Predictive engine maintenance apparatus, methods, systems, and techniques

Abstract

A method includes inputting used oil analysis data to a pre-trained predictive model, the used oil analysis data including values quantifying a plurality of chemical components measured in a sample of used oil taken from an engine being analyzed; determining, in response to the used oil analysis data, a probability of at least one fault code using the pre-trained predictive model, the at least one fault code corresponding to one of a plurality of predetermined engine fault types; providing the at least one fault code and the probability of the at least one fault code to an expert system; performing, using the expert system, a root cause analysis of the at least one fault code to determine a root cause indicative of a preventative maintenance action; and performing the preventative maintenance action on the engine being analyzed.

Description

Background Technology

[0001] This disclosure relates to predictive engine maintenance apparatuses, methods, systems, and techniques. Predictive engine maintenance apparatuses, methods, systems, and techniques seek to predict future engine failure events (e.g., failure or complete malfunction of one or more engine systems or components requiring maintenance, repair, and / or replacement to restore engine function) and identify and provide maintenance, repair, or replacement before such future engine failure events occur. Numerous proposals have been made for predictive engine maintenance apparatuses, methods, systems, and techniques; however, existing proposals have many drawbacks, defects, and unrealized potential. There remains a substantial need for the unique apparatuses, methods, systems, and techniques disclosed herein.

[0002] Disclosure of illustrative embodiments For the purpose of clearly, concisely, and accurately describing the illustrative embodiments of this disclosure, the manner and process of making and using this disclosure, and to enable the practice, making, and use of this disclosure, reference will now be made to certain exemplary embodiments, including those illustrated in the accompanying drawings, and these exemplary embodiments will be described using specific language. However, it should be understood that this is not intended to limit the scope of the invention, and that the invention includes and protects such changes, modifications, and further applications of the exemplary embodiments that would occur to those skilled in the art. Summary of the Invention

[0003] One embodiment is a unique predictive engine maintenance process. Another embodiment is a unique predictive engine maintenance system. The predictive maintenance system and process according to this disclosure may include system features and process operations related to data preparation, classification modeling, recommendation modeling, expert system analysis, and a web-enabled user interface providing predictive maintenance capabilities for the evaluated engine. Further embodiments, forms, objects, features, advantages, aspects, and benefits will become apparent from the following description and figures. Attached Figure Description

[0004] Figure 1 This is a schematic diagram illustrating certain aspects of an example predictive maintenance system.

[0005] Figure 2 This is a schematic diagram illustrating certain aspects of an example predictive maintenance system.

[0006] Figure 3 This is a flowchart illustrating some aspects of an example predictive maintenance process.

[0007] Figure 4 This is a diagram illustrating some aspects of an example user interface used in predictive maintenance systems. Detailed Implementation

[0008] refer to Figure 1 The illustration depicts certain aspects of a predictive maintenance system 100 according to an example embodiment. System 100 includes a combination of components including one or more classification models 104, one or more recommendation models 108, and an expert system 112. Each of these components, as well as other components of system 100, can be implemented in one or more computer systems, including one or more computers specifically configured to provide the configurations and functions described herein according to the techniques of this disclosure. In some forms, one or more of these components, as well as other components of system 100, can be provided on one or more dedicated computing systems, such as a web-accessible cloud computing platform, a local computing system like a maintenance facility, or partially distributed within a web-accessible cloud computing platform and partially distributed on one or more dedicated computing systems.

