Resource-aware performance and communication-optimized federated learning at the edge in industrial settings

Abstract

Current approaches to federated learning are expensive, unaware of client process priorities and lack efficiencies. A system can optimize performance and communications for federated learning and perform enhanced resource orchestration in various industrial settings. The system can define a resource-aware collaborative model-update federated learning framework that can work with streaming data in heterogenous environments. For example, various models can learn to estimate execution times on various types of devices. Furthermore, embodiments can define a performance-estimate-driven client selection policy that minimizes the total energy consumption of selected devices in communication and computing processes, while maximizing the convergence speed of drift handling.

Description

202415815RESOURCE-AWARE PERFORMANCE AND COMMUNICATION-OPTIMIZED FEDERATED LEARNING AT THE EDGE IN INDUSTRIAL SETTINGSBACKGROUND

[0001] Industrial systems and applications can be used to control or monitor the operation of machines and other components in a systematic manner. Industrial systems can include various domains such as factory automation (e.g., industrial control systems, assembly line monitoring, predictive maintenance), process automation, building automation (e.g., occupancy counter, energy efficiency, anomaly detection), healthcare applications (e.g., remote robotic surgery, collaborative diagnosis) transportation applications (e.g., occupancy counter, aggression detection, positive train control), and the like. Such industrial applications can define machine learning components that are configured to perform various functions. Such machine learning components can experience degradation in accuracy over time, for example, due to change in input data. To overcome model staleness, continuous learning can be employed, where a given model is updated based on recent changes in the input data. In some cases, due to performance, cost, privacy, regulatory considerations, and the like, model updates can be performed in a federated setting, where a given model updates locally using on-premises edge devices and then synchronizes with a global model.

[0002] The increased computational power of edge devices helps model updates in federated settings. It is recognized herein, however, that co-locating a resource intensive task of model training can take away resources from existing latency sensitive tasks and can result in deadline violations, among other technical drawbacks. Moreover, the communication cost for performing machine learning in federated settings, also known as federated learning (FL), can be significant for resource constrained edge devices, especially when frequent model updates are warranted.BRIEF SUMMARY

[0003] Embodiments of the invention address and overcome one or more of the described- herein shortcomings or technical problems by providing methods, systems, and apparatuses that optimize communications for federated learning and perform enhanced resource orchestration in various industrial settings. Embodiments define a system state-aware collaborative model-update federated learning framework that can work with streaming data in heterogenous environments. For example, various models can learn to estimate execution times on various types of devices. Furthermore, embodiments can define a performance-202415815 estimate-driven client selection policy that minimizes the total energy consumption of selected devices in communication and computing processes, while maximizing the convergence speed of drift handling.

[0004] In an example aspect, an industrial system or network includes a central server communicatively coupled to a plurality of client devices. The server can send a request to the plurality of client devices. Responsive to the request, the server can obtain a plurality of system metrics for each client device of the plurality of client devices. Based on the plurality of system metrics, the server can select a set of client devices of the plurality of client devices, so as to define a selected set of client devices. The server can distribute a global model to the selected set of client devices, such that each client device in the selected set can perform a re-training on the global model. Based on the global model and the respective re-training, the server can receive respective updates to the global model from each of the client devices in the selected set of client devices. The server can aggregate the updates into the global model, so as to generate a new global model with federated learning.

[0005] In another example aspect, based on the plurality of metrics for each client device, the server can detect data drift associated with each client device, wherein the plurality of metrics includes respective loss information associated with each client device. Furthermore, the server can compare the respective loss information, so as to determine a group of client devices having loss information greater than other client devices of the plurality of client devices. The sever can select the group of client devices for the selected set of client devices that perform re-training. In some examples, the central server determines an execution time for each of the client devices of the plurality of client devices. The execution time corresponds to the re-training. The server can select the selected set of the client devices further based on the respective execution time associated with each client device.BRIEF DESCRIPTION OF THE DRAWINGS

[0006] The foregoing and other aspects of the present invention are best understood from the following detailed description when read in connection with the accompanying drawings. For the purpose of illustrating the invention, there is shown in the drawings embodiments that are presently preferred, it being understood, however, that the invention is not limited to the specific instrumentalities disclosed. Included in the drawings are the following Figures:202415815

[0007] FIG. 1 is a block diagram of an example industrial internet of things (I loT) system configured to perform federated learning, wherein the system includes a central server communicatively coupled to a plurality of edge or client devices.

[0008] FIG. 2 is a flow diagram that includes various operations that can be performed by the central server during federated learning, in accordance with an example embodiment.

