Smart Sensors and Sensor Networks

1. Smart Sensors and Sensor Networks Lecture14 Reliability and fault tolerance

2. Smart Sensors and Sensor Networks Reliability in SNs • The sensors and the sensor network infrastructure are prone to failures, insufficient energy supply, high error rate and disconnection; • Sensors may be also mobile and deployed randomly in hostile environments; • In many cases, the sensor networks are ad hoc, with little or no fixed network support; • Sensing and processing of data must be performed reliably to ensure the correctness and accuracy of the applications’ results; • Critical real-time information must still be disseminated dynamically from mobile sensor data sources through the self-organizing network infrastructure; • Information fusion must be reliable in order for components to maintain correct control of dynamic replanning and reoptimization of the theater of operation, based on newly available information;

3. Smart Sensors and Sensor Networks • Reliability and distributed services: • Reliability and reconfiguration can be done through three main distributed services: service lookup, sensor node composition and dynamic adaptation; • Based on them, another system mechanism, connectors, is developed that supports reliable and reconfigurable communication between sensor nodes; • Using distributed lookup servers, composition servers, adaptation servers and connectors, remote surveillance and target tracking systems may adapt these services to device failure and degradation, movement of sensor nodes and changes in task and network requirements; new application-specific services may be deployed reliably to support existing distributes sensor applications while they are executing; • These mechanisms also enable sensor nodes to have capabilities for self-assembling networks that are extensible, sensor node mobility and changes in task and network requirements; although nodes are autonomous, they may cooperate with one another to disseminate information or assist each other in adapting to changes due to failures or degradation of some sensors;

4. Smart Sensors and Sensor Networks • Architecture of distributed sensor systems

5. Smart Sensors and Sensor Networks • There are three key system layers: • Application system: sensor information processing layer and collaborative signal processing; • Configurable distributed system: provides distributed services to the applications; • Physical device layer: routes messages through the ad hoc sensor network; • At the physical device layer, sensors and mobile devices may be assembled and reconfigured dynamically in an ad hoc wireless network; • Each sensor contains battery, wireless communication, multiple sensing modalities, computation unit, memory and localization module; message routing and query processing use the location information; • At the networking layer, routing protocols allow messages to be forwarded through multiple physical clusters of sensors; • Directed diffusion routing may be used because of its ability to adapt to changes in sensor network topology dynamically and its energy-efficient localized algorithms; • Applications and programs may use simpler communication interfaces and abstractions, such as subscribe/publish used in diffusion routing; • These distributed services may enhance overall performance; • At the application layer, distributed query processing and collaborative signal processing modules communicate with each other to support the surveillance and dynamic tracking functions;

6. Smart Sensors and Sensor Networks • Distributed services • To enhance the ability to reconfigure their networking, configuration and adaptation functionalities, sensors may take use of three main classes of distributed services: lookup, composition and adaptation services; • Reconfigurable sensors: • By exploiting these distributed services, sensors can be self-aware, self-reconfigurable and autonomous; they may adapt rapidly to abrupt changes in the sensors’ capabilities, events and new real-time information; • Reconfigurable sensors interact with other reconfigurable sensors through well-defined interfaces that maintain interaction states to allow nodes to be reconfigured dynamically; these explicit interaction states and behavior information allow localized algorithms with the adaptation servers to maintain consistency when autonomous nodes and clusters are reconfigured dynamically, move around or recover from failure; • When new reconfigurable sensors are added to the network, they register their services with a lookup server; other nodes that require a service will discover the services available in a cluster through the lookup servers that return the location and interface of the service nodes; • Distributed lookup server • New network and system services may be introduced by a sensor node for use by other nodes as the sensor network self-organize;

7. Smart Sensors and Sensor Networks • A sensor node that provides a service is called a service provider and a node that uses the service is called a service client; a sensor node may register a resource that it maintains or service that it can perform with a lookup server; • A lookup server may contain information on services or resources at multiple clusters; • Other nodes requiring the service may request it through a lookup server; if the service is recorded in the lookup server, it will return the location to the requesting node; otherwise, a discovery protocol will be used to locate the service through other lookup servers;

