For a PDF of the whitepaper, click here.

What is the Secret to Preparing Your Data for AI Success?

Sean Hunter, Eclipse Mining Technologies, Tucson, AZ

INTRODUCTION

Artificial Intelligence has a long history but is currently coming strongly into focus with new advancements. This leads to questions many people are asking: what exactly is AI, how can it be used effectively in mining operations, and what foundation is required to make AI implementations successful?

In mining, many AI applications exist, from drill pattern optimization to predictive maintenance. However, many implementations struggle due to data integration challenges. Mining operations generate large amounts of data from many sources, such as fleet management systems, drill sensors, process control systems, and laboratory results. Unfortunately, this data often exists in silos where it’s difficult to get it out, making AI applications unable to leverage it. Worse still, tools that use AI don’t always provide their data back to the system, creating even further data silos. All these problems make it hard to take advantage of the data at the mine, slowing down and reducing the effectiveness of AI.

This paper presents a comprehensive approach to organizing mining data using an AI-ready knowledge graph system. We will first explore the evolution of AI in mining, from early expert systems to modern machine learning applications. Then, we will detail how a knowledge graph can integrate business context, operational data, and AI models into a cohesive system that enables more effective AI implementations across mining operations.

The proposed method has been developed considering both the unique data challenges faced by the mining industry, while also considering best practices learned from the AI field at large. By following this approach, it’s possible to better utilize existing data, getting higher quality and more complete results faster, and in a way that’s reusable and composable.

HISTORY OF AI IN MINING

Artificial intelligence in mining is often viewed through the narrow lens of machine learning and predictive analytics. However, AI covers a much broader range as it has evolved over decades. Understanding the full breadth of AI can help assist in overcoming the data challenges behind successfully implementing an AI solution.

CURRENT VIEWS ON AI AND MACHINE LEARNING

In mining, there is a strong focus on AI/ML (Machine Learning.) This is the branch of AI which learns from historical data to build models to make future predictions. This can cover everything from predictive maintenance to recovery rate estimation. [1]

Some products are based on this with their own pre-trained models, drawing on their own data sets. For example, a pre-trained classification model could estimate an ore body bounds underground given photos.

These models can be created from simple techniques like linear regression all the way up to more complex approaches using things like neural networks and gradient-boosted trees. Over the years, a large toolkit of these methods has been created, which can handle a wide range of cases and data types.

While this is an important branch of AI, there are many other important cases to consider.

EARLY SYMBOLIC AI

Historically, AI was separate from ML and was more focused on exact answers rather than estimates and predictions. This covers ideas like deductive databases, knowledge graphs, and expert systems.

All these systems are built on the idea of creating new information from base data by following a set of predefined, static rules. A deductive database, for example, is a database where rules will take the base set of data in a database and define new data that’s well defined by that base data, recursively. Tasks here generally focused on understanding the relationships between certain “things,” such as understanding that the brother of your parent is your uncle. This sort of framing worked well for certain well-defined cases, including diseases and categorizing species.

While this type of AI struggles with uncertainty, it can be very effective in cases where things are well-known, and some of these systems are still used today. The OWL framework has been used to organize data on the web. Ontologies and knowledge graphs are created as part of a mapping step before doing potentially costly tasks like creating data warehouses[2] because it forces a rigorous understanding of an area before investing time and resources. It hasn’t had as many direct use cases in mining, though it can still be valuable in organizing and understanding complex mining data and its relationships.

DETERMINISTIC OPTIMIZATION

Operations Research (OR) is the discipline of using analytical methods to improve decision-making. The boundary of this and AI is often blurred, as methods that can generically be used to make smart decisions can be thought of as “artificial intelligence.” In mining, optimization methods from OR are used widely, though not often considered to be AI. Importantly, these also can solve these problems optimally, finding the mathematically best result.

One example of this is scheduling problems. Starting with the Simplex algorithm, which was invented by Dantzig in the 1940s, these sorts of problems can be solved with exact solutions using tools like mixed-integer linear programming. With a well-defined objective and constraints, optimal answers can be found. Work in this area has continued up to today, being able to handle large, complex cases like optimized direct block scheduling.

A related area is the Pseudoflow algorithm [3], an algorithm which can solve for optimal flows in a network, which is equivalent to some other optimization tasks. This algorithm is often used for pit optimization for example, but it’s useful across a range of tasks that require picking an optimal set given explicit constraints. It can be thought of as a binary integer programming method in this sense. This can be used to optimize underground workings, for example, where each piece of construction has a cost, but it unblocks a set of further things that can be built. Using Pseudoflow, you can find the optimal set of things to build where you’d make the most money.