[0009] In system 100, used oil analysis data 102 is input to and received by one or more classification models 104. The used oil analysis data 102 includes values ​​quantified from samples of used oil obtained from engines analyzed by the predictive maintenance system 100. Examples of such measured properties include measurements of viscosity (e.g., kinematic viscosity at 40°C, 100°C, or other predetermined temperatures), soot content, aluminum (Al), chromium (Cr), copper (Cu), iron (Fe), lead (Pb), tin (Sn), nickel (Ni), silicon (Si), sodium (Na), potassium (K), and additional and alternative values ​​indicating other physical, elemental, and / or chemical properties, components, or properties measured in samples of used oil obtained from engines analyzed by the predictive maintenance system 100. Some embodiments preferably utilize a set of measurement properties, including viscosity and measurements of the content of one or more components, wherein the component content is selected from the group consisting of soot content, Al content, Cr content, Cu content, Fe content, Pb content, Sn content, Ni content, Si content, Na content, and K content. Some embodiments preferably utilize a set of measurement properties selected from the group consisting of measurements of viscosity, soot content, Al content, Cr content, Cu content, Fe content, Pb content, Sn content, Ni content, Si content, Na content, and K content. Some embodiments preferably utilize a set of measurement properties consisting of measurements of viscosity, soot content, Al content, Cr content, Cu content, Fe content, Pb content, Sn content, Ni content, Si content, Na content, and K content.

[0010] Measurements of component content used to provide the oil analysis data 102 can be performed using optical emission spectroscopy (OES) techniques such as rotating disk electrode optical emission spectroscopy (RDE-OES) or inductively coupled plasma optical emission spectroscopy (ICP-OES), microscopy-based instruments (such as scanning electron microscopy / energy dispersive X-ray analysis (SEM / EDX)), or other component content measurements. It should be further understood that the results of component content measurements can be expressed in fractional or percentage terms, or in absolute terms such as parts per million (ppm), parts per billion (ppb), other fractional symbols, or other absolute terms based on mass, weight, or volume.

[0011] One or more classification models 104 are examples of components pre-trained using one or more machine learning techniques, such as those described below, before receiving oil analysis data 102. Figure 2 The described technique. One or more classification models 104 perform a classification operation in response to oil analysis data 102 and the pre-trained attributes of the classification models 104, thereby determining and outputting one or more fault code probabilities 106 (in...). Figure 2 The classification is denoted as P(fc1), P(fc2), ..., P(fcn). In some forms, the classification operation can be a two-class classification. In other forms, the classification operation can be a higher-order multi-class classification. Each of the one or more fault code probabilities indicates the probability that a predetermined unique engine failure event (such as a failure of one or more specific engine components or one or more specific failure types) will occur before a predetermined future time, for example, the probability that it will occur before the next scheduled oil change or other scheduled maintenance of the engine analyzed by system 100. In some embodiments, the classification model(s) 104 preferably includes multiple classification models selected during the training of the machine learning model, for example, in combination as follows. Figure 2 As described. The multiple classification models can represent those classifications for which a minimum probability threshold or criterion has been established during training. One or more fault code probabilities 106 are input to and received by the recommendation model 108.

[0012] Recommendation model 108 uses one or more machine learning techniques (such as, combined below) before receiving the used oil analysis data 102. Figure 2Examples of pre-trained components (described by the technique). Recommendation model 108 determines and outputs a relevant item dataset 110 in response to fault code probabilities 106 and its pre-trained attributes. The relevant item dataset 110 includes a set of one or more fault codes that are related to a given fault code probability in the fault code probabilities 106, for example, through causality or correlation. Recommendation model 108 is configured to perform an evaluation to identify fault codes that are adjacent to fault codes corresponding to one of the one or more fault code probabilities 106 input to and received by recommendation model 108. In some forms, this evaluation may use techniques such as collaborative filtering from pre-trained information to generate relevant items and may continue learning after initial training. The relevant item dataset 110 is input to and received by expert system 112.

[0013] Expert system 112 is an AI-based system configured to mimic the decision-making of human experts. Expert system 112 is configured to use an inference engine to identify the root causes of fault codes in a relevant project dataset 110. This inference engine operates on a knowledge base primarily represented by a set of if-then rules, rather than through conventional algorithms. Expert system 112 can be configured to combine the following... Figure 2 The described techniques are used for pre-training and configuration. Expert system 112, in response to a relevant project dataset 110 and its pre-trained attributes, determines and outputs one or more maintenance action datasets 114. The maintenance action datasets 114 define one or more maintenance actions according to root causes indicating failure, damage, or degradation of one or more specific components of the engine analyzed by predictive maintenance system 100, and then maintain, repair, or replace these components via the execution of the maintenance actions.

