Agentic Predictive Maintenance on OpenVINO: One Pipeline, Five Industrial Use Cases

Pipelines, utilities and other critical assets need constant inspection to head off failures, leaks and safety incidents. Maintenance that is purely reactive or calendar-driven means repeated truck rolls, early defect signals that slip past, and evidence that ends up scattered when an audit arrives. Agentic AI offers a way to watch continuously and act earlier, which saves on site visits, shortens downtime, improves safety and leaves an audit-ready trail.

This write-up covers a Google Summer of Code 2026 project that took the Intel Metro AI Suite predictive maintenance blueprint, built on OpenVINO, and extended it with new use cases and datasets.

The gap between a prediction and a decision

A model can emit thousands of predictions, yet the person on the receiving end still wants three answers: what happened, why, and what to inspect next. A label with a confidence score does not give those. A practical pipeline has to keep the evidence behind every result, accept different kinds of input, and let people investigate outputs without understanding the code underneath.

The project aimed at exactly that. It tied model inference to structured storage, coordinated AI agents, natural-language questions and evidence-based tickets, all inside one workflow.

How the pipeline is put together

Processing starts with an OpenVINO model whose results are written to JSONL files and a SQLite database. A LangGraph workflow then coordinates three agents, one for policy, one for analysis and one for evidence. Users then explore the output through chat or raise a ticket backed by the supporting evidence.

Four improvements shaped the work, aimed at usability, explainability, extensibility and breadth of application. They are different fixes, but all sit in the same end-to-end flow.

Four contributions

One box for every kind of question

Previously a user had to choose between analysis, evidence and SQL mode before typing anything. That meant knowing how the internals were split up, which is a poor thing to demand from an operator. The replacement detects intent and sends the question to the right backend automatically, in both the command line and the web interface. Nobody picks a mode any more.

Questions about raw sensor readings

The configurable SQLite schema was widened to keep raw sensor values, confidence per modality and metadata about the use case. That makes questions like this possible: which samples had higher sensor confidence than image confidence? The natural-language condition is turned into a database query that returns the matching readings, labels and source records. The same context also reaches the policy and analysis agents. What matters is traceability, since stored raw values and metadata tie each answer back to the records it came from.

New models without a rebuild

The original inference entry point carried too many jobs. It was broken up into distinct pieces: settings, routing, per-model handlers, the run step, writing results and plotting. A dispatcher now picks a small, focused handler, while shared components take care of the behavior common to every use case. Adding a new data type now means writing a handler and a configuration, and reusing the rest of the lifecycle.

Five industrial use cases on one runtime

Inputs, prediction targets and evidence differ across the five cases, but they share the same surrounding machinery. Four input patterns are covered: tabular data, images, audio with text, and raster data with text.

Oil and gas pipelines. Tabular operating and material data feeds an OpenVINO MLP that predicts two things at once, the condition of the pipeline and the loss of wall thickness. Inputs include geometry, material, operating and corrosion-related values. Each processed record yields a condition class and a thickness-loss estimate, and both stay attached to the original measurements. That link lets downstream agents explain a result using the real operating figures, and the workflow includes corrosion-focused analysis.

Solar panels. A YOLO-style OpenVINO detector looks for defects in single-channel electroluminescence images of solar cells and keeps the bounding boxes as evidence. In one sample, a red box marks a ground-truth star crack that the detector is expected to find. A reproduced evaluation over all 19,150 test images reached an F1 of 0.8277. The predicted box is stored with its class and confidence, so a user can see where the defect was found.

Power transmission. A fine-tuned YOLOv8s-cls model sorts transmission-line images into three circuit categories. It handles both synthetic and real-world images, which tests whether one classifier copes with rendered circuit views and field inspection photos. On 348 held-out images it got 334 right, an accuracy of 0.9598. Both image sources go through identical preprocessing and inference.

Operational environment maintenance. This case pairs an audio clip with a caption. The two are preprocessed separately, classified by two OpenVINO MLP branches and then combined through late fusion. A sample labeled rotating machinery came with the caption of a power tool vibrating as it runs. A video frame supplies context for the human reader, but the model input is the clip's audio plus the text. The branches stay individually inspectable until their scores are merged.

Water treatment. A 15-channel Sentinel-2 raster patch is paired with an environmental text prompt, for example one describing a county in Arizona with evapotranspiration, precipitation, groundwater depth and surface-water depth. Statistical features from the raster and TF-IDF features from the text pass through separate OpenVINO branches before late fusion, which keeps each branch's confidence. A JPEG preview is generated so people can look at the sample, but the model itself consumes the original multichannel NumPy array. The preview can also be attached to a ticket, giving a multimodal prediction visible evidence.

Running on Intel hardware

Every model was deployed in OpenVINO IR format, and the one configurable entry point handles images, tabular sensor vectors, audio features and paired multimodal inputs. The code was validated on an Intel GPU, and inference through the pipeline was also confirmed on an Intel NPU, including generation of the JSONL and SQLite outputs.

Testing more than accuracy

Validation went beyond model accuracy. Each workflow was exercised from model loading and structured output through the agents, the interfaces, evidence, tickets and execution on Intel hardware. The checks covered both the boundaries between components and the outputs of individual models.

What the demos show

Two views show the system working. The first is backend orchestration. It starts from inference results in the shared data layer. The policy agent filters them according to the configured policy, and the analysis agent summarizes what they mean operationally. The evidence agent then groups each result with its source data and metadata into a traceable record. Together they turn stored model output into structured analysis and evidence ready for review.

The second is the unified chat. Users review analysis summaries, inspect evidence and filter results, and query structured maintenance data. After investigating, they can create a maintenance ticket for follow-up. It demonstrates the whole path from exploring results to raising a ticket.

Lessons from the design

Balancing reuse against modality-specific behavior was the central design problem. All five cases share configuration, execution, storage and agent components, while small handlers preserve the preprocessing and postprocessing each data type needs. Duplication drops without forcing very different inputs through identical logic.

The results also argue against storing a bare label. When raw sensor values, bounding boxes, per-branch confidences, metadata and source previews travel with the prediction, every result stays connected to evidence that both people and agents can examine.

For the multimodal cases, late fusion keeps the audio, text and raster branches individually traceable. The current equal weighting is simple and reproducible, though learned or dynamically calibrated weights might better reflect how reliable each modality is.

The agent layer does not replace the predictive models. It organizes stored results, fetches supporting evidence and routes a question to the right analysis or database operation. Validation covers the evaluated datasets and Intel hardware, and real deployments would still need site-specific testing, monitoring and calibration.

From prototype to blueprint

What began as a pipeline built for a single use case became more modular, multimodal and user-oriented. OpenVINO supplies the common inference layer, and structured storage plus coordinated agents turn model output into information a person can examine and act on. The same route, from industrial data through inference and evidence to natural-language exploration and tickets, now repeats across several maintenance domains.