8. Smart Sensors and Sensor Networks • Applications use the lookup service for registering, finding and calling a service; the following service function calls may be used: • service_register():the function allows a service provider to register its service with a lookup server in the region; • service_deregister(): a service provider will remove its service from the lookup server registry; • lookup_service(): the function allows a service client to find the location or address of a service provider and/or the interface for using the service; service lookup can also be based on cluster or predicate matching; • service-exec(): it allows a sensor node to request the service and obtain the results from the service provider; the service provider performs the requested service or remote procedure call and returns the results; • service_call(): it allows a service client to find and make a call for a service when the service client does not know the location or address of the service provider and/or the interface for using the service; it is a implemented as a combination of the lookup_service()and service-exec()calls; • To search multiple lookup servers, a request message is propagated to all the lookup servers; the server that contains the service registration information will return the reply with the service location; it may also return the cluster name of that service; the lookup server that made the request will then cache that service location; at regular frequency, service and resource registration information may be disseminated from one lookup server to others;

9. Smart Sensors and Sensor Networks • Compositional server: • The compositional server manages clusters of sensor nodes by allowing various nodes to be added or removed; • It manages network abstractions and hierarchical composition of clusters; • The advantage of forming clusters is that failure and recovery of sensor nodes can be contained within a cluster maintaining also the effects of failure within the cluster; this simplifies the development of a large-organizing sensor network by allowing individual nodes and clusters to be specified and designed independently; • Adaptation server: • Adaptation servers utilize information from the compositional server, lookup server and analytical tools to control sensors during dynamic reconfiguration and failure recovery; • Adaptation servers monitor clusters of sensors during normal execution by probing the smart nodes or network management directives for reconfiguration and failure recovery; • When a runtime reconfiguration is requested or triggered by a failure event, the adaptation server will generate the correct schedule of reconfiguration and recovery operations;

10. Smart Sensors and Sensor Networks • Mechanisms and tools • Reliable remote service execution: • Sensor applications require reliable communication over unreliable sensor network layers; • The following mechanism allows sensors to locate the appropriate service provider and execute the service remotely and reliably: • When a node wants to execute a service from a service provider, it searches its local service table to get the information on the service and the service provider; • If it cannot find the relevant information, it calls the service lookup function to update its local service table; • Three types of service interfaces may be specified by the service provider’s information: • Location or address of the service provider with known interface; • Interface definition of the service; • Mobile code for the interface protocol; • Example: the service client uses remote procedure calls (RPC) to request remote services from the service provider; RPCs are implemented as follows: • The call to service_exec()will first send an interest to the service provider through the subscribe function; • The service provider sends to the client the permission to send the request;

11. Smart Sensors and Sensor Networks • The client then sends a request and all the input data; • If the input data are large, several packets are sent reliably through an automatic repeat request protocol with retransmission; • The service provider will then process the service remotely; • The client will request the result of the service through another subscription; • The provider then returns the result in response to the interest; • Connectors: • Sensors interact with each other through connectors; • They encapsulates the properties and states of the interaction and contain specification of the communication methods and the interfaces of its endpoints and the attached sensor nodes; • Connectors are units of adaptable communications; sensors use a connector without being aware on how and when changes are made in the communication methods or the sensors that are being connected; • For example, a sensor at one end of the connector could fail and be replaced with another; the connector will manage the replacement transparently so that the other sensor at the other end may not be aware of the replacement; • The runtime mechanisms enable active connectors to adapt dynamically as sensor nodes are deployed, replaced or removed; for that, the library code of the connector contains functions for detecting and managing changes;

12. Smart Sensors and Sensor Networks • Sensor task structures • Data flow and task structures of the sensor nodes may be specified using connectors for facilitating communication and reconfiguration; • The connector facility allows applications to assemble incrementally and reconfigure sensors to form task groups; • The advantage of specifying these task structures is that communication is more efficient and the system mechanisms automatically maintain the structure and reconfigure when failures occur in the nodes; • The group structures can also be changed on the fly in response to changes in the global task requirements; • During failure recovery and reconfiguration, the connector, task structure and state information stored in the composition server allow the sensors to be replaced and reconfigured automatically by the adaptation and composition services; • Adaptation is performed transparently to the sensor application; • Sensor applications need only deal with the simple communication interface provided by the connectors and do not need to be aware of communications between the adaptation and composition server for performing recovery from failure;