Like the symbolic AI, these methods work great where there’s a clear objective and no uncertainty. It’s important to realize that these methods are only as “optimal” as how correctly they model reality. Using an optimal method on estimations where there can also be unknown effects in operations means the result may not be the optimal method to execute.

DEEP LEARNING

Machine learning has expanded quite a bit in recent years, especially in deep learning. Deep Learning refers to neural networks with multiple layers, allowing them to learn more complex relationships and patterns from data. Advancements like convolutional neural networks have allowed images to be learned from efficiently, expanding neural networks from just working with numbers.

Different combinations of neural network layers are called architectures. For example, the transformers architecture[4], which is used in ChatGPT, has made it possible for AI to understand and respond in natural language. There are other architectures that allow for many other useful things, such as autoencoders, which, among other things, can help understand the similarity between abstract concepts by learning the most important patterns that appear in data sets.

At this point, audio, pictures, and video can also all be used. Neural networks can now efficiently do translation, transcription, image recognition, and more. In mining, this can be used for things like estimating fragmentation from blast face images or analyzing material on conveyor belts with just simple imaging.

REINFORCEMENT LEARNING AND STOCHASTIC OPTIMIZATION

Stochastic optimization is optimization that works even with uncertainty, unknowns, and randomness. This makes it very useful for mining, handling the real-world complications that come with mining operations.

Reinforcement learning (RL) is a key tool for doing stochastic optimization. It involves an agent which learns by taking actions and learning from their results, like how a person learns from experience. Over time, an agent can gradually learn how to handle complex problems, achieving superhuman performance.[5]

This approach is used in sequential decision problems, which are problems where each choice affects future results. A famous example is the AlphaGo AI [6], which plays the game Go at a superhuman level. Go is a complex board game, with more possible board combinations than atoms in the observable universe. Reinforcement learning allowed AlphaGo to defeat the best players in the world, whereas previously, it was difficult to make an AI that played the game well at all. In mining, this same approach can be applied to problems like truck dispatching, where each dispatch affects the state of the mining, affecting the entire shift [7]. Like Go, there’s a huge number of possibly configurations and subtle interactions, but this is an area that now can be tackled by AI.

This work is very new. Reinforcement learning techniques have recently been combined with techniques from deep learning to help agents understand complex data as well as gain new observations from things like images rather than simpler data sets. These approaches show a lot of promise for tackling complex problems in mining operations, full of unknowns and uncertainty.

GENERATIVE AI

Some AI architectures from deep learning can generate images, text, and more using the advancements in deep learning. This is a fast-moving area of AI right now, which is under a lot of active research.

Text generation especially is very powerful, and it can do a surprising number of tasks. The large pre-trained models like ChatGPT, which do these text output cases, are known as large language models (LLMs.) These models take input, which is usually text, referred to as prompts, and generate relevant responses. Despite the power of these models, the prompts need to be created intelligently. Prompts need to be clear and contain enough information for the model to be able to generate a meaningful response.

To solve this prompting problem, in-context learning is used, where data from specific datasets is injected into the prompt. In this area, people have added the ability to chat about the contents of PDF files, books, and more. This has created chat interfaces that can quickly search through huge amounts of data while providing citations to the injected data to be sure that the AI isn’t just hallucinating incorrect outputs. [8]

LLMs have also been extended with the ability to use tools. Code is itself just text, so it can be generated quite effectively by LLMs and then executed. Beyond programming languages, there are formats for transmitting data (like CSV files and JSON files, which use text to represent data), and using these LLMs can be hooked up to interact with programs or even the real world, enhancing their abilities. One case is to let an LLM query a database [9], which can be used to answer questions with accurate data even in a complex environment like a mine.

Having a tool that can do intelligent searches through large, unstructured data, provide answers, and more can be a large performance booster. Mining has a lot of documents, processes, and a lot of complex data, and LLMs can help users sort through it all.

REASONING

Reasoning in classic symbolic AI is defined as using the provided rules to figure out new, implied data. This is like how humans deduce new facts from known information.

Reasoning with LLMs is more general, involving the ability to think through complex multi-step problems. While LLMs can do some reasoning naturally due to the large number of problems they’ve seen in their training data, they’ve struggled with more complicated problems. One example of this is solving a Sudoku puzzle. Solving this would be extremely difficult for an LLM that hasn’t seen that specific Sudoku puzzle before, even though the rules are clear.

