Machine learning-based arc fault detection and nuisance trip avoidance

Abstract

Machine learning-based arc fault detection and nuisance trip avoidance includes (i) obtaining data representing properties of an electrical signal sensed by an arc-fault circuit interrupter, (ii) interpreting the properties of the electrical signal as indicating an arc fault, the interpreting being based on application of an initial artificial intelligence (AI) model, (iii) based on the interpreting, opening a switch of the arc-fault circuit interrupter to interrupt the conduction of a supply of power to a load output terminal, (iv) receiving an indication that the properties of the electrical signal do not reflect an arc fault, such indication being that the arc-fault circuit interrupter is to refrain from arc-fault-based opening of the switch, (v) obtaining an updated AI model, where the updated AI model undergoes training using the data representing the properties of the electrical signal as an example of the absence of an arc fault, and (vi) deploying the updated AI model in place of the initial AI model.

Description

MACHINE LEARNING-BASED ARC FAULT DETECTION AND NUISANCE TRIP AVOIDANCEBACKGROUND

[0001] Electrical loads (e.g., lights, motors, appliances) are connected to an Alternating Current (AC) power source in various manners, such as by hardwiring to branch circuit conductors or by connecting via plugs to electrical receptacles. When an electrical load is energized (e.g., operates, is turned ON), current is conveyed to the load through a plurality of conductors, such as phase and neutral wires. An electrical arc can occur when electrical current flows improperly outside of the conductors.Such an arcing condition is commonly referred to as an ‘arc fault’. Some causes of arc faults include faulty connections due to corrosion or faulty installation. There are varying types of arc faults. Two such types are ‘series’ arc faults and ‘parallel’ arc faults. A series arc fault is an arc that occurs in-series with the load and, thus, the arc has a current that is no greater than the load current. In contrast, a parallel arc fault is an arc that occurs between different conductive paths, for instance any pairing of phase and neutral conductive paths, phase and ground conductive paths, or neutral and ground conductive paths. The current in a parallel arc can exceed the load current and can potentially be much greater than the load current. Both series and parallel arc faults can potentially ignite surrounding combustible materials.SUMMARY

[0002] Shortcomings of the prior art are overcome and additional advantages are provided through the provision of a method. The method includes obtaining data representing properties of an electrical signal sensed by an arc-fault circuit interrupter, interpreting the properties of the electrical signal as indicating an arc fault, the interpreting being based on application of an initial artificial intelligence (Al) model, based on the interpreting, opening a switch of the arc-fault circuit interrupter to interrupt the conduction of a supply of power to a load output terminal, receiving an indication that the properties of the electrical signal do not reflect an arc fault, such indication being that the arc-fault circuit interrupter is to refrain from arc-fault-based opening of the switch, obtaining an updated Al model, wherein the updated Al model undergoes training using the data representing the properties of the electrical signal asPA-03004 ORIG WO Page 1 of 32an example of the absence of an arc fault; and deploying the updated Al model in place of the initial Al model.

[0003] In one or more embodiments, the method is performed by the arc-fault circuit interrupter, and the arc-fault circuit interrupter further performs: sensing the properties of the electrical signal to obtain the data representing the properties of the electrical signal, wherein the sensing uses a sensor of the arc-fault circuit interrupter, and the data represents properties of the electrical signal immediately prior to switch opening, storing the data representing the properties of the electrical signal, and providing the stored data for the training.

[0004] In one or more embodiments, the training includes retraining the initial Al model using the data representing the properties of the electrical signal, the retraining providing the updated Al model.

[0005] In one or more embodiments, the initial Al model is of a first model architecture, and the obtaining the updated Al model includes evaluating one or more alternative model architectures, different than the first model architecture, for model fit or performance with the data representing the properties of the electrical signal. In one or more embodiments, the evaluating identifies a second model architecture, of the one or more alternative model architectures, for which model fit or performance with the data representing the properties of the electrical signal is better than the first model architecture, and the obtaining the updated Al model further includes building a new Al model of the second model architecture, training the new Al model using the data representing the properties of the electrical signal, and providing the trained new Al model as the updated Al model.

[0006] In one or more embodiments, the method further includes pre-processing the data representing the properties of the electrical signal for use in the training, the pre-processing including verifying and filtering data values provided by sensors of the arc-fault circuit interrupter.

[0007] In one or more embodiments, the method further includes providing the data representing the properties of the electrical signal as a profile of a harmless arcing condition in a database of profiles of arc fault arcing conditions and profiles ofPA-03004 ORIG WO Page 2 of 32harmless arcing conditions, where the obtaining the updated Al model uses the database in the training to provide the updated Al model.

[0008] In one or more embodiments, the indication that the properties of the electrical signal do not reflect the arc fault is provided by a user input. In one or more embodiments, The user input is implemented by one or more physical inputs of the arc-fault circuit interrupter, or the user input is a trigger provided over a wired or wireless communication path.

[0009] In one or more embodiments, the properties of the electrical signal include properties of at least one of voltage or current of the electrical signal.

[0010] In one or more embodiments, the method further includes performing subsequent interpretation of electrical signal properties based on application of the updated Al model to provide arc-fault protection for the electrical circuit.

[0011] The varying embodiments presented herein may be separable / optional from each other.

[0012] Additional aspects of the present disclosure are directed to systems and computer program products configured to perform the methods described above and herein. The present summary is not intended to illustrate each aspect of, every implementation of, and / or every embodiment of the present disclosure. Additional features and advantages are realized through the concepts described herein.BRIEF DESCRIPTION OF THE DRAWINGS

[0013] By way of example, a specific embodiment of the disclosed device will now be described, with reference to the accompanying drawings, in which:

[0014] FIG. 1 depicts an example schematic representation of an environment to incorporate and / or use aspects described herein;

[0015] FIG. 2 depicts an example arc-fault circuit interrupter to incorporate and / or use aspects described herein;

[0016] FIG. 3 depicts an example process for nuisance trip avoidance in accordance with aspects disclosed herein; andPA-03004 ORIG WO Page 3 of 32

[0017] FIG. 4 depicts an example process for machine learning-based arc fault detection and nuisance trip avoidance, in accordance with aspects described herein.DETAILED DESCRIPTION

[0018] Described herein, in some aspects, are various embodiments of methods and devices for arc-fault circuit interruption, including arc-fault circuit interrupters (AFCI) to mitigate against the occurrence of tripping based on false-positive indications of arc-fault conditions. Properties of an electrical signal (including patterns, for instance) sensed by an AFCI can be indicative of arcing conditions. Some arcing conditions are normal, expected, and harmless, though they might be falsely interpreted as indicating an arc fault, which represents a hazard. An arc fault (also referred to herein as an ‘arc fault condition’) is an arcing condition in which potentially hazardous arcing occurs. This type of arcing condition is to be detected, and a switch is to be opened to interrupt the circuit. This is in contrast to a harmless / non-hazardous arcing condition that is not considered to be one in which to trigger an arc-fault-based trip because the arcing condition is not considered hazardous. So-called nuisance tripping in the context of arc-fault circuit interruption includes tripping of the circuit based on a signal (e.g., properties / pattems of the electrical signal) that is confirmed to be either a signal resulting from an arcing condition that is a harmless arcing condition, or a signal that is not actually arcing at all (a non-arcing signal). In other words, a nuisance trip occurs when the properties of the circuit’s electrical signal at the time of the trip were interpreted as indicating an arc fault, even though no arc fault was present. When it is determined that the properties reflect a harmless, faultless arcing condition, that is they do not reflect an arc fault, the AFCI is to avoid / refrain from arc-fault-based opening of the switch. Avoiding tripping based on an arc fault does not mean that no trip should ever occur at that time; there are various other types of tripping, and even though a harmless arcing condition may be present, another type of trip condition may or may not be applicable and inform a legitimate trip at that time.

