Federated Learning for Multi-Label Classification Models for Oil Pump Management

Abstract

A computer-implemented federated learning method for predicting asset failures includes generating a local model at a local site for each of a cohort and training the local model for each of the cohorts on local data for each failure type. The local models are shared with a central database. A global model is created based on an aggregation of multiple local models from multiple local sites. At each of the multiple local sites, one of the global model and the local model is selected for each cohort. The selected model operates on the local data to predict asset failures. The utilized features include partitioning asset features into static features, semi-static features, and dynamic features and forming asset cohorts based on the static features and semi-static features.

Description

[Technical Field]

[0001] The present disclosure relates generally to artificial intelligence and machine learning systems, and more particularly to methods and systems for distributed training of multi-label classification models for industrial equipment repair and maintenance, such as oil pump preventive maintenance. [Background technology]

[0002] Oil and other mining companies need to maintain their assets. In oil fields, companies need to maintain pump wells in good condition to produce high oil production. The current trend is to use a preventive maintenance approach, and the success of a cost-effective preventive maintenance approach depends on the accuracy of predicting the need for maintenance of specific components.

[0003] Large oil companies have sites around the world pumping oil in many different geographic regions. While each site is open to sharing aggregate knowledge to have better predictive models, sites are restricted from sharing detailed operational sensor data due to privacy considerations and management issues, for example.

[0004] Current general forecasting models have significant limitations. These limitations include a lack of data sharing between sites, a lack of failure examples, and scenarios for individual geographic sites. Different scenarios of operating conditions may require different models, and building one general-purpose forecasting model may not be suitable for local conditions at a particular geographic site.

[0005] Therefore, there is a need to have an approach to develop and improve prediction / classification models that are manageable both locally and globally. Summary of the Invention

[0006] According to various preferred embodiments, a computing device, a non-transitory computer-readable storage medium and a method for developing and improving predictive models for asset failure are provided, where only the models are shared between different sites.

[0007] In one embodiment, the present invention provides a computer-implemented method for predicting asset failures, the method including: dividing asset characteristics into static, semi-static, and dynamic characteristics; and forming cohorts of assets based on the static and semi-static characteristics. The method further includes generating a local model at a local site for each of the cohorts; and training the local model for each of the cohorts on local data for each failure type. The local models are shared with a central database. A global model is created based on an aggregation of multiple local models from multiple local sites. At each of the multiple local sites, a global model and one of the local models are selected from each cohort. The selected model operates on the local data to predict failure of one or more assets in the cohort.

[0008] Preferably, the present invention provides a method further comprising generating a template model for creating each of the local models for each of the cohorts.

[0009] Preferably, the present invention provides a method further comprising pooling local models from each of a plurality of local sites into a pool of local models, and determining the performance of the global model and a selected one of the local models from the pool of local models.

[0010] Preferably, the present invention provides a method further comprising updating the global model based on an average of each local model in the pool of local models for each cohort. In some preferred embodiments, each local model in the pool of local models is weighted based on the average number of assets at the local site that contributed the local model to the pool of local models.

[0011] According to various embodiments, a computer-implemented method for predicting asset failure includes partitioning asset characteristics into static, semi-static, and dynamic characteristics, and forming cohorts of assets based on the static and semi-static characteristics. A local model is generated at a local site for each cohort, and the local models are shared with a central database. A pool of local models is created from the local models from each of the local sites. A global model is created based on an aggregation of the local models from the local sites. At each of the local sites, one of the global models and one of the local models from the pool of local models is selected for each cohort. The selected model operates on the local data to predict asset failure.

[0012] According to various embodiments, a computer-implemented method for asset failure prediction includes partitioning asset characteristics into static, semi-static, and dynamic characteristics and forming cohorts of assets based on the static and semi-static characteristics. A local model is generated at a local site for each of the cohorts, and a global model is created for each of the cohorts. At each of the multiple local sites, one of the global model and the local model is selected for each cohort. The selected model operates on local data to predict failure of assets belonging to the cohort for which the model is developed.