13. Smart Sensors and Sensor Networks • Dynamic adaptation of distributes sensor applications: • Sensor failure can trigger distributed sensor applications to adapt to recover from the failure; • Dynamic adaptation may be initiated in two ways: • Explicit request by the human operator or the sensor application; • Triggered by sensor values above a threshold or other changes in sensor or distributed environments; • Adaptation operations may include methods for modifying services, replacing sensor nodes or introducing new sensor services; • Recovery from sensor failure and degradation is a special case of dynamic adaptation of some or all of the services of a sensor node; recovery operations may be used to restore those services or individual sensors; • However, sensors may interact with other sensors to perform coordinated services when the failure occurs; the distributed sensor application may be inconsistent when failure occurs or when restoring an intermediate sensor after a failure; • The adaptation server analyses the failure and determines the schedule of recovery operations that will restore global consistency of the service; • The recovery schedule analysis also requires information from the composition server to determine the cluster composition and the lookup server for services involved in the cluster.

14. Smart Sensors and Sensor Networks • Recovery from sensor failure and degradation: • When a failure occurs, at node E, the adaptation server will detect the failure using its monitoring facility; • Through the trigger mechanism it will automatically determine the sequence of procedures for initiating recovery of affected sensors and connectors; • The adaptation server will communicate with the composition server to determine the task structures and states of the connectors in order to recover the sensor group into an equivalent group task structure; • A new sensor node, F, with similar services as node E will be found through the lookup service and activated to replace node E;

15. Smart Sensors and Sensor Networks Fault tolerance in SNs • Reasons for research in fault tolerance in WSNs: • Technology and implementation aspects: • At least two components of a sensor node, sensors and actuators, will directly interact with the environment and be subjected to a variety of physical, chemical and biological forces; • They will have significantly lower intrinsic reliability than integrated circuits in fully enclosed packaging; • SNs must work with severe limitations; • The number of sensor nodes may be very large; • Complex applications: • SNs will often operate in autonomous mode without a human in the loop; security and privacy concerns will often prevent expensive testing procedures; this will adversely affect not only testing and fault tolerance but also related tasks, such as debugging, in which reproducing the fault conditions will be difficult; • Applications may require to deploy sensor nodes in uncontrolled even hostile environments; • Many applications will be safety critical and could have an adverse impact on humans and environments, particularly when actuators are used;

16. Smart Sensors and Sensor Networks • The fault tolerance approach must take into account the main problems of the WSN domain: • For example, if one considers power consumption, each fault tolerance solution will be judged accordingly; specifically, if communication energy is significantly higher that computation energy, it is important to develop localized algorithms that will require only a limited amount of energy; • Fault tolerance is necessary at the sensor node level as well as at the communication level too; • Fault tolerance is necessary during sensor fusion; • The goal is to design fault-tolerant techniques that do not significantly increase the communication overhead; • On the other hand, if the computation energy is significantly higher than the communication requirements, a good solution may be to support communication resources at one end with computation resources at other nodes; • It is preferable to develop fault-tolerant sensor fusion approaches that require little additional computation regardless of any additional communication requirements;

17. Smart Sensors and Sensor Networks • Example of fault tolerance in a SN system • Fault-tolerant multimodal sensor fusion: • A sensor recognition system is deployed in an office to identify people as they walk in through the main door; • Six people, named A, B, C, D, E and F work in the office; • The system consists of two different types of sensors: • A height sensor – a set of light sensors in series; • A voice recognition system that requires everybody entering in the room to speak a given phrase into a microphone;

18. Smart Sensors and Sensor Networks • Figures show the identification characteristics of people in the office; • The system can distinguish between two persons, P1 and P2, if they fall into different squares, when mapped in the chart from fig. b; • If all the sensors work properly, each person will naturally fit into a different square, according to the fig.; • For most of the cases, even if one of the height sensors or voice sensors fails, recognition of the right person is still possible; • This is accomplished using heterogeneous fault tolerance, in which a failed sensor of one type can be replaced by the functionality of a sensor of another type; • For the case of persons B and E, with the same height, voice is the only way to distinguish the two persons, so the system does not have any fault tolerance to the failure of sensor V3 that distinguishes between the objects B and E; • If the office had only five people, that is B or E, then it would be completely fault tolerant; • Complex sensor network systems can be designed in such a way that the heterogeneous fault tolerance scheme can make the system resilient of a specified number of sensors of specified modality;

