Predicting performance degradation with nonlinear characteristics
A failure prediction model effectively addresses nonlinear performance degradation challenges by clustering data with anomaly scores, improving prediction accuracy and reducing false positives in mechanical systems.
Patent Information
- Application Number
- JP2022562596
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2020-05-11
- Filing Date
- 2021-05-05
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2041-05-05
AI Technical Summary
Predicting nonlinear performance degradation in mechanical systems, such as wear-induced degradation in progressive cavity pumps, is challenging due to non-linear relationships, imbalanced training data, and the gradual nature of failures, which complicates accurate prediction and data labeling.
A computer-implemented method using a failure prediction model that clusters data samples into multiple clusters with normal and abnormal scores, calculates anomaly scores, and outputs classifications based on comparisons, without relying on ANNs or oversampling techniques, utilizing a lookback window and prediction range to create training samples.
Accurately predicts nonlinear performance degradation from imbalanced datasets, providing precise quantification of abnormality and reducing false positives, enabling proactive maintenance.
Smart Images

Figure 0007577127000007 
Figure 0007577127000008 
Figure 0007577127000009
Abstract
Description
[Technical field]
[0001] The present disclosure relates to predictive modeling, and more particularly, to predicting performance degradation with non-linear characteristics. [Background technology]
[0002] Nonlinear performance degradation may include phenomena associated with gradual degradation of one or more mechanical systems or elements, which may ultimately result in failure or persistent suboptimal performance of one or more mechanical systems or elements. Such gradual degradation associated with nonlinear performance degradation may be associated with any type of wear (e.g., weakening due to gradual removal or deformation of a component based on interaction with another material), fatigue (e.g., weakening resulting from cyclic loading), creep (e.g., deformation resulting from sustained mechanical stress), or other nonlinear phenomena, or combinations thereof. Nonlinear performance degradation may be caused by mechanical, chemical, thermal, or other stresses. For example, wear phenomena may include abrasive wear, erosive wear, corrosive wear, and other types of wear.
[0003] However, predicting nonlinear performance degradation such as wear-induced deterioration poses many challenges. As an example, wear-out is a gradual failure that progresses over a long period of time. Accordingly, the relationship between the healthy and worn-out state is nonlinear, making linear models (e.g., Naive Bayes, Support Vector Machines (SVMs), etc.) inapplicable. Furthermore, the progression of wear-out failures is highly variable among similar assets, for example, depending on the usage characteristics (e.g., some assets fail within 50 days, while others fail within 6 months). As an example, wear-out-related performance degradation in progressing cavity pumps (PCPs) may depend on multiple factors, such as, for example, the type of underground formation, the grain size of the sand, or the operational profile of any particular PCP, or a combination thereof. Thus, the level of performance degradation as a function of time and failure at the same point in time may vary among different assets based on various operational factors.
[0004] Another challenge in predicting nonlinear performance degradation is related to the imbalance between normal and faulty operating conditions of an asset experiencing nonlinear performance degradation, such as wear-out-related performance degradation. In other words, the training data set exhibiting normal performance dominates (e.g., more than 90%), while a very small percentage of data exhibits faulty performance. Making accurate predictions from imbalanced training data is notoriously difficult. For example, utilizing highly parameterized nonlinear methods (e.g., Artificial Neural Networks (ANNs)) is not a sensible solution insofar as an abundance of data (including sufficient examples of both classes) is required to be able to fine-tune and optimize the vast number of parameters in the ANN. However, when the data is highly imbalanced, the number of examples in abnormal conditions is insufficient to effectively train the ANN. To ameliorate the imbalanced data problem, oversampling and / or undersampling techniques can be employed, but these techniques may ultimately distort the original distribution in the data and thus bias the solution (e.g., reduced accuracy due to increased false positive indications).
[0005] An additional challenge associated with accurately predicting nonlinear performance degradation is that nonlinear performance degradation, such as wear-related performance degradation, is not necessarily a catastrophic failure. Returning again to the PCP example, a worn PCP will not necessarily cease functioning, but will function suboptimally due to rotor wear. This makes the "failure date" for replacing a pump at a particular level of performance degradation (as opposed to a catastrophic failure date) subjective to the operator's judgment. This further complicates data labeling (e.g., determining when to label data from a PCP as a failure when in fact the PCP continues to function at substandard performance).
