Learning and explaining solutions within the asset hierarchy

Abstract

Systems and methods described herein are directed to generating an asset hierarchy from a plurality of assets, the asset hierarchy indicating relationships between the plurality of assets from a lowest level to a highest level; performing a solution learning process to learn one or more model solutions for each of the plurality of assets based on the relationships between the plurality of assets in the asset hierarchy, where output of the one or more model solutions in a lower level of the asset hierarchy is used as input for the solution learning process to learn one or more model solutions for each of the plurality of assets in a higher level; and storing the asset hierarchy, the one or more model solutions for each of the plurality of assets, and knowledge for solution explanations for the one or more model solutions in storage.

Description

[Technical Field]

[0001] The present disclosure is directed generally to the Internet of Things (IoT) and operational technology (OT) fields, and more particularly to facilitating solution learning and explanation within an asset hierarchy. [Background technology]

[0002] An asset hierarchy is a logical and / or physical way of organizing all of the machines, equipment, and individual components owned by a company at one or more locations. The top-down structure allows maintenance professionals to understand the relationships between assets using "parent-child" relationships.

[0003] Figure 1 shows an example of an asset hierarchy. The example asset hierarchy in Figure 1 is Industry → Business Type → Plant / Unit → Section / System → Equipment Unit → Component → Part, where "Industry" is the highest level asset, "Business Type" is a sub-asset of "Industry," and so on. As shown in Figure 1, the asset hierarchy can organize assets in a logical and / or physical way and may include other sub-components of the asset (e.g., sensors mounted on the part) depending on the desired implementation.

[0004] When assets of interest are organized into an asset hierarchy, the asset hierarchy can facilitate understanding of varying degrees of physical and logical relationships between assets, asset management and smarter asset lifecycle decisions, more efficient (e.g., in terms of time, money, and cost of ownership) scheduling of asset maintenance and repairs, and faster identification of root causes.

[0005] Given a set of assets, the relationships between those assets need to be identified and used to construct an asset hierarchy. Various asset hierarchies can be constructed depending on the type and generality of the asset hierarchy relationships.

[0006] There are two types of relationships between assets. The first type of relationship can be a physical relationship, which describes how assets are attached or connected to each other. Physical relationships are usually "compositional," also known as parent-child relationships, where one asset is the child of another asset. In Figure 1, the bottom three levels ("Equipment Unit" → "Component" → "Part") represent the physical relationships between assets.

[0007] The second type of relationship can be a logical or functional relationship, which means which group of assets act as a logical or functional unit, and such a logical / functional unit can be defined as a logical / functional asset. Note that the logical or functional units themselves can also be at different levels, with smaller sets of functional units usually located at lower levels, while larger sets of functional units are located at higher levels. In Figure 1, the top five levels ("Industry" → "Business Type" → "Plant / Unit" → "Section / System" → "Equipment Unit") represent the logical relationships between assets.

[0008] Given a set of assets, an asset hierarchy can be constructed to address most of the tasks, for example, tasks may include fault detection, fault prediction, remaining useful life, operator skill assessment, etc. Such asset hierarchies are generic and can be based on physical relationships between assets or logical / functional relationships between assets.

[0009] Sometimes there is a business need to solve only one task, and only relevant assets and asset relationships need to be kept in the asset hierarchy. For example, in the task of operator skill assessment, only operational assets need to be considered in the asset hierarchy. Such asset hierarchies are task-oriented and may accordingly base their cost functions on physical relationships between assets and / or logical / functional relationships between assets.

[0010] Given an asset and a problem with that asset, physics-based modeling and / or machine learning techniques can be applied to solve the problem. For example, in a coal mining plant, there is a "thick-a-tail" asset, and anomaly detection and prediction techniques can be applied to solve the fault detection and prediction problem. Traditionally, solutions are built asset by asset and problem by problem, without considering or leveraging the relationships between assets and problems. Additionally, the corresponding optimization is locally constrained (e.g., optimizing efficiency with respect to time, cost, or ownership) and cannot be leveraged across solutions.

[0011] Another related task for solutions is to explain the models and results and generate prescriptive actions to prevent or fix asset problems, including but not limited to root cause analysis, repair recommendations, alert suppression, etc. Traditionally, work on solution explanations has been limited, and explanations are done on a per-asset, per-problem basis.

[0012] Some limitations and restrictions of conventional systems and methods are discussed below, and example implementations introduce techniques for solving these problems.

[0013] First, learning solutions to problems in the related art is typically done at the level of individual assets (i.e., component-based learning). A problem with an asset of interest is recognized, and a physics-based and / or machine learning solution is built for the asset separately. Because assets in a system function as a whole, and the functional status and / or performance of one asset can affect other assets (e.g., immediately upstream and downstream assets), solutions that do not incorporate relationships between assets will not work well. To address this issue, implementations described herein identify relationships between assets and utilize such relationships as part of the learning process to build more accurate and effective solutions.

[0014] Second, in the related art, explanations of solutions to problems and the results of the solutions are typically performed at the level of individual assets (i.e., component-based explanations). This relates to the first problem, in which solutions are built separately for each asset. Because assets in a system function as a whole and the functional status and / or performance of one asset can affect other assets (especially the assets immediately upstream and downstream), explaining solutions and results by focusing only on the asset that exhibits the problem may miss the root cause of the problem and may not work well. To address this issue, the implementations described herein identify relationships between assets and utilize such relationships as part of the explanation process to properly explain solutions and results.

[0015] Third, in the related art, the representation and storage of asset hierarchies, solutions, and information in expert data stores to describe the solutions is done by relational databases, and the use of relational databases is inefficient for representing relationship information and facilitating efficient querying of the results. To address this issue, implementations described herein identify relationships between assets, solutions, and expert data stores, utilize representations that can better capture the relationships, and store the representations for easier access and more efficient querying.

[0016] Several techniques are introduced to address the above problems of the related art. Summary of the Invention [Means for solving the problem]

[0017] One technique in an implementation example involves learning solutions within an asset hierarchy that includes several learning schemes, which are used for both supervised and unsupervised learning to learn and build solutions for assets within the asset hierarchy by utilizing hierarchical relationships between assets.

[0018] Another technique in an example implementation includes solution descriptions within the asset hierarchy, which include root cause attribution models that help explain outcomes at both the solution and individual outcome levels.

[0019] Another technique in an example implementation includes representing and storing solutions within an asset hierarchy, including techniques for representing and storing asset hierarchies, solutions, and information within an expert data store to describe the solutions.

[0020] Aspects of the present disclosure may include a method, the method may include generating an asset hierarchy from a plurality of assets, the asset hierarchy indicating relationships among the plurality of assets from a lowest level to a highest level; running a solution learning process to learn one or more model solutions for each of the plurality of assets based on the relationships among the plurality of assets in the asset hierarchy, wherein output of the one or more model solutions in a lower level of the hierarchy is used as input for the solution learning process to learn one or more model solutions for each of the plurality of assets in a higher level of the asset hierarchy; and storing the one or more model solutions for each of the plurality of assets, knowledge from the generated solution descriptions for the one or more model solutions, and the asset hierarchy in storage as a representation of a solution.

[0021] Aspects of the present disclosure may include a computer program, the computer program may include instructions including: generating an asset hierarchy from a plurality of assets, the asset hierarchy indicating relationships among the plurality of assets from a lowest level to a highest level; executing a solution learning process to learn one or more model solutions for each of the plurality of assets based on the relationships among the plurality of assets in the asset hierarchy, wherein output of one or more model solutions in a lower level of the hierarchy is used as input for the solution learning process to learn one or more model solutions for each of the plurality of assets in a higher level of the asset hierarchy; and storing the one or more model solutions for each of the plurality of assets, knowledge from the generated solution descriptions for the one or more model solutions, and the asset hierarchy as a representation of a solution in storage. The computer program and instructions may be stored on a non-transitory computer-readable medium and executed by one or more processors.

[0022] Aspects of the present disclosure may include a system, the system may include: means for generating an asset hierarchy from a plurality of assets, the asset hierarchy indicating relationships among the plurality of assets from a lowest level to a highest level; means for executing a solution learning process to learn one or more model solutions for each of the plurality of assets based on the relationships among the plurality of assets in the asset hierarchy, wherein outputs of one or more model solutions in lower levels of the hierarchy are utilized as inputs for the solution learning process to learn one or more model solutions for each of the plurality of assets in higher levels of the asset hierarchy; and means for storing the one or more model solutions for each of the plurality of assets, knowledge from the generated solution descriptions for the one or more model solutions, and the asset hierarchy in storage as a representation of a solution.