[0013] The concepts described herein provide systems and methods that improve upon current approaches to predicting asset failures. The systems and methods described herein can improve the accuracy of asset failure prediction models by sharing information between different sites without compromising privacy and security.

[0014] These and other features will become apparent from the following detailed description of illustrative embodiments, which should be read in connection with the accompanying drawings.

[0015] Preferred embodiments of the present invention will now be described, by way of example only, with reference to the following drawings: [Brief explanation of the drawings]

[0016] [Figure 1] According to one embodiment of the present disclosure, a representation of the evolution of a model based on feature groupings. [Figure 2] 1 is a representation of the system architecture of an asset failure prediction engine after deployment of a federated model, illustrating model tuning and model updating consistent with an illustrative embodiment. [Figure 3] Consistent with the exemplary embodiments, a method for selecting a local model for a cohort at a local site is presented. [Figure 4] Consistent with the exemplary embodiment, a method for selecting a local model without sharing local models from other sites is presented. [Figure 5] Consistent with the exemplary embodiment, a model template for prospective cohort multi-label prediction is shown. [Figure 6] Consistent with the illustrative embodiments, an architectural model for mismatch and tuning at a local site is presented. [Figure 7] 10 is a flowchart illustrating operations related to setting up an asset failure prediction engine consistent with an illustrative embodiment. [Figure 8]10 is a flowchart illustrating operations related to operating an asset failure prediction engine consistent with an illustrative embodiment. [Figure 9] Figure 2 is an example of a functional block diagram of a computer hardware platform that can be used to implement the asset failure prediction engine. DETAILED DESCRIPTION OF THE INVENTION

[0017] In the following detailed description, numerous specific details are set forth by way of example to provide a thorough understanding of the relevant teachings. However, it should be appreciated that the present teachings may be practiced without such detail. In other instances, well-known methods, procedures, components, or circuits, or combinations thereof, are described at a relatively high level and without detail to avoid unnecessarily obscuring aspects of the present teachings. In other instances, well-known methods, procedures, components, or circuits, or combinations thereof, are described at a relatively high level and without detail to avoid unnecessarily obscuring aspects of the present teachings.

[0018] Unless otherwise indicated, and as will be apparent from the description and claims that follow, throughout the specification, statements utilizing terms such as "processing," "computing," "calculating," "determining," and the like should be understood to refer to the actions and / or processes of a computer or computing system or similar electronic computing device that manipulate and / or transform data represented as physical quantities, e.g., electronic quantities, in the registers and / or memory of the computing system into other data similarly represented as physical quantities in the memory, registers, or other such information storage, transmission, or display device of the computing system.

[0019] As used herein, the term "distributed learning" refers to a learning model in which no local data is shared, and only the models (local and global) are shared among distributed sites.

[0020] As used herein, the term "multi-learning classification" refers to a classification problem in which multiple labels may be assigned to each case and are not exclusive.

[0021] As used herein, the term "cohort" refers to breaking down assets into different groups for better analysis and prediction, instead of grouping all assets into a single category.

[0022] As used herein, the term "global model" refers to a model for a particular cohort where local sites can use the global model due to the lack of a local model or because the local model is not as good as the global model.

[0023] As used herein, the term "local model" refers to the model used at the local site.

[0024] As used herein, the term "model aggregator" refers to a collection of local models and an updated global model based on their use.

[0025] As used herein, the term "model selection at a local site" refers to an algorithm used to select a local model for use at a local site, where the pool of candidate models comes from global models for the cohort or local models from other sites, or both.

[0026] As described in more detail below, the present disclosure generally relates to methods and systems for industrial asset management, e.g., oil pump management, through distributed learning. The methods and systems use a centralized approach for storing cohort models and predictive models and sending centralized models to local sites as a starting point for the local sites. The methods and systems can use a distributed approach to refine and improve predictive model performance at local sites, resulting in local improvements by retraining the model with the same model structure and initial features. Centralized management of cohort models can help ensure that learning from each site is aligned to common or similar scenario groups. The systems and methods of the present disclosure can provide significant improvements in asset failure predictive management models by sharing models across multiple sites in an effort to use information learned from one site at other sites without having to share the specific details of any particular asset failure.