[0006] Yet another challenge associated with predicting nonlinear degradation in real-world applications relates to properly identifying the fault signature. This challenge is two-fold. First, the data available to identify the fault signature must be evaluated. This evaluation may involve inferring information from data that is not necessarily directly related to the nonlinear degradation. For example, assets prone to wear-related degradation are rarely explicitly instrumented to directly measure wear. Returning again to the PCP example, the interaction between speed, production rate, torque, and casing pressure may be the only information available to infer the fault signature. Second, the exclusivity of the identified fault signature must be evaluated. In other words, the identified fault signature may be correlated with two or more phenomena, thus increasing false positives. In light of the above, it can be seen that nonlinear degradation poses the further challenge of (i) identifying the fault signature from available data on nonlinear degradation, and (ii) determining whether the identified fault signature is exclusively (or primarily) representative of the nonlinear degradation.
[0007] The combination of the aforementioned problems makes the problem of predictive modeling of nonlinear performance degradation difficult to solve. Thus, a need exists for techniques that accurately predict performance degradation due to nonlinear phenomena. Summary of the Invention
[0008] Aspects of the present disclosure are directed to a computer-implemented method that includes inputting a new data sample into a failure prediction model. The failure prediction model is trained using a labeled historical data set. Each data point is associated with a lookback window and a prediction range to create each training sample. Each training sample is clustered into a plurality of clusters, and each of the plurality of clusters is associated with a normal score and an anomaly score. The method further includes outputting a classification associated with the new data sample based on comparing a first anomaly score of a first cluster of the plurality of clusters that includes the new data sample to an average anomaly score of the clusters of the plurality of clusters having a normal score greater than the anomaly score.
[0009] The present invention preferably can accurately predict non-linear performance degradation from imbalanced training data sets. Additionally, the normality and abnormality scores preferably can quantify the degree of abnormality associated with a classification. A further advantage of the present invention is that it does not rely on ANNs (which tend to over-parameterize imbalanced training data) nor does it rely on over- or under-sampling techniques of imbalanced training data (which tend to bias classification).
[0010] According to a preferred embodiment, the present disclosure further includes a lookback window defining the number of consecutive previous data points to include in each training sample. The lookback window can advantageously adjust the size of a feature indicative of normal or abnormal behavior. For example, a shorter lookback window may be more sensitive to individual data points, while a longer lookback window may be less sensitive to individual data points.
[0011] According to a preferred embodiment, the method further includes a prediction horizon defining a predefined amount of time into the future, and each label of each data point the predefined amount of time into the future being associated with each training sample. The prediction horizon can advantageously link various features defined by the lookback window to corresponding future outcomes. For example, a shorter prediction horizon may result in a shorter warning for a particular prediction (e.g., one day before a wear-related failure), whereas a longer prediction horizon may result in a longer warning for a particular prediction (e.g., one month before a wear-related failure).
[0012] According to a preferred embodiment, the method further comprises clustering each training sample using K-means clustering, which is advantageously an efficient and scalable clustering technique.
[0013] According to one aspect, a system is provided comprising one or more processors and one or more computer-readable storage media having program instructions stored thereon, the program instructions, when executed by the one or more processors, configured to cause the one or more processors to perform a method including inputting a new data sample into a failure prediction model, the failure prediction model being trained using a labeled historical dataset, where each data point is associated with a lookback window and a prediction range to create each training sample, where each training sample is clustered into a plurality of clusters, where each of the plurality of clusters is associated with a normal score and an anomaly score, respectively; and outputting a classification associated with the new data sample based on comparing a first anomaly score of a first cluster of the plurality of clusters including the new data sample to an average anomaly score of the plurality of clusters having a normal score greater than the anomaly score.
[0014] According to one aspect, a computer program product is provided comprising one or more computer-readable storage media and program instructions collectively stored on the one or more computer-readable storage media, the program instructions including instructions configured to cause one or more processors to execute a method including inputting a new data sample into a failure prediction model, the failure prediction model being trained using a labeled historical data set, where each data point is associated with a lookback window and a prediction range to create each training sample, where each training sample is clustered into a plurality of clusters, where the plurality of clusters are associated with a normal score and an anomaly score, respectively; and outputting a classification associated with the new data sample based on comparing a first anomaly score of a first cluster of the plurality of clusters including the new data sample to an average anomaly score of the plurality of clusters having a normal score greater than the anomaly score.
