Towards zero touch - automating site inspection
The journey toward zero touch network operations is just beginning. We are confident concepts like machine reasoning will play a key role in its realization, helping us to automate every aspect of our network, including our field services operation for inspection of radio sites. In this post we present an example of how machine reasoning can be applied to simplify and scale network operations, increase human safety and reduce operational costs.
The example is an AI based system that automates radio site inspections by adding the capability of intelligent drones. We explain the principles of such AI reasoning-based system, including its Intents, Planner and Knowledge Model building blocks, capable of extracting domain expertise or intent from, for example, translating it to knowledge, generalizing it and creating AI based plans for newly specified inspection tasks.
Today, radio site inspections are continuously performed by teams of field engineers. Every field visit for preventive checks, maintenance or problem-solving aim to reduce operational costs (OPEX). , Every tower climbing operation exposes a team of engineers to risks.
If you ever worked in a team that saw an expert move to a new position you probably remember the time and effort it takes just to catch up with the knowledge the expert had. It would be great if we could continuously learn from the expert, document and share their knowledge and experiences. But how?
Thus, the key questions are:
- Can we automate repetitive steps of such highly specialized jobs as a field engineer performing radio site inspections?
- Can a system learn from experts, generalize knowledge and apply to different situations?
- Based on generalized knowledge and reasoning, can a system even plan for situations it was never exposed or trained before?
With these challenges in mind, we have created an AI reasoning-based system that automates radio site inspection, adding the capability of intelligent drones, but still keeping humans in the center of operations. The expertise of the field service personnel is leveraged, while ensuring their own physical safety. Therefore, the strengths of humans like higher order thought and creativity are combined with the strengths of AI, such as reasoning at scale and knowledge transfer.
Our AI reasoning-based system
Thus, the “How” is given by an expert, the “What” is captured by the system, which is further generalized into the bigger “WHAT”. The result is a system that learns by example, is transparent, can transfer its knowledge, and can scale well.
Further, we envision zero touch as an iterative process, where the drone can interact with humans and learn to perform tasks with better efficacy. Thus, the plan can be edited, with the human preferring to edit or delete particular actions.
An example of the interface seen by the Field Engineer is shown here. It’s taken from the proof of concept we demonstrated at MWC. In the first screen, the engineer can fly the drone to different points of interest and perform actions.
The training actions can be reviewed and expanded to include extra information, like which post-processing to run on the action results. In the example below the trainer annotated the photo action.
The next screen shows an automatically derived plan for a new radio site (never trained on before). The actions will also include the annotations from the previous step, where appropriate.
The plan is transparently editable and includes previous knowledge that the system has. In this case, the previous knowledge are simple facts about the battery levels for each drone and their energy use while flying – but just imagine working with a system that can easily fetch and reason about the knowledge of all products and services combined with expertise learned in the field.
With the plan approved we can finally see a drone dutifully executing it, step-by-step.
The following components constitute the intelligent system:
Intent Recognizer: takes a training action and determines the most likely component of the site for the action – detects that the action takes a photograph, e.g., of antenna A from its front left.
Intent Generalizer: generalizes the training actions so they apply to classes of components, taking into action the requested post-actions like which analyses to run on photographs.
Intent Instantiator: converts a generalized intent to a concrete goal adapted for a new site using the knowledge about the new situation to map the goal classes to components.
Planner: receives the description of the goal and the environment and searches for a feasible plan to achieve the goal. The search can be directed to optimize for different criteria, e.g., the least number of steps or lowest amount of energy spent.
Knowledge Model: contains the descriptions of the environment, actions, components, component classes, radio sites and the dynamically provided knowledge – training sample and its generalization.
The block diagram above shows the architecture of the system. The procedure followed by the engineer in training the system is denoted by ℘0. The intent recognizer captures the intents G0, which are generalized to G and stored in the knowledge model database KM. When a new Task T1 arrives, the generalized goals G are composed with T1 through logical inferencing. This creates an instantiated goal state G1. This goal is input to an automatic planner such as Metric-FF , along with the current state Init1 of the environment. The planner derives a plan to achieve T1, i.e. using AI planning techniques to find a sequence of actions from Init1 to G1 by searching the state space. The resulting plan can be reviewed and modified before it is sent to execution.
At Ericsson Research AI, we are constantly looking for new use cases on automating operations based on responsible AI principles. Our intent is to build a human centric operations engine, helping our experts to fully access complex network management data at scale.
AI learning and reasoning will permit us to automate complex tasks, perform network predictive maintenance, anticipate problems and take clearly explained actions to prevent them. Nevertheless, important decisions shall still be taken by experts, who can have a clear view over large scale network management operations, soon to be supported by explainable and responsible AI based operations.
Read more about our Zero touch research in these earlier blog posts:
Information about Ericsson's Operations Engine is available here.