[0027] Additionally, as described in more detail below, the present disclosure provides systems and methods that can perform model mismatch analysis to determine when to update local models and generate model refinements. The systems and methods can apply predictive models, including but not limited to traditional and deep learning model architectures for prediction, or classification models with sufficient model complexity to enable model integration from different sites and avoid a loss of accuracy for other sites after global model integration.

[0028] There are three main types of oil pumps: submersible pumps or electric submersible pumps (ESPs), eccentric screw pumps (PCPs), and reciprocating rod lifts (RRLs). Each of these pumps can represent a specific cohort. Machine types belong to the same cohort because they share similar structures and functions. There are three main types of component failures, including pump failures, rod failures, and pipe failures. These failure types can be defined as labels. While this disclosure defines cohorts as different pumps in the oil industry, it should be understood that other assets may be defined as cohorts depending on the desired application and industry.

[0029] Reference is now made in detail to the following description and illustrated in the accompanying drawings.

[0030] Referring to FIG. 1 , details about each oil pump model system 100 can be used to provide a set of features 102 for that model system 100. The set of features 102 can include static features 104, semi-static features 106, and dynamic features 108. A global database 110 can be maintained that contains data about each asset (e.g., each pump system) and each asset's characteristics. The static features 104 can typically include information that does not change for an asset, such as asset purchase year, model number, brand, geographic location, etc. The semi-static features 106 can include slowly changing information, such as date since purchase, date since last maintenance, number of scheduled maintenances since purchase, average days between repairs, etc. Certain data transformations, as described below, may be required to convert the static and semi-static data into usable features. The dynamic features 108 include information from monitoring sensors. Typically, the dynamic features 108 provide fault signals over a short time horizon. Certain data aggregations and transformations can be used to convert this data into usable features, such as hourly / daily averages, exponential smoothing, outlier identification, missing values ​​from the previous week, etc.

[0031] Assets can be divided into one or more cohorts 112. In the example of FIG. 1, the cohorts are based on pump type (pump type A, pump type B, pump type C, and pump type D), but depending on the asset of interest, cohorts may be established based on any given asset or division of assets. For example, a wind farm may divide cohorts into different power generation components, such as bearings, inverters, storage devices, etc. In some embodiments, cohorts 112 can be created by analyzing static and semi-static characteristics of assets, and a given cohort can have similar static and / or semi-static characteristics for a given asset.

[0032] As shown in FIG. 1 , a cohort 112 can include static features 104, semi-static features 106, and dynamic features 108 for that particular asset. A failure prediction model 114 (also referred to as a global model 114) can be constructed for each cohort. The global model 114 can be based on a previously established local model for a given asset or cohort, or it can be based on a model based on similar assets. Typically, the global model 114 can have a deep learning model structure with a fixed architecture based on the static, semi-static, and dynamic features of the asset. The global model 114 based on the cohort 112 can be deployed to each of multiple local sites 116.

[0033] Referring now to FIG. 2, each deployment site 200 (also referred to as each local site 200) can select a local model 202 from a model repository 204. The model repository 204 can include a global model for each cohort. In some embodiments, as described in more detail below, the model repository can include a global model and at least one local model for each cohort. A model consolidator / aggregator 206 may be provided to align models to cohorts and limit the total number of models by removing similar or identical local models.

[0034] A local mismatch analysis 208 may be performed at each deployment site 200, and selected local models 202 may monitor model performance. Discrepancies between model predictions and actual asset performance may result in a mismatch report 210 for federated analysis 212. Because mismatch reports 210 from various local sites 200 may be analyzed together, this analysis is referred to as a "federated analysis." In response to the mismatch data, the system 250 may provide a cohort update 214 or a model update 216 that can be sent to the model repository 204. Additionally, each local site 200 may provide local model tuning 216. Details of the model tuning 216 may be provided for the model update 216 and federated analysis 212 for the updated local model for the model repository 204.