[0015] A further aspect of the disclosure relates to a computer-implemented method for predicting wear-related deterioration of a PCP, the method including inputting new data samples of the PCP into a model configured to predict wear-related deterioration of the PCP. The model is trained using a labeled historical PCP dataset. Each data point is associated with a lookback window and a prediction range to create each training sample. Each training sample is clustered into a plurality of clusters, and the plurality of clusters are associated with a normal score and an anomaly score, respectively. The method further includes outputting a classification associated with the new data sample based on comparing a first anomaly score of a first cluster of the plurality of clusters that includes the new data sample to an average anomaly score of the clusters of the plurality of clusters having a normal score greater than the anomaly score, the classification being indicative of wear-related deterioration of the PCP.
[0016] The foregoing aspects of the present disclosure preferably enable accurate prediction of non-linear performance degradation of wear-related degradation in PCPs from an imbalanced training dataset of data related to the PCP. Further, the normality and abnormality scores preferably enable quantifying the degree of abnormality associated with the classification.
[0017] A further aspect of the present disclosure relates to a computer-implemented method for predicting wear-related degradation of a PCP, the method including generating labeled historical data by performing a binary labeling of historical data associated with one or more PCPs. The method further includes generating a plurality of training data samples by applying a lookback window and a prediction horizon to each data point of the labeled historical data. The method further includes clustering the plurality of training data samples into a plurality of clusters. The method further includes calculating a cluster score for each cluster of the plurality of clusters. The method further includes assigning a new data sample of the PCP to a first cluster of the plurality of clusters. The method further includes assigning a classification to the new data sample based on the cluster score associated with the first cluster, the classification being indicative of a likelihood of future wear-related degradation of the PCP.
[0018] The present disclosure advantageously enables accurate prediction of non-linear performance degradation of wear-related degradation in PCPs from an imbalanced training dataset of PCP-related data. Further, the cluster score advantageously enables quantification of the degree of anomaly associated with the classification.
[0019] According to a preferred embodiment, the method further includes labeling the labeled historical data as faulty for a predefined period of time prior to a known pump replacement date. This embodiment of the present disclosure advantageously provides a clear decision boundary between "normal" and "faulty" data, whereas with other methods, such a definitive decision boundary does not necessarily exist due to the nonlinearity and gradual nature of wear-related degradation in performance.
[0020] According to a preferred embodiment, the method further includes the historical data including pump speed data, pump torque data, casing pressure data, production rate data, and maintenance records associated with the PCP. This is advantageously data that can be used by the PCP. In other words, this data can be used to predict wear-related degradation in the performance of the PCP, thereby eliminating the need for additional data instruments.
[0021] According to a preferred embodiment, the method further includes calculating a cluster score for each cluster, the normality score being a first proportion of training data samples associated with a normal state in the first cluster divided by a second proportion of training data samples associated with a normal state in the plurality of training data samples. Calculating the cluster score further includes calculating an anomaly score for the first cluster, the anomaly score being a third proportion of training data samples associated with a degraded state in the first cluster divided by a fourth proportion of training data samples associated with a degraded state in the plurality of training data samples. Calculating the normality score and anomaly score advantageously quantifies the relative degree of anomaly in the predicted classification, thus providing additional accuracy and context to the classification.
[0022] According to a preferred embodiment, the method further includes generating a fault signal for the new data sample, the fault signal comprising an average anomaly score for the new sample over a predetermined number of previous data points. The fault signal advantageously represents a post-processed smoothing of the anomaly scores for the particular data sample, thereby reducing false positives and / or noise in the successive anomaly scores of the stream of data.
[0023] Additional aspects of the present disclosure are directed to systems and computer program products configured to perform the aforementioned methods. This Summary is not intended to describe every aspect of every implementation and / or embodiment of the present disclosure.
