Failure of electronic components

Electronic components have a wide range of failure modes. These can be classified in various ways, such as by time or cause. Failures can be caused by excess temperature, excess current or voltage, ionizing radiation, mechanical shock, stress or impact, and many other causes. In semiconductor devices, problems in the device package may cause failures due to contamination, mechanical stress of the device, or open or short circuits. Failures most commonly occur near the beginning and near the ending of the lifetime of the parts, resulting in the bathtub curve graph of failure rates. Burn-in procedures are used to detect early failures. In semiconductor devices, parasitic structures, irrelevant for normal operation, become important in the context of failures; they can be both a source and protection against failure. Applications such as aerospace systems, life support systems, telecommunications, railway signals, and computers use great numbers of individual electronic components. A

Optimization of the treatment of hydrocephalus by the non-invasive measurement of the intra-cranial pressure

Alec Ginggen

This thesis work aims to optimize the treatment of hydrocephalus. It consists of two distinct objectives: the optimization of the diagnosis of patients implanted with a shunt system for the derivation of cerebro-spinal fluid (CSF) and the optimization of the treatment by a better understanding of the fluidic characteristics of shunts. The diagnosis optimization was addressed by a pressure sensor allowing for the measurement of the intra-cranial pressure (ICP). It has been characterized experimentally in vitro and in vivo in animals. The pressure sensor is integrated along a shunt system for the derivation of CSF and it can be interrogated by telemetry. The non-invasive access to the ICP will allow the physician to verify that the implantation of a shunt system has successfully restored a physiological ICP level. It will also provide a method to verify that the shunt is not blocked and fully functional. Simple mathematical models of the hydrodynamics of shunt systems used clinically are presented to address the second objective. These simple models are verified experimentally by characterizing the shunts on a setup developed for that purpose. This setup allows the characterization of non-bijective flow-pressure characteristics. The integration of mathematical models of shunt systems in a broader model describing the circulation of CSF in a patient implanted with a shunt will permit to predict the dynamics of the CSF circulation. It will also be a useful tool for shunt manufacturers to identify shunt characteristics critical to successfully restore a physiological condition after implantation. The results of this thesis work are presented in the form of an introduction, three scientific publications and a short conclusion. The motivations for the thesis work are presented in the introduction chapter along with background information about hydrocephalus and its treatment. A mathematical model describing the circulation of CSF is described in the same chapter. This model illustrates the interactions between the fluidic constituents of the cranial vault that lead to a physiological CSF circulation and ICP level. This model has been adapted to account for the implantation of a shunt system. The key technical characteristics of the pressure sensor are exposed in the subsequent section along with the experimental techniques and a short summary of the publications and patents. The pressure sensor is described in the first publication with a focus on the pressure transducer, the original sensor encapsulation technique, the electronics embedded in the sensor and the telemetry technique. A capacitive pressure transducer, which provides an accurate measurement of the absolute pressure, has been developed. The transducer technology guarantees a minimal drift of its characteristics with time. The design and the key manufacturing steps of the transducer are exposed. A hermetically closed capsule protects the sensor electronics. The effect of the packaging components and the assembly techniques on the sensor characteristics are analyzed. An analysis of the water diffusion through the walls of the capsule is presented since resistance to water permeation is a critical characteristics of the packaging. Minimal drift with time can only be achieved if the atmosphere in the capsule is extremely stable. The packaging provides a barrier to water diffusion in order to keep a constant low humidity level in the sensor. The telemetric link between the sensor and the external reading unit is provided by an inductive coupling between two loop antennas. The coupling allows the simultaneous transmission of power to the sensor and of the pressure data to the external unit. The fundamental characteristics of the telemetric link are described. The interfacing electronic components are integrated in an ASIC that consumes a minimal amount of energy. Lastly, the experimental characterization of the accuracy of the sensor and its thermal stability are presented. The second publication is focused on the in vivo characterization of the pressure sensor in animals. An animal model of hydrocephalus has been developed in order to characterize the correlation between the pressure measured by the sensor integrated into a shunt system and a reference provided by a direct measurement of the ICP in the ventricles of the brain. Hydrocephalus was induced in dogs by injection of kaolin and glue in the cisterna magna. The time course of hydrocephalus induction was followed by the measurement of the size of the ventricles by an ultrasonic imaging technique. Approximately 50 days after the injection of kaolin and glue, a shunt system integrating a pressure sensor was implanted between the brain ventricles and the peritoneum. A reference pressure sensor was also implanted in the ventricles. The pressure was then measured for up to 48 hours. The correlation between the pressure in the shunt and the intra-ventricular pressure is documented in the publication. The advantages associated with the measurement of the pressure in the shunt versus an intra-ventricular measurement technique are discussed and the pressure sensor is compared with other published sensors. The third publication describes three commonly used shunt systems by simple hydrodynamic mathematical models. A setup allowing the characterization of shunt systems having a non-bijective pressure-flow characteristics is presented. The experimental characteristics of the three shunt systems is compared to the theoretical models. The "pressure-controlled" and "flow-controlled" terminologies, which are frequently used in the clinical field, are discussed in the context of the hydrodynamic characteristics of shunts. Four patent applications have been submitted in the course of this thesis work. The three first applications have been granted, and the last one is filed. The first patent protects the pressure sensor encapsulation technique. The second patent protects the concept of a shunt system allowing the non-invasive adjustment of the resistance of the shunt valve. This concept has been developed as a result of the analysis of the critical parameters to successfully restore a physiological ICP level by the mathematical model of the CSF circulation. The third patent protects the use of the pressure sensor packaging technique for a flow sensor integrated in the shunt system. The last application protects the mechanism that regulates the power of the radio-frequency wave emitted by the external reading unit on the basis of the voltage induced in the sensor. The last chapter, entitled "Conclusion and Perspectives" summarizes the outcome of this thesis work in reference to the optimization of the treatment of hydrocephalus. This work laid the foundation for the use of the sensor in humans through a clinical study. The ICP data gathered during such a study would then allow the validation of the mathematical model of CSF circulation presented in this thesis.