[0009] FIG. 3 illustrates example operations that can be performed by the central server, in accordance with an example embodiment.

[0010] FIG. 4 illustrates example operations that can be performed by the plurality of edge or client devices, in accordance with an example embodiment.

[0011] FIG. 5 shows an example of a computing environment within which embodiments of the disclosure may be implemented.DETAILED DESCRIPTION

[0012] As an initial matter, it is recognized herein that current approaches to implementing federated learning in an industrial setting fail to address various technical challenges, such as system and communication cost challenges that arise due to the heterogeneity of devices and the execution of co-located processes. It is further recognized herein that current approaches typically consider energy consumption or convergence time in federated learning systems in isolation, such that current approaches lack the ability to control both convergence time and energy consumption, among other technical shortcomings. For example, some current approaches to implementing federated reinforcement learning use CPU-cycle frequency to estimate the execution time for solving the problem of energy saving and convergence rate. Such approaches, however, only focus on offline environments instead of heterogeneous incoming data, resulting in the decrease of the accuracy of drift data.

[0013] By way of further background, federated learning (FL) refers to distributed machine learning in which a global model is trained collaboratively by multiple client devices, without the need to share the raw data from the client devices with a central server. The client devices can train the model on their local data and only share model updates with the central sever. The central server can aggregate the updates to improve the global model. Drift data in federated learning generally refers to gradual changes or shifts that can occur in the data distributions across the client devices that participate in the federated learning process. The presence of drift data can lead to performance degradation of the global model because the global model might not be able to generalize well to the changing data distributions on the client devices.202415815

[0014] Embodiments described herein define an optimization problem for federated learning (FL) on various Industrial Internet of Things (I loT) applications, so as to minimize the overall learning computation and communication cost via an adaptive client selection policy, while simultaneously maximizing the convergence speed under the non-stationery and drift data. I loT systems often need to transmit data between edge devices and a server, which can lead to various technical problems such as, for example, the exposure of edge device data during transmission that can render devices susceptible to various data security and privacy threats. In some examples, to safeguard data privacy, edge (client) devices in FL can share only machine learning model parameters with a server after training the global model using their own local data. This can effectively address various security and privacy issues stemming from direct data transmission from local devices to a server. Furthermore, when implementing FL on various I loT systems, data drift can present a significant technical challenge because the characteristics of data collected from sensors and devices can change over time in industrial environments. This data drift issue can have profound implications for the accuracy, reliability, and effectiveness of HoT applications in a FL network, which can lead to larger time and bandwidth consumptions and lower convergence rates, for example. To address these challenges, among others, embodiments described herein define a monitoring system that can determine the state of the data from each device and a client selection policy for FL training. To minimize the time consumption, embodiments can determine an estimated time cost for each device under their current condition.

[0015] Referring initially to FIG. 1 , an example industrial loT system or industrial control network 100 can include a central server 102 that defines various manager modules 110, such as a drift detector module 104, a device selector module 106, and a parameter aggregator module 108. The industrial loT system 100 can further include a plurality of edge or client devices 112 communicatively coupled to the central server 102 via a Representational State Transfer (REST) connection 101 and a gRPC (remote procedure call) connection 103. In various examples, the gRPC connection 103 defines a bi-directional connection that is used for federated learning operations and handles retrain requests. The REST connection 101 can define a bi-directional connection that is used for status updates and processes State-Retrieve requests. The edge devices 112 can include virtual programmable logic controllers (PLCs), industrial computers, management computing devices, or the like. The arrangement and number of edge or client devices 112 can vary as desired, and all such arrangements and numbers are contemplated as being within the scope of this disclosure.

[0016] Still referring to FIG. 1 , in some cases, the client devices 112 can include various production machines configured to work together to perform one or more manufacturing operations, such as physical processes 113. For example, the client devices 112 can define,202415815 without limitation, robots and other field devices that can be controlled by a respective PLC, such as sensors 114, actuators 116, or other machines. It will be understood that the industrial control network 100 is simplified for purposes of example. That is, the industrial control network 100 may include additional or alternative nodes or systems, for instance other network devices, that define alternative configurations, and all such configurations are contemplated as being within the scope of this disclosure.