[0024] Preferred embodiments of the present invention will now be described, by way of example only, with reference to the following drawings in which: [Brief description of the drawings]
[0025] [Figure 1] FIG. 1 is a block diagram illustrating an example computing environment, consistent with some embodiments of the present disclosure. [Diagram 2] FIG. 2 is a block diagram illustrating an example failure prediction model, in accordance with some embodiments of the present disclosure. [Diagram 3] FIG. 2 illustrates a flowchart of an example method for predicting performance degradation with non-linear characteristics, in accordance with some embodiments of the present disclosure. [Figure 4] FIG. 2 illustrates a flowchart of an example method for training a failure prediction model in accordance with some embodiments of the present disclosure. [Figure 5A] FIG. 2 illustrates a flowchart of an example method for generating results based on output from a failure prediction model, in accordance with some embodiments of the present disclosure. [Figure 5B] FIG. 2 illustrates a flowchart of an exemplary method for calculating a fault signal, in accordance with some embodiments of the present disclosure. [Figure 6A]FIG. 13 illustrates experimental results of a graph of cumulative anomaly score as a function of days to failure, in accordance with some embodiments of the present disclosure. [Figure 6B] FIG. 13 illustrates experimental results of a graph of failure label intensity as a function of days to failure, in accordance with some embodiments of the present disclosure. [Figure 7] FIG. 13 illustrates experimental results of confusion matrices for different time periods before a failure, in accordance with some embodiments of the present disclosure. [Figure 8] FIG. 1 is a block diagram illustrating an example computer according to some embodiments of the present disclosure. [Figure 9] FIG. 1 illustrates a cloud computing environment in accordance with some embodiments of the present disclosure. [Figure 10] FIG. 2 illustrates an abstract model layer according to some embodiments of the present disclosure. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0026] While the disclosure is susceptible to various modifications and alternative forms, details thereof have been shown by way of example in the drawings and will be described in detail. It is to be understood, however, that it is not intended to limit the disclosure to the particular embodiments described. On the contrary, the intention is to cover all modifications, equivalents, and alternatives falling within the spirit and scope of the disclosure.
[0027] SUMMARY OF THE DISCLOSURE
[0002] Embodiments of the present disclosure are directed to predictive modeling, and more particularly, to predicting performance degradation with non-linear characteristics. Although not limited to such applications, embodiments of the present disclosure may be better understood in light of the foregoing context.
[0028] Embodiments of the present disclosure are directed to techniques for detecting degradation in an asset's performance during the early stages of non-linear failure mechanisms (e.g., wear, fatigue, creep, etc.) to predict the asset's upcoming failure and recommend preventive maintenance of the asset prior to failure. Embodiments of the present disclosure utilize semi-supervised machine learning methods that take the asset's historical data as input and generate as output anomaly scores, classifications, and / or failure signals indicative of the likelihood of future failure or performance degradation of the asset.
[0029] In overcoming the aforementioned challenges in predictive modeling of nonlinear phenomena, embodiments of the present disclosure achieve features such as, but not limited to, (i) a nonlinear decision boundary that distinguishes between normal data and data indicative of faults; (ii) a computationally simple implementation (e.g., not overly parameterized as can be the case with ANNs, despite being nonlinear); (iii) no balancing of classes (thus not skewing the original distribution in the data and subsequently biasing the solution); and / or (iv) a generic solution that can be easily applied to any group of similar assets.
[0030] Referring now to the figures, FIG. 1 illustrates an exemplary computing environment 100 in accordance with some embodiments of the present disclosure. The computing environment 100 includes a failure prediction system 102 communicatively coupled to a sensor data recording system 104 and an endpoint application 106 via a network 108. The failure prediction system 102 can be configured to receive data from the sensor data recording system 104 and make predictions related to future degradations in the performance of the associated asset 110 due to non-linear phenomena (e.g., predicting performance degradation related to wear of the PCP at a future time interval). The failure prediction system 102 can further be configured to communicate with the endpoint application 106. For example, the failure prediction system 102 can receive a request from the endpoint application 106 to detect an indication of a failure from the sensor data recording system 104. As another example, the failure prediction system 102 can deliver updates to the endpoint application 106 regarding characteristics of possible failures identified from the sensor data recording system 104.
