Safety-critical Applications Research Articles

Certification of the proximal gradient method under fixed-point arithmetic for box-constrained QP problems

In safety-critical applications that rely on the solution of an optimization problem, the certification of the optimization algorithm is of vital importance. Certification and suboptimality results are available for a wide range of optimization algorithms. However, a typical underlying assumption is that the operations performed by the algorithm are exact, i.e., that there is no numerical error during the mathematical operations, which is hardly a valid assumption in a real hardware implementation. This is particularly true in the case of fixed-point hardware, where computational inaccuracies are not uncommon. This article presents a certification procedure for the proximal gradient method for box-constrained QP problems implemented in fixed-point arithmetic. The procedure provides a method to select the minimal fractional precision required to obtain a certain suboptimality bound, indicating the maximum number of iterations of the optimization method required to obtain it. The procedure makes use of formal verification methods to provide arbitrarily tight bounds on the suboptimality guarantee. We apply the proposed certification procedure on the implementation of a non-trivial model predictive controller on 32-bit fixed-point hardware.

The impact of uncertainty in hERG binding mechanism on in silico predictions of drug-induced proarrhythmic risk.

Drug-induced reduction of the rapid delayed rectifier potassium current carried by the human Ether-à-go-go-Related Gene (hERG) channel is associated with increased risk of arrhythmias. Recent updates to drug safety regulatory guidelines attempt to capture each drug's hERG binding mechanism by combining in vitro assays with in silico simulations. In this study, we investigate the impact on in silico proarrhythmic risk predictions due to uncertainty in the hERG binding mechanism and physiological hERG current model. Possible pharmacological binding models were designed for the hERG channel to account for known and postulated small molecule binding mechanisms. After selecting a subset of plausible binding models for each compound through calibration to available voltage-clamp electrophysiology data, we assessed their effects, and the effects of different physiological models, on proarrhythmic risk predictions. For some compounds, multiple binding mechanisms can explain the same data produced under the safety testing guidelines, which results in different inferred binding rates. This can result in substantial uncertainty in the predicted torsade risk, which often spans more than one risk category. By comparison, we found that the effect of a different hERG physiological current model on risk classification was subtle. The approach developed in this study assesses the impact of uncertainty in hERG binding mechanisms on predictions of drug-induced proarrhythmic risk. For some compounds, these results imply the need for additional binding data to decrease uncertainty in safety-critical applications.

Detecting and Processing Unsuspected Sensitive Variables for Robust Machine Learning

The problem of algorithmic bias in machine learning has recently gained a lot of attention due to its potentially strong impact on our societies. In much the same manner, algorithmic biases can alter industrial and safety-critical machine learning applications, where high-dimensional inputs are used. This issue has, however, been mostly left out of the spotlight in the machine learning literature. Contrary to societal applications, where a set of potentially sensitive variables, such as gender or race, can be defined by common sense or by regulations to draw attention to potential risks, the sensitive variables are often unsuspected in industrial and safety-critical applications. In addition, these unsuspected sensitive variables may be indirectly represented as a latent feature of the input data. For instance, the predictions of an image classifier may be altered by reconstruction artefacts in a small subset of the training images. This raises serious and well-founded concerns about the commercial deployment of AI-based solutions, especially in a context where new regulations address bias issues in AI. The purpose of our paper is, then, to first give a large overview of recent advances in robust machine learning. Then, we propose a new procedure to detect and to treat such unknown biases. As far as we know, no equivalent procedure has been proposed in the literature so far. The procedure is also generic enough to be used in a wide variety of industrial contexts. Its relevance is demonstrated on a set of satellite images used to train a classifier. In this illustration, our technique detects that a subset of the training images has reconstruction faults, leading to systematic prediction errors that would have been unsuspected using conventional cross-validation techniques.