[0017] The system 100 can transfer trained models between the central server 102 and the client devices 112 via the gRPC connections 103, rather than transferring raw data between the central server 102 and the client devices 112 due to, for example, data privacy and bandwidth limitation concerns. The client devices 112 can each further include a local database or datastore 118 and a local model repository 120. In various examples, a given device 112 can receive a global model, for instance the latest version of a global model from the central server 102. When the device 112 receives the latest global model, the device 112 can perform a re-trainer or re-training service 122 to validate the performance of the current model on the current refreshing dataset. Thus, the local model repository 120 can be renewed when the device 112 receives the latest global model. The re-training service 122 can use the local dataset from the local datastore 118 for training so as to generate an updated trained model. The devices 112 can respond to the central server 102, in particular to receiving the latest global model, with its respective updated trained model. The local dataset of the respective device 112 can include not only the retraining dataset, but also predicted labels and actual labels in addition to the retraining data set. Furthermore, each of the client devices 112 can actively monitor data drift over time. Alongside their local training efforts with incoming data, the devices 112 can synchronize with the server 102 by providing updated system information. Such system information can include, for example and without limitation, training information (e.g., batch size, dataset size, etc.) and device 112 information (e.g., type, etc.).

[0018] Referring also to FIGs. 2 and 3, the central server 102 can perform example operations 200 to optimize federated learning. The operations 200 can define a round of federated learning. A round of federated learning can include client selection at 206, in which the central server 102, in particular the device selector module 106, selects a subset of the available client devices 102 to participate in the current round of federated learning. A round of federated learning can further include model distribution at 212, in which the central server 102 sends the current global model to the selected client devices. At 208, the central server 102 waits for replies, in particular respective updated local models, from the selected clients 102. At 210, the central server 102 can rebuild the global model using the updated local models received from the clients devices 112.202415815

[0019] A round of federated learning can further include local model training, in which each selected client device 102 trains the received model on its local dataset. In various examples, the central server 112 can trigger each of the edge devices 102 to initiate a re-training process so that the edge devices 102 can learn the new global model that can handle the data drift through the implementation of federated learning. The server 112 can measure the time for each client device 102 to complete the local training in each global epoch, so as to determine execution time for the federated learning retraining task. After the local training, the client devices 102 can generate model updates, such as gradients or model parameter updates, based on their local training. The client devices can send their model updates to the central server 102. At 210, the central server can aggregate the updates that are received from the client devices 102, so as to update the global model. Thus, a round of federated learning can include client selection, model distribution, local training, update generation, and a global model update. In some cases, the federated learning operations continue for multiple rounds until the global model reaches a satisfactory level of performance or convergence. A global epoch refers to a complete round of federated learning, in which the global model is updated based on the aggregated updates from the participating client devices 102.

[0020] The central server 102 can be configured to oversee the entire state of the network 100. In some cases, the system 100 can implement a policy so as to schedule an appropriate time for the client devices 112 to start retraining. With respect to the initiation of drift detection, the central server 102 can use its comprehensive overview of the entire status graph of each client device 112. For example, with particular reference to FIG. 2, the server 102 can remain in an idle state 201 until a predetermined time passes. In some cases, the predetermined time period is determined by a pre-trained model. The predetermined time period can be based on the dataset size, batch size, the device type, and the like. After the predetermined time period, at 202, the server 102 can send a State-Retrieve request to the active clients 112 to obtain system information or metrics. In particular, for example, the server 102 can periodically send mandatory State-Retrieve requests that requires the active clients 112 to provide their most current (or up-to-date) loss information back to the server 102. At 204, based on the loss information that is received, the server 102 can determine or detect the data drift. In particular, in some cases, the central server 102 obtains losses at a first time and a second time after the first time, so as to obtain losses at two consecutive previous stable time stamps (t* t) to represent the concept drift from t*to t. Determination of the drift at 204 by the central server 112 can be represented by Equation 1 :202415815 where is the prediction loss value of current weight w from an unknown dataset 124, l2is the prediction loss value from a known dataset 126, 8 is the adaptive threshold to quantify the concept drift, N represents the number of client devices 112. In various examples, the unknown dataset 124 is created during the validation of the current model. Data that is not predicted correctly during the validation is forwarded to unknown dataset 124. The known dataset 126 can be selected by using the prototype. For each class, in some cases, a small amount of data can be selected according to the distance between data representation and its class prototype. In some examples, to find the best threshold 8 for both datasets 124 and 126, simulations with drift are run. In example, the adaptive threshold 8 set to 20 results in optimal performance.