[0019] Arcing conditions that are not arc faults may be observed when appliances, motors, and other types of loads draw power from the circuit, triggering the trip mechanism (or other type of interrupter) of the AFCI to break the circuit (be it a branch circuit or other circuit protected by the AFCI device). One such example ofPA-03004 ORIG WO Page 4 of 32resolution of this nuisance tripping includes manual recalibration of the AFCI for it to accept the harmless signal (e.g., appliance signal) that presents as a harmless arcing condition, and avoid triggering the breaker mechanism. Such manual aspect of recalibration may involve long wait times for the customers where they must ship the sample (AFCI device) back to the manufacturer and wait for it to be recalibrated and sent back.

[0020] FIG. 1 depicts an example schematic representation of an environment to incorporate and / or use aspects described herein. The environment includes an electrical circuit 102 having phase ( ) and neutral (N) connect! ons / conductors for supplying electrical power from a power source 104 (such as an electrical grid for instance). In this example, the electrical power flows through a load center and breakers (112), an Arc-fault circuit interrupter (AFCI) 106, and receptacle(s) 114 to load(s) 108. It should be understood that various embodiments are possible relative to the AFCI protection offered by AFCI 106. For instance, and as shown in FIG. 1, the AFCI 106 could sit between a load center or breakers thereof and receptacles into which loads plug. Alternatively, the AFCI 106 could be provided in / by a breaker of the load center 112 or a receptacle 114.

[0021] AFCI 106 protects circuit 102 against dangerous arc conditions (e.g., arc faults) by sensing properties of the electrical signal on the circuit and interrupting the circuit (tripping / breaking) by opening a switch of the AFCI. Opening the switch interrupts the conduction of the supply of power to a load output terminal of the AFCI, and therefore to load(s) 108, which are devices that sit on the circuit and take power therefrom.

[0022] As noted, the AFCI can take any of varying forms. In one example, the AFCI is implemented in a circuit breaker of load center 112. The circuit breaker provides protection for a branch circuit to which load(s) 108 connect. A load might, for example, plug into a receptacle 114 or might be hardwired to the circuit. The AFCI will trip the circuit breaker to open a switch of the breaker if an arc-fault condition is observed. In another example, the AFCI is included in a receptacle 112 into which one or more loads, such as appliances or other devices, are plugged. These are just two example embodiments of the AFCI; others exist.PA-03004 ORIG WO Page 5 of 32

[0023] The AFCI may optionally be in communication with other device(s) 110 over communication path(s) 112, for instance via a network interface of the AFCI or device into which the AFCI is provided. The other devices could be located anywhere. For instance, they could be or include devices co-located with the AFCI and in proximity with the circuit, such as other circuit breakers of a load panel, a local whole-home energy monitoring device, a server, or other device that is onsite or proximate the circuit 102 and AFCI 106. Additionally or alternatively, other devices could be or include remote devices, such as those that are not co-located with the circuit, for instance one or more cloud devices. Communication path(s) 112 can be any wired and / or wireless communications paths / links across which the AFCI communicates with one or more other devices. The path(s) 112 can accordingly include one or more networks, such as a local area network (LAN), a wide area network (WAN), and / or a public network (e.g., the Internet). The path(s) could be implemented using any wired and / or wireless standards, protocols, hardware, software, or the like. Accordingly, a remote device could be cloud or other network- accessible devices, such as systems accessible to the AFCI via the internet.

[0024] In some examples, the path(s) 112 are, or include, wired or wireless connections to other devices. Example wireless connections include cellular, Wi-Fi, Bluetooth®, proximity-based, near-field, or other types of wireless connections. More generally, and without limiting aspects described herein, communications path(s) may be any appropriate wireless and / or wired communication link(s), devices, or the like, for communicating data.

[0025] Arc-fault protection can include various functions, such as sensing electrical properties of the circuit, analyzing / interpreting data generated based on that sensing, making a determination as to whether to open a switch (and potentially triggering the opening of the switch), also referred to as tripping, and performing the action to open the switch. Sensing the electrical properties may be accomplished by leveraging sensor(s), which may be sensor(s) of the AFCI or other sensors which sense the properties and generate / produce data representing those properties. In some examples, the data includes sampled / sensed properties, such as current and / or voltage, sensed at various times and presented as time domain data (sampled values as a function of time). The data can undergo processing, such as cleaning, filtering, augmenting or other processing (examples of which are described in further detailPA-03004 ORIG WO Page 6 of 32below), and then a determination can be made as to whether to open the switch on account of an apparent arc fault.

[0026] The AFCI can perform all or just some of these functions. In some embodiments, the AFCI performs the sensing, interpreting of the data, determining to open the switch, and actually opening the switch. In other embodiments, the AFCI performs the sensing to generate the data representing the electrical properties and then sends relevant data to other device(s) for processing, analyzing, interpreting, etc.. Then based on the results of that processing, a determination whether to trip could be made. The determination could be made by the AFCI based on receiving results from another device, or could be made by the other device itself. In examples, the other device opens the switch by indicating / communicating to the AFCI that an arc fault is present or by explicitly instructing the AFCI to open the switch, and the AFCI opens the switch accordingly. In this manner, the actions to interpret the data, determine whether to interrupt the circuit, and open the switch directly or by communicating to another device depending on which device has the switch, may be performed solely by the AFCI, or may be performed in whole or in part by other device(s).

[0027] In some embodiments, artificial intelligence (Al) models, for instance those based on machine learning (ML) or other types of Al models, facilitate a determination of whether an arc-fault. For instance, an Al model can be used to determine whether an arcing condition, if detected, is an arc fault or is instead a non- hazardous arcing condition. In specific examples, a determination is made based on applying a ML model in real-time or near-real time (as the electrical properties are sensed). In these embodiments, the AFCI or other device periodically or aperiodically applies an Al model to observed properties of the electrical signal. For instance, the electrical properties of the circuit may be sensed periodically to generate new data that is input to the Al model. The model provides one or more outputs, such as prediction(s), probabilities, or other results, and a determination can be made based on this output. In some examples, data representing properties of an electrical signal may need to be processed into a form suitable for inputting to the model. An example of Al-model-based arc fault detection is disclosed by patent application Ser. No.17 / 778,524, entitled “Arc Fault Detection Using Machine Learning” filed on Aug. 25, 2022, by inventor Yazhou Zhang et al., and is hereby incorporated by reference herein in its entirety.PA-03004 ORIG WO Page 7 of 32

[0028] FIG. 2 depicts an example arc-fault circuit interrupter to incorporate and / or use aspects described herein. The AFCI in this example stores and applies an Al model to use in determining whether to trip. As noted, the AFCI could take any of various forms, for instance as part of a circuit breaker, a receptacle, an inline device (for instance in the plug / connector of a load), and others.