[0014] System 100 is an example of a machine learning system according to this disclosure, comprising at least three machine learning model components or layers. In the example of System 100, the provision of at least three machine learning model components (i.e., one or more classification models 104, one or more recommendation models 108, and expert system models 112) provides numerous unexpected benefits relative to other methods, including those related to accuracy, reliability, speed, and trainability. These at least three machine learning model components are trained individually to model separate aspects of the overall machine learning-based maintenance prediction system (i.e., classifying oil analysis data 102 to determine one or more fault code probabilities 106, determining a related item dataset 110 including one or more related fault codes in response to the one or more fault code probabilities 106, and determining one or more maintenance action datasets 114 in response to the related item dataset 110).

[0015] refer to Figure 2 The illustration shows certain aspects of an example predictive maintenance system 200 undergoing multiple example machine learning model training operations 210. It should be understood that the machine learning model training operations 210 described in conjunction with the predictive maintenance system 200 can be used to configure, define, and / or provide the above-described combination. Figure 1 The described corresponding system components include one or more classification models 104, one or more recommendation models 108, and expert system models 112. During training operations, one or more classification models 212 are trained using an oil analysis training dataset 202 and a fault training dataset 204. One or more other training datasets 206 may also be used to train the classification models 212, including, for example, engine control module (ECM) data, telematics data, engine oil sensor data, or other engine sensor data. Such data can provide a wide variety of information about engine performance, such as fuel economy, idle percentage, and other data.

[0016] Oil analysis training dataset 202 includes values ​​quantified from engine oil samples obtained during oil change maintenance events for multiple engines of a common type or model. In this example, oil analysis training dataset 202 includes approximately 248,000 data points and 38,000 records from engine oil changes of the same engine type or model over a three-year period. In this example, each row or record in the dataset includes values ​​for engine identification (e.g., engine serial number), oil miles (engine mileage at the time of a major oil change), soot content, viscosity (e.g., kinematic viscosity at 100°C), Al content, Cr content, Cu content, Fe content, Pb content, Sn content, Ni content, Si content, Na content, and K content, as well as additional and alternative values ​​for other physical, elemental, and / or chemical properties, components, or properties measured in samples of used oil obtained from the engine. In this example, the engine serial number is used as a unique identifier to link the oil data to other datasets. The data points in the oil analysis training dataset 202 can be represented as O i,j This corresponds to the j-th oil change of the i-th engine.

[0017] Fault training dataset 204 includes a set of engine fault codes indicating one of several predetermined fault types for the same plurality of engines and serving as part of oil analysis training dataset 202 over the corresponding time period. This information can be included in or determined from maintenance and / or warranty data. In this example, fault training dataset 204 contains approximately 45,000 records or rows corresponding to the same group of engines across the same three-year time period as oil analysis training dataset 204. Each record or row includes values ​​for engine identification (e.g., engine serial number), repair mileage (engine mileage at the time of the repair event), and fault code (e.g., an identifier indicating a specific component or system fault observed during an engine service or repair event associated with the fault event). Data points in fault training dataset 204 can be represented as R... i,k This corresponds to the k-th repair of the i-th engine.