[0021] Thus, in accordance with various examples, the central server 102 can perform drift detection, so as to define a centralized approach that shifts the decision-making process for retraining from individual devices 112 to the server 102. The drift can be determined by comparing the ratio of losses from unknown and known datasets 124 and 126, respectively, across all the clients 112 at two different time points. If the ratio of these aggregate loss ratios exceeds the threshold 8, the system 100 determines that significant drift has occurred, and in some cases, a model update is triggered. Thus, the model update can be triggered when the left side of Equation (1) is greater than the threshold. For example, because of the size of the network can vary and can large, the system 100 is configured to determine the optimal time to start re-training, so as to not waste unnecessary bandwidth and computation resources for retraining and transmission.

[0022] In various examples, at the beginning of the federated learning round (e.g., at 206), the server 102 can select a fraction or subset of the devices 112, for example, based on the loss value l2the server 102 collects at 202 (e.g., see 302 of FIG. 3). In an example, the server 102 selects a given client device 112 with a larger difference in loss between two global epochs, so as to lead to faster convergence as compared to a client device with a lesser difference in loss value. In various examples, a larger loss during local training results in a larger performance increase. Because the system 100 has multiple client devices 112 running re-training, the system 100 does not expend time to retrain on local models that reached convergence. That is, the server 102 can be configured to select client devices 102 that have the most potential to increase their performance.

[0023] By way of further example, with particular reference to 206 in FIG. 3, once the retraining starts, the server 112 can call upon the plurality of edge devices 102 to serve as clients in the federated learning process to train the latest model. An example objective of the server 112 is to speed up the convergence speed while limiting communication costs by202415815 selecting an optimal set of client devices 102 for each global epoch. In various examples, the server 112 selects client devices 102 based on the highest local loss difference provided from the system metrics. In some cases, however, the device 112 with the highest local loss consistently is also a less efficient device, such that the total training duration time is extended. Thus, in various examples, the server 112 performs a client selection strategy that prioritizes client devices 102 with higher local loss while considering lower retraining time, ultimately enhancing convergence speed. Furthermore, the central server 102 can perform measures to prevent a scenario in which certain devices 112 are never selected, so that those devices 112 are unable to contribute their local newest data information to construct the global model.

[0024] In some examples, at the beginning of the client selection (at 206), the server 102 first obtains the system metrics x for each device 102. The client device 112 can then run an estimation function using the pretrained model: T EstRuntime (x )■ The server 102 can obtain loss values that can be transmitted with the respective local model from previous epochs. Based on the loss values, the server 102 can select the client devices for re-training, for instance the client devices 102 that have a sufficiently large loss value I and a sufficiently small estimated run time T. Furthermore, in some examples, the server 102 selects client devices 112 that have the oldest updates (or have not re-trained for time threshold), so as to prevent a situation in which a given client device 112 is never selected.

[0025] Referring now to FIG. 4, the client devices 112 can perform example operations 400, 402, and 404. Each client device 112 can continuously receive incoming streaming data. Each client device 112 can define maintain a local data buffer. When a given client device 112 has not received a request, the client device 112 can perform a monitor role. The client device 112 can perform an evaluator role when the state retrieval request is received, and the client device can perform the client role when a retraining request is received from the central server 102. A given client device 112 can perform these three roles in parallel.

[0026] Referring in particular to operations 400, when the edge devices 112 operate as monitors, the edge devices 112 can continuously receive the incoming input at an uncertain rate. Once the size of new incoming data count has accumulated to a predetermined value or parameter, which can be configured and adjusted according to the cache size and incoming data speed, devices 112 can initiate the validation process to test the performance for the global model on newest ones.

[0027] Referring now to operations 402, the edge devices 112 can also continuously wait for the server 102 for the metrics update requests. When such an update request arrives, the edge devices 112 can act as an evaluator by using the latest retraining dataset and responding to the central server 102 with the state metrics. During operations 404, the edge202415815 devices 112 perform as clients in federated learning retraining, which includes training locally, and then updating to the server 102. Each client device 112 receives the latest global model at the end of each global epoch.

[0028] Embodiments described herein implement a performance-estimate-driven client selection policy that minimizes the overall energy consumption of selected devices 112 during communication and computation phases, while simultaneously maximizing the convergence speed of drift handling. Without being bound by theory, by dynamically prioritizing clients based on their estimated performance and loss value, embodiments can optimize the allocation of computational resources, leading to faster model convergence and reduced energy consumption, ultimately promoting sustainability and efficiency in various federated learning setups. Embodiments described herein also implement a novel learning approach that is tailored estimate the execution time of machine learning tasks on different types of devices under different situations. Without being bound by theory, by leveraging historical data and device characteristics, the server 102 can generate accurate predictions of execution times. This valuable information empowers the system 100 to make informed decisions when selecting devices for model updates, ensuring efficient resource allocation and timely convergence.