Recently, efforts have been made to combine LLMs with Reinforcement Learning to solve these sorts of problems, with OpenAI releasing a first version of this technology with their o1-preview model[10]. As this technology advances, these reasoning AIs will be able to do more complex optimization cases automatically. Combined with in-context learning and the correct data being provided, reasoning LLMs will be able to think through and give answers to complex mining problems where many elements interact.

SUMMARY

AI has had several distinct but complementary approaches. From solving explicit, well-defined problems up to handling uncertain, complex problems that require reasoning, different AI techniques and combinations of techniques can prove useful.

The common thread between all these techniques is having good data and understanding of that data. From the data used to train machine learning models to getting data to use with in-context learning to the understanding of the data to define rules and relationships in expert systems, this data work is required to work with AI.

As AI continues to evolve, proper organization and understanding of underlying data will remain key.

AI-READY KNOWLEDGE GRAPH

The previous section highlighted many different AI approaches and how they relate to mining. However, to effectively leverage those AI technologies, data needs to be organized in a way that makes it accessible. Data organization will allow for rapid development of AI solutions, which can create outsized effects. This section presents a knowledge graph architecture that achieves this type of accessible, organized data.

The proposed knowledge graph is made up of four components that work together synergistically: business definitions, a data lake, an integrated virtual database, and a model layer. Working together, they provide a robust, AI-ready solution.

BUSINESS CONTEXT

Traditional databases typically do not store business context about the data they are holding. This makes it hard to do reasoning in either the classical sense or the LLM sense. For generative AI, context helps generate higher-quality answers. Having business goals and business processes well documented and integrated can help provide this context. [11]

Business goals are the company’s high-level objectives. Being profitable is a key example of a business goal, but this will have many facets and other related goals. For example, a goal might be to improve safety, reduce energy use, better utilize resources, meet production targets, or reduce instances of unplanned maintenance.

Business processes are descriptions of how business is currently being done. These are descriptions of the current processes of one aspect of the business, such as drilling, hauling, loading, etc. There are established standards for this, such as Business Process Model and Notation (BPMN.)

Mining companies often have already invested time and energy to document these sorts of things, from process documentation to corporate objectives, but have not integrated it with their operational, AI, or analytical systems. By including it, AI systems can have a much clearer understanding of the broader implications of their decisions and generate more relevant answers.

DATA LAKE

Mines are constantly generating data from many different sources in many different formats at many different cadences. This can present a challenge, as databases and tools often require transforming data into their preferred format to bring it in. This can cause data loss if the two formats don’t share common definitions. On the other hand, a data lake keeps data in its original format, like how a file system can store files of all types. This means data isn’t lost as it’s brought into the system, and multiple rich views of the data can be created and updated over time, building solutions off this strong base.

Data Lakes are often hosted using Object Storage in the cloud, such as Amazon’s S3 storage or Azure’s blob storage, which is storage optimized for storing large amounts of rarely changing data. These systems can hold a huge amount of data, and access speeds have gotten quite fast, so they’re becoming a go-to solution for analytical systems. [12] Object storage works like a file system, so it can handle everything from CSV files to videos.

There are several proven techniques to load data into a data lake using data ingestion pipelines. For some data, it makes sense to do a one-off import, like exploration results, but some data may require continuous updates, such as sensor data. Data pipelines handle these different cases, ensuring data from all types of sources makes it into the data lake.

Having all the data accessible in a data lake is a good start towards being able to do machine learning, but it’s not quite ready in this format. There’s still a lack of organization at this point, which is covered in other parts of the system.

DATA INTEGRATION, CLEANING, AND AUGMENTATION

Data in a data lake doesn’t have to follow any rhyme or reason. Different sources might refer to the same thing with different names, units of measure can be different, and so on. This can be seen with grade measurements, where one system might use percentages while another uses parts per million or with different naming conventions for the same mining locations.

Fixing these problems is a data integration task. Traditionally this is handled by having a series of data pipelines that take data from a data lake into a database, transforming it and standardizing it into one shared format. There are downsides to this approach, however. The database must now be maintained as a new copy of the data. Instead, it’s much more useful to have a system that’s an enriched view of the data lake.

A “View” in the database sense has a lot of benefits. It can be thought of as a window into the data rather than a copy. The original data is used instead, with strategic caching done to maintain performance on the fly. Views update automatically as underlying data changes, which for sensor data or data that changes frequently in operations can help a
lot. The view can enrich the data with new calculations as well, such as aggregations like shift key performance indicators (KPIs) or different equivalent calculations for different units of measure.