[0029] Referring to FIG. 2, the disclosed AFCI 200 includes a controller 202 and memory 204 (which may be in the form of one or more memories) that stores firmware 212 and machine learning model 214, both which may be retained in memory 204 and loaded for execution by the controller 202. AFCI 200 also includes interrupter 206, user input device(s) 208, indicator 210, sensors 216, and network interface 218. AFCI 200 is in communication with a remote entity 240 across network(s) 230, such as the internet, by way of network interface 218 of AFCI 200 and communication path(s) 242.

[0030] Controller 202 may include any type of computing device, computational circuit, or any type of processor or processing circuit capable of executing a series of instructions that are stored in a memory, for instance memory 204. The controller may include multiple processors and / or multicore central processing units (CPUs) and may include any type of processor, such as a microprocessor, digital signal processor, microcontroller, programmable logic device (PLD), field programmable gate array (FPGA), or the like. The controller 202 may also itself include a memory (e.g., a cache) to store data and / or instructions that, when executed by one or more processors, causes the one or more processors to perform one or more methods and / or algorithms.

[0031] Examples of a memory (e.g., memory 204, memory of controller 202, main or system memory, or any other memory of AFCI 200) that may be a non- transitory computer readable or machine-readable storage medium, may include any tangible media capable of storing electronic data, including volatile memory or nonvolatile memory, removable or non-removable memory, erasable or non-erasable memory, writeable or re-writeable memory, and so forth. The memory may be used in the execution of program instructions. Non-limiting examples of computer executable instructions may include any suitable type of computer program having a set (e.g., at least one) of program modules, code, such as source code, compiled code, interpretedPA-03004 ORIG WO Page 8 of 32code, executable code, static code, dynamic code, object-oriented code, visual code, firmware and the like that is / are configured to carry out functions of embodiments described herein when executed by one or more processors. The memory may be one or more memory chips capable of storing data and allowing any storage location to be directly accessed by the processor such as any type or variant of random-access memory (RAM), Static random-access memory (SRAM), Dynamic random-access memory (DRAM), electrically erasable programmable read-only memory (EEPROM), Ferroelectric RAM (FRAM), Flash, NAND Flash, and NOR Flash and Solid State Drives (SSD). Other storage device(s) include optical fiber, a portable computer disk / diskette, such as a compact disc read-only memory (CD-ROM) or Digital Versatile Disc (DVD), a magnetic storage device, hard drive(s), flash media, or optical media / storage devices as examples, and / or cache memory, as examples, or any combination of the foregoing. Memory can include, for instance, a cache, such as a shared cache, which may be coupled to local caches (examples include LI cache, L2 cache, etc.) of processor(s).

[0032] Interrupter 206 includes any type of switch or trip device for interrupting the supply of power on the circuit. In examples, the interrupter 206 includes a switch between two or more contacts / terminals, where opening (tripping) the switch interrupts the supply of power. The interrupter 206 may comprise various hardware elements. In some examples, the interrupter 206 includes a solenoid and / or an energy storage element to trip the solenoid and cause the line-side connection to decouple from load-side connection. In further examples, the interrupter 206 may include a solid-state switching device (e.g., one or more transistors, thyristors, relays, or other solid state switching devices to interrupt current). Examples of transistors include Bipolar Junction Transistors (BJT’s), Insulated-Gate Bipolar Transistors (IGBT’s), and Metal-Oxide-Semiconductor Field-Effect Transistors (MOSFET’s). Examples of thyristors include Gate Turn-Off Thyristor (GTO), Silicon-controlled Rectifiers (SCRs), and Triodes for Alternating Current (TRIACs). Examples of relays include electromechanical relays and solid-state relays.

[0033] User input device(s) 208 include any inputs which a user can provide input to the AFCI 200. Examples of user input devices 208 include physical devices / hardware / actuators such as switches, toggles, push-buttons, touch input devices, or the like.PA-03004 ORIG WO Page 9 of 32

[0034] Indicator 210 includes any sensory (audio, visual, haptic, etc.) device for providing indications to a user, common examples of which are light emitting diodes (LEDs) or other lights, displays, buzzers, and speakers.

[0035] Sensor(s) 216 could be any type of sensors, for instance sensor(s) to sense current, voltage, temperature, humidity, light, proximity, resonant frequency motion, or other conditions or properties. More specific examples of such sensors are differential transformer / coils, current transformer / coils, shunts, high frequency sensors, low frequency sensors, and Rogowski coils, though others are possible.

[0036] Network interface 218 may be a wired interface, a wireless interface, or a combination thereof. The network interface 218 may provide an interface through which the AFCI 200 communicates via wired and / or wireless communication with other entities, for instance remote entity 240 via one or more intervening devices (routers, gateways, access points, etc.) and networks 230, such as a local area network (LAN) and / or a wide area network (WAN), such as the internet. Communications paths 242 may be any wired and / or wireless paths for communications, as explained herein.

[0037] Examples of a wired network interface include an ethemet interface (local area network, LAN), a fiber interface, a Recommended Standard 232 (RS-232) interface, or a universal serial bus (USB) interface, as examples.

[0038] A wireless interface is a network interface controller for communicating wirelessly via a wireless local area network (WLAN), personal area network (PAN), near field communication (NFC), low-power mesh (e.g., Thread, Z-Wave), ultranarrow band (Sigfox®, a trademark of UnaBiz SAS), a Low Power, Wide Area networking protocol (LoRaWAN™, a trademark of the LoRa Alliance), or cellular, as examples. An example of a WLAN is the Institute of Electrical and Electronics Engineers (IEEE) 802.11 standards (e.g., Wi-Fi®, a trademark of the Wi-Fi Alliance). Examples of PANs are Bluetooth® (a trademark of Bluetooth SIG, Inc.), Bluetooth® Low Energy (BLE), Bluetooth Smart, IPv6 over Low-Power Wireless Personal Area Networks (6L0WPAN), and Zigbee (a trademark of the Connectivity Standards Alliance). Examples of cellular networks are fourth generation (4G), long-term evolution (LTE), and fifth generation (5G) cellular networks.PA-03004 ORIG WO Page 10 of 32

[0039] Additionally, network interface 218 may be configured to employ any suitable network topology such as client / server, mesh, or peer-to-peer (P2P). The use of the term wireless is not intended to be limiting so as to exclude devices which also employ wired communication or are coupled to wires or physical conductors / conductive paths.

[0040] Examples of remote entities 240 are cellular phones, laptops, computers, servers, and any other form of computer systems / devices, external networks (such as the internet), or any services or offerings thereon, such as cloud services, for instance. Network interface 218 may receive information from a remote entity 240 through a transceiver, in examples.

[0041] Firmware 212 includes firmware for the AFCI, for instance instruct! ons / program code embedded in the device’s hardware to control core functionality of the AFCI 200. It may be used by the controller 202 of AFCI 200 and may be modified (e.g., calibrated, recalibrated, updated) by one or more processes. In some examples, it is modified to facilitate minimizing nuisance tripping as described herein. In some examples, the firmware could be periodically or aperiodically updated by a third-party ‘over-the-air’ by sending a firmware update from a remote entity to the AFCI 200 for storing (‘flashing’) to update the AFCI device 200. In other examples, firmware of the device 200 can be modified / provided by plugging another device into the AFCI and transferring firmware files from the other device to the AFCI.