[0029] Thus, as described herein, an industrial system or network includes a central server communicatively coupled to a plurality of client devices. The server can send a request to the plurality of client devices. Responsive to the request, the server can obtain a plurality of system metrics for each client device of the plurality of client devices. Based on the plurality of system metrics, the server can select a set of client devices of the plurality of client devices, so as to define a selected set of client devices. The server can distribute a global model to the selected set of client devices, such that each client device in the selected set can perform a re-training on the global model. Based on the global model and the respective re-training, the server can receive respective updates to the global model from each of the client devices in the selected set of client devices. The server can aggregate the updates into the global model, so as to generate a new global model with federated learning.

[0030] In another example aspect, based on the plurality of metrics for each client device, the server can detect data drift associated with each client device, wherein the plurality of metrics includes respective loss information associated with each client device. Furthermore, the server can compare the respective loss information, so as to determine a group of client devices having loss information greater than other client devices of the plurality of client devices. The sever can select the group of client devices for the selected set of client devices that perform re-training. In some examples, the central server determines an execution time for each of the client devices of the plurality of client devices. The execution time corresponds202415815 to the re-training. The server can select the selected set of the client devices further based on the respective execution time associated with each client device.

[0031] FIG. 5 illustrates an example of a computing environment within which embodiments of the present disclosure may be implemented. A computing environment 300 includes a computer system 310 that may include a communication mechanism such as a system bus 321 or other communication mechanism for communicating information within the computer system 310. The computer system 310 further includes one or more processors 320 coupled with the system bus 321 for processing the information. The industrial control network 100 may include, or be coupled to, the one or more processors 320.

[0032] The processors 320 may include one or more central processing units (CPUs), graphical processing units (GPUs), or any other processor known in the art. More generally, a processor as described herein is a device for executing machine-readable instructions stored on a computer readable medium, for performing tasks and may comprise any one or combination of, hardware and firmware. A processor may also comprise memory storing machine-readable instructions executable for performing tasks. A processor acts upon information by manipulating, analyzing, modifying, converting or transmitting information for use by an executable procedure or an information device, and / or by routing the information to an output device. A processor may use or comprise the capabilities of a computer, controller or microprocessor, for example, and be conditioned using executable instructions to perform special purpose functions not performed by a general purpose computer. A processor may include any type of suitable processing unit including, but not limited to, a central processing unit, a microprocessor, a Reduced Instruction Set Computer (RISC) microprocessor, a Complex Instruction Set Computer (CISC) microprocessor, a microcontroller, an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), a System-on-a- Chip (SoC), a digital signal processor (DSP), and so forth. Further, the processor(s) 320 may have any suitable microarchitecture design that includes any number of constituent components such as, for example, registers, multiplexers, arithmetic logic units, cache controllers for controlling read / write operations to cache memory, branch predictors, or the like. The microarchitecture design of the processor may be capable of supporting any of a variety of instruction sets. A processor may be coupled (electrically and / or as comprising executable components) with any other processor enabling interaction and / or communication there-between. A user interface processor or generator is a known element comprising electronic circuitry or software or a combination of both for generating display images or portions thereof. A user interface comprises one or more display images enabling user interaction with a processor or other device.202415815

[0033] The system bus 321 may include at least one of a system bus, a memory bus, an address bus, or a message bus, and may permit exchange of information (e.g., data (including computer-executable code), signaling, etc.) between various components of the computer system 310. The system bus 321 may include, without limitation, a memory bus or a memory controller, a peripheral bus, an accelerated graphics port, and so forth. The system bus 321 may be associated with any suitable bus architecture including, without limitation, an Industry Standard Architecture (ISA), a Micro Channel Architecture (MCA), an Enhanced ISA (EISA), a Video Electronics Standards Association (VESA) architecture, an Accelerated Graphics Port (AGP) architecture, a Peripheral Component Interconnects (PCI) architecture, a PCI-Express architecture, a Personal Computer Memory Card International Association (PCMCIA) architecture, a Universal Serial Bus (USB) architecture, and so forth.