DATA FUSION OF ULTRASONIC SENSORS FOR DISTANCE MEASUREMENT

Distance measurement plays a fundamental role in contemporary technology, exerting a significant influence on applications across a myriad of sectors, ranging from autonomous navigation and industrial automation to robotics. Ultrasonic sensors, celebrated for their simplicity, cost-efficiency, and adaptability, have grown into essential instruments for gauging distances. Nevertheless, individual ultrasonic sensors are not exempt from constraints, which encompass inaccuracies stemming from environmental conditions, sensor noise, and measurement errors. This article delves into the advanced practice of data fusion, a method that entails amalgamating data from multiple ultrasonic sensors, bolstered by intricate algorithms like artificial intelligence (AI) and statistical techniques. The fusion of ultrasonic sensor data surpasses mere amalgamation; it harnesses the potency of AI to refine data integration. Consequently, this doesn't only enhance measurement precision but also equips systems to function dependably in dynamic and demanding environments. The article explores the importance of ultrasonic sensors, the necessity for data fusion, procedures for data preprocessing, the extraction of relevant features, the application of fusion algorithms, and the handling of uncertainties. It investigates the utilization of ultrasonic sensor data fusion in diverse domains such as autonomous vehicles, robotics, industrial automation, healthcare, and environmental monitoring. Moreover, it confronts issues associated with data quality, environmental variations, privacy concerns, and ethical deliberations. The article concludes by shedding light on the encouraging avenues that lie ahead in the field, which include the development of advanced algorithms, the seamless integration of different sensor types, real-time data processing, and a strong emphasis on safety-critical applications. Data fusion from ultrasonic sensors signifies an evolving and transformative technology with profound consequences for various sectors. The responsible advancement and deployment of this technology are of paramount importance to ensure the realization of its full potential, while maintaining the principles of safety, ethics, and reliability in a continuously evolving technological landscape. Keywords: Complex measurement, Multi-sensor distance measurement, Ultrasonic sensors, Data fusion, Measurement accuracy.

Simulation of coaxial time-of-flight measurements using SiPM as detector

Modern LiDAR sensors find increasing use in safety–critical applications. Therefore, highly accurate modeling of the system’s behavior under demanding environmental conditions is necessary. In this paper, we present a modular structure to accurately simulate the amplified raw detector signal of a direct time-of-flight LiDAR system for coaxial transmitter–receiver optics using a silicon photomultiplier (SiPM) as a detector. The SiPM causes strong non-linearities of the response due to its response to subsequent light pulses. To verify the model’s predictions, single-point measurements for targets of different reflectivity at defined distances were performed. Statistical analysis shows an R-squared value greater than 0.990 for simulated and measured signal amplitude levels. Noise modeling shows good accordance with the performed measurements for different target irradiance levels. The presented results have a guiding significance in the modeling of the complex signal processing chain of LiDAR systems based on SiPM detectors, as it enables the prediction of key parameters of the system early in the development process. Hence, unnecessary costs by design flaws can be mitigated.

TC-SEPM: Characterizing soft error resilience of CNNs on Tensor Cores from program and microarchitecture perspectives

As an architectural CNN accelerator integrated into NVIDIA’s GPUs, existing research mainly focuses on improving the performance of Tensor Cores. However, the highly integrated Tensor Cores are vulnerable to transient faults (i.e., soft errors), causing catastrophic consequences in safety-critical applications like automatic driving. Thus, it is imperative to estimate the reliability of CNNs on Tensor Cores. However, obtaining a statistically significant resilience profile of CNNs on Tensor Cores with the existing fault injection (FI)-based reliability estimation methods is expensive. To this end, we build TC-SEPM to predict the error resilience of CNNs on Tensor Cores instead of FI methods. To ensure the accuracy of TC-SEPM, we first investigate resilience-related features from program and microarchitecture perspectives. Then, leveraging these heuristic features, we train machine learning models to learn the hidden relationship between error resilience and the investigated features, enabling us to predict the impact of soft errors in Tensor Cores on CNN output. Experimental results show that TC-SEPM achieves high accuracy for individual soft error resiliency prediction and overall program resilience estimation while its overhead is only 1/27 of FI methods. Additionally, TC-SEPM can provide valuable insights for programmers or architects to design more robust CNN models on Tensor Cores.

Preventive and Predictive CNN Based Solution for Pipeline Leak, Blockage and Corrosion Detection

Using Artificial Intelligence (AI) in Internet of Things (IoT) environments has shown great potential to transform many industries. As oil and gas industries continue to push for digital transformation, safety critical applications such as pipeline leak detection, corrosion and blockage detection could also benefit from the adoption of AI and IoT. In our day-to-day life, we come across pipeline accidents and their after-effects. Structural deterioration of pipes in urban distribution network has presented great challenges to the utilities all over the world. When the deterioration exceeds its resiliency, the pipes leak or burst, leading to significant socio-economic overhead. The harsh operating environment, extreme temperature and weather conditions increases the potential for corrosion induced damages to the pipelines. The extreme conditions where the pipelines are located also make it difficult to rely on human operators to physically monitor these pipelines and respond to observed leaks. We propose a preventive and predictive AI-based solution that can continuously monitor the health of a complex pipeline infrastructure. It is a complete solution where pipelines will be scrutinized for blockage, leakage, corrosion, and defects. We propose a novel approach in which video images from IoT cameras installed across various locations on the pipelines are continuously analyzed and a convolutional neural network model is implemented for detecting leaks from the pipeline. Flow sensors will be mounted on the hardware to measure the fluid flow rate. This approach is proposed to deliver greater benefits in terms of accuracy and efficiency.