[0018] Because each oil analysis data point O i,j Not the initial and fault data point R i,k Correlated, therefore each oil analysis data point O i,j It can be obtained through fault data point R of the same engine. i,k The pairing process is expanded, and each paired data point (oil data point) can be provided with a fault code f according to Equation 1 below. o i,j Additional information: f o i,j = fr i,k , where min (m r i,j - m o i,k ), (m r i,j - m o i,k ) ≥ 0 (Equation 1) In equation 1, m r Indicates repair mileage, m o This represents the oil mileage, and min(m r i,j - m o i,k ) and (m r i,j - m o i,k ≥ 0 is a pairwise selection criterion for choosing repair mileage values, where the difference between repair mileage and oil mileage is minimal, and repair mileage is greater than oil mileage. It should be understood that this is an example of identifying the relationship between a given fault code and oil or other data described over time by comparing oil mileage to fault code engine mileage. A further example could be to correlate all oil data prior to the fault code occurrence to generate this model, rather than only correlating the oil data closest to the fault code occurrence.

[0019] One or more of the classification models 212 can be trained to provide a two-class classification model, wherein f o i,j Oil data that is not null will be included in Category 1 (faults observed using fault codes), while data with f o i,j Oil data with a value of null will be included in category 0 (no fault observed). It is also conceivable that one or more of the classification models 212 can be trained to provide a higher-order multi-class classification model. Many types of binary classification models can be utilized, including, for example, logistic regression, support vector machines, neural networks, and augmented decision tree models. For some applications, such as this example, using a binary augmented decision tree (BDT) classification model provides unexpectedly superior results. In this example, each of the classification models 212 classifies a unique fault code, although other examples may utilize higher-order multi-class classification models.

[0020] The fault training dataset 204 comprises multiple instances of fault codes from an overall set of 500 possible unique fault codes. The most frequently occurring fault codes (e.g., the top 30, or another number or percentage defined as a cutoff or threshold) are selected and modeled using a two-class classification model. The area under the curve (AUC) of the resulting probability distribution is used to further select certain fault codes and their associated classification models. In this example, fault code-based models with an AUC greater than 80% are selected for use in subsequent testing and evaluation operations, as shown in Table 1 below. Fault Codes #Example AUC FC008 1048 0.977 FC022 6482 0.960 FC133 3629 0.928 FC025 3561 0.891 FC406 596 0.891 FC073 1064 0.890 FC014 238 0.880 FC066 1051 0.863 FC207 8319 0.857 FC442 538 0.855 FC058 1692 0.832 FC109 666 0.819 FC278 413 0.809 FC328 591 0.807 FC081 371 0.803 FC209 460 0.801 Table 1 .

[0021] Machine learning model training operation 210 may also include training a recommendation model 214. Based on the classification results, the most probable fault codes (or multiple fault codes) that may occur before the next oil change can be identified. Furthermore, each fault code may have a correlation or causal relationship, and they may occur sequentially or in a clustered manner. Recommendation model 214 can be configured to model the evaluation using large-scale online Bayesian recommendation techniques to identify relevant items. To train and evaluate such a model, a new dataset is derived from the warranty dataset using the mapping defined in Equation 2. Engine ID → User ID, Fault Code → Project ID, # (Fault Code) for each (Engine ID, Fault Code) → Rate (Equation 2) Using the training techniques described above, recommendation model 214 is trained and configured to generate a table of related items with 500+ rows. Each row has six columns, which correspond to an item and its related items 1 through 5. The model's mean normalized depreciation cumulative gain (NDCG) is 0.95. Machine learning model training operation 210 may also include training recommendation model 214.

[0022] The machine learning model training operation 210 may also include the training of an expert system 216, which is a computer-implemented system that mimics the decision-making abilities of human experts. The expert system 216 is configured to solve complex problems by applying inference engine-based reasoning to a knowledge base with a set of if-then rules, rather than through conventional procedural codes. To configure and train the expert system 216, knowledge from several engine experts is collected regarding the root causes of the first 16 fault codes discussed above in Table 1. The expert knowledge of the root causes is represented in the format of if-then rules. For example, for the asserted preceding fault codes (f... i ), where i is an element of I (iєI), for matching Rule kThe inference engine, where k is an element of K (kєK), Rule k Assertions about consequences (causes) c j Where j is an element of J (jєJ), Rule k It can be expressed according to Equation 3. Rule k :f i →c j (Equation 3) In Equation 3, I, J, and K ≤ 1. This process can also be called forward chaining. To obtain a set of fault codes associated with the same cause, expert system 216 can use reverse chaining, which declares that if the system attempts to determine c j If it is true, then it will find the Rule among all k (k ⊆ K). k And query the knowledge base to see if there is any f i (iєI) is true.