[0023] Aspects of the disclosure may include an apparatus, the apparatus may include a processor configured to: generate an asset hierarchy from a plurality of assets, the asset hierarchy indicating relationships among the plurality of assets from a lowest level to a highest level; perform a solution learning process to learn one or more model solutions for each of the plurality of assets based on the relationships among the plurality of assets in the asset hierarchy, wherein outputs of the one or more model solutions in lower levels of the hierarchy are utilized as inputs for the solution learning process to learn one or more model solutions for each of the plurality of assets in higher levels of the asset hierarchy; and store the one or more model solutions for each of the plurality of assets, knowledge from the generated solution descriptions for the one or more model solutions, and the asset hierarchy in storage as representations of solutions. [Brief explanation of the drawings]

[0024] [Figure 1] Figure 1 shows an example of an asset hierarchy.

[0025] [Figure 2] FIG. 2 illustrates a solution architecture for solution learning, solution explanation, and solution representation and storage for assets in an asset hierarchy, according to one implementation.

[0026] [Figure 3] FIG. 3 illustrates the system architecture on which the solution is built, implemented, and described, according to one implementation.

[0027] [Figure 4] FIG. 4 shows an example of how a bottom-up learning scheme works within an asset hierarchy, according to one implementation.

[0028] [Figure 5] FIG. 5 shows a flowchart for a learning algorithm in a bottom-up learning scheme according to one implementation.

[0029] [Figure 6] FIG. 6 shows an example of how a reactive learning scheme works within an asset hierarchy, according to one implementation.

[0030] [Figure 7] FIG. 7 shows a flowchart for a learning algorithm within a reactive learning scheme according to one implementation.

[0031] [Figure 8] FIG. 8 shows an example of how a deep learning scheme works within an asset hierarchy, according to one implementation.

[0032] [Figure 9] FIG. 9 shows an example of how the solution description scheme works within an asset hierarchy, according to one implementation.

[0033] [Figure 10] FIG. 10 illustrates an example of data types in a knowledge graph according to one implementation.

[0034] [Figure 11] FIG. 11 illustrates an example process for building a knowledge graph according to one implementation.

[0035] [Figure 12] FIG. 12 illustrates an example of the information stored within each node of a knowledge graph according to one implementation. DETAILED DESCRIPTION OF THE INVENTION

[0036] The following detailed description provides details of the drawings and implementation examples of the present application. Redundant element reference numbers and descriptions between drawings are omitted for clarity. Terms used throughout the description are provided by way of example and are not intended to be limiting. For example, the use of the term "automatic" may include a fully automatic implementation or a semi-automatic implementation that includes user or administrator control over certain aspects of the implementation, depending on the desired implementation of one skilled in the art practicing the implementation of the present application. Selections may be made by a user via a user interface or other input means, or may be implemented by a desired algorithm. The implementations described herein may be utilized alone or in combination, and the functionality of the implementations may be implemented by any means depending on the desired implementation.

[0037] FIG. 2 illustrates a solution architecture for solution learning, solution explanation, and solution representation and storage for assets in an asset hierarchy, according to one implementation. This solution architecture may include the following elements: To build a solution, sensor data 200 is collected and preprocessed. Asset hierarchy 201 contains, captures, represents, and stores relationships between assets. The asset hierarchy may capture physical and / or logical relationships between assets. Solution learning 202 employs several learning schemes to build a model solution (or "solution" for short) based on the sensor data and asset hierarchy. The model solution is a trained model used as a solution for solution explanation 203, including a model from which the solution and solution output can be explained by utilizing the asset hierarchy. Explaining the solution and solution output may include identifying critical factors for predicted results from the trained model in solution learning 202. For example, identifying root causes or critical factors for a predicted failure may be part of the functionality in solution explanation 203. Solution representation and memory 204 is stored in a manner to manage and efficiently query knowledge from solution explanation 203, such as a representation of the asset hierarchy, the solutions from solution learning 202, and a mapping between root causes to recommended actions. Thus, solution representation and memory 204 storage stores knowledge for asset hierarchy 201, one or more model solutions for each of a plurality of assets from solution learning 202, and the solution explanation for one or more model solutions from solution explanation 203.

[0038] 3 illustrates a system architecture upon which the solution is built, implemented, and described, according to one implementation. The system may include the following elements: Assets 300 may include physical assets in an asset hierarchy. Sensors 301 may include sensors mounted on assets. In the implementation, each sensor is attached to a corresponding asset.

[0039] Storage 302 may include a storage device or system that stores asset hierarchy 320, sensor data 321, and solutions and solution results 322.

[0040] The computing unit 303 may be a computing device including a central processing unit (CPU) 330 and memory 331 on which solutions are built, executed, and explained. The computing unit 303 retrieves data from a storage device, builds, executes, and explains solutions, and stores the solutions and results in the storage device. Depending on the amount of data, the number of solutions, and their complexity, the computing unit may include various numbers of CPUs 330 and / or graphics processing units (GPUs), various memory capacities, various numbers and types of computing machines, etc. Parallel computing devices and edge computing devices may also be incorporated herein. Depending on the desired implementation, any hardware configuration that facilitates the functionality of the computing unit 303 may be utilized. The CPU 330 may include one or more processors, such as a hardware processor or a combination of a hardware processor and a software processor.

[0041] Input 304 may include devices used by a data scientist, engineer, operator, etc. to interact with computational unit 303, such as a mouse, keyboard, etc., depending on the desired implementation.

[0042] User interface 305 is software capable of displaying data and results. Output 306 includes devices capable of displaying a user interface, such as a monitor, printer, etc.

[0043] Each element in the solution architecture is discussed in detail below.

[0044] Asset hierarchy 320 captures the relationships between assets in a hierarchical tree structure. As described herein, asset hierarchy 320 may be built based on physical relationships between assets and / or logical or functional relationships between assets. Asset hierarchy 320 may be general purpose for multiple tasks or specific to just one task, depending on the desired implementation.

[0045] The asset hierarchy 320 may be visualized using a hierarchical tree structure. To build solutions using the information in the asset hierarchy 320, implementations represent and store the asset hierarchy 320 in a manner that can be easily and efficiently queried. In the implementations described herein, a knowledge graph is used to represent and store the asset hierarchy 320, as discussed in more detail herein.

[0046] The sensor data 321 may include data from Internet of Things (IoT) sensors mounted on the asset of interest and used to collect data to monitor the asset's health and performance. Different types of sensors are designed to collect different types of data in different industries, assets, and tasks. In this context, it is not necessary to distinguish between sensors, and most sensors can be considered relevant to the solutions presented herein.

[0047] As described herein, in a first aspect, as shown by the flow of FIG. 2, CPU 330 may be configured to: generate an asset hierarchy from a plurality of assets, the asset hierarchy indicating relationships between the plurality of assets from a lowest level to a highest level; execute a solution learning process to learn one or more model solutions for each of the plurality of assets based on the relationships between the plurality of assets in the asset hierarchy, wherein outputs of the one or more model solutions in lower levels of the asset hierarchy are used as inputs for the solution learning process to learn one or more model solutions for each of the plurality of assets in higher levels of the asset hierarchy; and store in storage the asset hierarchy, the one or more model solutions for each of the plurality of assets, and knowledge for solution explanations for the one or more model solutions.

[0048] In a second aspect, as shown with respect to the flow of FIG. 2, the CPU 330 may include the first aspect and may be further configured to generate a solution description for each of the outputs of the model solution for each of the plurality of assets from the top level to the bottom level.

[0049] In a third aspect, the CPU 330 may include any of the above aspects and may be further configured to: generate a first knowledge graph, as shown in Figures 2 and 10-12, the first knowledge graph including a plurality of first nodes and a plurality of first edges, each of the plurality of first nodes representing an asset from the plurality of assets and associated with one or more model solutions for the asset from the plurality of assets, and each of the plurality of first edges representing a relationship between the plurality of assets; generate a second knowledge graph, the second knowledge graph including a plurality of second nodes and a plurality of second edges, each of the plurality of second nodes including knowledge for a solution explanation of a model solution for each of the plurality of assets, and each of the plurality of second edges representing a relationship between the knowledge for the solution explanation for the one or more model solutions; and store the first knowledge graph and the second knowledge graph as a solution representation, thereby storing the asset hierarchy, the one or more model solutions for each of the plurality of assets, and the knowledge for the solution explanation for the one or more model solutions in storage.