[0031] The sensor data recording system 104 may include one or more data collection systems configured to obtain data directly or indirectly related to the function of the asset 110. For example, if the asset 110 is a PCP in an oil well production system, the sensor data recording system 104 may collect data such as, but not limited to, speed, torque, casing pressure, or production rate, or a combination thereof. As another example, if the asset 110 is a mechanical component of a vehicle (e.g., a fuel pump, wheel bearings, head gaskets, etc.), the sensor data recording system 104 may collect data such as, but not limited to, mileage, speed, engine error codes, etc. As can be seen from these two non-limiting examples, the sensor data recording system 104 does not necessarily collect data directly associated with the asset 110. Conversely, in some embodiments, the sensor data recording system 104 collects data associated with other components related to the asset 110, but the collected data may still be useful in providing indirect inferences regarding the function of the asset 110. This can be beneficial to the extent that it is economically infeasible and / or technically difficult to implement to instrument certain components of a system to collect data for predicting nonlinear degradation. Thus, in many real-world applications, predicting nonlinear degradation involves the challenges of (i) identifying fault signatures from available data on nonlinear degradation, and (ii) determining whether the identified fault signatures are exclusively (or primarily) representative of nonlinear degradation.
[0032] The endpoint application 106 can be an application that runs on a user workstation, such as, for example, a desktop, laptop, tablet, smartphone, or other endpoint device. The endpoint application 106 can provide a user with an interface for interacting with the failure prediction system 102. For example, a user can request a predictive failure analysis of the asset 110 based on data from the sensor data recording system 104. As another example, the failure prediction system 102 can deliver updates, notifications, or alerts to the endpoint application 106 based on failure characteristics detected from data from the sensor data recording system 104 that are associated with the asset 110. Additionally, in some embodiments, the endpoint application 106 provides a mechanism by which a user can configure a trained failure prediction system 102 to receive streaming data, which can be for assets similar to the asset 110 used to train the failure prediction system 102 (e.g., a well operator may stream his or her own PCP data to a failure prediction system 102 that has already been trained on similar PCP data from one or more other PCPs).
[0033] The failure prediction system 102, in some embodiments, can be virtually provisioned in a cloud computing architecture. In some embodiments, the failure prediction system 102 can reside on a computer, such as a mainframe, a compute node, a desktop, a laptop, a tablet, or another system that includes one or more processors and one or more computer-readable storage media.
[0034] The failure prediction system 102 can include a data warehouse 112, a model container 114, and a calculation engine 116. The data warehouse 112 can include aggregated sensor data 118, which can be data collected from the sensor data recording system 104 and can include one or more samples of the data.
[0035] The model container 114 may include data preparation / engineering utilities 126 that may be performed on the aggregated sensor data 118 to at least partially generate the data formatted sensor data 120. For example, the data preparation / engineering utilities 126 may be configured to remove outliers, correct data formatting issues, resolve null values, etc. when converting the aggregated sensor data 118 to the data formatted sensor data 120. In some embodiments, the data formatted sensor data 120 may include, for example, a lookback window applied to the aggregated sensor data 118.
[0036] The formatted sensor data 120 can be input to a failure prediction model 124. In some embodiments, the formatted sensor data 120 and the failure prediction model 124 are loaded into a deployment resource 130 of the computation engine 116. After execution of the failure prediction model 124 using the formatted sensor data 120 as input, the computation engine 116 can generate results 122 and store the results 122 in the data warehouse 112. The results 122 can include, for example, an anomaly score, a classification, or a failure signal, or a combination thereof. The results 122 can indicate the likelihood of future degraded performance due to non-linear performance degradation (e.g., the likelihood of performance degradation associated with wear in the PCP).
[0037] The computation engine 116 may further include a prediction service 128, which may be configured to receive requests from or deliver notifications to the endpoint applications 106. The prediction service 128 may coordinate the functioning of the failure prediction system 102. For example, in some embodiments, the prediction service 128 may cause the data preparation / engineering utilities 126 to be executed on the aggregated sensor data 118 using the deployment resources 130 to generate the data formatted sensor data 120. Continuing with the above example, the prediction service 128 may be further configured to deploy the failure prediction model 124 to the deployment resources 130 using the data formatted sensor data 120 as an input to generate the results 122. The prediction service 128 may be further configured to send the results 122 to the endpoint applications 106.
[0038] 2, a block diagram of the failure prediction model 124 is shown in accordance with some embodiments of the present disclosure. The failure prediction model 124 can include, for example, a training mode 200 and a deployment mode 228. In the training mode 200, the failure prediction model 124 is trained to accurately predict nonlinear performance degradation, such as failure or degradation due to wear. In the deployment mode 228, the failure prediction model 124 is configured to receive input data, format the input data, and make predictions related to nonlinear phenomena associated with the data (e.g., predicting future suboptimal performance due to wear in a PCP).