[0023] refer to Figure 3 and 4 The diagram illustrates a flowchart depicting an example predictive maintenance process 300 and certain aspects of a user interface display 400 that can be used in conjunction with process 300. Process 300 is initiated at operation 301 and proceeds to operation 302, in which used oil analysis data is input into and received by a pre-trained predictive model. The used oil analysis data may include values ​​quantified in a sample of used oil taken from the engine being analyzed, such as those described above in conjunction with system 100 or system 200. Figure 4 As illustrated, the oil analysis data used can be displayed in Table 410 of the user interface display 400, or, additionally or alternatively, can be displayed as a graph, chart or other graphic on the user interface display 400.

[0024] At operation 304, the pre-trained predictive model invokes and initiates one or more classification models (represented as N classifiers, where N≥1), such as one or more of classification models 104. These one or more classification models can be invoked, initiated, and executed in parallel; it should be understood that sequential or serial operations, as well as partially sequential or partially serial operations, are also possible. Operation 302 determines the probabilities of one or more fault codes input to and received by condition 306, which evaluates whether the received fault codes are true and whether their probabilities are greater than an established threshold. The probabilities of fault codes for which condition 306 is true can be provided as a classification result set 308, which can be displayed in table 420 of the user interface display 400, or, additionally or alternatively, can be displayed as a graph, chart, or other graphic on the user interface display 400.

[0025] One or more fault code probabilities included in the classification result set 308 are provided to operation 310. The fault code probabilities provided to operation 310 may be selected based on predetermined rules (such as maximum probability) or based on user input (such as selection of probabilities on user interface display 400). Operation 310 invokes and initiates the operation of a recommendation model (such as recommendation model 108). The recommendation model performs an evaluation to identify fault codes that are related to or adjacent to the one or more fault codes provided to operation 310. The fault codes and their related or adjacent fault codes may be provided as a related fault code set 312, which may be displayed in table 430 of user interface display 400, or, additionally or alternatively, may be displayed as a graph, chart, or other graphic on user interface display 400.

[0026] A set of relevant fault codes 312 is provided to operation 314, which invokes and initiates the operation of an inference engine (such as expert system 112 or the inference engine of another expert system). Expert system 112 identifies one or more root causes in response to received input. These root causes are provided as root causes and a set of relevant fault codes 316, and may be displayed in tables 440 and 450 of user interface display 400, or, additionally or alternatively, may be displayed as graphs, charts, or other graphics on user interface display 400. Expert system 112 may utilize a relevant item dataset to determine the root causes instructing preventative maintenance actions, for example, by determining the root cause corresponding to an item in the relevant item dataset, or the root cause associated with an item in the relevant item dataset.

[0027] Further descriptions of several example embodiments are provided below. A first example embodiment is a method comprising: inputting used oil analysis data into a pre-trained predictive model, the used oil analysis data including values ​​of a plurality of chemical components measured in a sample of used oil obtained from an engine being analyzed; in response to the used oil analysis data, using the pre-trained predictive model to determine the probability of at least one fault code, the at least one fault code corresponding to one of a plurality of predetermined engine fault types; providing the at least one fault code and the probability of the at least one fault code to an expert system; using the expert system to perform a root cause analysis of the at least one fault code to determine a root cause indicative of preventive maintenance action; and performing the preventive maintenance action on the engine being analyzed.