19. Smart Sensors and Sensor Networks • Fault tolerance basics • Reliable system design is approached in the following life stages of a product: design, manufacturing and operational; • Error is the manifestation of a fault inside the program; error can occur not only at the fault site but also at some distance; • Fault is the incorrect state of hardware or a program as a consequence of a component’s failure; • Permanent faults are continuous and stable in time; they can be consequences of irreversible physical alteration within a component; • An intermittent fault is one that has only occasional manifestation due to unstable characteristics of the hardware or as a consequence of a program being in a particular subset of space; • A transient fault is the consequence of temporary environmental impact on otherwise correct hardware; • Fault tolerance considers three main types of concerns: • Fault models; • Fault detection and diagnosis; • Resiliency mechanisms;

20. Smart Sensors and Sensor Networks • Each level of abstraction has its own types of faults; for example, at the gate level: several fault models have been used in the testing phase: • “Stuck”: the logical value on interconnect gate or pin is permanently set to a value stuck at one or zero; • “Bridging”: two or more neighboring lines are physically connected introducing wired AND or OR functions, depending on the logic family used; • “Shorts and opens”: correspond to missing or introduced connections; • Reliability techniques include the following phases: • Fault confinement: establishes limits of fault effects over a particular area; therefore, contamination of other area is prevented; • Fault detection: it is recognized that an unexpected event has occurred; • Traditionally fault detection techniques are classified into offline and online detections • Most often, for offline detection, special diagnostic programs are employed during idle periods of time or using multiplexing with a regular mode of operation; • Online detection targets real-time fault identification and is performed simultaneously with a real work load; typical online detection techniques are checking, duplication and triplication; • Fault latency: the time that passes between the fault occurrence and the moment when the fault is detected;

21. Smart Sensors and Sensor Networks • Diagnosis: is a stage at which the exact occurrence of a fault is attributed to a specific atomic piece; • Reconfiguration: is the stage entered after diagnosis at which the system is reconstructed in such a way that faults do not have impact on the correct output; • Graceful degradation is a reconfiguration technique in which performance of the system is reduced but the correct functionality is preserved; • Recovery: is a stage at which an attempt is made to eliminate the effects of faults; the most widely used recovery techniques are fault masking and retry; • The fault masking approach uses redundant correct information to eliminate the impact of incorrect information; • In retry, after the fault is detected, a new attempt to execute a piece of program is made in the hope that the fault is transient; • Restart: is the stage invoked after the recovery of correct, undamaged information; in cold restart, a complete resetting of the system is conducted; • Repair: is the stage during which the failed component is substituted with the operational component; • Almost all testing approaches assume a single fault model, regardless of which type of fault is considered;

22. Smart Sensors and Sensor Networks • Fault tolerance at different sensor levels • Physical layer • Is responsible for establishing communication in a given medium between two nodes; • Typical tasks: modulation – demodulation and encoding – decoding; • Traditionally, fully hardwired solutions have been used in order to minimize cost and maximize energy efficiency; • A software radio is a wireless communication device in which parts or all of the physical layer functions are realized in software; • Software radios are a way to extend programmability into the physical layer and to enable adaptation to channel conditions; • The primary reason for deployment of hardware and software radios has been to solve interoperability problems (for ex. incompatibility between cellular and PCs communication standards) and to enhance performances in noisy media; • They are also ideally suited for realization of a variety of fault tolerance techniques at the physical layer; for ex. if some components of the software radio fail, one can switch to modulation and encoding schemes that can be realized with operational hardware resources; • Adaptation to noise characteristics – fault tolerance;

23. Smart Sensors and Sensor Networks • Hardware • Components can be divided in two groups: • Computation, storage and power supply; • Wireless communication, sensors and actuators; • First group: • Very reliable, includes sophisticated fault tolerance techniques; • Microprocessors, microcontrollers and DSPs are very reliable circuits with very low rates of malfunctioning; • This does not necessarily imply that computational systems of sensor nodes are very reliable; reasons: • Sensor nodes are very cost sensitive and not always the highest quality components were used for designing them; • Energy constraints imply that repeated computation solutions are not realistic; use of application-specific design can have up to two orders of magnitude less energy consumption for the same functionality; for these subsystems, heterogeneous BISR fault-tolerant schemes will provide the targeted level of fault tolerance and low energy consumption; • Sensor nodes are often deployed in harsher environments than those in which computers function; • Storage components are very reliable; MRAMs, SRAMs and DRAMs have simple and regular structures; flash limits the number of writes in the same cell;