[0042] Also provided in memory 204 is machine learning model 214, for instance one applied by the AFCI 200 as described above to determine whether an arcing condition exists and / or whether an arcing condition is an arc fault. The AFCI 200 may then take appropriate action based on applying the model.

[0043] Aspects described herein provide intelligent recalibration of an AFCI device through updates to a machine learning (ML) model used in arc fault detection processing. In some examples, the updating may be performed wholly or partly onboard, meaning on the AFCI device, such as the breaker or receptacle, as examples. In some embodiments, machine learning-based arc fault detection is provided, and the AFCI device is conveniently recalibrated for nuisance trip avoidance. For instance, a process can reconfigure / retrain an existing ML model and / or train a new ML modelPA-03004 ORIG WO Page 11 of 32based on confirming that a pattern represents an arcing condition that is not an arc fault, in which the pattern is used as an example set of properties representing a harmless arcing condition. The pattern may be used in training to configure the model to properly identify the pattern (any patterns that the Al model might classify the same or similarly) as an example of the absence of an arc fault, i.e. a harmless arcing condition. In examples, a pattern that indicates an arcing condition that is not an arc fault, also referred to herein as a nuisance trip signal, is added to a database, such as a database maintained on-board or maintained remotely. Aspects can provide data representing properties of electrical signals as profiles in a database of profiles. The profiles can be profiles of arc fault arcing conditions and profiles of harmless arcing conditions. A process can use the database in retraining an existing ML model and / or in training a new ML model of a different algorithm / architecture to avoid future nuisance tripping in a breaker mechanism, for instance. In some examples, as described herein, automatic steps may be initiated by a user button input (or other input) to a controller / microcontroller in the device.

[0044] As noted, training / retraining of an Al model could be performed on or by the AFCI, as is provided in some examples described herein, or elsewhere, for instance on or by one or more other local or remote devices. For instance, in some examples the Al model is built and trained on a cloud system. The model could be trained by a device that physically or wirelessly connects to the AFCI, such as a plug that plugs into a receptacle-based AFCI or a computer system or other device that connects to a breaker-based AFCI, as examples. Many variations are possible.

[0045] If the AFCI is to apply the Al model, then the model can be retrieved by or pushed to the AFCI device for application of the model on the AFCI. In other examples, the model is applied by a device other than the AFCI, and this other device could be the same or different than the device(s) that perform model building / training.

[0046] One aspect of nuisance trip avoidance in accordance with the present disclosure is receipt of an indication that properties of an electrical signal as sensed by the AFCI and that were initially interpreted as indicating an arc fault in fact do not reflect an arc fault and therefore that the AFCI is to avoid / refrain from arc-fault-based opening of the switch. In this regard, some embodiments include an on-board process (e.g., performed locally by the AFCI), optionally without any third-party interventionPA-03004 ORIG WO Page 12 of 32or remote connectivity, in order to receive and process this indication. In some embodiments, the on-board process is performed with or without user intervention, and may be independent of whether the AFCI is also the device applying the Al model for arc-fault detection. As explained herein, the indication that the properties do not reflect an arc fault can cause the process to collect the data representing those properties and / or store them to a database. In examples, after a trip event in which properties were interpreted as an arc fault, the data representing the properties of the electrical signal are stored at least temporarily. If at some point thereafter, and perhaps within a given timeframe or other parameter, an indication is received that it was not an arc fault and therefore the AFCI is to avoid opening of the switch based on a (false) interpretation of that signal reflecting an arc fault, then the data can be regarded as new training data and added to a training database (optionally with some preprocessing thereof, such as filtering and cleaning). From there, the data in the database can be used for training, which includes (re)training an existing model or training a new model. All or some of the above may be performed as part of the onboard process on the AFCI.

[0047] In this manner, nuisance trip signatures and any other arc or other signatures (e.g., signatures indicative of conditions that are not faulting conditions) may be maintained and used for future training, be it immediately or at a later time. This can mitigate future occurrences of nuisance tripping.

[0048] In one implementation, an on-board process is initiated by a user input. Examples of such user input are the user pressing a button on the AFCI or a user issuing a command via an application running on a mobile or other computing device that is in communication with the AFCI. Other embodiments include the user using a mobile device to scan a tag, such as a near-field communication (NFC) or barcode tag, to initiate the on-board process. In further embodiments, a process for sample retention and retraining can be initiated automatically based on criteria (a trip event) or periodically without any user intervention. For example, a cloud system might maintain a more sophisticated Al model than the model deployed on the AFCI. The more sophisticated model may be trained based on additional and / or more robust arc profiles than the Al model on the AFCI, for instance. The two Al models might therefore perform slightly differently from each other. It might be desired to run the more sophisticated Al model through training based on the data representingPA-03004 ORIG WO Page 13 of 32properties of the electrical signal (under trip and / or non-tripping scenarios). In this arrangement, there is a model used in real-time at a client location (the AFCI for instance) and then a more robust background model also used, perhaps in real-time or perhaps not. The more robust model could eventually be for replacing the deployed model and / or could be used as a backup / b ackground model on which to call upon in cases where the deployed model does not provide a sufficiently confident determination. In any case, an on-board process of the AFCI can include interfacing with, including providing data to, and receiving results of processing performed by, a remote entity.

[0049] Thus, the on-board process may provide for a convenient and quick resolution for nuisance tripping. In addition, the on-board process may allow for incremental improvements, via updated training of an Al model, to avoid further nuisance trips, for instance a trip on a branch circuit interrupted by the AFCI. Aspects can reduce or eliminate an alternative in which samples (for instance the AFCI devices) are sent back and forth for recalibration, and instead allow the user to fix a confirmed nuisance trip in a fraction of the time. In addition, the updated model can avoid similar nuisance trip events from occurring in the future.

[0050] In example operation of the AFCI, the AFCI obtains data representing properties of an electrical signal sensed by the AFCI, and interprets the properties onboard the AFCI and / or sends the data away for interpretation. In any case, the properties are interpreted by the AFCI and / or another device as indicating an arc fault, and the interpretation is based on application of one (or more) Al models. The AFCI (or the other device) then opens a switch. For instance, the AFCI opens a switch of the AFCI to interrupt the conduction of the supply of power to a load output terminal. Also upon the occurrence of the trip (e.g., a trip event), the AFCI stores trip event data in memory (e.g., memory 204). An example of trip event data is signals / measurements that are received / sensed by one or more sensors (e.g., differential transformer / coil, current transformer / coil, shunt, high frequency sensor, low frequency sensor, or Rogowski coil, as examples). The data includes, in some examples, time domain data, which is sampled / sensed data over time, for instance sensed current and / or voltage periodically or aperiodically sampled across a duration of time. In some examples, the data could include frequency domain data - for example data indicating frequency versus amplitude, response, or other variables.PA-03004 ORIG WO Page 14 of 32