[0050] In a fourth aspect, the CPU 330 may include any of the above aspects and may be further configured to generate a solution explanation of the output of the one or more model solutions for each of the plurality of assets from the top level to the bottom level by determining root causes for each output of the one or more model solutions based on one or more of: executing a trace-down process from a top level to a bottom level for each of the plurality of assets; executing an explanation scheme that traverses multiple levels of relationships; and executing a learning scheme by using root causes derived from the output of the one or more model solutions for each of the plurality of assets as targets; and incorporating the root causes as knowledge for the solution explanation for the one or more model solutions, as shown in Figures 4-5 and 10-12.

[0051] In a fifth aspect, the CPU 330 may include any of the above aspects, and as shown in Figures 4 and 5, the solution learning process includes learning one or more model solutions for each of a plurality of assets, initially starting from a lowest level, and outputs of the one or more model solutions at the lower levels are used as inputs to learn one or more model solutions in higher levels of the asset hierarchy, iteratively from the lowest level to the highest level.

[0052] In a sixth aspect, the CPU 330 may include any of the above aspects, and as shown in FIGS. 6 and 7, the solution learning process includes calculating model performance metrics and weights of each input for each of the plurality of assets for one or more model solutions; if the model performance metrics meet the success criteria, continuing the solution learning process to the next asset in each of the plurality of assets; if the model performance metrics do not meet the success criteria, traversing from each of the plurality of assets to assets in lower levels in the asset hierarchy in descending order of weights at each level; for each asset in the traversed plurality of assets, running a broader set of model algorithms and parameter sets to generate a plurality of model solutions; and applying hyperparameter optimization to the plurality of model solutions to select a model solution.

[0053] In a seventh aspect, the CPU 330 may include any of the above aspects, and as shown in FIG. 8, the solution learning process includes generating a deep neural network to represent an asset hierarchy, the deep neural network including an input layer representing sensors associated with a plurality of assets, an output layer representing an asset from a plurality of assets at a top level in the asset hierarchy, and one or more hidden layers representing assets at other levels in the asset hierarchy, wherein connections between layers of the neural network represent one or more physical or logical relationships in the asset hierarchy.

[0054] In an eighth aspect, the CPU 330 may include any of the above aspects, and as shown in FIGS. 4-12 , the solution learning process further includes using model performance metrics from one or more model solutions at a lower level as input to learn one or more model solutions at a higher level, each of the plurality of assets is associated with one or more model solutions for one or more tasks, each of the plurality of assets is associated with one or more versions of the one or more model solutions for each of the one or more tasks, the one or more model solutions are based on one or more machine learning model algorithms or physics-based models, the one or more model solutions are configured to identify and exploit fault-tolerant relationships between one or more of the sensors or assets in the asset hierarchy, some of the one or more of the sensors or assets are configured to have similar functions or roles in the system, the one or more model solutions are configured to capture and exploit cross-level relationships between the assets, the input for the one or more model solutions for the assets is from one or more of the assets or sensors at a different lower level, and the asset hierarchy is refined by removing connections based on feature importance in the one or more model solutions.

[0055] In a ninth aspect, the CPU 330 may include any of the above aspects, and as shown in Figures 5 to 9, the solution learning process further includes traversing assets of the plurality of assets below the current asset by following a descending order of weights using a traversal algorithm (e.g., a depth-first search (DFS) algorithm, a breadth-first search algorithm, etc.), and using weights on connections to restrict the assets of the plurality of assets to be traversed.

[0056] In a tenth aspect, the CPU 330 may include any of the above aspects, and as shown in Figures 4-9, the solution learning process further includes constructing a deep neural network to generate multiple outputs, with each output related to an asset from a plurality of assets in an asset hierarchy, each of the plurality of assets being associated with one or more model solutions related to the one or more tasks, the one or more model solutions being configured to capture and exploit cross-level relationships between the plurality of assets, and links connecting pairs of assets in non-adjacent layers in the deep neural network.

[0057] In an eleventh aspect, the CPU 330 may include any of the above aspects and may be further configured to generate the asset hierarchy using a deep learning scheme, as shown in FIGS. 4 and 8, where generating the asset hierarchy includes identifying an asset among the plurality of assets at each level; generating a fully-connected neural network including a plurality of nodes, where each node among the plurality of nodes at each level is connected to another node among the plurality of levels at a higher level via a plurality of connections; training the fully-connected neural network to obtain a weight for each of the plurality of connections; and pruning the plurality of connections in the fully-connected neural network by removing connections among the plurality of connections having a weight below a predetermined threshold.

[0058] In a twelfth aspect, an asset hierarchy represents one or more of a physical hierarchy or a logical hierarchy of a plurality of assets.

[0059] In a thirteenth aspect, there is a method for carrying out any of the above aspects.

[0060] In a fourteenth aspect, there is a computer program storing instructions for carrying out any of the above aspects. The computer program and instructions can be stored in a non-transitory computer-readable medium and executed by one or more processors.

[0061] In a fifteenth aspect, there is a system having means for carrying out any of the above aspects.

[0062] There are situations where data collected by one sensor S1 is closely related to data collected by another sensor S2. In this case, S1 can be used as a proxy for S2, and vice versa. For example, the wind turbine shaft torque can be approximately represented by the amount of vibration caused by the generator, and vice versa. Such alternative or redundant relationships can be obtained based on domain knowledge and / or data analysis (e.g., correlation analysis). Redundant sensors provide fault tolerance, i.e., if one sensor is not functioning, the other sensor can be used as a proxy to build a solution. Furthermore, if such redundant relationships no longer hold, the faulty sensor can be recognized.

[0063] In solution learning 202, there are various learning schemes for assets in the asset hierarchy. In a first example, there is a bottom-up learning scheme, in which data and results from solutions at lower levels in the asset hierarchy can be used as input to solutions at higher levels. Figure 4 shows an example of how a bottom-up learning scheme works in an asset hierarchy, according to one implementation. In this example, "Asset 11 " is the asset at the highest level (i.e., the root asset), and "Asset 21 ","assets 22 " and "Assets 23 " is the next higher level asset, and so on. Direct relationships between assets are shown by arrows. For example, "Asset 11 " is "Assets 21 ","assets 22 " and "Assets 23 " has a direct relationship with the ". The relationship can be physical and / or logical.

[0064] FIG. 5 illustrates a learning algorithm within a bottom-up learning scheme, according to one implementation. At 501, the algorithm creates an asset hierarchy based on business needs. Depending on the desired implementation, the asset hierarchy can be physical and / or logical / functional, and can be general-purpose or task-oriented. At 502, the algorithm identifies sensors applicable to each asset at the lowest level (i.e., leaf assets). Each asset can be associated with multiple sensors, and each sensor can be associated with multiple assets. At 503, the algorithm first builds a solution for each leaf asset. At 504, the output of the solution at the lower level is used as input to build a solution at the next higher level by traversing the asset hierarchy by the algorithm. The output of the solution at the lower level can be considered as derived features for building a solution at the next higher level. At 505, the process continues until the top-level asset is reached.

[0065] Furthermore, there are several variations of the above bottom-up learning scheme that can be utilized depending on the desired implementation. In one example, model performance metrics can be used as inputs, such that model performance metrics from one or more model solutions at a lower level can be used as inputs to train one or more model solutions at a higher level. When building a solution for an asset, model performance metrics from a solution at a lower level can be used as part of the inputs to build a solution at the next higher level. Intuitively, this indicates the “confidence” of the output from the solution at the lower level. Metrics include, but are not limited to, model-based metrics (e.g., accuracy, precision, recall, etc.) and / or business-based metrics (e.g., time between failures, production yield, etc.). One or more such model performance metrics from a solution at a lower level can be used as inputs to build a solution at the next higher level. Such metrics can be calculated based on historical data, where ground truth target data is available, and can be compared to derived / predicted target values ​​and used to calculate the model-based metrics and / or business-based metrics.

[0066] In another example of a variation, there may be multiple outputs: a solution may be built for each asset in the asset hierarchy, and each asset solution may generate an output to solve a business problem for the asset of interest.

[0067] Another example of a variation can be cross-problem multitasking. There can be multiple business problems in the system. Depending on the desired implementation, each of multiple assets can be associated with one or more model solutions for one or more tasks. With multiple bottom-up learning schemes, multiple tasks can be performed together to solve those problems. To that end, each asset can be associated with several solutions for different tasks. One example is that each asset can be associated with multiple solutions (e.g., anomaly detection, clustering, fault detection, remaining useful life, failure prediction, etc.). The output from each solution for a lower-level asset can be used as input to build a solution for a higher-level asset. A solution for one task for a lower-level asset can help solve all tasks for the higher-level asset.