24. Smart Sensors and Sensor Networks • Regarding power supply: • The traditional energy sources (batteries) are fault tolerant if a back-up battery is foreseen; • The fuel cells are the future but, for the moment, some energy scavenging subsystems, such as the ones converting light in energy, can have very volatile performance; • Second group: • The standard way to enhance performance of radios is to use aggressive error correction schemes and retransmissions; they are examples of time redundancy; • In addition, other solutions with two or more radios were proposed; • Sensors and actuators are the subsystems most prone to malfunctioning; • In the case of sensors, three types of errors can be delimitated: • Calibration systematic error; • Random noise error; • Complete malfunctioning; • The first two can be addressed through time redundancy and the second with hardware redundancy too; • For actuators, the hardware redundancy is the solution; • In addition to the schemes that operate on the physical and link layers, techniques that operate at the network layer are important; they are implemented at the system software level;

25. Smart Sensors and Sensor Networks • System software • System software consists of the operating system and the utility programs; • Fault tolerance at the system software level can be addressed in several ways; • The most promising is through software diversity: each program is implemented in n different versions hopping that different versions will not have identical bugs; • The subsystem that can most benefit from fault tolerance at the system software level is the communication unit; • For example, one can reroute messages using different paths in the multihop network; with respect to sensors and actuators, the most important piece of system software is the one related to calibration; • A very important component of system software is the one that supports distributed and simultaneous execution of localized algorithms; • For example, in the case of energy minimization under functionality constraints requirements, several protocols have been developed for the coordination of distributed actions; • When communication protocols are considered, there is a clear trade-off between complexity and effectiveness;

26. Smart Sensors and Sensor Networks • Middleware • Starting with the middleware level, emphasis is shifted toward data aggregation, data filtering and sensor fusion; • These are tasks mainly related to sensor readings; • Because it is difficult to provide fault tolerance in an economic way at the level of a single sensor, numerous fault tolerant approaches were included in the middleware level; • N-versioning is a solution at this level too; • Heterogeneous approaches that can substitute the readings of one type of sensor with the readings of another type are important; • Another middleware-connected issue is how many sensors of each type should be placed on which positions on a particular node; if error resiliency of communication is much higher than the error resiliency of sensors, solutions with sensors of the same type placed on the same node will be favored; • Application • For example, to identify a particular person, one can use sensors that measure several biometric features of that person; each feature and possibly a combination of features will be sufficient to identify that person; • Fault tolerance techniques will be application-specific;

27. Smart Sensors and Sensor Networks • Case studies • Discrepancy-based fault detection • The solution does an online detection of sensor faults; it can be applied to a broad set of fault models; • A fault is defined as an arbitrary type of inconsistent measurement by a sensor that cannot be compensated systematically; in particular, faults are associated with incorrect measurements that cannot be corrected using calibration techniques; • The approach is based on two ideas: • Comparing the results of multisensor fusion with and without each of the sensors involved; • Using nonparametric statistical techniques to identify measurements that are not correctable, regardless of the mapping function used between the measured and accepted values; • Sensor measurements are inevitably subject of errors of two types: • Random fluctuations in data due to a noise in a sensor or in a sensed phenomenon; • Gross errors; • A practical method to distinguish a random noise is to run maximum likelihood on the multisensor fusion measurements; • A random noise would exist if running these procedures improves the accuracy of final results of multisensor fusion;

28. Smart Sensors and Sensor Networks • In multisensor fusion, measurements from different sensors are combined in a model for consistent mapping of sensed phenomena; • Although the new fault detection technique is generic and can be applied to an arbitrary type of data fusion, for the sake of brevity and clarity the focus was on equation-based sensor fusion; • The technique: • Assume a set of sensors si, 0 ≤ I ≤ n, each measuring a value xi at time t; • The multimodal sensor fusion model equations are f1, …, fp and are typically nonlinear functions with the following forms: fj (x1, x2, …, xn) = 0, 0 ≤ j ≤ p • The system of equations is overconstraint and solves the system n + 1 times; • First, they solve with all the equations in the original format; • Then, they ignore each variable and solve a least-constrained system with n – 1 variables (n times); • They compare the values for each variable xn in all n + 1 scenarios; • In order to improve accuracy of fault detection, the system can be solved for m measurements by each sensor; • At last, they conduct statistical analysis on the data for each sensor; • If the obtained values for a sensor are not consistent within a confidence interval calculated by the percentile method, that sensor is considered faulty;