Enhancing Semantic Web Technologies Using Lexical Auditing Techniques for Quality Assurance of Biomedical Ontologies

Semantic web technologies (SWT) represent data in a format that is easier for machines to understand. Validating the knowledge in data graphs created using SWT is critical to ensure that the axioms accurately represent the so-called “real” world. However, data graph validation is a significant challenge in the semantic web domain. The Shapes Constraint Language (SHACL) is the latest W3C standard developed with the goal of validating data-graphs. SHACL (pronounced as shackle) is a relatively new standard and hitherto has predominantly been employed to validate generic data graphs like WikiData and DBPedia. In generic data graphs, the name of a class does not affect the shape of a class, but this is not the case with biomedical ontology data graphs. The shapes of classes in biomedical ontology data graphs are highly influenced by the names of the classes, and the SHACL shape creation methods developed for generic data graphs fail to consider this characteristic difference. Thus, the existing SHACL shape creation methods do not perform well for domain-specific biomedical ontology data graphs. Maintaining the quality of biomedical ontology data graphs is crucial to ensure accurate analysis in safety-critical applications like Electronic Health Record (EHR) systems referencing such data graphs. Thus, in this work, we present a novel method to create enhanced SHACL shapes that consider the aforementioned characteristic difference to better validate biomedical ontology data graphs. We leverage the knowledge available from lexical auditing techniques for biomedical ontologies and incorporate this knowledge to create smart SHACL shapes. We also create SHACL shapes (baseline SHACL graph) without incorporating the lexical knowledge of the class names, as is performed by existing methods, and compare the performance of our enhanced SHACL shapes with the baseline SHACL shapes. The results demonstrate that the enhanced SHACL shapes augmented with lexical knowledge of the class names identified 176 violations which the baseline SHACL shapes, void of this lexical knowledge, failed to detect. Thus, the enhanced SHACL shapes presented in this work significantly improve the validation performance of biomedical ontology data graphs, thereby reducing the errors present in such data graphs and ensuring safe use in the life-critical applications referencing them.

X-Masking for Deterministic In-System Tests

Deterministic in-system tests begin to play an essential role in safety-critical applications, in large data centers, or in monitoring silicon aging, to name just a few. All of these ecosystems require periodic, high-quality tests to assure required test coverage and short test application, especially in designs that must test themselves during system operations. In order for deterministic tests to be in-system applicable, they should compact multi-million-bit test responses with unknown (X) values to small signatures. This, in turn, allows for a much faster input-only streaming and a simultaneous reduction of the on-chip-stored test data volume, a system memory, and test time. Typically, the unknown states, whose sources vary from uninitialized memories to unpredictable last-minute timing violations, render signatures unusable. Hence, test response compaction requires some form of protection. This paper presents a user-tunable X-masking scheme. It works synergistically with on-chip test compression logic by employing encoded test data to completely filter out unknown values that otherwise might reach a test response compactor such as a MISR or test result sticky-bits used by the on-chip compare framework. It makes the proposed scheme a very versatile of its kind. Experimental results obtained for several industrial cores show feasibility and efficiency of the proposed scheme altogether with actual impact of X-masking on various test-related statistics.

AUDIT: Functional Qualification in Additive Manufacturing Via Physical and Digital Twins

Abstract Additive manufacturing (AM) has revolutionized the way we design, prototype, and produce complex parts with unprecedented geometries. However, the lack of understanding of the functional properties of 3D-printed parts has hindered their adoption in critical applications where reliability and durability are paramount. This paper proposes a novel approach to the functional qualification of 3D-printed parts via physical and digital twins. Physical twins are parts that are printed under the same process conditions as the functional parts and undergo a wide range of (destructive) tests to determine their mechanical, thermal, and chemical properties. Digital twins are virtual replicas of the physical twins that are generated using finite element analysis (FEA) simulations based on the 3D shape of the part of interest. We propose a novel approach to transfer learning, specifically designed for the fusion of diverse, unstructured 3D shape data and process inputs from multiple sources. The proposed approach has demonstrated remarkable results in predicting the functional properties of 3D-printed lattice structures. From an engineering standpoint, this paper introduces a comprehensive and innovative methodology for the functional qualification of 3D-printed parts. By combining the strengths of physical and digital twins with transfer learning, our approach opens up possibilities for the widespread adoption of 3D printing in safety-critical applications. Methodologically, this work presents a significant advancement in transfer learning techniques, specifically addressing the challenges of multi-source (e.g., digital and physical twins) and multi-input (e.g., 3D shapes and process variables) transfer learning.

