plans/legacy/jetengine_plan.json

{
    "sub_tasks": [
        {
            "bullet_points": [
                "Load the simulated jet engine health monitoring dataset (assume CSV format).",
                "Split the data into training (80%) and validation (20%) sets.",
                "Simulate a target domain dataset by adding small Gaussian noise and scaling to selected sensor readings.",
                "Inject synthetic anomalies into the target dataset (e.g., by perturbing sensor values or introducing out-of-range values) and record ground truth anomaly labels for evaluation.",
                "Apply StandardScaler: fit on source training data, transform all datasets (training, validation, target).",
                "Output summary statistics of the processed datasets."
            ],
            "sub_task": "Data Preparation, Preprocessing, and Synthetic Anomaly Injection",
            "sub_task_agent": "engineer"
        },
        {
            "bullet_points": [
                "Select a Gaussian (RBF) kernel and determine kernel bandwidth (e.g., median heuristic).",
                "Compute MMD between source training data and target data using all features.",
                "Compute and compare MMD values for different feature subsets to identify which sensor readings contribute most to domain shift.",
                "Output the MMD values and a ranked list of features by their contribution to domain shift."
            ],
            "sub_task": "Maximum Mean Discrepancy (MMD) Calculation and Feature Subset Analysis",
            "sub_task_agent": "engineer"
        },
        {
            "bullet_points": [
                "Select exhaust gas temperature (EGT) as the target variable for next-step prediction.",
                "Train a simple feedforward neural network (one hidden layer) on source training data to predict EGT at t+1.",
                "Train a simple baseline model (e.g., linear regression) for the same prediction task.",
                "Evaluate both models on source validation data (MSE).",
                "Extract hidden layer activations from the neural network as feature representations for anomaly detection."
            ],
            "sub_task": "Predictive Coding and Baseline Model Training",
            "sub_task_agent": "engineer"
        },
        {
            "bullet_points": [
                "Compute reconstruction (prediction) error (MSE) for both source validation and target data using both models.",
                "Experiment with different thresholding methods (mean + k*std, percentile-based, Gaussian mixture models) to determine the optimal anomaly threshold.",
                "Identify anomalies in target data based on each thresholding method.",
                "Output confusion matrix and metrics (precision, recall, F1-score) using ground truth labels for each method."
            ],
            "sub_task": "Anomaly Detection and Threshold Optimization",
            "sub_task_agent": "engineer"
        },
        {
            "bullet_points": [
                "Generate plots: MMD value comparison (all features and subsets), reconstruction error distributions, ROC or precision-recall curves, and confusion matrix.",
                "Generate a feature importance plot showing which sensor readings are most predictive of anomalies (e.g., using permutation importance or model coefficients).",
                "Summarize key statistics and findings in tables.",
                "Prepare all results, plots, and statistics for interpretation."
            ],
            "sub_task": "Visualization, Feature Importance, and Results Compilation",
            "sub_task_agent": "engineer"
        },
        {
            "bullet_points": [
                "Provide a comprehensive results section, including:",
                "Interpretation of MMD alignment and feature subset analysis.",
                "Analysis of predictive coding and baseline model performance.",
                "Discussion of anomaly detection results (quantitative and qualitative), including thresholding method comparison.",
                "Reference and interpret all generated plots and statistics.",
                "Discuss feature importance findings and implications for maintenance.",
                "Discuss computational efficiency, how the approach meets the \"few minutes on a laptop\" constraint, and trade-offs between model complexity and performance.",
                "Summarize insights and implications for lightweight domain adaptation in predictive maintenance."
            ],
            "sub_task": "Results Interpretation, Discussion, and Computational Efficiency Analysis",
            "sub_task_agent": "researcher"
        }
    ]
}