[0034] Continuing with reference to FIG. 5, the computer system 310 may also include a system memory 330 coupled to the system bus 321 for storing information and instructions to be executed by processors 320. The system memory 330 may include computer readable storage media in the form of volatile and / or nonvolatile memory, such as read only memory (ROM) 331 and / or random access memory (RAM) 332. The RAM 332 may include other dynamic storage device(s) (e.g., dynamic RAM, static RAM, and synchronous DRAM). The ROM 331 may include other static storage device(s) (e.g., programmable ROM, erasable PROM, and electrically erasable PROM). In addition, the system memory 330 may be used for storing temporary variables or other intermediate information during the execution of instructions by the processors 320. A basic input / output system 333 (BIOS) containing the basic routines that help to transfer information between elements within computer system 310, such as during start-up, may be stored in the ROM 331. RAM 332 may contain data and / or program modules that are immediately accessible to and / or presently being operated on by the processors 320. System memory 330 may additionally include, for example, operating system 334, application programs 335, and other program modules 336. Application programs 335 may also include a user portal for development of the application program, allowing input parameters to be entered and modified as necessary.

[0035] The operating system 334 may be loaded into the memory 330 and may provide an interface between other application software executing on the computer system 310 and hardware resources of the computer system 310. More specifically, the operating system 334 may include a set of computer-executable instructions for managing hardware resources of the computer system 310 and for providing common services to other application programs (e.g., managing memory allocation among various application programs). In certain example embodiments, the operating system 334 may control execution of one or more of the program modules depicted as being stored in the data storage 340. The operating system 334 may202415815 include any operating system now known or which may be developed in the future including, but not limited to, any server operating system, any mainframe operating system, or any other proprietary or non-proprietary operating system.

[0036] The computer system 310 may also include a disk / media controller 343 coupled to the system bus 321 to control one or more storage devices for storing information and instructions, such as a magnetic hard disk 341 and / or a removable media drive 342 (e.g., floppy disk drive, compact disc drive, tape drive, flash drive, and / or solid state drive). Storage devices 340 may be added to the computer system 310 using an appropriate device interface (e.g., a small computer system interface (SCSI), integrated device electronics (IDE), Universal Serial Bus (USB), or FireWire). Storage devices 341 , 342 may be external to the computer system 310.

[0037] The computer system 310 may also include a field device interface 365 coupled to the system bus 321 to control a field device 366, such as a device used in a production line. The computer system 310 may include a user input interface or GUI 361 , which may comprise one or more input devices, such as a keyboard, touchscreen, tablet and / or a pointing device, for interacting with a computer user and providing information to the processors 320.

[0038] The computer system 310 may perform a portion or all of the processing steps of embodiments of the invention in response to the processors 320 executing one or more sequences of one or more instructions contained in a memory, such as the system memory 330. Such instructions may be read into the system memory 330 from another computer readable medium of storage 340, such as the magnetic hard disk 341 or the removable media drive 342. The magnetic hard disk 341 and / or removable media drive 342 may contain one or more data stores and data files used by embodiments of the present disclosure. The data store 340 may include, but are not limited to, databases (e.g., relational, object-oriented, etc.), file systems, flat files, distributed data stores in which data is stored on more than one node of a computer network, peer-to-peer network data stores, or the like. The data stores may store various types of data such as, for example, skill data, sensor data, or any other data generated in accordance with the embodiments of the disclosure. Data store contents and data files may be encrypted to improve security. The processors 320 may also be employed in a multi-processing arrangement to execute the one or more sequences of instructions contained in system memory 330. In alternative embodiments, hard-wired circuitry may be used in place of or in combination with software instructions. Thus, embodiments are not limited to any specific combination of hardware circuitry and software.

[0039] As stated above, the computer system 310 may include at least one computer readable medium or memory for holding instructions programmed according to embodiments of the invention and for containing data structures, tables, records, or other data described202415815 herein. The term “computer readable medium” as used herein refers to any medium that participates in providing instructions to the processors 320 for execution. A computer readable medium may take many forms including, but not limited to, non-transitory, nonvolatile media, volatile media, and transmission media. Non-limiting examples of non-volatile media include optical disks, solid state drives, magnetic disks, and magneto-optical disks, such as magnetic hard disk 341 or removable media drive 342. Non-limiting examples of volatile media include dynamic memory, such as system memory 330. Non-limiting examples of transmission media include coaxial cables, copper wire, and fiber optics, including the wires that make up the system bus 321 . Transmission media may also take the form of acoustic or light waves, such as those generated during radio wave and infrared data communications.

[0040] Computer readable medium instructions for carrying out operations of the present disclosure may be assembler instructions, instruction-set-architecture (ISA) instructions, machine instructions, machine dependent instructions, microcode, firmware instructions, statesetting data, or either source code or object code written in any combination of one or more programming languages, including an object oriented programming language such as Smalltalk, C++ or the like, and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The computer readable program instructions may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection may be made to an external computer (for example, through the Internet using an Internet Service Provider). In some embodiments, electronic circuitry including, for example, programmable logic circuitry, field-programmable gate arrays (FPGA), or programmable logic arrays (PLA) may execute the computer readable program instructions by utilizing state information of the computer readable program instructions to personalize the electronic circuitry, in order to perform aspects of the present disclosure.