These views can be created using rules in the classic symbolic AI sense. These are logical statements that define relationships – Data A in the data lake implies the existence of Data B available in a view. These rules can standardize things, fixing issues like the naming problem, where the same concept might be referred to differently in multiple places due to differing languages, standards, limitations from a source program, or using jargon. Rules can also be created that define relationships between views and other views, enhancing them, such as calculating KPIs for a shift from the underlying data. These rules implement the classical reasoning discussed earlier, producing clean, integrated data ready for machine learning.

These views are like the idea of data virtualization, which is usually implemented by making a virtual database that’s a view of multiple source databases. By basing the virtual database on top of a known data lake, problems with virtualization can be avoided, such as difficulty in managing permissions.

Defining the schema of these views, which can be thought of as what data belongs where, and creating these rules may seem difficult, but an approach is possible here from the work we did previously. Generative AI can help in defining the schema of this virtual database, using in-context learning of the business-side data and tool usage to make suggestions. Similarly, it can help define the logical rules of how data from the data lake gets represented in this virtual database.

This virtual database, defined by deductive rules, ends up being a bit different than a traditional database. Rather than being focused on the physical schema of data storage, the virtual database’s schema is an integrated definition of the data lake’s contents. The rules, virtual database schema, and the context provided from the business side together create an ontology, not just a database. An ontology is a formal representation of a domain and its relationships, which emerges naturally from this definition.

MODELING

A mathematical model uses calculations and logic as a substitute for a real-world object or concept. For example, a model of a road network can be used to calculate and understand haul truck cycle times, without having to go and physically use the original roads to do calculations.

In practical terms, for the knowledge graph, a model acts like a function that takes some set of inputs to calculate some set of outputs. This can be a trained machine learning model, like the example of calculating mining process recovery for different material types and processes, or other model types, such as reinforcement learning models giving dispatch recommendations, or simple statistical calculations on the input parameters, like averages and variance in production.

Machine learning models can be trained using data from the virtual database since it’s already known to be integrated, cleaned, and enriched. When these models are stored inside the same system, then rules defined in the virtual database can incorporate models, for example, adding predicted production values to each shift, or extracting features from images, creating a further enriched virtual database.

Models can also be built on top of other models, which can allow tackling complex problems from the bottom up. For example, haul cycle time could be created by combining models managing load time, spot time, and travel time, which could all be separate sophisticated models themselves. Once defined, these allow easy reuse, including over department boundaries, where the difficult calculations can be encapsulated.

Model creation can be assisted by generative AI in the same way that it can be used to help generate virtual database ontology. Domain experts can help guide the creation of models, using AI to fill in the gaps. Once created, the domain expertise contained within the model will be usable by the whole company.

PUTTING IT ALL TOGETHER

Together, these components define a powerful knowledge graph that can leverage all the various types of AI discussed previously.

Symbolic AI is used to define the virtual database ontology, creating enriched, integrated data from the raw data in the data lake. The relationships defined here by the rules help provide data for everything else, standardizing the complex set of mining data.

Machine Learning works using this enriched virtual database, where data is known to be cleaned and integrated. This gives a very strong base for training machine learning models and doing statistical analysis.

Deterministic optimization can be done on the integrated data, including testing multiple scenarios. Models allow ways to store and reference some of these what-if scenarios in a clean way a database alone would struggle with.

Deep learning is also enhanced by the data lake’s ability to hold any type of file. This makes it easy to efficiently store and reference data like pictures, video, audio, or whatever other rich data types are present at the mine.

Reinforcement learning can be done in a variety of ways, from running simulations on old data to doing real-time learning as new data gets ingested into the data lake from data pipelines. Reinforcement learning can recommend optimized actions to take, which will also then bring new data to be learned from back into the system.

Generative AI helps tie everything together. Guided by information from the business context, it can help enrich the knowledge graph using the knowledge graph’s own metadata. Generative AI makes interaction with the system possible with natural language, and it makes all written descriptions and metadata into executable instructions. Descriptions of rules, models, and more can aid in creating a truly intelligent system.

The future lies in reasoning AI, which will help the system build itself and perform its own analysis. While generative AI is at the point where it can generate useful suggestions, reasoning will take this much further, planning complex implementations and suggesting sophisticated optimizations across mining operations. This is a natural extension of how generative AI already works inside this system, so it will naturally work as AI continues to evolve.

CONCLUSION

This paper has presented a comprehensive approach to organizing mining data using a knowledge graph architecture. This knowledge graph approach addresses common data integration issues and provides a strong basis for incorporating AI, both old and new. As AI also continues to evolve with reasoning, this knowledge graph will show more and more benefits.

From the flexibility of the data lake to the structure of the ontology, the knowledge graph can adapt to the myriad challenges presented by the mining industry and be ready for future technology to come.