[0028] The second example embodiment includes features of the first example embodiment and includes: determining a relevant item dataset in response to the probability of the at least one fault code, the relevant item dataset including one or more other fault codes that are related or causally related to the at least one fault code; and providing the relevant item dataset to the expert system, wherein the expert system uses the relevant item dataset to determine the root cause instructing the preventive maintenance action.

[0029] The third example embodiment includes features of the first example embodiment, wherein the pre-trained predictive model comprises one or more classification models. An additional form of the third example embodiment further includes features of the second example embodiment.

[0030] The fourth example embodiment includes features of the third example embodiment, wherein the one or more classification models include an augmented decision tree model.

[0031] The fifth example embodiment includes features of the first example embodiment, wherein the pre-trained predictive model is trained using an oil analysis training dataset, which includes values ​​of multiple chemical components measured using engine oil samples obtained during oil change maintenance events for multiple engines of a common type or model. Additional forms of the fifth example embodiment further include features of the second example embodiment. Additional forms of the fifth example embodiment further include features of the third example embodiment. Additional forms of the fifth example embodiment further include features of the second and third example embodiments. Additional forms of the fifth example embodiment further include features of the third and fourth example embodiments. Additional forms of the fifth example embodiment further include features of the second, third, and fourth example embodiments.

[0032] The sixth example embodiment includes features of the fifth example embodiment, wherein the pre-trained predictive model is trained using an engine failure training dataset, the engine failure training dataset comprising a set of engine failure codes indicating one of a plurality of predetermined failure types of the plurality of engines and serving as an oil analysis training dataset within a corresponding time period.

[0033] The seventh example embodiment includes features of the first example embodiment, wherein the pre-trained predictive model is trained using an engine failure training dataset, the engine failure training dataset comprising a set of engine failure codes indicating one of a plurality of predetermined failure types for a plurality of engines.

[0034] The eighth example embodiment includes features of any one of the first to sixth example embodiments, wherein the expert system root cause analysis is configured to use an inference engine to identify the root cause of the fault code, the inference engine operating on a knowledge base represented by a rule set of if-then rules.

[0035] The ninth example embodiment includes features of any one of the first to sixth example embodiments, wherein the expert system root cause analysis is performed using a forward linking operation.

[0036] The tenth example embodiment includes features of any one of the first to sixth example embodiments, wherein the expert system root cause analysis is performed using a backlink operation.

[0037] The eleventh example embodiment is a system for predictive engine maintenance, the system comprising: a pre-trained predictive model component configured to: receive input including used oil analysis data, and in response to the used oil analysis data determine the probability of at least one fault code, the used oil analysis data including values ​​of a plurality of chemical components measured in a sample of used oil obtained from the engine being analyzed, the at least one fault code corresponding to one of a plurality of predetermined engine fault types; and an expert system model component configured to: receive input including the at least one fault code and the probability of the at least one fault code, perform root cause analysis of the at least one fault code to determine a root cause, and in response to the root cause indicate preventive maintenance actions for the engine being analyzed.

[0038] The twelfth example embodiment includes features of the eleventh example embodiment and includes a recommendation model component configured to: determine a relevant item dataset in response to the probability of the at least one fault code, the relevant item dataset including one or more other fault codes that are correlated or causally related to the at least one fault code, wherein the expert system component is configured to receive the relevant item dataset and determine the root cause instructing the preventive maintenance action.

[0039] The thirteenth example embodiment includes features of the eleventh example embodiment, wherein the pre-trained predictive model component comprises one or more classification models. An additional form of the thirteenth example embodiment further includes features of the twelfth example embodiment.

[0040] The fourteenth example embodiment includes the features of the thirteenth example embodiment, wherein one or more classification models include an augmented decision tree model.

