Photo 1: Example 1, Circumferential crack

Toronto Assesses Its Assets

Toronto embarked on a comprehensive sewer inspection and condition assessment program in 2014

The City of Toronto — the fourth largest city in North America — maintains and operates an extensive sanitary, storm and combined sewer network with more than 10,600 km of mains and 160,000 manholes. As part of Toronto’s best practices, the City runs an on-going Sewer Inspection and Condition Assessment Program to maintain a long-term, accurate database of network condition and to help plan rehabilitation needs.

Andrews.engineer (A.E) was retained by the City to manage the program for a five-year period, commencing in 2014. As the designated program manager, A.E is responsible for delivery of the inspection, assessment and rehabilitation planning for 2,500 km of sewers and 40,000 manholes over the five-year duration of the program. In 2019, A.E completed Phase 1 of the five-year program and initiated Phase 2 of the five-year program.

// ** Advertisement ** //

The members of the team include A.E, two sub-contractors, and a subconsultant. The sub-contractors are required to carry out CCTV inspection as per NASSCO PACP standards. The condition assessment team is required to review all the videos, correct any errors in the defect coding, and provide condition assessment and rehabilitation recommendations. The results are submitted to the client on a monthly basis, typically consisting of 50 to 80 km of inspected length.

Photo 2: Example 1, Joint
Photo 2: Example 1, Joint

Project Challenges

As a team that always strive to deliver high quality results, A.E implemented rigorous QA/QC procedures in every step. However, due to the large volume of the inspection data generated and processed every month, there are challenges in delivering consistent and quality results:

// ** Advertisement ** //

The project employs standard pan-tilt-zoom cameras with a resolution of 720 x 480. Operators must keep the crawler’s speed below 9 m per minute and stop at key defects to capture the details. Failure to adhere to these guidelines may result in insufficient video quality, leading to potential rejection of videos and slowing down the condition assessment process.

Despite training and certification by NASSCO, there are subjective differences among operators in coding defects. Similar defects may be coded differently, and minor issues are sometimes overlooked or grouped together. For instance, minor encrustation at joints is often disregarded or combined into a single defect.

// ** Advertisement ** //
Photo 3: Example 2, Defective liner with roots-like fabric
Photo 3: Example 2, Defective liner with roots-like fabric

Opportunities for AI

Advancements in AI in the recent years have provided potential solutions to help meet these challenges. In theory, a well-trained AI model can automate defect coding, provide more consistent and more objective results and enhance efficiency and accuracy in sewer inspection projects. However, building effective AI models requires careful consideration of various factors, including size of training data, diversity in defect types and pipe materials.
Understanding of the requirements of building good AI and knowing the current state of AI development in the sewer industry, A.E developed the following criteria to evaluate AI performance:

  • Capable of recognizing defects in many different common pipe materials
  • Identify ALL severe defects but accept some minor mis-labeling
  • Coding accuracy must equal or exceed 85 percent, as this is the minimum NASSCO requirement for a human operator
Photo 4: Example 2, Defective replacement with offset joints
Photo 4: Example 2, Defective replacement with offset joints

Toronto Pilot Project

As a pilot study, A.E selected five videos covering various pipe materials and with a wide range of pipe defects and provided them to three different AI vendors. Two of the vendors fell short in meeting the stated criteria in their review. One AI vendor met the criteria.

// ** Advertisement ** //

Examples of the AI results are discussed below. Overall the A.E team was pleased to see the accuracy of the AI results.

Example 1: In this inspection, some of the circumferential cracks are so uniform they almost appear as joints (Photo 1 and 2), which made it difficult for the operator and the engineers to interpret. AI was able to correctly and consistently identify the circumferential cracks and not confuse them with joints.

// ** Advertisement ** //

Example 2: AI was able to recognize different types of spot repairs. In Photo 3, AI coded the defective spot repair correctly, even though the loose fabric looked like roots at first glance. In Photo 4, AI correctly identified this as a defective replacement with offset joints.

Example 3: In a lined sewer, AI was able to identify minor features and use the correct codes to depict them, such as wrinkles (LFW), abandoned connections (LFAC) and overcut connections (LFOC).

// ** Advertisement ** //
Photo 5: Example 3, Dimple indicating possible abandoned connection at 2 o’clock
Photo 5: Example 3, Dimple indicating possible abandoned connection at 2 o’clock

The Pilot Project Continues

To further verify AI performance, another 5.6 km of sewer inspection videos were analyzed by the successful vendor using AI. These videos were originally coded by experienced certified operators. The purpose of this stage of the Pilot Test was to compare the original coding with the AI results.
Overall, AI generated more unique codes than the operators. The codes used by AI were typically more descriptive. For example, AI-coded IRB (Infiltration Runner Barrel) which indicated the location of the infiltration, while operator coded IR (Infiltration Runner).

AI also coded significantly more defects than the operators. When broken down into categories, it is obvious that AI identified individual minor defects that most operators would omit, such as minor cracks (operators coded 88 while AI coded 258), deposits and infiltration. For severe defects, AI captured similar number as the operators, and when examined, AI codes were typically more appropriate. The numbers of connections were identical between AI and operators, which was reassuring since the operators were required to capture all connections.

// ** Advertisement ** //
Photo 6: Example 3, Overcut service connection
Photo 6: Example 3, Overcut service connection

Trial Implementation

Following successful evaluation of the pilot testing results, AI was integrated into our project workflow. A phased approach was adopted, by carefully comparing AI coding to the operators’ coding for a three-month trial period. As the A.E team gained more and more confidence in AI coding, the team was able to rely on the AI results.

Conclusions and Recommendations

The Toronto project shows that AI can detect sewer defects consistently and accurately.
It is foreseeable that AI can make transformative impact in the sewer inspection industry. In the near future, operators may no longer be required to perform defect coding in the field, allowing them to focus better on the camera work; Engineers no longer need to correct the coding and can spend more time on other engineering work, such as designing the proper rehabilitation solutions.

// ** Advertisement ** //

Cloud Zhang, P.Eng. is general manager at Andrews.engineer.

// ** Advertisement ** //