A unified and practical user-centric framework for explainable artificial intelligence

Adoption of artificial intelligence (AI) is causing a paradigm change in many fields. Its practical utilization, however, especially in safety-critical applications like medicine, remains limited, mainly due to the black-box nature of most advanced AI models, which creates difficulties understanding why and how a model produces a particular output or decision. To overcome this issue, various methods and techniques have been proposed within the emerging field of explainable artificial intelligence (XAI). In this paper, we introduce a user-centric and interactive framework that enables a holistic understanding of AI systems. The proposed framework is designed to aid the development of more explainable AI systems by promoting transparency and trust in their use and allow different stakeholders to better understand and evaluate AI decisions. To illustrate the effectiveness of the framework, we implement a case study of an AI system analyzing optical coherence tomography (OCT) images. The development of this example case is reported using the proposed framework.

Fast adversarial attacks to deep neural networks through gradual sparsification

Deep learning networks, emerging machine learning models that present beyond human-level performance in terms of accuracy, are critically vulnerable to adversarial attacks. This vulnerability limits the utilization of deep learning architecture in many real-world safety-critical applications such as autonomous vehicles, medical diagnosis, sensitive industrial systems, etc. Adversarial attacks are methods to measure the robustness of different architectures and can be used to evaluate the suitability of a model to be used in a safety-critical situation. White-box sparse adversarial attacks can unhide interesting features of deep learning networks by identifying critical elements in the input pattern to design black-box attacks. This motivated us to design a new algorithmic procedure to design sparse adversarial attacks to feed-forward neural networks based on sparsity regularization. The proposed method comes with the gradual sparsification which starts by designing a dense attack and prone it until a desired level of sparsity is attained. We evaluate the performance of the proposed algorithm in designing attacks on convolutional neural networks and attention-based architectures for image classification task using three non-smooth sparsity-promoting regularizers. Compared to the state-of-the-art sparse attack schemes, we show that the proposed method can significantly decrease the time needed to design the attack, while the perturbation distortion is unchanged or reduced in some cases.

DSMC Evaluation Stages: Fostering Robust and Safe Behavior in Deep Reinforcement Learning – Extended Version

Neural networks (NN) are gaining importance in sequential decision-making. Deep reinforcement learning (DRL), in particular, is extremely successful in learning action policies in complex and dynamic environments. Despite this success, however, DRL technology is not without its failures, especially in safety-critical applications: (i) the training objective maximizes average rewards, which may disregard rare but critical situations and hence lack local robustness; (ii) optimization objectives targeting safety typically yield degenerated reward structures, which, for DRL to work, must be replaced with proxy objectives. Here, we introduce a methodology that can help to address both deficiencies. We incorporate evaluation stages (ES) into DRL, leveraging recent work on deep statistical model checking (DSMC), which verifies NN policies in Markov decision processes. Our ES apply DSMC at regular intervals to determine state space regions with weak performance. We adapt the subsequent DRL training priorities based on the outcome, (i) focusing DRL on critical situations and (ii) allowing to foster arbitrary objectives. We run case studies on two benchmarks. One of them is the Racetrack, an abstraction of autonomous driving that requires navigating a map without crashing into a wall. The other is MiniGrid, a widely used benchmark in the AI community. Our results show that DSMC-based ES can significantly improve both (i) and (ii).