[0051] The indication that a trip event is a nuisance trip (an example of arc-fault- based tripping in the absence of an actual arc fault) could be provided in any of various ways. One way a trip event may be designated as a nuisance trip event is by a user actuating a user input device (e.g., 208). Once this occurs, the on-board process may be initiated. In other words, this might be initiated by the user actuating the user input. During the on-board process, various processing of the trip event data can occur. For instance, the trip event data can be confirmed to have sensor data within upper and / or lower bounds and, if necessary, the data may be cleaned to be compliant with any application requirements for model input - that is, any constraints or other requirements for the data presented to the model as input. These data cleaning processes can include any of various pre-processing activities, such as those that are common to training data preparation. Examples include stemming, extrapolation / interpolation, replacement with average values, or targeted data removal, as examples. Additionally or alternatively, the data may pass through digital filters on the microcontroller (of the AFCI) or other processing circuit that includes programmable high-pass, low-pass and / or band-pass filters. Filtering can include setting one or more frequencies as threshold(s) internally on the integrated circuit as part of the AFCI. A low pass filter is one that retains values lower than some threshold and filters-out values at or above that threshold. A high pass filter is one that retains values higher than some threshold and filters-out values at or below that threshold. A band pass filter filters-out / ignores values lower than a first threshold and higher than a second threshold that is higher than the first threshold. Filters can be useful in situations where it is determined that true positives (arc faults) all always have values falling above or below certain points. Filters can help with efficiency, for instance reduction in downstream processing.

[0052] It is noted that any cleaning / filtering or other processing activity performed relative to the data captured for a confirmed harmless arcing condition (the properties do not reflect an arc fault) can be the same or different than cleaning / filtering being performed during the real-time detection of arc faults, as explained above.

[0053] Logged or captured trip event data could encompass a wide variety of data types from various sensors, detectors, or the like. The data that may be relevant for arc-fault detection could be just a subset of that captured data, and thereforePA-03004 ORIG WO Page 15 of 32cleaning / filtering may be particularly useful to reduce strain on computational and other types of resources.

[0054] For trip event data, such as nuisance trip data that reflects either nonarcing conditions or harmless arcing conditions (and therefore not arc faults), a process performs any desired processing of the data (cleaning, filtering, etc.) to put the data into a form suitable for re-training the model and / or training a new model of a same or different architecture. The machine learning model may be, or employ, any of various types of model s / architectures, such as statistical feature-based models, support vector machine models, deep learning neural network models, or tree-based classification models. In example retraining, the trip event data may be used to recalculate model coefficients or adjust decision weights, as examples, and thus retrain the existing model. Alternatively or in addition, the software performing model building / training may concurrently search for alternate optimized model architecture(s) to check for better fit / performance, for instance efficiency, speed, and / or accuracy in correctly distinguishing between arc-fault conditions and harmless arcing conditions, with the new nuisance trip data. If a better alternate architecture is found, then a new machine learning model based on the alternative architecture may be built and trained.

[0055] Al models can provide different types of outputs and / or outputs that are to be interpreted in different ways. Some Al models might output a positive or negative indication of an arc fault. Other Al models might provide an output that is to be interpreted in order to determine whether an arc fault is present. For example, neural networks tend to output probabilities for each of one or more classifications of the input data. There may therefore be an additional component that uses the model output to make a determination as to whether the inputs reflect an arc fault. That additional component might rely on some parameters / thresholds. Training or retraining may include adjustment of such parameters / thresholds. For instance, there may be a parameter that requires the confidence of the Al model’s prediction of an arc fault to be 78% or higher in order to interpret the inputs as reflecting an arc fault and open the switch. Such parameter is not a parameter of the model but is a parameter of the determination whether to open the switch. The training based on captured nuisance trip event data may include modification to this parameter, in examples.PA-03004 ORIG WO Page 16 of 32

[0056] A trained model, whether it be a retrained existing model or a trained new model, can undergo a performance verification process. The performance verification process begins with a first round of the process - an Internal Performance evaluation, which may be performed on the AFCI itself or whatever device applied the model. This evaluation is to ensure that evaluation metrics fall within an appropriately tight range of acceptable performance, for instance one defined by the manufacturer’s requirements. In general, this may be an assessment of how well the model performs. An aspect of this may be achieving a desired accuracy in classifying inputs, for instance an accuracy of 90% or greater. Often an initial labeled training dataset may be split into training data and testing data, where testing data is used to test this accuracy based on the training performed using the training data. Example evaluation techniques may include: k-fold cross-validation, holdout validation, bootstrapping, stratified sampling (dividing data into homogeneous subgroups before sampling), leave-p-out cross-validation (using p data points as the validation set and the remaining observations as the training set), receiver operating characteristic curve (or ROC which is a graphical plot that illustrates the performance of a binary classifier model at varying threshold values and is the plot of the true positive rate against the false positive rate at each threshold setting), F-score (e.g., Fl), Root Mean Squared Error (RMSE), and / or confusion matrix, as examples.

[0057] In k-fold cross-validation, the data points are randomly partitioned into k equal sized samples / folds. A single sample is then retained as the validation data for testing the model, and the remaining samples are used as training data. The cross- validation process is then repeated k times, with each of the k samples used exactly once as the validation data. The k results can then be averaged to produce a single estimation.

[0058] In the holdout method, data points are randomly assigned to two sets (e.g., one set as a training set and another set as the a test set). The model is then trained on the one set and tested / evaluated on the other set.

[0059] Bootstrapping estimates the distribution of an estimator by resampling the data or a model estimated from the data.

[0060] The Fl -score is a measure of predictive performance. It is calculated from the precision and recall of the test, where the precision is the number of true positivePA-03004 ORIG WO Page 17 of 32results divided by the number of all data points predicted to be positive, including those not identified correctly, and the recall is the number of true positive results divided by the number of all samples that should have been identified as positive. Precision is also known as positive predictive value, and recall is also known as sensitivity in diagnostic binary classification. The Fl score is the harmonic mean of the precision and recall. It thus symmetrically represents both precision and recall in one metric.

[0061] The root mean square error (RMSE) measures the average difference between a statistical model’s predicted values and the actual values. Mathematically, it is the standard deviation of the residuals. Residuals represent the distance between the regression line and the data points.

[0062] A confusion / error matrix is a specific table layout that allows visualization of the performance of an algorithm. Each row of the matrix represents the instances in an actual class while each column represents the instances in a predicted class, or vice versa. The diagonal of the matrix therefore represents all instances that are correctly predicted. The name stems from the fact that it makes it easy to see whether the system is confusing two classes (i.e. commonly mislabeling one as another).

[0063] The retrained machine learning model may then undergo a second round of the performance verification process, which includes blind testing / simulation. For instance, the retrained model is blind-tested using, for instance, comprehensive databases for both known arc or arc fault signatures (for instance arcs defined under the ULI 699 standard) and known signatures that are not arcs (e.g., noise) or arc faults. Such databases may be stored in the microcontroller’s (e.g., 202) flash memory or in external flash memory on the board or in a cloud database, as examples.

[0064] The retrained model may be accepted only if it meets satisfactory evaluation requirements during data simulation by ensuring the evaluation metrics fall within a tight range of acceptable performance as defined by the manufacturer’s requirements, for instance.

