Automating Wooden pole inspections and damage assessments using AI and robotics
By Dr. Sujatha J | Senior Consultant — Robotics Practice, Wipro
Energy transmission- distribution and telecommunication segments are the largest users of utility poles all over the world. These are predominantly made of wood and need to be monitored for their condition regularly and predict their future condition accurately to operate their network reliably and safely. In this context, automated wooden pole inspection and maintenance, specifically with non-destructive techniques, is of current interest to companies, as well as researchers/academic institutions working in related areas.
2. Problem statement
The inspection and treatment of wooden utility poles is necessary for determining the strength and serviceability of a pole. The defects to be identified typically are visual (tip rot or any damage visible on the surface), internal (holes or fibrous rot inside the pole due to fungal, insect, woodpecker or any other microorganism attacks) and decay below the ground. Among these, internal defects can be life threatening to pole inspectors who climb up to 15 meters high during the inspection process. The crucial challenge here is to estimate the strength of the wooden pole in terms of stability/extent of internal defects present — to ensure the pole is safe enough to climb for any inspection or servicing.
Automated wooden pole inspection and maintenance can eliminate the fall hazard for pole inspectors. In addition, it can reduce the overall pole inspection time. It can also standardize the process with relevant reports for audit trail.
An effective pole inspection program strikes a balance between accurately identifying poles that put both system reliability and human life at risk while minimizing the number of still serviceable poles being replaced. The current methods used in identification of defective pole are subjective in nature and a lot depends on the intuition and experience of the inspector/pole tester. However due to the complex combination of several variables viz. Wood species, preservation methods and material, soil and climate conditions, insect and mechanical damage, inspection methodology as well as human errors involved, no foolproof inspection method exists today that can guarantee the condition of a standing wood pole with high accuracy. As a result, a large number of poles are replaced unnecessarily, and a significant number of poles continue to fail unexpectedly in-service, causing service disruptions and damage to assets as well as human lives.
We have built a solution that helps assess the life of In-Service wooden utility pole. The estimation of the remaining life of pole is carried out by an operator with substantiation provided by our system in terms of presence or absence of defects and related analytics.
3. Our Solution
Automated pole inspection consists of a robotic climber that climbs UP/DOWN larger size Utility poles as well as rotates around them — as shown in figure 1 below.
The ultrasonic sensing based non-destructive testing (NDT) system as depicted in figure 2 below, is used for detection and analysis of internal defects to assess the health of the pole using AI and ML algorithms deployed for identification and analysis of both external and internal defects.
Ultrasonic non-destructive testing uses the transmission of high-frequency waves through the wooden pole. A transmitter and a receiver are used at the point of measurement to capture the velocity of the pulse that has traversed. Based on this as well as other derived features, an assessment is performed about the presence or absence of internal defects at the point of measurement.
The system consists of 4 modules as shown in figure 3 below.
1. Ultrasonic Data Capture and Recording Module
2. Feature extraction module
3. Internal Defect Analysis Module
4. Wood Health Record Generation Module
Ultrasonic Data Capture and Recording Module
The ultrasonic pulses are propagated through the cross section of the pole on which the inspection is done using an ultrasonic device. This device will calculate the transit time which ultrasonic pulses takes to travel between the transmitter and receiver while propagating in radial direction through the cross section of wooden pole. This module captures the ultrasonic pulse velocity, transit time and distance travelled by the pulse which are fed to feature extraction module which in turn derives relative features for further processing.
Feature extraction module
This module derives a set of relative features from the captured data. A combination of these features and the data captured in the first module are used for training different machine learning (ML) modules for further analysis.
Internal Defect Analysis Module
This module has 3 submodules namely Defect detection submodule, Defect position determination submodule and Defect severity determination submodule as described below.
§ Defect Detection Submodule
This module provides the test data to a trained ML model. The model will determine whether an internal defect is present or not at the given cross section of wood, corresponding to the test data provided. If no defect is detected, the system moves on to the wood health record generation module. On the other hand, if an internal defect is found to be present at this point, the following submodules analyze the defect further.
§ Defect Position Determination Submodule
As a defect is found to be present at this cross section of wood by the earlier module, this module determines the position of the defect. The same is represented visually with a color coded 3 x 3 matrix.
§ Defect Severity Index Determination Submodule
This module provides the test data to a trained ML model for quantifying the defect present. The model will determine the extent of the internal defect present and categorize the same as one of the four — “Insignificantly-Defective”, “Mildly-Defective”, “Defective” or “Significantly-Defective”.
§ Wood Health Record Generation Module
This module provides a record of health for the given specimen of wood under test. This health record includes the total number of internal defects detected, their position, severity and related statistics/visual representations.
3.2 Defect analysis process
The entire flow of internal defect analysis solution is described through functional description of the different modules as below:
A particular cross section of a given pole is identified for testing and the transducers are placed at specified alignments as shown in Fig below.
A set of transmitter and receiver are placed at an alignment, say A-B (ref to figure 4), to pass ultrasonic pulses across the cross section of the wood. With the transmitter and receiver positioned at each of the above alignments, the pulse velocity, respective transit time and the distance covered are captured as 3 data points.
This module determines whether an internal defect is present at the given cross section by using a trained ML model. Based on measured 3 data points as indicated earlier, two other relative features are computed. These are organized as the feature set matrix for the given cross section of wooden pole.
Data set will contain lot many such matrices and the entire data set is split into 80% for training and 20% for testing. SVM or equivalent classification techniques are used to classify the data into 2 classes — “Defective” and “non-defective”. An accuracy of 92 % has been established for such cross-section testing, which can be enhanced further with more relevant data used for training the ML classifier.
Defect position determination
The entire cross section of wood is divided into say T regions, which is represented by a matrix of size M x N, where T=M x N. The same is illustrated for T=8, M=3, N=3 as in figure 5 below.
The output of the Defect Position Determination submodule is to visually represent the mentioned M x N regions with a color code across the range of values. The color that represents the highest value in the given range indicates the presence of defect. A feature set that consists of measured data values as well as some of their derivatives helps such analysis and visual representation.
It is to be noted that we were able to locate the defect in one of the 8 sections accurately. The results can be enhanced further with techniques like tomography, for even better visualization.
Determining the severity of the defect
In continuation of the internal defect analysis, it is possible to quantify the severity of an internal defect in wood. A severity value is computed which is based on a set of derived feature values from the data captured. This severity value can represent the size/ amount of the defect in proportion.
Utilizing this information, a ML model is trained for quantifying the defect present into different severity classes — the number and range of which can be defined by the user. The trained model will determine the extent of the internal defect present in the given test piece of wood and categorize the same as one of the four classes — “Insignificantly-Defective”, “Mildly-Defective”, “Defective” or “Significantly-Defective”.
Wood Health record
Documenting the outcomes of testing through the previous steps as a “wood health record” for a specimen under test is very important. This essentially helps for further analysis in future at any time and could be a major source for data collection for related purposes. More importantly, this leads to standardized testing procedures. The Wood Health Record Generation Module thus generates a health -record at the end of each testing cycle. This record is cumulatively updated through different test-cycles of the given wooden specimen under test.
An automated wooden pole inspection and evaluation solution which includes both internal and external defect analysis, would lead to safer, reliable and standardized inspections. In addition, there are several other uses of combined robotics, sensing and AI/ML algorithmic solutions with a high potential — e.g., heritage structure maintenance, wood product quality testing as well as live tree testing.