[0039] In the training mode 200, the failure prediction model 124 may receive historical data 202 from one or more sensor data recording systems 104 associated with one or more assets 110, and the historical data 202 may be segmented into a training subset 204-1 and a testing subset 204-2. The training subset 204-1 may be configured to train and validate the failure prediction model 124, and the testing subset 204-2 may be used to test the failure prediction model 124, and the testing may be used to quantify properties of the failure prediction model 124, such as, for example, accuracy, precision, recall, etc.
[0040] The training mode 200 can further include labeled historical data 206, which can be derived from time-aligned historical fault records (e.g., maintenance records indicating replacement of a PCP) including sensor readings (e.g., speed, torque, production rate, casing pressure, etc.). The labeled historical data 206 includes normal condition data 208-1, which is indicative of normal operation of the asset 110, and degraded condition data 208-2, which is indicative of suboptimal, degraded, or failed performance of the asset 110 (e.g., worn out condition of a PCP). The decision boundary between the normal condition data 208-1 and the degraded condition data 208-2 can be subjectively determined by a subject matter expert (SME), objectively determined by a statistical measure (e.g., one standard deviation outside the mean value for normal operation), inferred from machine learning, or determined using other strategies or techniques. As previously discussed, data indicative of a degraded state 208-2 need not necessarily be data relating to a non-functioning asset 110. Rather, data indicative of a degraded state 208-2 may be indicative of sub-optimal performance of an asset 110 even if the asset 110 remains functioning. For example, in some embodiments, if an asset 110 experiences an explicit failure or is otherwise replaced on x day, data from a predefined period prior to x day may be automatically labeled as data indicative of a degraded state 208-2.
[0041] The training mode 200 may further include time-framed historical data 210. The time-framed historical data 210 may include a look-back window 212 and a forecast horizon 214. The look-back window 212 may define the number of data points (e.g., D x ) The prediction range 214 can refer to a number (e.g., D y) may refer to the label of a future data point only. The lookback window 212 and the prediction horizon 214 may be used to convert each data point in the training subset 204-1 into a data sample 216-1 associated with a normal state and a data sample 216-2 associated with a degraded state (collectively referred to as data sample 216). In other words, the failure prediction model 124 is trained to use the performance history of the asset 110 (equal to the lookback window 212) to make decisions regarding the possible future state of the asset 110 in terms of the level of degradation in performance. Each data sample 216 may include a vector, matrix, or tensor of data points corresponding to a current data point and a number of previous data points per data stream (based on the lookback window 212), and each data sample 216 may be further associated with a label of a future data point by a prediction range 214 (the label of the future data point may be added to a predefined position in the vector, matrix, or tensor or otherwise associated with the vector, matrix, or tensor). In other words, a data sample 216-1 associated with a normal state may be identified by a data point in the prediction range 214 that is data 208-1 indicative of a normal state. Similarly, a data sample 216-2 associated with a degraded state may be a data sample that includes a data point in the prediction range 214 labeled as data 208-2 indicative of a degraded state.
[0042] For example, the lookback window 212 can be 10 days and the forecast horizon 214 can be 20 days. In this situation, for a data point at an initial time from a single stream of data in the training subset 204-1, the data point can be converted into a vector containing the previous 10 data points (e.g., the lookback window 212), and this vector can be associated with a label of a future data point 20 days from now (e.g., the forecast horizon 214, where the label is either data 208-1 indicating a normal state or data 208-2 indicating a degraded state). In this example, if the label of the future data point 20 days from now is data 208-2 indicating a degraded state, then the current data point and the vector of the data points of the previous 10 data points can be considered as a predictive failure feature for training purposes. Conversely, if the label of the future data point 20 days from now is data 208-1 indicating a normal state, then the current data point and the vector of the data points of the previous 10 data points can be assumed to be a predictive normal feature for training purposes.