Explainable Predictive Maintenance is Not Enough: Quantifying Trust in Remaining Useful Life Estimation

Machine learning (ML)/deep learning (DL) has shown tremendous success in data-driven predictive maintenance (PdM). However, operators and technicians often require insights to understand what is happening, why it is happening, and how to react, which these black-box models cannot provide. This is a major obstacle in adopting PdM as it cannot support experts in making maintenance decisions based on the problems it detects. Motivated by this, several researchers have recently utilized various post-hoc explanation methods and tools, such as LIME, SHAP, etc., for explaining the predicted RUL from these black-box models. Unfortunately, such (post-hoc) explanation methods often suffer from the \emph{disagreement problem}, which occurs when multiple explainable AI (XAI) tools differ in their feature ranking. Hence, explainable PdM models that rely on these methods are not trustworthy, as such unstable explanations may lead to catastrophic consequences in safety-critical PdM applications. This paper proposes a novel framework to address this problem. Specifically, first, we utilize three state-of-the-art explanation methods: LIME, SHAP, and Anchor, to explain the predicted RUL from three ML-based PdM models, namely extreme gradient boosting (XGB), random forest (RF), logistic regression (LR), and one feed-forward neural network (FFNN)-based PdM model using the C-MAPSS dataset. We show that the ranking of dominant features for RUL prediction differs for different explanation methods. Then, we propose a new metric \emph{trust score} for selecting the proper explanation method. This is achieved by evaluating the XAI methods using four evaluation metrics: fidelity, stability, consistency, and identity, and then combining them into a single \emph{trust score} metric through utilizing Kenny and Borda rank aggregation methods. Our results show that the proposed method effectively selects the most appropriate explanation method from a set of explanation methods for estimated RULs. To the best of our knowledge, this is the first work that attempts to address and solve the disagreement problem in explainable PdM.

Investigation of transmission characteristic for harmonic magnetic gears considering dynamic eccentricity using magnetic equivalent circuit

AbstractHarmonic magnetic gears (HMGs) have the advantage of lubrication‐free, high speed ratio and high torque density, which make it an attractive solution for safety critical applications. Due to eccentricity caused by machining and assembly, HMG suffers from dynamic eccentricity (DE) in operation, however, its effect on HMG performance is still unknown. Transmission characteristic of HMG under DE is studied. First, a magnetic equivalent circuit (MEC) model of HMG is proposed to build the magnetic coupling torque analytically, and the geometry of air gap is analysed parametrically to derive its equivalent reluctances. Flux density and coupling torque can be acquired by solving MEC equations. The accuracy of the MEC model is verified by finite element method. To study the transmission characteristic, an electromechanical coupling simulation framework for HMG is constructed, motion trajectories of rotors are investigated in case of DE, the output torque in locked‐rotor condition and speed response in continuous operation can be derived by simulation. It is found that torque ripples that have the same frequency with input rotor are induced by DE; the results are then verified in the experiment. This paper provides a theoretical guidance for the design and condition monitoring of HMG.

Characterizing and Improving Resilience of Accelerators to Memory Errors in Autonomous Robots

Motion planning is a computationally intensive and well-studied problem in autonomous robots. However, motion planning hardware accelerators (MPA) must be soft-error resilient for deployment in safety-critical applications, and blanket application of traditional mitigation techniques is ill-suited due to cost, power, and performance overheads. We propose Collision Exposure Factor (CEF), a novel metric to assess the failure vulnerability of circuits processing spatial relationships, including motion planning. CEF is based on the insight that the safety violation probability increases with the surface area of the physical space exposed by a bit-flip. We evaluate CEF on four MPAs. We demonstrate empirically that CEF is correlated with safety violation probability, and that CEF-aware selective error mitigation provides 12.3 ×, 9.6 ×, and 4.2 × lower dangerous Failures-In-Time rate on average for the same amount of protected memory compared to uniform, bit-position, and access-frequency-aware selection of critical data. Furthermore, we show how to employ CEF to enable fault characterization using 23, 000 × fewer fault injection (FI) experiments than exhaustive FI, and evaluate our FI approach on different robots and MPAs. We demonstrate that CEF-aware FI can provide insights on vulnerable bits in an MPA while taking the same amount of time as uniform statistical FI. Finally, we use the CEF to formulate guidelines for designing soft-error resilient MPAs.

Shear and delamination behaviour of basal planes in Zr3AlC2 MAX phase studied by micromechanical testing

The mechanical properties of layered, hexagonal-structured MAX phases often show the combined merits of metals and ceramics, making them promising material candidates for safety critical applications. While their unique mechanical performance largely arises from the crystal structure, the effect of chemistry on the properties of these materials remains unclear. To study this, here we employed two in situ electron microscope small-scale testing approaches to examine the micromechanical properties of Zr3AlC2, and compared the results with the properties of Ti3SiC2: we used micropillar compression tests to measure basal slip strength, and double cantilever beam splitting tests to evaluate fracture energy for basal plane delamination. We observed distinct and systematic differences in these measured properties between Zr3AlC2 and Ti3SiC2, where Zr3AlC2 appeared to be stronger but more brittle at the microscale, and discussed the implications of the results in the selection, design, and engineering of MAX phases for targeted engineering applications.