[0041] The fifteenth example embodiment includes features of the eleventh example embodiment, wherein the pre-trained predictive model component is trained using an oil analysis training dataset, which includes values ​​quantified from multiple chemical components measured using engine oil samples obtained during oil change maintenance events for multiple engines of a common type or model. An additional form of the fifteenth example embodiment further includes features of the twelfth example embodiment. An additional form of the fifth example embodiment further includes features of the thirteenth example embodiment. An additional form of the fifth example embodiment further includes features of both the twelfth and thirteenth example embodiments. An additional form of the fifth example embodiment further includes features of both the thirteenth and fourteenth example embodiments. An additional form of the fifth example embodiment further includes features of the twelfth, thirteenth, and fourteenth example embodiments.

[0042] The sixteenth example embodiment includes features of the fifteenth example embodiment, wherein the pre-trained predictive model component is trained using an engine failure training dataset, the engine failure training dataset comprising a set of engine failure codes indicating one of a plurality of predetermined failure types of the plurality of engines and serving as an oil analysis training dataset within a corresponding time period.

[0043] The seventeenth example embodiment includes features of the eleventh example embodiment, wherein the pre-trained predictive model component is trained using an engine failure training dataset, the engine failure training dataset comprising a set of engine failure codes indicating one of a plurality of predetermined failure types of a plurality of engines.

[0044] The eighteenth example embodiment includes features of any one of the eleventh to sixteenth example embodiments, wherein the expert system component is configured to use an inference engine to identify the root cause of the fault code, the inference engine operating on a knowledge base represented by a rule set of if-then rules.

[0045] The eighteenth example embodiment includes features of any one of the eleventh to sixteenth example embodiments, wherein root cause analysis is performed using a forward linking operation.

[0046] The eighteenth example embodiment includes features of any one of the eleventh to sixteenth example embodiments, wherein root cause analysis is performed using a backlink operation.

[0047] While illustrative embodiments of the present disclosure have been detailed and described in the accompanying drawings and foregoing description, they should be considered illustrative and not restrictive. It is to be understood that only certain exemplary embodiments have been shown and described, and protection is intended for all changes and modifications falling within the spirit of the claimed invention. It should be understood that while the use of terms such as preferred, preferably, or more preferred in the foregoing description indicates that a feature so described may be more desirable, it may not be necessary, and embodiments lacking such features may be contemplated within the scope of the invention, defined by the following claims. When reading the claims, it is intended that the use of terms such as “a,” “an,” “at least one,” or “at least a portion” is not intended to limit the claims to only one item, unless specifically stated otherwise in the claims. When the language “at least a portion” and / or “a portion” is used, the item may include a portion and / or the entire item, unless specifically stated otherwise.

Claims

1. A method for predictive engine maintenance, comprising: A computing system is provided with a combination of models, the combination of models including a pre-trained classification model pre-trained by machine learning to output one or more probabilities of one or more fault codes in response to input including oil analysis data; a recommendation model configured to receive the output of the pre-trained classification model and pre-trained by machine learning to output a relevant project dataset in response to one or more probabilities of one or more fault codes; and an expert system model configured to receive the output of the recommendation model and pre-trained by machine learning to output an indication of the root cause of preventive maintenance actions in response to the relevant project dataset. Used oil analysis data is input into a pre-trained classification model, the used oil analysis data including values ​​of multiple chemical components measured in a sample of used oil obtained from the engine being analyzed; In response to the oil analysis data already used, the probability of at least one fault code is determined using the pre-trained classification model, the at least one fault code corresponding to one of a plurality of predetermined engine fault types; In response to at least one fault code, a recommendation model is used to determine the relevant item dataset; The relevant project dataset, the at least one fault code, and the probability of the at least one fault code are provided to the expert system model; The root cause analysis of the at least one fault code is determined by using the expert system model in response to the relevant project dataset, the at least one fault code, and the probability of the at least one fault code to determine the root cause that indicates preventive maintenance actions. as well as Perform the preventative maintenance action on the engine being analyzed.

2. The method of claim 1, wherein the related project dataset includes one or more other fault codes that are related to the at least one fault code.

3. The method of claim 2, wherein the related project dataset includes one or more other fault codes that are causally related to the at least one fault code.

