High Quality Meta Data Improves Machine Learning for Better Outcomes

1. Data Collection: The first step is to gather a large dataset of images or video footage, using the camera, that accurately represents the normal operation as well as the various stages and indicators of a CSO event. This might include visual cues such as water level, color changes due to pollutants, or increased flow velocity.
    

2. Annotation: The collected data can then be annotated, meaning that each image or video frame is labelled with information indicating whether it shows a normal state or a CSO event, this can be done during collection and after time this need to intervene is rendered obsolete. This step is crucial for an effective end product, as it teaches the algorithm what to look for.
    

3. Model Training: With the labelled dataset, a machine learning model is trained to recognise patterns that correspond to CSO events. This training involves feeding the data into a neural network or another ML model, which adjusts its internal parameters to minimize the difference between its predictions and the actual labels.
    

4. Algorithm Optimisation: After the initial training, the model is refined to improve its accuracy. This might involve tuning hyperparameters, using more complex models, or incorporating additional features such as weather conditions.
    

5. Edge Computing Iteration: The trained model is then integrated into an edge computing device — this is the AW-108, a camera with onboard processing capabilities that can analyze data locally without needing to send it to a remote server. This allows for real-time processing and immediate response as well as eliminating the need for high bandwidth communications and heavy energy consumption.
    

6. Physical Water Parameters: Integration of real time water quality data gathered and delivered in real time by the AquaWatch waka.
    

7. Deployment and Monitoring: The edge device is deployed in the field, where it continuously monitors the sewer system, applying the trained model to the live video feed.
    

8. Alert System: When the model detects patterns that indicate a CSO event, it triggers an alert. This alert system can be customized to notify relevant personnel via text, email, or an integrated management system.
    

9. Continuous Learning: The system supports continuous learning, where it can refine its predictive capabilities over time based on new data and feedback from the performance of its alerts.
    

By leveraging machine learning and edge computing, the AquaWatch AW-108 can provide a proactive tool in managing and mitigating the environmental impact of CSO events, ensuring timely intervention and contributing to the health of urban waterways.

An example we are currently working on is data from the cameras is processed continuously to judge If an issue is detected. It would then query a large language model (LLM) with vision perception capabilities to create a text description of the scene. An email with a link to the image viewer for this camera and time and containing the alert condition and the text description of the scene can be sent to stakeholders.