Lightbox: Sensor Attack Detection for Photoelectric Sensors via Spectrum Fingerprinting

Photoelectric sensors are utilized in a range of safety-critical applications, such as medical devices and autonomous vehicles. However, the public exposure of the input channel of a photoelectric sensor makes it vulnerable to malicious inputs. Several studies have suggested possible attacks on photoelectric sensors by injecting malicious signals. While a few defense techniques have been proposed against such attacks, they could be either bypassed or used for limited purposes. In this study, we propose Lightbox, a novel defense system to detect sensor attacks on photoelectric sensors based on signal fingerprinting. Lightbox uses the spectrum of the received light as a feature to distinguish the attacker’s malicious signals from the authentic signal, which is a signal from the sensor’s light source. We evaluated Lightbox against (1) a saturation attacker, (2) a simple spoofing attacker, and (3) a sophisticated attacker who is aware of Lightbox and can combine multiple light sources to mimic the authentic light source. Lightbox achieved the overall accuracy over 99% for the saturation attacker and simple spoofing attacker, and robustness against a sophisticated attacker. We also evaluated Lightbox considering various environments such as transmission medium, background noise, and input waveform. Finally, we demonstrate the practicality of Lightbox with experiments using a single-board computer after further reducing the training time.

Automated SC-MCC Test Case Generation using Bounded Model Checking for Safety-Critical Applications

Modified Condition/Decision Coverage (MC/DC) is an important criterion to test the safety-critical applications because it generates test cases in a linear manner from N+1 to 2N, where N is the number of Atomic Conditions in a given predicate. It is desirable in comparison to the exponential test cases produced for Multiple Condition Coverage (MCC), i.e., 2N. However, MCC with Short-Circuit (SC-MCC) is recommended since most of the safety-critical applications are built on the high-level languages that use Short-Circuit evaluation property. Our goal is to demonstrate that, despite the added overhead, the SC-MCC coverage-based test cases have a high error-detection probability compared to that of MC/DC. In this work, we have considered the CBMC tool to generate both the SC-MCC and MC/DC test cases for 80 RERS benchmark programs. Then, by optimizing the traditional Mutation Testing (using the GCOV tool), we computed the Mutation Score (%) to assess the quality of the generated test cases. As per the Mutation analysis, the proposed SC-MCC outperformed MC/DC for over 75% of the 80 RERS programs considered. Additionally, as per the Execution time analysis, the average total time taken to evaluate the MC/DC part of the proposed framework is 3065.98 s, whereas SC-MCC framework evaluation took 2167.70 s. This proves the efficiency of the proposed (SC-MCC) work over the traditional criterion (MC/DC) based work.

In-Situ early anomaly detection and remaining useful lifetime prediction for high-power white LEDs with distance and entropy-based long short-term memory recurrent neural networks

High-power white light-emitting diodes (LEDs) have demonstrated superior efficiency and reliability compared to traditional white light sources. However, ensuring maximum performance for a prolonged lifetime use presents a significant challenge for manufacturers and end users, especially in safety–critical applications. Thus, identifying functional anomalies and predicting the remaining useful lifetime (RUL) is of enormous importance in the operational longevity of the device. To address such challenges, this study proposes a combination of distance-based Mahalanobis distance (MD), entropy generation rate (EGR), and deep learning models for improved anomaly detection and RUL prediction accuracy. Unlike conventional health indicators based on luminous flux data that are challenging to monitor relevant optical performance, the MD and EGR methods are employed to extract in-situ monitored thermal and electrical data as new health indicators. Long short-term memory recurrent neural networks (LSTM-RNN) and convolutional neural networks (CNN) are established to detect anomalies and predict the RUL. The accelerated degradation tests of 3 W high-power white LED have been conducted, and the online and offline collected experimental data are deployed for model development and performance evaluation. The performance of the proposed methods is compared against the Illuminating Engineering Society of North America (IESNA) TM-21 method. The results indicate that LSTM-RNN, when combined with either MD or EGR, can detect anomalies with significantly fewer data (70 %) than is typically required. Furthermore, a significant improvement in prediction accuracy in RUL prediction based on MD and EGR-constructed time series health indicators and employed with the LSTM-RNN model demonstrates the effectiveness of the proposed methods.