[0035] Referring to Figure 3, a diagrammatic representation of local model selection for a given cohort at a local site is depicted. A local site 300 can receive one or more local models 302 and a global model 304 for the given cohort. If no model (either global or local) is available for the given cohort (e.g., when a new cohort is established at the local site), a new local model is generated at the local site. This new local model may be based, for example, on models for similar assets at this site or other local sites.

[0036] Two models can be selected for comparison and refinement. One of the selected models can be a global model for a given cohort, and the other model can be a local model from the local model pool. Various criteria, such as the similarity of the site providing the local model to the local site 300, the proximity of the site providing the local model to the local site 300, and performance metrics of the local model, can be used to select a local model from the local model pool. In some embodiments, the local model can be selected from the local models at the local site 300 itself. Data can be applied to the two selected models, and performance can be analyzed. The best-performing model can be selected as the new local model 306 for a given cohort at the local site 300. If there is a performance mismatch between the two models, tuning can be performed to generate a new local model, which can be shared to the global model repository 204 (see FIG. 2).

[0037] Referring to FIG. 4 , in some embodiments, the local model must not be shared due to site management or privacy requirements. In this embodiment, a global model 402 may be provided to the local site 400 for a given cohort. Two models may also be selected: one model is the global model 204, and the other is a current version of the local model generated at the local site 400. If a local model is not currently at the local site 400, the local site 400 may create a local model in a manner similar to that described above with respect to FIG. 3 . The models may be compared and refined, as described above, to generate an updated local model 404. When the global model 402 is updated, for example, as described below, the local model selection process shown in FIG. 4 may then be repeated to generate an updated local model 404 based on the refined global model.

[0038] Referring to FIG. 5, an abstract model template 500 can be defined for every cohort. This abstract model template 500 can be used, for example, to generate a model for a new cohort, such as an original global model. The template 500 can input the static features 104 and semi-static features 106 to a multi-layer perceptron neural network 502. The template 500 can also input the dynamic features 108 to a long-short-term memory neural network 504. The outputs from both networks 502, 504 can provide an aggregate model 506 for multi-label prediction. The output of the template 500 can be a multi-label output, and the template 500 can be used for all future cohort multi-label predictions.

[0039] Referring now to FIG. 6, a model 600 for tuning and model mismatch is provided. A local model store 602 may contain information about at least the current local model at the local site. The local model store 602 may also contain information about other local models at other sites for a given cohort. A local data store 604 may be used to aid in cohort identification 606 or new cohort creation 608. As described above, two models may be run with data from the local data store 604 to generate a mismatch analysis 610. Model tuning and / or updates 612 may be performed based on the mismatch analysis 610, and the updates may be provided to a master model store 614.

[0040] In addition to updating the local models, the system may provide updates to the global model based on one or more evaluation methods. For example, one method may include averaging the weights of each local model for a given cohort (e.g., by a weighted average based on the number of assets in each local site) to access the weights and generate an updated global model. In some embodiments, the global model may be updated by accessing and averaging only the weights of the last layer of each model to generate the global model. In this embodiment, the weights of the lower layers are identical for each local model. In other embodiments, the global model may be updated using an ensemble approach to create a new global model based on the individual local models for a given cohort. The global model may be updated periodically or when new or updated local models are provided from one or more local sites.