[0068] Another example of a variation is heterogeneous models. Each task may have several versions of a solution per asset. For example, multiple model algorithms may be applied to each task for an asset. Thus, several solutions are obtained per task for an asset. The output from each lower-level solution may be used as input to build a higher-level solution. For tasks, the solution for a lower-level asset may help the solution for the higher-level asset.

[0069] Another example of a variation is a semi-empirical solution. The solution may be based on a machine learning model algorithm and / or a physics-based model. A machine learning model is a data-driven approach that applies a machine learning model algorithm to available data to build a solution. On the other hand, a physics-based model attempts to capture the underlying relationships between limited variables and form a formula or equation between them, with the target variable based on domain knowledge. Physics-based models are typically validated and fine-tuned through a simulation process. To leverage both solutions, they can be combined into a learning scheme. For each asset, a machine learning solution and a physics-based solution can be built. The output from the machine learning solution and / or the physics-based solution for the lower-level asset can be used as input to build a machine learning solution and / or the physics-based solution for the higher-level asset.

[0070] Another example of a modification is fault tolerance. Some sensors and / or assets in the asset hierarchy have the same or similar functions or roles in the system. This includes substitute sensors as described herein. This promotes fault tolerance, meaning that if one asset does not function, the entire system will continue to function if another asset with the same function or role does. Redundancy relationships can be reflected and captured in the asset hierarchy. Therefore, the solution also reflects such relationships. For example, "asset 21 " and "Assets 22 " has the same function in the system, and "Asset 21 " failure is predicted, but "asset 22 If the failure of the asset is not predicted, 11 " is "Assets 21 " failures. Thus, one or more model solutions may be configured to identify and utilize fault-tolerant relationships between one or more of the sensors or assets in the asset hierarchy, some of which are configured to have similar functions or roles within the system.

[0071] Another example of a variation is cross-level relationships. When a solution is constructed for an asset, the inputs for the solution may come from assets or sensors at different lower levels. Such an asset hierarchy is essentially represented by a graph. There are two cases, as described below:

[0072] In the first case, data from sensors can be used as input to build a solution for each asset in the asset hierarchy. 21 When building a solution for "Asset", sensors related to this asset may be input directly to this asset. For example, "Sensor 1", "Sensor 2", "Sensor 3", and "Sensor 4" may be input directly to "Asset" along with existing inputs from assets at the next lower level. 21 " can be used as input to build a solution for "

[0073] In the second case, the output from a solution for another asset at a lower level can be used as input to build a solution for an asset at a higher level in the asset hierarchy. 11 When building a solution for 231 The output from the solution for " can be used directly as input to build a solution for this asset.

[0074] Another example of a modification is asset hierarchy refinement. When building solutions for assets, feature selection techniques in machine learning and / or explainable artificial intelligence (AI) techniques can be used to determine which features are important or critical for building the solution. Such information can be used to refine the structure of the asset hierarchy, i.e., connections for less important features (corresponding to lower-level sensors or assets) are removed. Note that removing connections in this manner within the asset hierarchy may be problem or task specific and may not apply to another task.

[0075] A second example of a learning scheme is a reactive learning scheme, in which data and results from solutions at lower levels in the asset hierarchy can be used as input to build higher-level solutions, and results from higher-level solutions can be used as feedback to improve the lower-level solutions.

[0076] Figure 6 shows an example of how a reactive learning scheme works within an asset hierarchy, according to one implementation. 11 " is the asset at the highest level (i.e., the root asset), and "Asset 21 ","assets 22 " and "Assets 23 " is the next highest level asset, and so on. Direct relationships between assets are shown with solid arrows. For example, "Asset 11 " is "Assets 21 ","assets 22 " and "Assets 23 " has a direct relationship with

[0077] Each solid arrow is associated with a weight that indicates the contribution of the solution for the asset at the lower level to the solution for the asset at the higher level. The weight is measured per solution for the asset at the higher level. For example, 11 The solution for " contains three weights, "w 21 " is "Assets 21 "Assets" of solutions for 11 ” and contribute to the solution for “w 22 " is "Assets 22 " Assets of the solution 11 ” and contribute to the solution for “w 23 " is "Assets 23 "Assets" of solutions for 11" indicates its contribution to the solution for a model. The contribution can be calculated based on techniques for determining feature importance for a model. For example, in a random forest model, one can calculate how much each feature contributes to reducing weighted impurity or variance based on its position in the tree. Each feature in the model is associated with a feature importance value to indicate its importance in the model or solution. There are various techniques for determining feature importance for different types of model algorithms.

[0078] Feedback from solutions at higher levels to solutions at lower levels is represented by dashed arrows. 11 Feedback from the solution for "Assets" 21 ""assets 22 " and "Assets 23 This can be used to improve the solution for

[0079] FIG. 7 shows a flowchart for a learning algorithm within a reactive learning scheme according to one implementation.

[0080] At 701, the algorithm creates an asset hierarchy based on business needs. The asset hierarchy can be physical and / or logical / functional, and can be general-purpose or task-oriented. At 702, the algorithm identifies sensors applicable to each asset at the lowest level (i.e., leaf assets). Each asset can be associated with multiple sensors, and each sensor can be associated with multiple assets. At 703, the algorithm starts at the lowest level of the asset hierarchy and proceeds to determine whether there are any assets without solutions at 704. If there are (yes), the algorithm selects an asset without a solution at 705 and builds a solution for that asset at 706. By traversing the asset hierarchy, the outputs of lower-level solutions serve as inputs for building the next-higher level solutions. Intuitively, the outputs of the model can be thought of as derived features. Solutions are built up to the assets at the highest level (i.e., the root leaves).

[0081] At 707, the algorithm uses explainable AI techniques to calculate model performance metrics and per-input weights for the current asset. At 708, if the metric values ​​meet predefined success criteria (yes), the algorithm proceeds to 704 to continue with the next asset without a solution within the current level. Otherwise (no), the flow proceeds to 709 to traverse the assets below the current asset by starting with the current asset A and walking the asset hierarchy using a traversal algorithm, such as a breadth-first search (BFS) algorithm.

[0082] In the BFS algorithm, at each level, the algorithm follows the descending order of weights to the assets at the next higher level:

[0083] (a) For each asset traversed, try a broader set of model algorithms and parameter sets and apply hyperparameter optimization to select the best solution.

[0084] (b) Compute model performance metrics for the best solution. If the metric value meets the predefined success criteria, If the current asset is asset A, continue with the next asset without a solution within the current level. Otherwise, rebuild the solution for the parent asset of the current asset, and continue this until asset A is reached. ● Otherwise, go back to (a) for each child asset of the current asset.

[0085] At 710, a determination is made whether a next higher level exists. If so (yes), flow proceeds to 711 to the next level and repeats again from 704. In this manner, the process continues until the top-level assets are reached, and a solution for each asset is constructed with model performance metrics meeting predefined success criteria.

[0086] Furthermore, the modified form of the bottom-up learning scheme can also be applied as a modified form of the reactive learning scheme. In addition, the following two modified forms can also be applied to the reactive learning scheme.

[0087] In one variation, a depth-first search (DFS) algorithm may be used instead of BFS to traverse the assets below the current asset, with the child assets of the current asset being traversed in descending order of weight.

[0088] In one variation, weights can be used to limit the assets that are traversed. When using a traversal algorithm (e.g., BFS algorithm and DFS algorithm) to traverse assets below the current asset, only child assets of the current asset with weights above a predefined threshold are considered. Since weights represent the contribution of the corresponding child assets to the solution of the current asset, the rule here is to consider only important child assets to optimize the solution.

[0089] A third example of a learning scheme is a deep learning scheme, in which a deep neural network is constructed to represent an asset hierarchy, and the neural network is then trained to solve a business problem involving the assets in the asset hierarchy.

[0090] Figure 8 shows an example of how a deep learning scheme works within an asset hierarchy, according to one implementation. 11 " is the asset at the highest level (i.e., the root asset), and "Asset 21 ","assets 22 " and "Assets 23 " is the next highest level asset, and so on. Direct relationships between assets are indicated by arrows. For example, "Asset 11 " is "Assets 21 ","assets 22 " and "Assets 23 " has a direct relationship with

[0091] Below is a description of the learning algorithm within the deep learning scheme.