4. The method of claim 3, wherein the pre-trained classification model comprises an augmented decision tree model.

5. The method of claim 1, wherein the pre-trained classification model is trained using an oil analysis training dataset, the oil analysis training dataset comprising values ​​of multiple chemical components quantified using engine oil samples obtained during oil change maintenance events for multiple engines of a common type or model.

6. The method of claim 5, wherein the pre-trained classification model is trained using an engine fault training dataset, the engine fault training dataset comprising a set of engine fault codes indicating one of a plurality of predetermined fault types of the plurality of engines and serving as an oil analysis training dataset within a corresponding time period.

7. The method of claim 1, wherein the pre-trained classification model is trained using an engine fault training dataset, the engine fault training dataset comprising a set of engine fault codes indicating one of a plurality of predetermined fault types of a plurality of engines.

8. The method according to any one of claims 1-6, wherein the expert system model is configured to use an inference engine to identify the root cause of the fault code, the inference engine operating on a knowledge base represented by a rule set of if-then rules.

9. The method according to any one of claims 1-6, wherein the expert system model is configured to use forward linking operations.

10. The method according to any one of claims 1-6, wherein the expert system model is configured to use a backlink operation.

11. A system for predictive engine maintenance, the system comprising: One or more computers configured to implement a combination of model components, the combination of model components including a pre-trained classification model component pre-trained by machine learning to output one or more probabilities of one or more fault codes in response to input including oil analysis data; a recommendation model component configured to receive the output of the pre-trained classification model and pre-trained by machine learning to output a relevant project dataset in response to one or more probabilities of one or more fault codes; and an expert system model component configured to receive the output of the recommendation model component and pre-trained by machine learning to output an indication of the root cause of preventive maintenance actions in response to the relevant project dataset. A pre-trained classification model component is configured to receive input including used oil analysis data, and in response to the used oil analysis data, determine the probability of at least one fault code, the used oil analysis data including values ​​of multiple chemical components measured in a sample of used oil obtained from the engine being analyzed, the at least one fault code corresponding to one of multiple predetermined engine fault types; The recommendation model component is configured to determine the relevant item dataset in response to at least one fault code; as well as An expert system model component is configured to: receive input including a relevant project dataset, the at least one fault code, and the probability of the at least one fault code; perform root cause analysis of the at least one fault code in response to the relevant project dataset, the at least one fault code, and the probability of the at least one fault code; and indicate preventive maintenance actions for the analyzed engine in response to the root cause.

12. The system of claim 11, wherein the related project dataset includes one or more other fault codes that are related to the at least one fault code.

13. The system of claim 11, wherein the related project dataset includes one or more other fault codes that are causally related to the at least one fault code.

14. The system of claim 13, wherein the pre-trained classification model component includes an augmented decision tree model.

15. The system of claim 11, wherein the pre-trained classification model component is trained using an oil analysis training dataset, the oil analysis training dataset comprising values ​​of multiple chemical components quantified using engine oil samples obtained during oil change maintenance events for multiple engines of a common type or model.

16. The system of claim 15, wherein the pre-trained classification model component is trained using an engine fault training dataset, the engine fault training dataset comprising a set of engine fault codes indicating one of a plurality of predetermined fault types of the plurality of engines and serving as an oil analysis training dataset within a corresponding time period.

17. The system of claim 11, wherein the pre-trained classification model component is trained using an engine fault training dataset, the engine fault training dataset comprising a set of engine fault codes indicating one of a plurality of predetermined fault types of a plurality of engines.

18. The system according to any one of claims 11-16, wherein the expert system model component is configured to use an inference engine to identify the root cause of the fault code, the inference engine operating on a knowledge base represented by a rule set of if-then rules.

19. The system according to any one of claims 11-16, wherein root cause analysis is performed using forward linking operations.

20. The system according to any one of claims 11-16, wherein root cause analysis is performed using a reverse link operation.

Images