[0041] With the above overview of example system 250 (see FIG. 2 ), it may be useful to consider a high-level discussion of example processes. To that end, FIG. 7 presents an example process 700 associated with establishing system 250, including the generation of local and global models. FIG. 8 presents an example process 800 associated with local site selection for a local model. Processes 700 and 800 are illustrated in logical flowcharts as a group of blocks, which represent a sequence of operations that can be performed in hardware, software, or a combination thereof. In the software context, the blocks represent computer-executable instructions that, when executed by one or more processors, perform the recited operations. Generally, computer-executable instructions may include routines, programs, objects, components, data structures, etc. that perform a function or implement an abstract data type. In each process, the order in which operations are described is not intended to be construed as a limitation, and any number of the described blocks can be combined in any order and / or executed in parallel to perform the process.

[0042] Referring to FIG. 7, a process 700 for establishing a system for asset failure prediction includes operation 710, partitioning asset characteristics into static, semi-static, and dynamic characteristics related to addressing asset dynamics characteristics over different time scales. Operation 720 can generate a template model to ensure a common model structure. FIG. 5, discussed above, provides an example of template model generation. Operation 730 can include forming cohorts based on the static and semi-static characteristics to group similar assets and enable cohort-specific models. In some embodiments, the system can automatically generate cohorts based on the static and semi-static characteristics. Operation 740 can include training local models with local knowledge for each cohort based on various failure types for the asset. Operation 750 can include sharing the local knowledge to a central database and creating a global model that anchors or aggregates the local knowledge into global knowledge for each cohort. In operation 760, the central database can distribute the cohort definitions, global model, and local models to each local site.

[0043] Referring to FIG. 8, process 800 may include operation 810, in which the local site may select a local model from a pool of local models. At operation 820, the local site may also select a global model. At operation 830, the local site may select a best model as the new model based on the performance of the two selected models. At operation 840, the local site may perform a performance and mismatch analysis of the best model. At operation 850, the local site may tune the selected model and return the data to the central database.

[0044] 9 provides an example functional block diagram of a computer hardware platform 900 that may be used to implement a computing device particularly configured to host an asset failure prediction engine 950. The asset failure prediction engine 950 may include a cohort generation module 952, a global model database 954, a local model pool 956, and a model tuning module 958, as described above.

[0045] The computer platform 900 may include a central processing unit (CPU) 910, a hard disk drive (HDD) 920, a random access memory (RAM) and / or read-only memory (ROM) 930, a keyboard 950, a mouse 960, a display 970, and a communication interface 980 connected to a system bus 940.

[0046] In one embodiment, HDD 920 has functionality that includes storing programs capable of performing various processes in the manner described herein, such as asset failure prediction engine 950.

[0047] Although the above description describes a method for managing asset failure prediction by sharing models between different sites, a similar system may be utilized within a single site, where multiple uses of the same asset (cohort) may be realized across a single site. In this embodiment, sharing models without sharing data may be beneficial for developing and improving failure prediction models within a single site.

[0048] The description of various embodiments of the present teachings has been presented for purposes of illustration and is not intended to be exhaustive or limited to the disclosed embodiments. Many modifications and variations that do not depart from the scope of the described embodiments will be apparent to those skilled in the art. The terms used herein have been chosen to best explain the principles of the embodiments, practical applications, or technical improvements over techniques found in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

[0049] While the foregoing describes what is believed to be the best mode and / or other examples, it should be understood that various modifications thereof may be made, that the subject matter disclosed herein may be embodied in a variety of forms and examples, and that the present teachings are applicable to numerous applications, only a few of which have been described herein. It is intended by the following claims to claim any and all applications, modifications, and variations that fall within the true scope of the present teachings.

[0050] The components, steps, features, objects, benefits, and advantages described herein are merely exemplary. None of them, nor the discussion related thereto, are intended to limit the scope of protection. While various advantages have been described herein, it should be understood that not all embodiments necessarily include all advantages. Unless otherwise specified, all measurements, values, ratings, positions, dimensions, sizes, and other specifications set forth in this specification, including the following claims, are approximate rather than precise. They are intended to have a reasonable range consistent with the relevant function and what is customary in the relevant technical field.

[0051] Numerous other embodiments are also contemplated, including embodiments having fewer, additional, or different or different combinations of components, steps, features, objects, benefits, and advantages, including embodiments in which the components and / or steps are arranged and / or ordered differently.