[0065] Once the training / retraining process is complete and the trained / retrained model has been accepted per the above process and placed into production on the AFCI or whatever device is applying the model for arc-fault circuit interruptionPA-03004 ORIG WO Page 18 of 32determinations, the user, for instance the owners of the AFCI or the user who initiated the on-board process to initiate the training / retraining to obtain the updated model, can be provided a notification / confirmation that the on-board process is complete and that the nuisance trip has been mitigated. In this manner, the model now placed into production has been trained to avoid / refrain from nuisance-tripping on the collection of signals that previously caused the nuisance trip. Examples of such indication are an indicator provided on the AFCI (e.g., LED, buzzer, speaker), a notification on a mobile device (e.g., push notification, text message), a web services message, AFCI tripping or clearing a reset lockout mechanism. The on-board process and training / retraining can be repeated as needed after any nuisance trip event identified by the user.

[0066] A user input to initiate the on-board process may be activated by any suitable user, such as a homeowner. Alternatively, or in addition, a user input may only be activated by trained personnel (e.g., an electrician or an authorized representative of the manufacturer). Restricting the activation of the user input can help to ensure that trip events are designated as nuisance trip events only after trained personnel have properly evaluated the safety of the electrical installation. In other words, it may not be desirable for a homeowner to be able to activate the user input to confirm a nuisance trip in the event of dangerous conditions actually being present. For example, trained personnel may first inspect the electrical load and / or wiring to ensure no dangerous conditions exist (e.g., nicks / damage to the wiring / load, charring, evidence of overhearing). Once the safety of the electrical installation is confirmed, the trained personnel may contact the AFCI manufacturer by phone, email, by an application on a mobile device, or use cloud services to be authenticated to initiate the on-board process. In examples, the product instruction sheet has information on how to contact the manufacturer for authentication to initiate the on-board process. Once authenticated, the on-board process for recalibrating the AFCI (e.g., retraining to model to provide an updated model) can be initiated by activating the user input as discussed previously. For example, if the user input is implemented by a physical button on the AFCI, the process could be initiated via the physical button. More specifically, once authenticated, the manufacturer may provide the trained personnel with a set of sequences using the button to initiate the on-board process. An example of such a sequence may be for the trained personnel to press the button five times inPA-03004 ORIG WO Page 19 of 32three seconds, wait one second, and then press the button twice more in rapid succession. Alternatively, the user input could be a particular combination entered on a DIP switch on the AFCI or a code entered in an application running on a mobile device. For additional security, such a code entered on a mobile device could be a time-based one-time password (TOTP).

[0067] Thus, in an example embodiment of aspects described herein, after a trip event, the AFCI stores sensor data in the microcontroller’s flash memory. In one example, the AFCI continually measures what is sensed and keeps the generated data in volatile storage, for instance in a circular buffer thereof. When the device trips, the data that was previously recorded (e.g., during a window of time immediately prior and leading up to the trip) can be maintained / stored in flash or other memory so that it is not lost. From there, the data could be transmitted to a remote entity, processed locally, etc.

[0068] The AFCI is equipped with a user input, for instance a mechanical push button or a wireless trigger such as one provided via an NFC tag, a barcode tag, Bluetooth communication, or Wi-Fi communication. Thus, once the user actuates this input after a confirmed nuisance trip from the branching circuit, the microcontroller of the AFCI can capture the latest trip data from flash or other memory and run it through an on-board recalibration program of (meaning software existing in and executing on) the AFCI, as an example. As noted, the data could be processed as described, for instance to remove noise, clearly erroneous data, or the like. The trip data could additionally be verified for accuracy. In any case, the data may be provided for retraining the existing model and / or training a new model to obtain an updated model. The trip data can be used in a retraining process, for instance to recalculate model coefficients and retrain the model. The machine learning model may be statistical feature-based model(s), support vector machine model(s), deep learning neural network model(s) or tree-based classification model(s), as examples. Once the retraining process is completed, one or more notifications may be provided, e.g., to a user. Example notifications include an LED indication, a mobile app notification, and / or an audible notification. In any case, the notification(s) provided can confirm the end of the nuisance trip resolution. The foregoing process can be repeated after any nuisance trip event, at the user’s discretion.PA-03004 ORIG WO Page 20 of 32

[0069] Referring to FIG. 3, an example process for nuisance trip avoidance as disclosed herein is shown. The process begins with, or upon occurrence of, an AFCI trip 302. In an example, the AFCI is, or is part of, a circuit breaker. The process continues with the user inspecting (304) the electrical installation, and confirming (306) the nuisance trip. In an example, the user is an approved operator, such as a licensed contractor or approved representative. In an example, the user is a homeowner. At this point, after load / wiring inspection (304) and establishing / confirming that the occurrence was a nuisance trip (306), the user can trigger the recalibration. As part of this, the user can authenticate and / or activate a user input (308). In the case of authentication, the user could contact a customer service of the device manufacturer, or use cloud services to authenticate and establish contractor / administrator intervention to initiate the recalibration feature. Then once authenticated, the recalibration program in the device’s microcontroller can be triggered, for instance using user input(s), such as a mechanical push button, Wireless trigger such as NFC tag, barcode tag, Bluetooth, or Wi-Fi command, as examples. In alternative embodiments, there is no authentication, and the user simply initiates the recalibration using a user input to the device, such as a button.

[0070] The process continues with on-board machine learning model recalibration (310) as described herein. This includes triggering and data collection, in which, when recalibration is triggered, the microcontroller collects from its flash memory the signal data that caused the nuisance tripping. Then, a retraining pipeline process is performed that includes verification and filtering, in which the signal data is verified for accuracy and filtered for retraining. The filtered signal data is fed as training data for obtaining an updated Al model, for instance by retraining a version of the initial machine learning model or training a new model, for instance one based on a different model architecture than the initial model. As noted, the machine learning model can be one of several types: statistical feature-based models, support vector machines, deep learning neural networks, or tree-based classification models. The process performs a recalculation or adjustment of the existing ML model, or an optimization to result in a new model. For instance, the model building / training software calculates model coefficients and adjusts decision weights based on the new training data. Optionally, the algorithm concurrently searches for alternate optimized model architectures (alternative to the architecture of the existing ML model) to check forPA-03004 ORIG WO Page 21 of 32better fit / performance with the new nuisance trip data, and if a better alternate architecture is found, then the model can switch to the new model architecture. For instance, a cloud offering could be leveraged that has pretrained models that take uploaded training data and process it to determine accuracy metrics for various model architectures and determining an optimized model / model type for the desired application. In any case, a process can retrain the existing model or training a new model of an optimized architecture. It is noted that training data could train on profiles that reflect any of varying types of conditions, including arcing and nonarcing conditions, harmless / non-hazardous arcing conditions that are not arc faults, and / or arcing conditions that reflect arc faults.

[0071] Performance verification can include an internal performance evaluation (as a first portion of the performance verification), in which the trained / retrained model undergoes an internal performance evaluation on-board (i.e., device, e.g. AFCI) for instance. As noted, evaluation techniques may include: K-fold cross- validation, holdout validation, bootstrapping, stratified sampling, leave-one-out cross- validation, ROC curve analysis. Fl -score, Root Mean Squared Error (RMSE / MSE), and / or general confusion matrix evaluation parameters, as examples. Then the performance verification includes blind testing / simulation (as round 2 of performance verification), in which the retrained model is blind-tested using comprehensive databases for true arcs / arc faults (such as those defined by the Underwriters Laboratories (UL) Standard for Arc Fault Circuit Interrupters - UL 1699) and various noise “no-arcs”. These databases may be stored in the microcontroller’s flash memory or in external flash memory on the board (e.g. circuit board of the AFCI) or at a remote entity, such as in a cloud database, and be accessible to the AFCI via a network connection through its network adapter. At this point, new machine learning model confirmation may be performed, in which the retrained model is accepted only if it meets evaluation requirements during data simulation. Confirmation methods may include visual indicators (e.g., LED), audible notifications, push notifications via mobile app / web services, physical actions (e.g., device tripping or clearing the reset lockout mechanism on the breaker), or others.