[0092] First, the learning algorithm builds a deep learning neural network to represent the asset hierarchy. The input layer contains all the sensors on the assets. The input layer is the data layer where all the sensor data in this layer is used. The output layer contains the assets at the highest level in the asset hierarchy. The output layer is the solution layer, which is the assets at the highest level in the asset hierarchy, i.e., "Asset" in Figure 8. 11 The solution for "Asset" in Figure 8 is built within this layer. 11 The " boxes represent smaller neural networks that have an input layer that gets data from lower levels in the asset hierarchy, multiple hidden layers to perform calculations, and an output layer to output the results.

[0093] The hidden layers contain the assets at other levels. Hidden Layer 1 contains the assets at the second lowest level in the asset hierarchy, Hidden Layer 2 contains the assets at the third lowest level in the asset hierarchy, and so on. The hidden layers are solution layers where solutions for each asset in the intermediate levels of the asset hierarchy are constructed. Each asset box in the hidden layer of Figure 8 represents a smaller neural network with an input layer that receives data from sensors and / or assets at lower levels in the asset hierarchy, multiple hidden layers for computation, and an output layer for outputting results to the next level in the asset hierarchy. The connections (or links) between units in one layer to units in the next layer are determined by the physical and / or logical relationship between the two units. For example, if "Sensor 1" is connected to "Asset 2", then the hidden layer is connected to the asset. 211 If only used to build a solution for "Sensor 1" and "Asset 211 There is a connection between

[0094] A learning algorithm then trains the neural network on the given business problem. The entire neural network is trained to solve the business problem for each asset, and a model is built for each asset in the hidden and output layers. Standard training processes for neural networks are applied, including forward and back propagation. In addition, all techniques for deep learning in neural networks can be applied to the training process.

[0095] The learning algorithm then makes predictions based on the neural network model and the sensor data, which is collected and preprocessed, and the neural network is applied to the preprocessed sensor data to generate results from the output layer.

[0096] Additionally, there are several variations on the above deep learning scheme. In one variation, there are multiple outputs. A solution can be built for each asset in the asset hierarchy, and the solution for each asset can result in an output to solve the business problem of interest. In Figure 8, each box in the hidden and output layers can be a solution to a business problem for an asset. An alternative solution is to build a smaller neural network up to the level where the output is expected. For example, if the output of layer 2 is desired, the algorithm builds a model only up to layer 2 and uses the problem in layer 2 as the target.

[0097] Another example of a modification is multitasking. There may be multiple business problems in the system, and in this learning scheme, multiple solutions may be built together to solve those problems. Each of multiple assets may be associated with one or more model solutions for one or more tasks. To that end, each asset may be associated with several solutions for various tasks. One example is that each asset is associated with multiple solutions: anomaly detection, clustering, fault detection, remaining useful life, failure prediction, etc. The output from each solution for a lower-level asset may be input to a solution for a higher-level asset. The solution of one task for a lower-level asset may aid in the solution of another task for a higher-level asset. In FIG. 8, each asset in the hidden and output layers has multiple outputs.

[0098] Another example of a variation is a deep neural network (DNN). Each layer in the neural network maps to a level in the asset hierarchy, and there may be several levels in the asset hierarchy. Furthermore, each asset has a smaller neural network that also contains several layers. The whole is a deep network. Depending on the desired implementation, the DNN can be constructed to generate multiple outputs from multiple assets in the asset hierarchy, with each output being related to an asset.

[0099] Another example of a modification is cross-level relationships. In standard neural networks, connections are between the current layer and the immediately next layer. However, some deep learning neural networks also include cross-layer connections applicable here. In other words, the neural network proposed in this learning scheme allows connections between non-adjacent layers. Thus, in an example implementation, one or more model solutions are configured to capture and utilize cross-level relationships between assets, and inputs for one or more model solutions related to an asset come from one or more assets or sensors at different lower levels. Furthermore, one or more model solutions can be configured to capture and utilize cross-level relationships between multiple assets, with links connecting pairs of assets in non-adjacent layers in the deep neural network.

[0100] Another example of a modified form may be a fully connected neural network. FIG. 8 shows a neural network structure in which connections are based on the relationships between assets in the asset hierarchy. However, sometimes such relationships may not be clear or accurate in advance. Once the sensors and assets in each level are determined in the asset hierarchy, each sensor or asset in a level may be connected to each asset in the next level. Thus, a level or layer is fully connected to the next layer or level. Then, during training, the weights of the connections are determined accordingly. A large weight indicates a strong connection, and a small weight indicates a weak connection. If the weight of a connection is close to 0, it indicates that the connection may not be necessary. A predetermined threshold may be used for the weight to determine whether the corresponding connection is necessary. This relaxes the requirements on the relationships for building a solution for assets in the asset hierarchy. As a by-product, the modified form of the fully connected neural network helps build the asset hierarchy once the assets in each level are determined.

[0101] An alternative hybrid solution is to start with a fully connected neural network and remove as many connections as possible based on high confidence domain knowledge, then train such a neural network and remove connections with weights below a predetermined threshold. As a result, the asset hierarchy can be refined by removing connections based on feature importance within one or more model solutions.

[0102] With respect to the solution description 203, once a solution is constructed, it is used to generate results for the business problem. However, those results may not directly translate or map to some actionable conclusion. For example, a solution may generate predicted failures for an asset of interest. The predicted failures are presented with a failure score that indicates the severity of the failure, but do not include any information about the cause of the failure and therefore recommended actions to repair the failure. To repair the predicted failures, the root causes of the predicted failures need to be derived, and repair actions can be generated based on the root causes and domain knowledge.

[0103] Explainable AI can be used to derive the root causes of each outcome from a solution for an asset. Several libraries are available for implementing explainable AI. For example, ELI5 or SHAP can be used to derive the root causes for the outcomes of a trained model. Such libraries are designed to explain outcomes from one model or solution each time. Such libraries can be used to explain solutions and outcomes from each solution with respect to assets in the asset hierarchy. Furthermore, a scheme can be introduced to explain solutions and outcomes from solutions for assets in the asset hierarchy by incorporating relationships between assets.

[0104] FIG. 9 illustrates an example of how an explainable AI scheme works within an asset hierarchy, according to one implementation. In particular, FIG. 9 illustrates the execution of an explanation scheme across multiple relationship levels, and the execution of a learning scheme by targeting root causes derived from the failure output of solutions for each of multiple assets. In this example, "Asset 11 " is the asset at the highest level (i.e., the root asset), and "Asset 21 ","assets 22 " and "Assets 23 " is the next highest level asset, and so on. Direct relationships between assets are indicated by arrows. For example, "Asset 11 " is "Assets 21 ","assets 22 " and "Assets 23 " has a direct relationship with

[0105] An example of an explainable AI algorithm scheme for solution explanation 203 is shown below as a top-down tracing scheme.

[0106] First, the algorithm starts with the asset at the highest level in the asset hierarchy. Based on the solution for this asset, a result is derived for the data instance, and the root cause of the result is derived based on a root cause analysis model. For example, ELI5 or SHAP can be used to derive the root cause of the result.

[0107] Second, based on the derived root cause, identify the assets at the lower level that lead to the root cause. Root cause analysis is performed on the identified assets, and the root cause is derived. This step is repeated until the lowest level of the asset hierarchy (i.e., the level with sensors) is reached.

[0108] For example, in Figure 9, 11 " based on the solution for ", the root cause of the predicted results is "Asset 21 ", then the route "p 21 " is followed. Then, "Asset 21" based on the solution for ", the root cause of the predicted results is "Asset 213 ", then the route "p 213 " is traced. Finally, "Asset 213 Based on the solution for ", if the root cause is due to "sensor 4", then path "ps4" is followed.

[0109] This explanation scheme applies to both bottom-up and reactive learning schemes. In deep learning schemes, the root analysis model can be applied directly to the entire neural network to infer root causes at the sensor level.

[0110] Variations on the above top-down tracing explanation scheme may include the following.

[0111] An example of a variation is solution description at different levels. Solutions and results from the solutions can be described for each asset in the asset hierarchy. An implementation can start with any asset in the asset hierarchy and apply a tracedown scheme to that asset. Furthermore, root causes can be attributed to assets or sensors at different levels in the asset hierarchy. For example, in the example above, an implementation can describe the path "p 21 " and then "Assets 21 " and the implementation stopped. The root cause was "Asset 21 " can be concluded.