EPFL2007

Genuinely distributed Byzantine machine learning

El Mahdi El Mhamdi, Rachid Guerraoui, Arsany Hany Abdelmessih Guirguis, Le Nguyen Hoang, Sébastien Louis Alexandre Rouault

Machine learning (ML) solutions are nowadays distributed, according to the so-called server/worker architecture. One server holds the model parameters while several workers train the model. Clearly, such architecture is prone to various types of component failures, which can be all encompassed within the spectrum of a Byzantine behavior. Several approaches have been proposed recently to tolerate Byzantine workers. Yet all require trusting a central parameter server. We initiate in this paper the study of the “general” Byzantine-resilient distributed machine learning problem where no individual component is trusted. In particular, we distribute the parameter server computation on several nodes. We show that this problem can be solved in an asynchronous system, despite the presence of 13 Byzantine parameter servers (i.e., nps>3fps+1) and 13 Byzantine workers (i.e., nw>3fw ), which is asymptotically optimal. We present a new algorithm, ByzSGD, which solves the general Byzantine-resilient distributed machine learning problem by relying on three major schemes. The first, scatter/gather, is a communication scheme whose goal is to bound the maximum drift among models on correct servers. The second, distributed median contraction (DMC), leverages the geometric properties of the median in high dimensional spaces to bring parameters within the correct servers back close to each other, ensuring safe and lively learning. The third, Minimum-diameter averaging (MDA), is a statistically-robust gradient aggregation rule whose goal is to tolerate Byzantine workers. MDA requires a loose bound on the variance of non-Byzantine gradient estimates, compared to existing alternatives [e.g., Krum (Blanchard et al., in: Neural information processing systems, pp 118-128, 2017)]. Interestingly, ByzSGD ensures Byzantine resilience without adding communication rounds (on a normal path), compared to vanilla non-Byzantine alternatives. ByzSGD requires, however, a larger number of messages which, we show, can be reduced if we assume synchrony. We implemented ByzSGD on top of both TensorFlow and PyTorch, and we report on our evaluation results. In particular, we show that ByzSGD guarantees convergence with around 32% overhead compared to vanilla SGD. Furthermore, we show that ByzSGD’s throughput overhead is 24–176% in the synchronous case and 28–220% in the asynchronous case.

2022

Optimization of the treatment of hydrocephalus by the non-invasive measurement of the intra-cranial pressure

Alec Ginggen

EPFL2007

Genuinely distributed Byzantine machine learning

El Mahdi El Mhamdi, Rachid Guerraoui, Arsany Hany Abdelmessih Guirguis, Le Nguyen Hoang, Sébastien Louis Alexandre Rouault

2022