[0052] 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, can be implemented by computer-readable program instructions.

[0053] These computer-readable program instructions may be provided to a processor of a suitably configured computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for performing the functions / acts specified in one or more blocks of the flowcharts and / or block diagrams. These computer-readable program instructions may also be stored on a computer-readable storage medium capable of directing a computer, programmable data processing apparatus, or other device, or combination thereof, to function in a certain way, such that the computer-readable storage medium on which the instructions are stored comprises an article of manufacture containing instructions that perform aspects of the functions / acts specified in one or more blocks of the flowcharts and / or block diagrams.

[0054] The computer-readable program instructions may also be loaded into a computer, other programmable data processing apparatus or other device, causing the computer, other programmable apparatus or other device to perform a series of operational steps to generate a computer-implemented process, such that the instructions, which execute on the computer, other programmable apparatus or other device, perform the functions / operations specified in one or more blocks of the flowcharts and / or block diagrams.

[0055] The call flows, flowcharts, and block diagrams in the figures herein 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 a flowchart or block diagram may represent a module, segment, or portion of instructions, which comprises one or more executable instructions for implementing a specified logical function(s). In some alternative implementations, the functions shown in the blocks may occur out of the order shown 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 on the required functionality. It should also be noted that each block in the block diagrams and / or flowchart diagrams, and combinations of blocks in the block diagrams and / or flowchart diagrams, may be implemented by a dedicated hardware-based system that performs specified functions or operations or executes a combination of dedicated hardware and computer instructions.

[0056] Although the foregoing has been described in connection with exemplary embodiments, it should be understood that the term "exemplary" does not mean best or optimal, but merely an example. Except as immediately above, nothing described or illustrated is intended or should be construed as conferring to the public any components, steps, features, objects, benefits, advantages, or equivalents, whether claimed or not. Terms and expressions used herein should be understood to have the ordinary meanings ascribed to such terms and expressions in relation to their respective fields of inquiry and study, unless a specific meaning is otherwise defined herein. Relationship terms such as "first" and "second" may be used only to distinguish one entity or operation from another and do not necessarily require or imply any actual relationship or ordering between such entities or operations. The terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus comprising a list of elements does not include only those elements, but may also include other elements not expressly listed or elements inherent to such process, method, article, or apparatus. An element following "a" or "an" does not, absent further constraints, exclude the presence of additional identical elements in the process, method, article, or apparatus that comprises the element.

[0057] The Abstract of the Disclosure is provided to allow the reader to quickly grasp the nature of the technical disclosure. It is submitted with the understanding that it will not be used to interpret or limit the scope or meaning of the claims. In addition, in the foregoing Detailed Description, it can be seen that various features are grouped together in various embodiments for the purpose of streamlining the disclosure. This method of disclosure is not to be interpreted as reflecting an intention that the claimed embodiments have more features than are expressly recited in each claim. Rather, as the following claims reflect, inventive subject matter lies in less than all features of a single disclosed embodiment. Accordingly, the following claims are hereby incorporated into the Detailed Description, with each claim standing on its own as separately claimed subject matter.

Claims

1. A computer-implemented method for predicting asset failures, the computer-implemented method comprising: dividing the characteristics of the asset into static characteristics, semi-static characteristics, and dynamic characteristics; forming a cohort of the asset based on the static characteristics and the semi-static characteristics; generating a local model at a local site for each of the cohorts; training the local models for each of the cohorts for each failure type on local data; sharing the trained local models with a central database; creating a global model based on an aggregation of the local models from multiple local sites; at each of the multiple local sites, selecting the global model or the local model for each of the cohorts; operating the selected model on local data to predict one or more failures of one of the assets belonging to one of the cohorts; A computer-implemented method comprising the above steps.

2. The computer-implemented method further includes generating a template model for creating each of the local models for each of the cohorts, The computer-implemented method according to claim 1.