[0112] In another example of modification, there is also a cross-level explanation. If the learning scheme includes cross-level relationships, the explanation of the cross-level relationships can be enabled accordingly. For example, in Figure 9, "Asset 11 Based on the solution for "," if the root cause of the result is due to "sensor 7," the example implementation follows path "ps7" to determine that the root cause is related to "sensor 7."

[0113] In another example of a variant, the root causes are also within the outcome of the solution. When a solution is built for an asset, the implementation directly includes the root causes as targets. For example, when predicting failures for an asset, in addition to the failure score, the implementation also includes the failure modes (i.e., the root causes for the failures) as targets. This variant requires only one solution for the asset, instead of one solution to generate the outcome and another solution to generate the root causes.

[0114] Regarding solution representation and storage 204, there is representation and storage of information within the asset hierarchy. An implementation includes a knowledge graph, which is an efficient way to represent and store information related to solutions within the asset hierarchy. The information includes mappings between root causes to recommended actions from the asset hierarchy 201, solutions from solution learning 202, and solution descriptions 203.

[0115] FIG. 10 illustrates an example of data types in a knowledge graph according to one implementation. Depending on the type of data that needs to be represented in the knowledge graph, different processes (i.e., manual or automated) can be used. Once the knowledge graph is constructed, it is stored in a graph database known in the art. For example, the graph database can be Neo4J, ArangoDB, or the like, depending on the desired implementation. For example, solution learning 202 can be represented by generating a first knowledge graph including a plurality of first nodes and a plurality of first edges, where each of the plurality of first nodes represents an asset from a plurality of assets and is associated with one or more model solutions for the asset from the plurality of assets, and each of the plurality of first edges represents a relationship between the plurality of assets. Solution learning 203 can be represented by generating a second knowledge graph, where the second knowledge graph includes a plurality of second nodes and a plurality of second edges, where each of the plurality of second nodes includes knowledge for a solution explanation of a model solution for each of the plurality of assets, and each of the plurality of second edges represents a relationship between the knowledge for the generated solution explanation for the one or more model solutions. Such knowledge graphs are stored as solution representations in solution representation and storage 204 .

[0116] As shown in the data types for knowledge graphs in Figure 10, the data that needs to be represented and stored can be of different types. For example, in structured data types, data is stored in a table structure where columns represent variables and rows represent instances. Examples of structured data may include relational table formats or CSV data. In semi-structured data types, data does not follow a table structure but is represented by a structured data format that includes tags or other markers that separate semantic elements and enforce a hierarchy of records and fields within the data. For example, "XML" and "JSON" are two types of semi-structured data. In unstructured data types, data does not have a predefined data model or is not organized in a predefined way. For example, text, images, video, or audio data are unstructured data.

[0117] The knowledge graph can be built by manual or automated processes for all data types, except for unstructured data, which can only be built by automated processes.

[0118] Regarding the components in a knowledge graph, a knowledge graph is a directed graph that may contain nodes and edges. Nodes represent items or entities (physical or logical), such as staff, electrical systems, assets, etc. Edges capture the relationship of interest between two nodes and connect them in the graph, such as an asset. 11 is an asset 21 or problem A is caused by root cause B.

[0119] Figure 11 shows an example process for building a knowledge graph according to one implementation. Building a knowledge graph is essentially understanding the entities and relationships between them, and representing the entities as nodes and the relationships as edges. As shown in Figure 11, there are two processes or paths for building a knowledge graph.

[0120] Manual Process: An implementation can follow steps 1100 to 1104 manually as shown in "Manual 1", "Manual 2", "Manual 3", and "Manual 4" to build the knowledge graph. In the final step of 1104, the ontology graph is populated with data to obtain the knowledge graph. Entities and relationships between entities need to be manually identified and represented by the knowledge graph.

[0121] Automated Process: An implementation can follow "Auto 1" to build a knowledge graph from 1100 directly to 1104, which is done by a language model based on natural language processing techniques. Entities and relationships between entities can be automatically identified by natural language processing techniques and represented by a knowledge graph. Optionally, "Auto 2" from 1104 to 1103 is for extracting the knowledge graph into an ontology graph.

[0122] In the following description, the example implementation lists the information that is stored in the knowledge graph and used during the solution learning and explanation process.

[0123] Regarding storing the asset hierarchy in the knowledge graph, given a system, it is first necessary to identify the assets and their relationships based on domain knowledge. To build the knowledge graph for the system through a manual process, assets can be represented as nodes and relationships as edges.

[0124] First, an implementation identifies all assets at 1100. Then, the implementation identifies a list of unique assets in the system. At 1102 and 1103, asset types and relationships (hierarchical and non-hierarchical) between asset types are identified. This step includes building a taxonomy tree 1102 for hierarchical relationships and / or building an ontology graph 1103 for non-hierarchical relationships. Finally, an implementation may populate the ontology graph at 1104 with data to build a knowledge graph. If there are not many assets in the system, the implementation may directly identify relationships between assets. Thus, the implementation would directly go through processes 1100 through 1101-1104 to build the knowledge graph.

[0125] With respect to storing solutions for assets in a knowledge graph, each asset is associated with one or more model solutions for different tasks with different solution versions. In a knowledge graph that represents and stores an asset hierarchy, an implementation may add one or more entries to represent a solution. Thus, an implementation stores solutions together with assets in the asset hierarchy.

[0126] FIG. 12 illustrates an example of information stored within each node of a knowledge graph according to one implementation. Depending on the desired implementation, each asset in the asset hierarchy may have several tasks (e.g., anomaly detection, failure prediction, remaining useful life), and for each task there may be several versions of a solution (e.g., for the task of failure prediction, there may be a solution based on random forest classification, a solution based on sequence prediction from a recurrent neural network (RNN), a physics-based model based on domain knowledge and simulation), or any other machine learning model algorithm or physics-based model. As a result, for each node in the knowledge graph, such information is stored accordingly for that asset, as shown in FIG. 12. Thus, each of a plurality of assets may be associated with one or more versions of one or more model solutions for each of one or more tasks.

[0127] Regarding using a knowledge graph for solution explanation 203, root causes need to be mapped to recommended actions to explain the solution and outcome. An implementation may use a knowledge graph to store all domain knowledge for explaining the solution and outcome.

[0128] Building a knowledge graph can be done by a manual or automated process on structured or unstructured data. Each entity can be represented by a node and relationships can be represented by edges. For example, for a diagnostic problem, an implementation can store the following information: "Symptom" → "Problem" → "Root Cause" → "Repair Action," where each "Symptom," "Problem," "Root Cause," and "Repair Action" is represented by a node in the knowledge graph, and the relationship between each pair of entities is represented by an edge in the knowledge graph.

[0129] To query results given some keywords, an implementation can use fuzzy queries, lucence index queries, or sentence embedding queries. Fuzzy queries and lucence index queries are typically provided as part of a graph database. In a sentence embedding query, an implementation considers a keyword phrase as a sentence and obtains a sentence embedding. The query calculates the similarity between the sentence embedding of the data in the knowledge graph and the sentence embedding of the keyword phrase, and outputs the most similar results.

[0130] Further, the implementation extracts the knowledge graph into an ontology graph and a taxonomy tree. A tag is assigned to each entity in the knowledge graph. This assignment can be done manually based on domain knowledge or automatically based on some tagging algorithm. The implementation can then cluster the entities based on the tags assigned to each entity. Clusters of entities can be thought of as entity types. The implementation adds further connections between entity types based on existing connections between entities. Entity types with hierarchical relationships form a taxonomy tree. Entity types with non-hierarchical relationships form an ontology graph.

[0131] Furthermore, the implementation introduces an explanation methodology for solutions and results based on asset hierarchy, which can help explain results at various levels (from coarse to fine). Additionally, the implementation introduces the use of knowledge graphs to represent and store asset hierarchies and information needed to explain solutions. The implementation helps resolve relationships between assets and build asset hierarchies accordingly, and the proposed solutions can help refine and optimize the asset hierarchy.

[0132] 13 illustrates a system including multiple systems with connected sensors and a management device, according to one implementation. One or more sensor systems 1301-1, 1301-2, 1301-3, and 1301-4 are communicatively coupled to a network 1300 that is connected to a management device 1302 that facilitates functionality for an Internet of Things (IoT) gateway or other management system. The management device 1302 manages a database 1303 that contains historical data collected from the sensor systems 1301-1, 1301-2, 1301-3, and 1301-4, which may include labeled and unlabeled data received from the systems 1301-1, 1301-2, 1301-3, and 1301-4. In an alternative implementation, data from sensor systems 1301-1, 1301-2, 1301-3, 1301-4 may be stored in a central repository or database, such as a proprietary database that captures data from an enterprise resource planning system, or the like, and management equipment 1302 may access or retrieve the data from the central repository or database. Such systems may include robotic arms with sensors, turbines with sensors, lathes with sensors, etc., depending on the desired implementation. Management equipment 1302 may take the form of a system as indicated by computing unit 302 and storage 303.

