Mantenimiento predictivo

2 Questions about Asset Health Monitoring at Cigre Paris 2024

02
Sep 2024

Partial Discharge Expert answers two Asset Health Monitoring questions during Cigre Paris 2024 panel discussion

1) Are there new systems for evaluating PD measurements? Can the cost-effectiveness of such monitoring systems be confirmed by the experience of operators?

Since 2020, Ampacimon in collaboration with Redeia group (TSO of Spain) has been working to develop an automatic PD diagnostic platform. The platform has been designed to centralize PD pulse information (raw data) collected by multiple sensors installed on different assets and to deliver optimized alert list including recommended preventive actions. As shown in figure 1, inside this SW platform, three different steps including Artificial Intelligent (AI) have been developed for clustering of PD from different defects, recognition of phase resolved PD pattern, optimization of alert list and aggregation of evolution.

The platform analyses all the information acquired by monitoring devices with the AI to avoid any human expert interpretation reducing drastically the cost of implementing this monitoring solutions. Maintenance team of the utility can spend the minimum time to follow critical alerts detected automatically.

The TSO has started since 2019 a deployment of permanent PD monitoring installing HFCT and UHF sensors in GIS substations of critical lines (see figure 2). Nowadays, TSO counts with more than 405 sensors from 27 different HV installations, already connected to the AI platform and it is planned to follow up the deployment until the 3000 sensors covering the most critical assets of the transmission grid.

During the first months of production using the automatic PD diagnostic platform 78 alerts have been detected automatically, having only 7 corresponding to internal defects in the insulation of the assets. In table 1 it is shown detailed list of alerts generated automatically by the AI.

In one of the 220kV GIS substation, an internal cavity in solid was detected by the permanent monitoring system on one of the lines. The line with the internal void was not including sensors on the other end to perform automatic localization by permanent monitoring sensors. Further investigation using portable monitoring devices synchronized by GPS determined that defect was located at the GIS termination on the other end of the line, 1.3km away from the sensors detecting the problem initially.

Later inspection done by cable and GIS manufacturers determined that the problem was an oil leak inside the cable termination. Termination was starting to have a cavity due to the leak and oil was coming inside the cable. Finally, a programed repair was done been able to avoid a catastrophic failure and allowed the cost-effectiveness of this deployment to be demonstrated.

2) The approaches presented for utilising AI, online asset monitoring and power quality measurements in principle are becoming increasingly important. What are the challenges to improve the quality of measurement parameters?

During more than 10 years Ampacimon has been workingto develop online Partial Discharge (PD) monitoring solutions in collaborationwith TSOs, DSOs and HV Manufacturers. R&D activities during the last years havebeen focus on automatization of diagnosis utilizing Artificial Intelligent (AI).Developed process from sensor acquisition until alert list is shown in nextfigure.

AI is used in two critical steps, the clustering of multiple defects mixed in measurement and the recognition of the phase resolved PD pattern. After recognition of insulation defects present in the installation with AI, last step is an expert system that optimizes the alert list generated by multiple sensors and aggregates the evolution of the activity defining preventive actions.

There are three key factors in measurement to improve AI results:

- Noise rejection

- Sensitivity level

- Extraction of parameters for clustering from:

      *Pulse wave form (frequency content, steepness, polarity…)

      *Synchronized sensors (ratio between 3 phases, time of fly)

Figure 2 shows three samples where raw data was collected with different measuring techniques. First one has been collected with a technique to suppress the noise mathematically with a filter based on wavelet transform that removes noise and get the best sensitivity level (highest quality). Second sample has been obtained using a band pass filter combined with a low threshold cut to let pass as much information as possible. Third sample has been collected removing the background noise with a threshold cut over the blind area. In this three different scenarios same AI tools applied after acquisition will generate different results.

Sample 2 has two big disadvantages applying AI tools. First problem is found during clustering due to the size of resources needed in terms memory and processing workload. Due to the high number of pulses stored in the noise level, the size of the sample is about 1000 times higher, and clustering algorithms use to have logarithmic cost (O(N×log(N))) or exponential cost (O(N2)). In this case, efficiency of clustering could be compromised resulting a mixed pattern, having PD pulses of defect overlapped with noise that prevents pattern shape recognition using the convolutional neural network (CNN).

Technique used in Sample 3 avoids the problem with high number of pulses, but losing this area of the information will generate other problems. PD pulses that are traveling from other phases or far away from the sensor are attenuated and reach the sensor with very low magnitude. Removing the lowest area of the pattern prevents application of clustering with parameters that are based on correlation between multiple synchronized sensors, like ratio between the three phases or time of fly analysis. Additional trivial problem is that losing the lower area of the pattern the CNN tool will have less accuracy in the recognition of the physical phenomena.

What does it need to receive higher quality results for asset condition evaluation?

Due to different needs of utilities, two approaches of this automatic diagnosis process have been implemented. As shown in figure 3, first one is done sending raw data from all the acquisition units to a central server where correlation between sensors is calculated before applying AI tools for clustering and recognition of defects. Second approach is done implementing embedded AI into the acquisition device that can analyze data locally and send only alerts to the central server.

In first approach, sending all raw data to the central server request a good communication channel and a powerful server to analysis all the sensors. But moving AI to the server side permits to have clustering parameters from synchronized sensors and much more flexibility to increase computing resources for the AI tools.

In second approach, data transfer reduction is in a ratio of 10,000 times, improving implementation and integration of the solution as an IoT device. On the other hand, it is necessary to consider that computing resources for the AI tools in the embedded device are limited, requesting to reduce the data model in a ratio of 68 times, that provokes an impact on the accuracy of the AI recognition tool.

A validation procedure with three different tests has been implemented to evaluate the accuracy with both models. First test is done using a validation set of 100 samples for each type of PD phenomena recognized by the CNN, that have been collected in real installations from different locations. Second test is perform including additional stress in the validation adding background noise to simulate more complex situation where noise is overlapping the shape of the PD pattern. Third and last test is including high stress for AI tools including noise and requesting to cluster four different PD phenomena detected simultaneously by the sensor. Results are show in Table 1.

The accuracy of booth models is high in low stress conditions, but in high stress the embedded AI model has a lower accuracy of 75%. That impact in the quality of the results for the embedded AI model is due to the clustering step where only pulse wave form parameters are used. At installations where noncritical external PD are present all the time, like corona or external surface PD in air or floating potential in air, it is mandatory to have good clustering tools to find the internal defects mixed in the raw data. In conclusion to the analysis, embedded AI could be useful in cases where only internal PD defects are expected.

PARTICIPATE

Access the Recording of this Webinar

You can no longer participate in this event but you can access its recording