29. Smart Sensors and Sensor Networks • Heterogeneous fault detection • An object O is moving along its trajectory, which include points pi in an sensor network; the nodes are represented by shaded circles ni; • Four types of sensors are used to measure the angle in two-dimensional physical space: RSSI-based distance discovery, speedometer, accelerometer and compass; • Three RSSI-based measurements are used to locate the object O in any particular moment; • Euclidian space, Newton mechanics and trigonometry laws can be used to establish relationships between measurements; • Specifically, trilateration equations, Newton law equations and trigonometric equations can be obtained;

30. Smart Sensors and Sensor Networks

31. Smart Sensors and Sensor Networks • There are more equations than sensors; thus, if one sensor is not functioning, it can be calculated from the system of equations; • One way to identify and correct sensor measurements is to try all scenarios in which exactly one type of sensor measurement is not taken into account and compare the maximal error in the system; • Another important observation is that, by sampling all operational sensors more often, one can compensate for faulty sensors;

32. Smart Sensors and Sensor Networks • Future research directions: • The development of theoretically attractable and realistic fault models is one of the key prerequisites for development of fault tolerance techniques for SNs; • Very little has been published about fault models for sensors and actuators; at the same time, these components are the most important for overall system fault tolerance; • The development of fault models for sensors is particularly difficult due to the great variety of their types, environments in which they will be deployed and requirements in terms of fault tolerance of various applications; • For example it is clear that electromagnetic and mechanically based sensors will have fault characteristics very different from those of biological and chemical sensors; • Also, sensors deployed in harsh environments, such as nuclear plants, will have very different characteristics from those of sensors deployed in friendly environments, such as offices or residential areas;

33. Smart Sensors and Sensor Networks • In the VLSI domain, only simple fault models, such as stuck at one and stuck at zero were extensively been used; • In addition to these models, more complex models are needed for sensors; in particular, the fault models for biological and chemical sensors that can be used only once for reading are very complex; • In SNs, testing needs to be addressed not only on component and individual node levels, but also at network and distributed system levels; • Closely related to testing is calibration, which can be defined as the process of mapping row sensor data to a new set of data that is, according to more statistical measure, more accurate than the initial readings; • Calibration can be done offline or online; • In the former case, the emphasis will be on the accuracy and strict interval of confidence; in the latter, the focus will be on the localized mode of operation; • Calibration, not only for sensor readings but also for other parameters relevant to operation of SNs, including timing and the available energy level at each node, should be conducted;

34. Smart Sensors and Sensor Networks • At the application level, fault resiliency mechanisms required for common applications such as sensor fusion, data filtering and data aggregation will be of primary importance; • For each of these applications, a variety of approaches can be used; for example, in addition to equation-based sensor fusion, sensor fusion based on graphs, statistics and stochastic will be possible; • Each of these techniques has a number of unique peculiarities; fault resiliency techniques that can be applied to multiple classes of approaches will be developed; • Fault tolerance is interconnected with other fields; • One of the main constraints in the deployment and operation of WSNs is energy; • The most effective way to prolong the lifetime of the network is to place a subset of nodes in sleep mode; • For example, power consumption of the software radio in three operational modes (transmission, receiving and idle) rarely differs for more than a factor of two;

35. Smart Sensors and Sensor Networks • At the same time, energy consumption in sleep mode is most often lower by two orders of magnitude or even more; • A simple and powerful observation is that a node in sleeping mode can be treated as faulty and vice versa; • It will be possible to retarget theoretical results and algorithms and even software for one objective to the other relatively easily; • Security and privacy are a major concern; • For example, a key question concerns the extend to which one can trust results obtained using sensor fusion or data aggregation in a particular scenario in a particular sensor network, assuming that one or more nodes are compromised; • In order to study this problem, it will be necessary to develop a threat model and models of attacks in one or more nodes in a sensor network; these attacks can be modeled as worst-case fault models; • Another problem is that any technique that is resilient against nonintentional faults could also be retargeted to intentional faults.