[0041] Aspects of the present disclosure are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, may be implemented by computer readable medium instructions.

[0042] The computing environment 300 may further include the computer system 310 operating in a networked environment using logical connections to one or more remote202415815 computers, such as remote computing device 380. The network interface 370 may enable communication, for example, with other remote devices 380 or systems and / or the storage devices 341 , 342 via the network 371 . Remote computing device 380 may be a personal computer (laptop or desktop), a mobile device, a server, a router, a network PC, a peer device or other common network node, and typically includes many or all of the elements described above relative to computer system 310. When used in a networking environment, computer system 310 may include modem 372 for establishing communications over a network 371 , such as the Internet. Modem 372 may be connected to system bus 321 via user network interface 370, or via another appropriate mechanism.

[0043] Network 371 may be any network or system generally known in the art, including the Internet, an intranet, a local area network (LAN), a wide area network (WAN), a metropolitan area network (MAN), a direct connection or series of connections, a cellular telephone network, or any other network or medium capable of facilitating communication between computer system 310 and other computers (e.g., remote computing device 380). The network 371 may be wired, wireless or a combination thereof. Wired connections may be implemented using Ethernet, Universal Serial Bus (USB), RJ-6, or any other wired connection generally known in the art. Wireless connections may be implemented using Wi-Fi, WiMAX, and Bluetooth, infrared, cellular networks, satellite or any other wireless connection methodology generally known in the art. Additionally, several networks may work alone or in communication with each other to facilitate communication in the network 371.

[0044] It should be appreciated that the program modules, applications, computerexecutable instructions, code, or the like depicted in FIG. 5 as being stored in the system memory 330 are merely illustrative and not exhaustive and that processing described as being supported by any particular module may alternatively be distributed across multiple modules or performed by a different module. In addition, various program module(s), script(s), plug-in(s), Application Programming Interface(s) (API(s)), or any other suitable computer-executable code hosted locally on the computer system 310, the remote device 380, and / or hosted on other computing device(s) accessible via one or more of the network(s) 371 , may be provided to support functionality provided by the program modules, applications, or computer-executable code depicted in the figures and / or additional or alternate functionality. Further, functionality may be modularized differently such that processing described as being supported collectively by the collection of program modules depicted in the figures may be performed by a fewer or greater number of modules, or functionality described as being supported by any particular module may be supported, at least in part, by another module. In addition, program modules that support the functionality described herein may form part of one or more applications executable across any number of systems or devices in accordance with any suitable202415815 computing model such as, for example, a client-server model, a peer-to-peer model, and so forth. In addition, any of the functionality described as being supported by any of the program modules depicted in the figures may be implemented, at least partially, in hardware and / or firmware across any number of devices.

[0045] It should further be appreciated that the computer system 310 may include alternate and / or additional hardware, software, or firmware components beyond those described or depicted without departing from the scope of the disclosure. More particularly, it should be appreciated that software, firmware, or hardware components depicted as forming part of the computer system 310 are merely illustrative and that some components may not be present or additional components may be provided in various embodiments. While various illustrative program modules have been depicted and described as software modules stored in system memory 330, it should be appreciated that functionality described as being supported by the program modules may be enabled by any combination of hardware, software, and / or firmware. It should further be appreciated that each of the above-mentioned modules may, in various embodiments, represent a logical partitioning of supported functionality. This logical partitioning is depicted for ease of explanation of the functionality and may not be representative of the structure of software, hardware, and / or firmware for implementing the functionality. Accordingly, it should be appreciated that functionality described as being provided by a particular module may, in various embodiments, be provided at least in part by one or more other modules. Further, one or more depicted modules may not be present in certain embodiments, while in other embodiments, additional modules not depicted may be present and may support at least a portion of the described functionality and / or additional functionality. Moreover, while certain modules may be depicted and described as sub-modules of another module, in certain embodiments, such modules may be provided as independent modules or as sub-modules of other modules.