[0072] Thus, a disclosed AFCI avoids / minimizes nuisance tripping and distinguishes between (i) arcing conditions that are arc faults, which may bePA-03004 ORIG WO Page 22 of 32dangerous, for instance due to a damaged conductors, on which it is desired to trip, and (ii) arcing conditions that are harmless arcing conditions and examples of the absence of an arc fault, for instance those due to a fan motor or drill, on which faulting / tripping is to be avoided. Harmless arcing conditions could be experienced during normal operation of certain common electrical devices found in brushed motors, switching power supplies, ballasts, and dimmers, as examples.

[0073] An AFCI described herein can include a controller (such as a processor or processing circuit) and memory / storage, where the memory is to store instructions for execution by the controller to perform actions described herein. In this regard, the AFCI may be regarded as a computer system capable of executing program instructions. The AFCI may be in communication with external device(s).

[0074] An AFCI can include one or more processor(s), for instance central processing unit(s) (CPUs) and / or microprocessors. A processor can include functional components used in the execution of instructions, such as functional components to fetch program instructions from locations such as cache or main memory, decode program instructions, and execute program instructions, access memory for instruction execution, and write results of the executed instructions. A processor can also include register(s) to be used by one or more of the functional components. An AFCI also includes memory. In embodiments, an AFCI includes input / output (VO) devices, and VO interfaces, which may be coupled to the controller and each other via one or more buses and / or other connections. Bus connections represent one or more of any of several types of bus structures, including a memory bus or memory controller, a peripheral bus, an accelerated graphics port, and a processor or local bus using any of a variety of bus architectures.

[0075] Examples of VO devices include but are not limited to microphones, speakers, lights, buttons, switches, toggles, and sensor devices. An VO device may be incorporated into the AFCI, though in some embodiments an VO device may be regarded as an external device coupled to the AFCI through one or more VO interfaces.

[0076] An AFCI may communicate with one or more external devices via one or more VO interfaces. Example external devices include any devices that enable a user to interact with the AFCI. Other example external devices include any device thatPA-03004 ORIG WO Page 23 of 32enables the AFCI to communicate with one or more other systems or peripheral devices. A network interface / adapter is an example I / O interface that enables AFCI to communicate with one or more networks, such as a local area network (LAN), a general wide area network (WAN), and / or a public network (e.g., the Internet), providing communication with other computing devices or systems, storage devices, or the like. Ethernet-based (such as Wi-Fi) interfaces and Bluetooth® adapters are just examples of the currently available types of network adapters.

[0077] The communication between I / O interfaces and external devices can occur across wired and / or wireless communications link(s), such as Ethernet-based wired or wireless connections.

[0078] Particular external device(s) may include one or more data storage devices, which may store one or more programs, one or more computer readable program instructions, and / or data, etc. An AFCI may include and / or be coupled to and in communication with (e.g. as an external device of the computer system) removable / non-removable, volatile / non-volatile computer system storage media.

[0079] FIG. 4 depicts an example process for machine learning-based arc fault detection and nuisance trip avoidance, in accordance with aspects described herein. The process may be executed, in one or more examples, by a processor or processing circuitry of one or more devices, such as those described herein, and more specifically an AFCI and / or a computer system as described herein. In some examples, aspects are performed wholly by an AFCI, wholly by a device proximate an AFCI, wholly by a remote entity in communication with an AFCI or device proximate the AFCI, or a combination of the forgoing.

[0080] The process of FIG. 4 includes obtaining (402) data representing properties of an electrical signal sensed by an arc-fault circuit interrupter. In examples, the AFCI senses the properties of the electrical signal to obtain the data representing the properties of the electrical signal. The sensing uses a sensor of the arc-fault circuit interrupter, and the data represents properties of the electrical signal immediately prior to switch opening. In some examples, the properties of the electrical signal include properties of at least one of voltage or current of the electrical signal.PA-03004 ORIG WO Page 24 of 32

[0081] The process continues by interpreting (404) the properties of the electrical signal as indicating an arc fault. The interpreting is based on application of an initial artificial intelligence (Al) model. The process could optionally perform preprocessing of the data representing the properties of the electrical signal, for instance verifying and filtering data values provided by sensor(s) of the arc-fault circuit interrupter to provide the data to the model in an acceptable manner for inputting to the model.

[0082] Based on the interpreting, the process opens (406) a switch of the arc-fault circuit interrupter to interrupt the conduction of a supply of power to a load output terminal. In examples where the AFCI itself performs this aspect of the process, the AFCI opens the switch. In examples where another device performs this aspect of the process, the other device sends, and the AFCI receives, a communication which triggers the AFCI to open the switch.

[0083] The process continues by receiving (408) an indication that the properties of the electrical signal do not reflect an arc fault, the indication being that the arc-fault circuit interrupter is to avoid / refrain from arc-fault-based opening of the switch. The indication could be provided by or through any device, including the AFCI or another device. In examples, the indication that the properties of the electrical signal do not reflect the arc fault is provided by a user input. The user input is implemented by one or more physical inputs of the arc-fault circuit interrupter, or the user input is a trigger provided over a wired or wireless communication path, for example.

[0084] The process then obtains (410) an updated Al model, where the updated Al model undergoes training using the data representing the properties of the electrical signal as an example of the absence of an arc fault. The AFCI can store the data representing the properties of the electrical signal and provide the stored data for the retraining. Any desired pre-processing of the data representing the properties of the electrical signal can be performed for use in training the updated Al model. The pre-processing could include verifying and filtering data values provided by sensor(s) of the arc-fault circuit interrupter, for instance, to provide the data to the software / processing responsible for training. In some examples, the process provides the data representing the properties of the electrical signal as a profile of a harmless arcing condition in a database of profiles of arc fault arcing conditions and profiles ofPA-03004 ORIG WO Page 25 of 32harmless arcing conditions, and obtaining the updated Al model uses the database in the training to provide the updated Al model.

[0085] In some examples, the training includes retraining the initial Al model using the data representing the properties of the electrical signal, where the retraining provides the updated Al model. In addition, or in other examples, the initial Al model is of a first model architecture, and the obtaining the updated Al model includes evaluating one or more alternative model architectures, different than the first model architecture, for model fit or performance with the data representing the properties of the electrical signal. The initial model may be retained and continued in use, even when other architectures are evaluated. However, in some examples, the evaluating identifies a second model architecture, of the one or more alternative model architectures, for which model fit or performance with the data representing the properties of the electrical signal is better than the first model architecture, and the obtaining the updated Al model further includes building a new Al model of the second model architecture, training the new Al model using the data representing the properties of the electrical signal, and providing the trained new Al model as the updated Al model.