[0043] As will be appreciated by those skilled in the art, the lookback window 212 and the forecast horizon 214 can be various numbers according to various scales. For example, in some embodiments, the units of the lookback window 212 and the forecast horizon 214 can be seconds, minutes, days, weeks, months, etc. In various embodiments, the lookback window 212 is smaller, larger, or equal to the forecast horizon 214. In some embodiments, the lookback window 212 is sized to address a tradeoff between usefulness and computational overhead. For example, a relatively large lookback window 212 increases the information used to accurately detect the fault signature, but also increases the computational power required to implement. Conversely, a relatively small lookback window 212 reduces the information used to accurately detect the fault signature, but also reduces the computational power required to implement. In some embodiments, the forecast horizon 214 is sized based on the fault signature. For example, if the asset 110 is associated with a 20 day time frame from when indications of a failure begin to manifest themselves in the data, then the forecast horizon 214 should be 24 days or less (in other words, a forecast horizon 214 greater than 20 days will result in an increase in false positives). Additionally, in some embodiments, the units of the lookback window 212 and forecast horizon 214 do not necessarily have to be in time-based increments at all, but can simply be referenced by a number of previous or subsequent data points, and the spacing of the data points may be based on non-temporal characteristics. Additionally, while the data samples 216 described in the above examples are in vector format, in other embodiments, matrices or tensors can be used to represent multi-dimensional or multi-modal data streams. As one example, for an asset 110 associated with three data streams, the data samples 216 may include input and output data pairs, where the input portion includes a three-dimensional tensor made up of samples from each lookback window 212 for each of the three data streams, and the output portion includes a binary indicator (e.g., 0 for normal, 1 for failure) of future performance in the forecast horizon 214.
[0044] The training mode 200 may further include clustering the data 218. The clustering 218 may cluster the data samples 216 using any testable clustering technique now known or later developed. In other words, the clustered data 218 may include multiple clusters 220, each cluster including at least one data sample 216. In particular, the number of clusters 220 is configurable to accurately fit (without overfitting) the data samples 216.
[0045] In some embodiments, the clusters 220 are determined by using K-means clustering. K-means clustering is advantageously a computationally efficient clustering technique that is scalable to large data sets. More generally, the type of clustering technique used, the number of clusters used, and the parameters of the clusters used (e.g., shape, size, etc.) are all tunable parameters that can be adjusted as needed to improve the performance of the failure prediction model 124 according to some embodiments of the present disclosure. For example, it may be beneficial to have a sufficient number of clusters to capture a variety of normal and degraded operating profiles.
[0046] The clustered data 218 may further include a cluster score 222 assigned to each of the clusters 220. The cluster scores 222 may include a normal score 224 and an abnormal score 226. In some embodiments, the normal score 224 may be calculated according to Equation 1. Equation 1:
number
[0047] Similarly, in some embodiments, the anomaly score 226 may be calculated according to Equation 2. Equation 2:
number
[0048] Regarding Equation 1,
number
number
number
[0049] In other words, the normality score 224 can be the proportion of data samples 216-1 associated with a normal state in a particular cluster divided by the proportion of data samples 216-1 associated with a normal state in the entire time-framed historical data 210. Similarly, the abnormality score 226 can be the proportion of data samples 216-2 associated with a degraded state in a particular cluster divided by the proportion of data samples 216-2 associated with a degraded state in the entire time-framed historical data 210.
[0050] After creating the clusters 220 and generating the cluster scores 222, the failure prediction model 124 may be considered trained. In some embodiments, after training the failure prediction model 124 using the training subset 204-1, the failure prediction model 124 may be tested using the testing subset 204-2 and a deployment mode 228. Although the description of the deployment mode 228 is provided with respect to the testing subset 204-2, the description of the deployment mode 228 is equally applicable to receiving streaming real-time data for the purpose of predicting future nonlinear phenomena associated with the new data (e.g., wear-related performance degradation in a PCP).
[0051] In the deployment mode 228, the fault prediction model 124 can format the test subset 204-2 into time-boxed data samples 230. The time-boxed data samples 230 can be similar to the data samples 216, but do not include indications of normal or degraded states (insofar as indications of normal and degraded states are information predicted by the fault prediction model 124 and are therefore hidden from the fault prediction model 124 while testing the performance of the fault prediction model 124). Thus, each sample in the time-boxed data samples 230 can include a data point from the test subset 204-2 and multiple previous data points according to the look-back window 212. As previously mentioned, this set of data can be stored in a vector, matrix, or tensor format depending on the complexity, dimensionality, and modality of the data in the test subset 204-2. In some embodiments, the time-boxed data samples 230 correspond to the formatted sensor data 120.