[0046] Although specific embodiments of the disclosure have been described, one of ordinary skill in the art will recognize that numerous other modifications and alternative embodiments are within the scope of the disclosure. For example, any of the functionality and / or processing capabilities described with respect to a particular device or component may be performed by any other device or component. Further, while various illustrative implementations and architectures have been described in accordance with embodiments of the disclosure, one of ordinary skill in the art will appreciate that numerous other modifications to the illustrative implementations and architectures described herein are also within the scope of this disclosure. In addition, it should be appreciated that any operation, element, component, data, or the like described herein as being based on another operation, element, component, data, or the like can be additionally based on one or more other operations,202415815 elements, components, data, or the like. Accordingly, the phrase “based on,” or variants thereof, should be interpreted as “based at least in part on.”

[0047] Although embodiments have been described in language specific to structural features and / or methodological acts, it is to be understood that the disclosure is not necessarily limited to the specific features or acts described. Rather, the specific features and acts are disclosed as illustrative forms of implementing the embodiments. Conditional language, such as, among others, “can,” “could,” “might,” or “may,” unless specifically stated otherwise, or otherwise understood within the context as used, is generally intended to convey that certain embodiments could include, while other embodiments do not include, certain features, elements, and / or steps. Thus, such conditional language is not generally intended to imply that features, elements, and / or steps are in any way required for one or more embodiments or that one or more embodiments necessarily include logic for deciding, with or without user input or prompting, whether these features, elements, and / or steps are included or are to be performed in any particular embodiment.

[0048] The flowchart and block diagrams in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of instructions, which comprises one or more executable instructions for implementing the specified logical function(s). In some alternative implementations, the functions noted in the block may occur out of the order noted in the Figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and / or flowchart illustration, and combinations of blocks in the block diagrams and / or flowchart illustration, can be implemented by special purpose hardware-based systems that perform the specified functions or acts or carry out combinations of special purpose hardware and computer instructions.

Claims

202415815CLAIMSWhat is claimed is:1 . A method performed by a central server communicatively coupled to a plurality of client devices in an industrial system, the method comprising: sending a request to the plurality of client devices; responsive to the request, obtaining a plurality of system metrics for each client device of the plurality of client devices; based on the plurality of system metrics, selecting a set of client devices of the plurality of client devices, so as to define a selected set of client devices; distributing a global model to the selected set of client devices, such that each client device in the selected set can perform a re-training on the global model; based on the global model and the respective re-training, receiving respective updates to the global model from each of the client devices in the selected set of client devices; and aggregating the updates into the global model, so as to generate a new global model with federated learning.

2. The method as recited in claim 1 , the method further comprising: based on the plurality of metrics for each client device, detecting data drift associated with each client device, wherein the plurality of metrics includes respective loss information associated with each client device.

3. The method as recited in claim 2, the method further comprising: comparing the respective loss information, so as to determine a group of client devices having loss information greater than other client devices of the plurality of client devices; and selecting the group of client devices for the selected set of client devices that perform re-training.

4. The method as recited in claim 1 , the method further comprising: determining an execution time for each of the client devices of the plurality of client devices, the execution time corresponding to the re-training.

5. The method as recited in claim 4, the method further comprising:202415815 selecting the selected set of the client devices further based on the respective execution time associated with each client device.

6. An industrial system comprising: a plurality of client devices and a central server communicatively coupled to the plurality of client devices, the central server comprising: a processor and a memory storing instructions that, when executed by the processor, cause the processor to: send a request to the plurality of client devices; responsive to the request, obtain a plurality of system metrics for each client device of the plurality of client devices; based on the plurality of system metrics, select a set of client devices of the plurality of client devices, so as to define a selected set of client devices; distribute a global model to the selected set of client devices, such that each client device in the selected set can perform a re-training on the global model; based on the global model and the respective re-training, receive respective updates to the global model from each of the client devices in the selected set of client devices; and aggregate the updates into the global model, so as to generate a new global model with federated learning.

7. The system as recited in claim 6, the memory further storing instructions that, when executed by the processor, further cause the processor to: based on the plurality of metrics for each client device, detect data drift associated with each client device, wherein the plurality of metrics includes respective loss information associated with each client device.

8. The system as recited in claim 6, the memory further storing instructions that, when executed by the processor, further cause the processor to: compare the respective loss information, so as to determine a group of client devices having loss information greater than other client devices of the plurality of client devices; and select the group of client devices for the selected set of client devices that perform retraining.2024158159. The system as recited in claim 6, the memory further storing instructions that, when executed by the processor, further cause the processor to: determine an execution time for each of the client devices of the plurality of client devices, the execution time corresponding to the re-training.

10. The system as recited in claim 9, the memory further storing instructions that, when executed by the processor, further cause the processor to: select the selected set of the client devices further based on the respective execution time associated with each client device.

Images