INSIDE THE BRAIN: THE HSI ROBOT
Empowering Vertical Autonomy in HMLV Surface Inspection
The world's consumption behaviour towards customisation and hyper-personalisation is leading to increasing demand for high-mix, low-volume (HMLV) production. This shift presents unique challenges for traditional inspection systems, which are often designed for mass production. At DAINVI, we are committed to addressing these challenges by drawing advanced cross-disciplinary engineering principles from the patented technology and doctoral research conducted at Imperial College London, United Kingdom.
The DAINVI HSI Robot is meticulously designed for HMLV Surface Inspection based on Human-centric, Sparse and Inverse techniques. It (i) mirrors the way humans identify and categorise defects and (ii) utilises the concept of sparse controlling points with inverse methods in engineering. The HSI is designed with hardware and software built around our proprietary technology to address the unique challenges in implementing autonomous HMLV inspection, ensuring unparalleled flexibility and efficiency.
i. Human: Identifying Anomalies First
When it comes to surface inspection, we humans don't begin by memorising the names of each defect from a textbook and identifying them during an inspection session. Ironically, although deep learning is inspired by the workings of the human brain, the current non-human approach has become a widely adopted technique for defect detection.
The DAINVI HSI Robot is designed to mimic how skilled human inspectors learn over their lifetime, ensuring a more flexible, intuitive, and efficient inspection process, integrated with the mindset of lifelong learning. We achieve this by first identifying deviations from the norm (anomalies). Over time, we accumulate sufficient knowledge to form categories and then assign names to each category—only then is a known defect type formed.
Factory workers sorting and categorising components into different boxes.
ii. Sparse and Inverse: Advanced Engineering Principles
The use of sparse controlling points is a powerful tool derived from the advanced aerospace research conducted by our founder at Imperial College London [1]. This concept allows for the creation of flexible and lightweight tools using only small sets of controlling points, thereby reducing the complexity and cost of computation.
One such application is in aerospace engineering, where structures are associated with high peripheral costs due to their extra-large physical size, and reducing the number of fixtures to holding these extra-large structures at the smallest possible number of points, or sparsely spaced controlling points, can significantly lower project costs.
A SpaceX Falcon 9 booster being controlled at merely 2 sparse controlling points during transportation [3].
The question is, how do we determine the smallest number of control points required with the most appropriate spacing between them so that the sparsely supported structure maintains a desired shape? Consider a simply supported beam with length L bent under a distributed load P such that:
A simply supported beam bent under a randomly distributed load [4].
and let's assume we can calculate the beam's verticle deflection (y(x)) at every position (x) along its entire length (L) using the equation:
A conceived equation describing the verticle deflection of a simply supported beam under load.
where E is a constant determined by the material, I is a constant determined by the structure of the beam and P describes the load on the beam. Unfortunately, the constants E, I and P are not always available as an accurate information in our imperfect physical world. The only information we can accurately measure is the vertical deflection along the length of the beam. In reality, it is also impossible for us to measure every point along the beam, especially for the extra-large structures in concern. An ideal equation therefore, would enable us to measure only the deflections y(x) at a few spraesely spaced points along the length of the beam, and reach a solution for the verticle deflection of the entire beam without the need for any knowledge to the material, structure and force-related constants.
In fact, this can be achieved by rearranging the equation into:
w = Kc
where w is a vector that contains the deflections y(x). K is a matrix that contains the matrix of equations f(x) with the unknown constants removed and c is a vector that contains all the unknowns constants. The unknowns can now be obtained by inversing the matrix K such that:
c = K-1(x)w(y)
we can now calculate c, by plugging in the postiions of the sparse controlling points into K-1(x) and w(y) to compute the deflections for the entire beam such that:
w = K‧K-1(x)w(y)
which would work if the equation can be solved, but mathematically, this is an underdetermined system with too many unknowns for a single equation. This is where engineering principles and boundary conditions come into play. Consider if we perform a test to clamp down on one end of the beam, and we find that this action of clamping does not actually affect the curvature of the entire beam, then we can reduce a lot of the unknowns by simply zeroing them out, effectively conditioning a well-determined system that can now be solved, whilst still being able to solve for a good estimate of the beam's curvature.
A simply supported beam bent under a randomly distributed load and clamped to zero deflections on one end [4].
These are the core concepts behind "Sparse and Inverse". In the context of deep learning, the curvature of the beam represents the groundtruth (curvature) of the problem (beam) that we are trying to model (yellow line), and with the sparse controlling points (the weights, wooden blocks and clamping etc.) each representing a single data point that we assume we have collected and accurately annotated. Real-world problems are certainly much more complex than a singly curved beam; they can take the form of a plate bending with multiple curvatures.
The shape of the universe in space-time continuum [5].
Thus, modelling real-world problems is the human pursuit of understanding the complete picture (the direction and the degree of bending at every point) of the problem at hand (the plate) by collecting the smallest possible amount of data points (using sparse controlling points). This is the core concept behind the design of the brain of our HSI robot. Whilst various techniques have evolved in recent years due to the rapid advancement of AI, the core concept remain the same.
Various ways to bending a plate [4].
Interested in our technology? Contact us at info@dainvi.com.
References
[1] Lam ACL, Shi Z, Lin J, Huang X, Zeng Y and Dean TA (2015) A method for designing lightweight and flexible creep-age forming tools using mechanical splines and sparse controlling points. International Journal of Advanced Manufacturing Technology, 80(1 – 4), 361 – 372. [doi / url].
[2] Lam ACL, Shi Z, Huang X, Zeng Y, Li Z and Lin J (2020) Die mechanism, apparatus, and method for shaping a component for creep-age forming. United States Patent and Trademark Office. US10,875,074B2. [url].
[3] Tech Vision (2021) How SpaceX Transports Its Rockets. YouTube. [url].
[4] Engineering Models (2018) Plate Bending. YouTube. [url].
[5] Cottier C (2021) What Shape Is the Universe? Discover Magazine. [url].