[0133] The implementations described herein employ several learning schemes for asset hierarchies by utilizing the physical and / or logical relationships between assets. This can help achieve better performance of the solution for a given task for each asset, provide a comprehensive health status of the entire system, and prioritize tasks accordingly. This can also help fine-tune the solution for each asset. The implementations described herein further address the problems of the related art outlined herein.

[0134] Some portions of the detailed description are presented in terms of algorithms and symbolic representations of operations within a computer. These algorithmic descriptions and symbolic representations are the means used by those skilled in the data processing arts to convey the essence of their innovations to others skilled in the art. An algorithm is a sequence of defined steps leading to a desired end state or result. In one implementation, the steps performed require physical manipulations of tangible quantities to achieve a tangible result.

[0135] Unless otherwise specified, as will be apparent from the description, the description utilizing terms such as "processing," "computing," "calculating," "determining," "displaying," and the like throughout the description will be understood to include the actions and processes of a computer system or other information processing device that manipulates and converts data represented as physical (electronic) quantities in the registers and memory of the computer system into other data similarly represented as physical quantities in the memory or registers of the computer system, or other information storage, transmission, or display device.

[0136] Implementations may also relate to apparatuses for performing the operations described herein. This apparatus may be specially constructed for the required purposes, or it may comprise one or more general-purpose computers selectively activated or reconfigured by one or more computer programs. Such computer programs may be stored in a computer-readable medium, such as a computer-readable storage medium or a computer-readable signal medium. Computer-readable storage media may include tangible media, such as, but not limited to, optical disks, magnetic disks, read-only memory, random-access memory, solid-state devices and drives, or any other type of tangible or non-transitory medium suitable for storing electronic information. Computer-readable signal media may include media such as carrier waves. The algorithms and displays presented herein are not inherently related to any particular computer or other apparatus. A computer program may include a pure software implementation containing instructions for performing the operations of a desired implementation.

[0137] Various general-purpose systems may be used with the programs and modules according to the examples herein, or it may prove convenient to construct more specialized apparatus to perform the desired method steps. Additionally, the implementations are not described with reference to any particular programming language. It will be understood that a variety of programming languages ​​may be used to implement the techniques of the implementations described herein. Instructions in the programming language may be executed by one or more processing units, such as a central processing unit (CPU), processor, or controller.

[0138] As is known in the art, the operations described above may be performed by hardware, software, or some combination of software and hardware. Various aspects of the implementations may be implemented using circuits and logic (hardware), while other aspects may be implemented using instructions stored on a machine-readable medium (software), which, when executed by a processor, cause the processor to perform methods for carrying out the implementations of the present application. Furthermore, some implementations of the present application may be performed exclusively in hardware, while other implementations may be performed exclusively in software. Furthermore, the various functions described may be performed in a single unit or distributed among several components in any number of ways. When performed by software, the methods may be executed by a processor, such as a general-purpose computer, based on instructions stored on a computer-readable medium. If desired, the instructions may be stored on the medium in compressed and / or encrypted format.

[0139] Additionally, other implementations of the present application will be apparent to those skilled in the art upon consideration of this specification and practice of the present techniques. Various aspects and / or components of the described implementations may be used alone or in any combination. It is intended that the specification and implementations be considered as examples only, with the true scope and spirit of the present application being indicated by the following claims.

Claims

1. A method performed by an apparatus, comprising: generating an asset hierarchy from a plurality of assets, the asset hierarchy indicating relationships among the plurality of assets from a lowest level to a highest level; executing a solution learning process to learn one or more model solutions for each of the plurality of assets based on the relationships between the plurality of assets in the asset hierarchy, wherein outputs of the one or more model solutions in lower levels of the asset hierarchy are used as inputs for the solution learning process to learn the one or more model solutions for each of the plurality of assets in higher levels; storing the asset hierarchy, the one or more model solutions for each of the plurality of assets, and solution explanation knowledge for the one or more model solutions in a storage; Including, generating the solution description for each of the outputs of the model solution for each of the plurality of assets from the top level to the bottom level; Storing the asset hierarchy, the one or more model solutions for each of the plurality of assets, and the knowledge for the solution explanations for the one or more model solutions in storage includes: generating a first knowledge graph, the first knowledge graph including a plurality of first nodes and a plurality of first edges, each of the plurality of first nodes representing an asset from the plurality of assets and associated with the one or more model solutions for the asset from the plurality of assets, and each of the plurality of first edges representing the relationship between the plurality of assets; generating a second knowledge graph, the second knowledge graph including a plurality of second nodes and a plurality of second edges, each of the plurality of second nodes including the knowledge for the solution description of the model solution for each of the plurality of assets, and each of the plurality of second edges representing a relationship between the knowledge for the solution description for the one or more model solutions; storing the first knowledge graph and the second knowledge graph as the solution description; A method comprising:

2. A method performed by an apparatus, comprising: generating an asset hierarchy from a plurality of assets, the asset hierarchy indicating relationships among the plurality of assets from a lowest level to a highest level; executing a solution learning process to learn one or more model solutions for each of the plurality of assets based on the relationships between the plurality of assets in the asset hierarchy, wherein outputs of the one or more model solutions in lower levels of the asset hierarchy are used as inputs for the solution learning process to learn the one or more model solutions for each of the plurality of assets in higher levels; storing the asset hierarchy, the one or more model solutions for each of the plurality of assets, and solution explanation knowledge for the one or more model solutions in a storage; Including, generating the solution description for each of the outputs of the model solution for each of the plurality of assets from the top level to the bottom level; generating the solution description for the outputs of the one or more model solutions for the each of the plurality of assets from the top level to the bottom level, determining the root cause for each output of the one or more model solutions based on one or more of: executing a tracedown process for each of the plurality of assets from a higher level to a lower level to derive a root cause from an output of a model solution for the asset based on a root cause analysis model; executing an explanation scheme for deriving a root cause across levels of relationships of the plurality of assets from the output of a model solution for the asset based on a root cause analysis model; and executing a learning scheme for constructing the one or more model solutions for each of the plurality of assets to include a root cause as an output, and for outputting the root cause from the model solution; incorporating the root causes as the knowledge for the solution explanation for the one or more model solutions; A method comprising:

3. In the method of claim 1 or claim 2, The solution learning process includes: learning the one or more model solutions for said each of said plurality of assets starting from said lowest level; Including, the outputs of the one or more model solutions at lower levels of the asset hierarchy are used as inputs to train the one or more model solutions in higher levels, iteratively from the lowest level to the highest level. method.

4. In the method of claim 1 or claim 2, The solution learning process includes: calculating model performance metrics for the one or more model solutions and weights for each of the inputs for each of the plurality of assets; If the model performance metrics meet success criteria, continuing the solution learning process to the respective next asset of the plurality of assets; and If the model performance metrics do not meet the success criteria, from said each of said plurality of assets, traversing assets of said plurality of assets at said lower levels in said asset hierarchy in order of decreasing weight at each level; for each of the assets in the traversed plurality of assets, executing a broader set of model algorithms and parameter sets to generate a plurality of model solutions, and applying hyperparameter optimization to the plurality of model solutions to select the model solution; A method comprising:

5. In the method of claim 1 or claim 2, The solution learning process includes: generating a deep neural network to represent the asset hierarchy, the deep neural network comprising: an input layer representing sensors associated with the plurality of assets; an output layer representing assets of the plurality of assets at the highest level in the asset hierarchy; one or more hidden layers representing assets at other levels in the asset hierarchy; connections between layers of the neural network represent one or more physical or logical relationships within the asset hierarchy; method.

6. The method of claim 3, The solution learning process includes: using model performance metrics from the one or more model solutions at the lower level as the input for training the one or more model solutions at the higher level. Further comprising: each of the plurality of assets is associated with the one or more model solutions for one or more tasks; each of the plurality of assets is associated with one or more versions of the one or more model solutions for each of the one or more tasks; the one or more model solutions are based on one or more of a machine learning model algorithm or a physics-based model; the one or more model solutions are configured to identify and utilize fault-tolerant relationships between one or more of the sensors or assets in the asset hierarchy, some of the one or more of the sensors or assets being configured to have similar functions or roles within the system; the one or more model solutions are configured to capture and utilize cross-level relationships between assets, and the inputs for the one or more model solutions for an asset are from one or more of the assets or sensors at different sub-levels; the asset hierarchy is refined by removing connections based on feature importance within the one or more model solutions. method.