3. The computer-implemented method further includes: pooling the local models from each of the multiple local sites into a pool of local models; determining the performance of one local model selected from the pool of the global model and local models; The computer-implemented method according to claim 1, further including the above steps. The computer-implemented method according to claim 1.

4. The selected model is selected based on the performance of the global model and the selected one local model The computer-implemented method according to claim 3.

5. The computer-implemented method further includes determining a mismatch between the selected one of the global model and the local model, The computer-implemented method according to claim 3.

6. The computer-implemented method further includes improving the performance of the local model for tuning the selected model based on the local data of the asset, The computer-implemented method according to claim 3.

7. The computer-implemented method further includes providing the tuned selected model to the pool of local models. The computer-implemented method according to claim 6. Claim 8 The computer-implemented method further includes updating the global model based on an average of each of the local models in the pool of local models for each of the cohorts. The computer-implemented method according to claim 3. Claim 9 Each of the local models in the pool of local models is weighted based on an average number of assets of the local site that provided the local model to the pool of local models. The computer-implemented method according to claim 3. Claim 10 A computer-implemented method for predicting asset failures, the computer-implemented method comprising: segmenting the characteristics of the asset into static characteristics, semi-static characteristics, and dynamic characteristics; forming a cohort of the assets based on the static characteristics and the semi-static characteristics; generating a local model at a local site for each of the cohorts; sharing the local model with a central database; forming a pool of local models from the plurality of local models from each of the plurality of local sites; creating a global model based on an aggregation of the plurality of local models from the plurality of local sites; at each of the plurality of local sites, selecting one of the global models and one of the plurality of local models from the pool of local models for each of the cohorts; operating the selected model on local data to predict one or more of the failures of one of the assets belonging to one of the cohorts; A computer-implemented method including. Claim 11 The computer-implemented method further includes generating a template model for creating each of the local models for each of the cohorts. The computer-implemented method according to claim 10. Claim 12 The computer-implemented method further includes determining a mismatch between the global model and the selected model. The computer-implemented method according to claim 10. Claim 13 The computer-implemented method further includes improving the performance of the local model that tunes the selected model based on local data. The computer-implemented method according to claim 10.

14. The computer-implemented method further includes providing the tuned selected model to the pool of local models. The computer-implemented method according to claim 13.

15. The computer-implemented method further includes updating the global model based on the average of each of the local models in the pool of local models for each of the cohorts. The computer-implemented method according to claim 10.

16. Each of the local models in the pool of local models is weighted based on the average number of assets of the local site that provided the local model to the pool of local models. The computer-implemented method according to claim 15.

17. A computer-implemented method for predicting asset failures, the computer-implemented method comprising: segmenting the characteristics of the asset into static characteristics, semi-static characteristics, and dynamic characteristics; forming a cohort of the assets based on the static characteristics and the semi-static characteristics; generating a local model at a local site for each of the cohorts; creating a global model for each of the cohorts; at each of a plurality of local sites, selecting one of the global model and the local model for each of the cohorts; operating the selected model on local data at the local site to predict one or more failures of one or more of the assets belonging to one of the cohorts; A computer-implemented method comprising.

18. The computer-implemented method comprises: storing each of the local models for the plurality of local sites in a central database; updating the global model based on the average of each of the local models in the pool of local models for each of the cohorts; further comprising The computer-implemented method according to claim 17.

19. Each of the local models within the pool of local models is weighted based on the average number of assets of the local site that provided the local model to the pool of local models. The computer-implemented method according to claim 18. **Claim 20** The computer-implemented method comprises at each of the plurality of local sites selecting, for each of the cohorts, one of the updated global model and the local model operating the selected model on local data to predict a failure of the asset and further comprising The computer-implemented method according to claim 18. **Claim 21** A computer program for causing a computer to execute the method according to any one of claims 1 to 20. **Claim 22** A non-transitory computer-readable storage medium recording the computer program according to claim 21.

Images