Autonomous fault and anomaly detection is critical for ensuring the safety and success of space missions, addressing the limitations of ground-based analysis due to bandwidth constraints and operational delays. The conventional approach in Space Operations involves using Out-of-Limits (OOL) alarms for anomaly detection, which may prove insufficient in identifying and responding to complex anomalies or unforeseen novelties within the range of nominal values. In our previous work (Voldrich, Luschykov, Harwot, & Manilla, 2023), we proposed a machine learning approach for on-board telemetry anomaly detection that addresses these limitations. We demonstrated a proof-of-concept integration of a TensorFlow model onto a radiation-tolerant LEON 3 processor and benchmarked various unsupervised and semi-supervised techniques with respect to their performance, memory footprint, and runtime. Our experiments indicated that a two-phase model comprising a Siamese convolutional encoder for dimensionality reduction followed by an outlier detector such as K-Nearest Neighbors (KNN) or Isolation Forrest is the most promising approach for our use case. We reimplemented these outlier detectors and developed a custom tool to export trained models from the PyOD library into a C++ environment for integration with our inference code. We developed separate models for each subsystem (batteries, reaction wheels, and solar panels), as this approach empirically yielded better performance. While a unified model theoretically offers advantages in capturing interconnected relationships and identifying related anomalies, it did not empirically outperform the specialized models in practice. Our recent advancements focus on bridging the gap between a proof-of-concept solution and a fully production-ready system. Mainly, we focused on development of a semi-automatic pipeline to simplify the end-to-end development and integration of these models, enabling rapid iteration once real mission data becomes available, rather than relying solely on simulators. This pipeline automates model training, big-endian conversion, and import into TensorFlow Lite For Microcontrollers (TFLM) framework for inference. Furthermore, we investigated the application of Explainable AI (XAI) and Uncertainty Quantification techniques to enhance model transparency and allow alarm rate calibration even in online settings. However, we found that the benefits of both XAI and Uncertainty Quantification were somewhat limited by our two-stage model architecture and the inherent nature of our task which relies on sequential data. This is only one of the drawbacks of our two stage approach. Another drawback was that the neural network in the first stage was trained without an explicit objective to differentiate between features of nominal and anomalous samples, which complicated the anomaly detection task in the subsequent stage. To address this, we implemented and evaluated two new models: Deep Support Vector Data (Deep SVDD) and Autoencoder with One-Class Support Vector Machine (AE-1SVM). These models integrate a neural network for dimensionality reduction with a one-class SVM for outlier detection, but critically, they are trained with a unified anomaly detection objective. This single-stage approach offers several advantages: they are unsupervised, eliminating the need for anomalous training data (though a contamination parameter is still required); and their objective functions have connected computational graphs, enabling gradient computation with respect to input data, which can be leveraged for tasks like adversarial sample generation or estimating feature importance. Both Deep SVDD, which maps data into a compact sphere in latent space, and AE-1SVM, which applies a one-class SVM in the autoencoder’s latent space, were adapted using convolutional and recurrent layers to handle our high-dimensional, time-dependent telemetry data effectively. This also creates opportunities to apply more advanced XAI techniques in the next phase of the project, which is expected to follow. First data from HERA also became available during this period. Transitioning from synthetic and historical MEX data involved training models with actual telemetry from the HERA mission, collected between October 2024 and February 2025. This move presented new challenges, including the absence of clear subsystem delineations, missing solar array data, and ambiguous variables within the chemical propulsion system. Our initial focus for HERA data was the reaction wheel subsystem, leveraging available bearing temperature, current, rotation speed, and torque measurements recorded at 16-second intervals. Rigorous data pre-processing was essential: temperature readings were excluded due to external influences, and known error-induced spikes and insufficient data segments related to rotation speed changes were carefully removed. For this HERA-specific application, we trained a single Deep SVDD model incorporating all twelve reaction wheel variables, using 36-measurement subsequences sampled every 40th interval. While the model achieved promising F1, Precision, and Recall scores (0.822, 0.853, 0.793 respectively), it initially misclassified known nominal spikes caused by commands as anomalies. To mitigate this, a unique approach was employed, iteratively retraining the model by prioritizing a percentage of samples with the highest anomaly scores from previous runs. This method significantly reduced false positives for commanded events, although it introduced considerations regarding potential data leakage due to the sequential nature of the data and shared samples across training, validation, and test sets, suggesting further improvements like oversampling or weighted sampling of spike-containing subsequences. Finally, we have significantly enhanced the robustness of our on-board deployment by developing a comprehensive patch for the TensorFlow Lite Micro (TFLM) framework. This patch, combined with a custom tool we created to manipulate the model’s flatbuffer file by swapping underlying buffers to change endianness, enables big-endian support. The custom tool specifically addresses the endianness of buffers and subgraph metadata within the Flatbuffer model, performing byte-level swaps on operator inputs/outputs and tensor shapes. The patch leverages TFLM’s memory management classes to ensure compatibility with the latest TFLM versions while maintaining strict static memory allocation, which is of great importance as the other core of the processor is dedicated to run the flight software. The patch uses the same memory pool as TFLM, making it an efficient solution that adds very little computational overhead. Additionally, the byte swapping is performed offline, so this step does not consume any resources on the target system, which is critical given its limited capacity. This approach specifically targets the LEON processor, a big-endian system, by disabling usage of reinterpret_cast for Flatbuffer vectors and instead allocating and copying data element by element, thereby resolving potential endianness and memory alignment issues. This novel integration of an ML-driven anomaly detection framework on big-endian, resource-constrained space processors represents a crucial step towards more resilient and autonomous spacecraft operations, ultimately enabling advanced on-board FDIR capabilities and reducing reliance on ground intervention.