Across diverse industries and verticals, there are numerous business processes that are often repetitive and mundane.
These tasks do not often require critical thinking, but in fact, calls for a detail-oriented approach to minimise errors and increase productivity. However, these tasks require a certain amount of human intelligence in order to validate the feasibility of the task or ensure the accurate execution of a task that may be present on a document, image or dataset (in a proposed quality-control related work process).
In the BFSI vertical, the process of filling up customer paperwork and its subsequent verification is deemed repetitive. In the case of Manufacturing, the practice of optimizing Supply-Chains and inventory management has been automated with the help of Artificial Intelligence (AI). These systems are designed to be more precise than the human eye, whilst mimicking human intelligence. Machine Learning (ML) models allow the system to ‘humanise’ their approach by constantly learning how to identify details that were previously only spotted by human interference.
AI systems have even managed to make their mark in niche industries like Green Energy, Real Estate and National Defence.
The ability of AI to successfully navigate repetitive tasks, with the least amount of error, has helped industries overcome bottlenecks and utilise their resources more effectively.
AI in the Telecom industry?
The Telecom industry is a pioneer, with respect to the adoption of innovative and smart technologies. The usage of Virtual Assistants, ChatBots and Smart Interfaces helped in improving the overall customer journey for the telecom segment. Armed with a horde of AI-based systems, the telecom industry has managed to reduce customer friction by streamlining processes pertaining to Maintenance, Troubleshooting and Problem Resolution.
But what about one of the most painstaking tasks, which is the installation of cables and wires?
The process of laying telecommunication cables and wires is often outsourced to third-party contractors. These contractors are not only tasked with the work of laying cables but also with the ordeal of verifying the right usage of pipes, cables types and routing channels.
Contractors on the field click images of routing channels, cable types and pipes, which are then used to verify against existent data. In order to lay a single cable, dozens of images are captured and verified.
The sheer scale of the underground cable system, coupled with human error, leads to mistakes. And these mistakes inadvertently translate into costly mistakes for an organisation. The entire cost associated with re-wiring a cable from a customer’s home to the central hub is monumental.
The laying of cables and wires has come with its fair share of business challenges, to say the least. The most visible challenges included:
• Time taken to verify all the images collected by the third-party contractors. This also included the process of cross-referencing existent data with hundreds of new images
• Investment into human resources and effort to verify every single image, by matching the right text, serial number etc to the existent data. Also, the level of scrutiny for a single image is cumbersome
• Delays in payment towards these third-party contractors in lieu of a backlog of images that are to be verified
• Overall cost of operations
Since the entire operation rested on the practice of verifying and analysing images, it became clear early on that the solution would involve a Computer Vision (CV) model. Over the last decade, the efficiency of Computer Vision learning models has increased drastically.
These systems are capable of absorbing input in the form of videos or images and provide aggregated insights or models.
In the case of laying cables and wires for a telecom company, the CV model analyses images fed in by the third-party contractors, verifies them against the existing database and provides the desired feedback. All of this happens in real-time, based on the efficiency of the Computer Vision model.
By partnering with Amazon Web Services (AWS), TheDataTeam built a scalable system on the AWS cloud to automate the entire validation process.
Our solution consisted of various engines running on the cloud, which processed images submitted by the contract workers. This could either be through a web portal or mobile application. The solution architecture included OCR (Optical Character Recognition) text extraction, which extracted data from the images and then matched them against the company’s database.
The overall success of this exercise predominantly depends on the quality of data being fed into the system. In order to bypass inevitable gaps that could arise due to the quality of the images being captured by the third-party contractors, TheData Team’s solution included a robust feedback mechanism to spot errors in the detection model being identified by human reviewers. This feedback is then aggregated, analysed and fed back to the AI models for regular improvements.
Frequently arising problems
The information seen on the cable wires has to be matched manually with the corresponding codes in the junction box. These were some of the manual tasks that required endless cross-referencing, which resulted in a loss of productive man-power hours.
Furthermore, a single image takes about 60 seconds to verify. And in the case of laying a single cable, around 30 images are captured to be verified.
And on a single day, assuming there are 50 connections to be provided, the total time taken for verifying just the images amounts to approximately 25 hours (1,500 minutes). Since a team of reviewers are required to be hired in order to scrutinize the images, this adds on another layer of costs associated with salaries.
Architecture of the Solution
The solution proposed by TheDataTeam incorporated key features from the AWS cloud, in order to integrate the platform more efficiently into the existent legacy system.
- Rekognition: Which was used for Computer Vision + Text OCR
- Sagemaker: For advanced processing and custom Deep Learning models
- Lambda Functions: To process the output in the required format
Apart from these primary tools, other AWS services were also incorporated into the system. Some examples include S3 for storage and Cognito for managing user profiles.
What are the advantages of this cloud-based solution?
The advantages of deploying an intelligent cloud-based solution are:
- High scalability and visibility
- Ease of model training. Here, it only takes a few 1000 images (augmented) to train a vision model with acceptable accuracy since the transfer learning from the base model’s help.
- Sagemaker Custom Labels interface helps in creating labelled datasets, by drawing bounding boxes or by classifying images. This module is tightly integrated with Rekognition and Sagemaker Services.
- Ability to run several hyper-parameter tuning jobs for deep learning models in fast clusters, and thus, reduce the experimentation time drastically.
The flow-chart, depicted below, gives us a broad understanding of the functionalities that are performed by the cloud-based platform.
The image enhancement stage is key in the standardization of the images captured from different mobile phones and camera configurations. This stage helps in limiting the variation of the inputs for the deep learning model to train better.
The Sagemaker image detection API is implemented with the idea of Single Shot Learning, which is usually faster than most other object-based detection models. The algorithm used in Rekognition is not revealed. However, few 1000 images are all that it takes to train a model with acceptable accuracy. Training is sped up by the type of computing power that we select in order to run these training jobs.
The post-processing stage essentially removes the outliers in the vision models, by matching the probability scores against a configurable threshold. This stage also computes the necessary outputs and performs a pattern-match against the metadata for the OCR output.
The final step involves the matching of the output, returned by the previous step against the ground truth available within the Company’s database. If it is a perfect match, then the image is passed. If there are any doubts or if the image is a clear mismatch, then it is then sent to the human reviewer for final validation.
By deploying a cloud-based platform, layered with intelligent AI systems, significant productivity gains were achieved amongst the workforce. With the help of AI, organisations can now perform cumbersome tasks with ease. This has also helped companies reduce overhead costs and minimize errors, which arise due to human fatigue.
TheDataTeam(TDT) is an AI solutions company helping businesses achieve unparalleled agility by building AI solutions. TheDataTeam specializes in building scalable turnkey AI solutions with quick turnaround and ease in business process integration. TheDataTeam has engineered Cadenz, the market’s first customer intelligence platform that provides enterprises with live behaviour intelligence about its customers using its data automatically