[0086] Returning to FIG. 4, the process deploys (412) the updated Al model in place of the initial Al model. In examples, the updated Al model is for use in subsequent interpretation of electrical signal properties to provide arc-fault protection for the electrical circuit. The deploying could be done by the device (such as the AFCI or another device) that applied the initial model replacing that initial model with the updated model, or could be an action by a remote device (such as one that trained the updated model) providing the updated model to the applying device. Thus, in this context deploying could be providing, sending, replacing, etc. the updated model on / to the device that is to apply the updated model for arc-fault circuit interruption.

[0087] Aspects disclosed herein may be a system, a method, and / or a computer program product, any of which may be configured to perform or facilitate aspects described herein.

[0088] In some embodiments, aspects may take the form of a computer program product, which may be embodied as computer readable medium(s). A computer readable medium may be a tangible storage device / medium having computer readablePA-03004 ORIG WO Page 26 of 32program code / instructions stored thereon. The computer readable medium may be readable by a processor, processing unit, or the like, to obtain data (e.g. instructions) from the medium for execution. In a particular example, a computer program product is or includes one or more computer readable media that includes / stores computer readable program code to provide and facilitate one or more aspects described herein.

[0089] As noted, program instruction contained or stored in / on a computer readable medium can be obtained and executed by any of various suitable components such as a processor of a system / device to cause the system / device to behave and function in a particular manner. Such program instructions for carrying out operations to perform, achieve, or facilitate aspects described herein may be written in, or compiled from code written in, any desired programming language. In some embodiments, such programming language includes object-oriented and / or procedural programming languages such as C, C++, C#, Java, etc.

[0090] Program code can include one or more program instructions obtained for execution by one or more processors. Computer program instructions may be provided to one or more processors of, e.g., one or more devices, to produce a machine, such that the program instructions, when executed by the one or more processors, perform, achieve, or facilitate aspects of the present invention, such as actions or functions described in flowcharts and / or block diagrams described herein. Thus, each block, or combinations of blocks, of the flowchart illustrations and / or block diagrams depicted and described herein can be implemented, in some embodiments, by computer program instructions.

[0091] Although various embodiments are described above, these are only examples. For example, computing environments of other architectures can be used to incorporate and use one or more embodiments.

[0092] This disclosure is intended to provide exemplary embodiments of the disclosed system and method, and these exemplary embodiments should not be interpreted as limiting. One of ordinary skill in the art will understand that the steps and methods disclosed may easily be reordered and manipulated into many configurations, provided they are not mutually exclusive.PA-03004 ORIG WO Page 27 of 32

[0093] The features disclosed in the foregoing description, or the accompanying drawings, expressed in their specific forms or in terms of a means for performing the disclosed function, or a method or process for attaining the disclosed result, as appropriate, may, separately, or in any combination of such features, be utilized for realizing the disclosure in diverse forms thereof.

[0094] Although various embodiments are described above, these are only examples.

[0095] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting. As used herein, the singular forms “a”, “an” and “the” are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms “comprises” and / or “comprising”, when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.

[0096] The corresponding structures, materials, acts, and equivalents of all means or step plus function elements in the claims below, if any, are intended to include any structure, material, or act for performing the function in combination with other claimed elements as specifically claimed. The description of one or more embodiments has been presented for purposes of illustration and description, but is not intended to be exhaustive or limited to in the form disclosed. Many modifications and variations will be apparent to those of ordinary skill in the art. The embodiment was chosen and described in order to best explain various aspects and the practical application, and to enable others of ordinary skill in the art to understand various embodiments with various modifications as are suited to the particular use contemplated.PA-03004 ORIG WO Page 28 of 32

Claims

CLAIMSWhat is claimed is:

1. A method including: obtaining data representing properties of an electrical signal sensed by an arc-fault circuit interrupter; interpreting the properties of the electrical signal as indicating an arc fault, the interpreting being based on application of an initial artificial intelligence (Al) model; based on the interpreting, opening a switch of the arc-fault circuit interrupter to interrupt the conduction of a supply of power to a load output terminal; receiving an indication that the properties of the electrical signal do not reflect an arc fault, such indication being that the arc-fault circuit interrupter is to refrain from arc-fault-based opening of the switch; obtaining an updated Al model, wherein the updated Al model undergoes training using the data representing the properties of the electrical signal as an example of the absence of an arc fault; and deploying the updated Al model in place of the initial Al model.

2. The method of claim 1, wherein the method is performed by the arcfault circuit interrupter, and wherein the arc-fault circuit interrupter further performs: sensing the properties of the electrical signal to obtain the data representing the properties of the electrical signal, wherein the sensing uses a sensor of the arc-fault circuit interrupter, and the data represents properties of the electrical signal immediately prior to switch opening; storing the data representing the properties of the electrical signal; and providing the stored data for the training.PA-03004 ORIG WO Page 29 of 323. The method of claim 1 or 2, wherein the training includes retraining the initial Al model using the data representing the properties of the electrical signal, the retraining providing the updated Al model.

4. The method of claim 1, 2, or 3, wherein the initial Al model is of a first model architecture, and wherein the obtaining the updated Al model includes evaluating one or more alternative model architectures, different than the first model architecture, for model fit or performance with the data representing the properties of the electrical signal.

5. The method of claim 4, wherein the evaluating identifies a second model architecture, of the one or more alternative model architectures, for which model fit or performance with the data representing the properties of the electrical signal is better than the first model architecture, and wherein the obtaining the updated Al model further includes building a new Al model of the second model architecture, training the new Al model using the data representing the properties of the electrical signal, and providing the trained new Al model as the updated Al model.

6. The method of claim 1, 2, 3, 4, or 5, further including pre-processing the data representing the properties of the electrical signal for use in the training, the pre-processing including verifying and filtering data values provided by sensors of the arc-fault circuit interrupter.

7. The method of claim 1, 2, 3, 4, 5, or 6, further including providing the data representing the properties of the electrical signal as a profile of a harmless arcing condition in a database of profiles of arc fault arcing conditions and profiles of harmless arcing conditions, wherein the obtaining the updated Al model uses the database in the training to provide the updated Al model.

8. The method of claim 1, 2, 3, 4, 5, 6, or 7, wherein the indication that the properties of the electrical signal do not reflect the arc fault is provided by a user input.

9. The method of claim 8, wherein the user input is implemented by one or more physical inputs of the arc-fault circuit interrupter, or the user input is a trigger provided over a wired or wireless communication path.PA-03004 ORIG WO Page 30 of 3210. The method of claim 1, 2, 3, 4, 5, 6, 7, 8, or 9, wherein the properties of the electrical signal include properties of at least one of voltage or current of the electrical signal.

11. The method of claim 1, 2, 3, 4, 5, 6, 7, 8, 9, or 10, further including performing subsequent interpretation of electrical signal properties based on application of the updated Al model to provide arc-fault protection for the electrical circuit.

12. An arc-fault circuit interrupter (AFCI) comprising: a memory; and a processing circuit in communication with the memory, wherein the AFCI is configured to perform a method according to any of claims 1-11.

13. A computer program product compri sing : a computer readable storage medium readable by a processing circuit and storing instructions for execution by the processing circuit to perform a method according to any of the claims 1-11.PA-03004 ORIG WO Page 31 of 32

Images