • Forecasting
  • Language Modelling and Generating Text.
  • Sentiment Analysis.
  • Video Tagging.
  • Generating Image Descriptions.
  • Text Summarization.

For operation and maintenance schedules, the prognostic information about the possibility of the system failure in different time windows are required. LSTM networks have the advantage of learning over long time sequences and retaining memory. Therefore, when applying LSTM for system diagnosis, it allows to looking back to the history of degradation processes.

Additionally, CNNs have been successful in finding patterns in 2-dimensional signals. Sequences processed by LSTM can be seen as 2D input (N vector size x M sequence size).

A hybrid model is a good and powerful approach but different architectures can perform equally if they are properly designed

Lastly, we want to point out the importance of quality data and a proper preprocessing. Even robust architectures perform poorly if we lack of enough quality data. 

DMD Solutions offers a wide variety of Reliability engineering services, including Reliability Prediction Analysis. You may also check out the Robin RAMS Reliability Prediction module, which offers a platform for RPA calculations including all major industry standards. Reach out to find out more about our activities in Reliability & related engineering fields.