7. 5. The method of claim 4, The solution learning process includes: using the model performance metrics from the one or more model solutions at the lower level as the input for training the one or more model solutions at the higher level. Further comprising: each of the plurality of assets is associated with the one or more model solutions for one or more tasks; each of the plurality of assets is associated with one or more versions of the one or more model solutions for each of the one or more tasks; the one or more model solutions are based on one or more of a machine learning model algorithm or a physics-based model; the one or more model solutions are configured to identify and utilize fault-tolerant relationships between one or more of the sensors or assets in the asset hierarchy, some of the one or more of the sensors or assets being configured to have similar functions or roles within the system; the one or more model solutions are configured to capture and utilize cross-level relationships between assets, and the inputs for the one or more model solutions for an asset are from one or more of the assets or sensors at a different sub-level; the asset hierarchy is refined by removing connections based on feature importance within the one or more model solutions. method.

8. 5. The method of claim 4, The solution learning process includes: traversing assets of the plurality of assets below a current asset by following the descending order of the weights using a traversal algorithm; using the weights for connections to restrict which assets of the plurality of assets are traversed; The method further comprises:

9. 6. The method of claim 5, The solution learning process includes: constructing the deep neural network to generate a plurality of outputs, with each output relating to an asset from the plurality of assets in the asset hierarchy; Further comprising: each of the plurality of assets is associated with the one or more model solutions for one or more tasks; the one or more model solutions are configured to capture and exploit cross-level relationships between the plurality of assets, with links connecting pairs of assets in non-adjacent layers within the deep neural network; method.

10. In the method of claim 1 or claim 2, generating the asset hierarchy through a deep learning scheme; generating the asset hierarchy includes: identifying an asset from the plurality of assets at each level; generating a fully connected neural network including a plurality of nodes, wherein a node of the plurality of nodes at each level is connected to another node of the plurality of levels that is at a higher level by a plurality of connections; training the fully connected neural network to obtain a weight for each of the plurality of connections; pruning the plurality of connections in the fully connected neural network by removing connections among the plurality of connections having a weight below a predetermined threshold; A method comprising:

11. In the method of claim 1 or claim 2, The method, wherein the asset hierarchy represents one or more of a physical hierarchy or a logical hierarchy of the plurality of assets.

12. 1. A computer program storing instructions for carrying out a process, the instructions comprising: generating an asset hierarchy from a plurality of assets, the asset hierarchy indicating relationships among the plurality of assets from a lowest level to a highest level; running a solution learning process to learn one or more model solutions for each of the plurality of assets based on the relationships between the plurality of assets in the asset hierarchy, wherein outputs of the one or more model solutions in a lower level of the asset hierarchy are used as inputs for the solution learning process to learn the one or more model solutions for each of the plurality of assets in a higher level; storing the asset hierarchy, the one or more model solutions for each of the plurality of assets, and solution explanation knowledge for the one or more model solutions in a storage; Including, generating the solution description for each of the outputs of the model solution for each of the plurality of assets from the top level to the bottom level; Storing the asset hierarchy, the one or more model solutions for each of the plurality of assets, and the knowledge for the solution explanations for the one or more model solutions in storage includes: generating a first knowledge graph, the first knowledge graph including a plurality of first nodes and a plurality of first edges, each of the plurality of first nodes representing an asset from the plurality of assets and associated with the one or more model solutions for the asset from the plurality of assets, and each of the plurality of first edges representing the relationship between the plurality of assets; generating a second knowledge graph, the second knowledge graph including a plurality of second nodes and a plurality of second edges, each of the plurality of second nodes including the knowledge for the solution description of the model solution for each of the plurality of assets, and each of the plurality of second edges representing a relationship between the knowledge for the solution description for the one or more model solutions; storing the first knowledge graph and the second knowledge graph as the solution description; a computer program comprising:

13. A computer program storing instructions for carrying out a process, the instructions comprising: generating an asset hierarchy from a plurality of assets, the asset hierarchy indicating relationships among the plurality of assets from a lowest level to a highest level; executing a solution learning process to learn one or more model solutions for each of the plurality of assets based on the relationships between the plurality of assets in the asset hierarchy, wherein outputs of the one or more model solutions in lower levels of the asset hierarchy are used as inputs for the solution learning process to learn the one or more model solutions for each of the plurality of assets in higher levels; storing the asset hierarchy, the one or more model solutions for each of the plurality of assets, and solution explanation knowledge for the one or more model solutions in a storage; Including, generating the solution description for each of the outputs of the model solution for each of the plurality of assets from the top level to the bottom level; generating the solution description for the outputs of the one or more model solutions for the each of the plurality of assets from the top level to the bottom level, determining the root cause for each output of the one or more model solutions based on one or more of: executing a tracedown process for each of the plurality of assets from a higher level to a lower level to derive a root cause from an output of a model solution for the asset based on a root cause analysis model; executing an explanation scheme for deriving a root cause across levels of relationships of the plurality of assets from the output of a model solution for the asset based on a root cause analysis model; and executing a learning scheme for constructing the one or more model solutions for each of the plurality of assets to include a root cause as an output, and for outputting the root cause from the model solution; incorporating the root causes as the knowledge for the solution explanation for the one or more model solutions; a computer program comprising:

14. 1. An apparatus comprising: generating an asset hierarchy from a plurality of assets, the asset hierarchy indicating relationships among the plurality of assets from a lowest level to a highest level; executing a solution learning process to learn one or more model solutions for each of the plurality of assets based on the relationships between the plurality of assets in the asset hierarchy, wherein an output of the one or more model solutions in a lower level of the asset hierarchy is used as an input for the solution learning process to learn the one or more model solutions for each of the plurality of assets in a higher level; and storing the asset hierarchy, the one or more model solutions for each of the plurality of assets, and solution explanation knowledge for the one or more model solutions in storage. and generating the solution description for each of the outputs of the model solution for each of the plurality of assets from the top level to the bottom level; Storing the asset hierarchy, the one or more model solutions for each of the plurality of assets, and the knowledge for the solution explanations for the one or more model solutions in storage includes: generating a first knowledge graph, the first knowledge graph including a plurality of first nodes and a plurality of first edges, each of the plurality of first nodes representing an asset from the plurality of assets and associated with the one or more model solutions for the asset from the plurality of assets, and each of the plurality of first edges representing the relationship between the plurality of assets; generating a second knowledge graph, the second knowledge graph including a plurality of second nodes and a plurality of second edges, each of the plurality of second nodes including the knowledge for the solution description of the model solution for each of the plurality of assets, and each of the plurality of second edges representing a relationship between the knowledge for the solution description for the one or more model solutions; storing the first knowledge graph and the second knowledge graph as the solution description; a processor configured to include 1. An apparatus comprising:

15. An apparatus comprising: generating an asset hierarchy from a plurality of assets, the asset hierarchy indicating relationships among the plurality of assets from a lowest level to a highest level; executing a solution learning process to learn one or more model solutions for each of the plurality of assets based on the relationships between the plurality of assets in the asset hierarchy, wherein an output of the one or more model solutions in a lower level of the asset hierarchy is used as an input for the solution learning process to learn the one or more model solutions for each of the plurality of assets in a higher level; and storing the asset hierarchy, the one or more model solutions for each of the plurality of assets, and solution explanation knowledge for the one or more model solutions in storage. and generating the solution description for each of the outputs of the model solution for each of the plurality of assets from the top level to the bottom level; generating the solution description for the outputs of the one or more model solutions for the each of the plurality of assets from the top level to the bottom level, determining the root cause for each output of the one or more model solutions based on one or more of: executing a tracedown process for each of the plurality of assets from a higher level to a lower level to derive a root cause from an output of a model solution for the asset based on a root cause analysis model; executing an explanation scheme for deriving a root cause across levels of relationships of the plurality of assets from the output of a model solution for the asset based on a root cause analysis model; and executing a learning scheme for constructing the one or more model solutions for each of the plurality of assets to include a root cause as an output, and for outputting the root cause from the model solution; incorporating the root causes as the knowledge for the solution explanation for the one or more model solutions; a processor configured to include 1. An apparatus comprising:

Images