[0052] Each sample of the time-boxed data samples 230 may then be associated with a respective cluster 220. A classification 232 may then be associated with each of the time-boxed data samples 230 based on the associated cluster of the clusters 220. For example, if the time-boxed data sample 230 is located within a respective cluster 220 having a normal score 224 that is greater than the anomaly score 226, then the time-boxed data sample 230 may be considered normal (e.g., a score of "0"). Conversely, if the time-boxed data sample 230 is located within a respective cluster 220 having an anomaly score 226 that is greater than the normal score 224, then the time-boxed data sample 230 may be considered abnormal or predictive of a future failure (e.g., a score of "1"). In some embodiments, the classification 232 includes comparing the anomaly score 226 of the cluster 220 capturing the time-boxed data sample 230 to the average anomaly score of all clusters 220 having a normal score 224 that is greater than the anomaly score 226. In this way, the anomaly score 226 of a cluster 220 capturing a time-framed data sample 230 can be compared to the baseline level of anomaly seen in other clusters 220 that are otherwise considered normal.
[0053] The failure prediction model 124 may further include a failure signal 234. The failure signal 234 may be configured to smooth the anomaly scores 226 or classifications 232 by taking an average value of the anomaly scores for a previous predetermined number of data points. For example, the failure signal 234 may be calculated according to Equation 3: Equation 3:
number
[0054] In Equation 3, the term x may refer to a predefined number of consecutive previous data points, for which the average anomaly score of the particular time-framed data sample 230 is determined. In some embodiments, x may be an adjustable parameter subject to design considerations of the fault prediction model 124. For example, a relatively large x may reduce the sensitivity of the fault prediction model 124 to a particular anomaly score indicative of a fault (thereby reducing false positives), while a relatively small x may increase the sensitivity of the fault prediction model 124 to each anomaly score indicative of a fault (thereby reducing false negatives). Additionally, the term AS in Equation 3 may n may refer to the anomaly score 226 of one of the clusters 220 that contains the data point n, although in other embodiments, classification 232 may be used. The fault signal 234 is described in further detail with respect to FIG. 5B.
[0055] 3, a flowchart of an example method 300 for utilizing the failure prediction model 124 is shown, in accordance with some embodiments of the present disclosure. The method 300 can be implemented, for example, by the failure prediction model 124, the failure prediction system 102, a computer, a computational node, a processor, or another combination of hardware or software or both.
[0056] Operation 302 includes training the failure prediction model 124. Training the failure prediction model 124 may include aspects described above with respect to the training mode 200 of the failure prediction model 124. Operation 302 is described in further detail below with respect to FIG.
[0057] Operation 304 includes formatting the aggregated sensor data 118 into formatted sensor data 120. In some embodiments, the formatted sensor data 120 corresponds to the time-boxed data samples 230. Operation 304 may include applying a lookback window 212 to each data point in the aggregated sensor data 118 to generate the formatted sensor data 120. In some embodiments, operation 304 further includes other data cleansing and / or formatting operations, such as removing outliers, removing null values, etc.
[0058] Operation 306 includes inputting the formatted sensor data 120 into a failure prediction model 124. Operation 308 includes generating a result 122 based on an output from the failure prediction model 124. In some embodiments, the result 122 includes a classification 232 (e.g., an indication of normality or future failure) or a failure signal 234 (e.g., a quantification related to the likelihood of a future failure) or both.
[0059] The operation 310 includes performing a mitigating action based on the result. For example, the mitigating action may include sending the result 122 to the endpoint application 106 in the form of a notification, an alert, a report, or another transmission. In some embodiments, the operation 310 includes triggering a scheduling event related to maintenance of the asset 110, such as replacing, rebuilding, or otherwise maintaining the asset 110. In some embodiments, the scheduling event is based on the result 122. For example, a fault signal 234 exceeding a threshold may trigger a maintenance event to be scheduled within a time window (e.g., within the forecast horizon 214). In some embodiments, the mitigating action may be related to a logistics activity, such as ordering a needed replacement part and shipping the ordered replacement part to a location associated with the asset 110.
[0060] 4, a flowchart of an example method 400 for training the failure prediction model 124 is shown in accordance with some embodiments of the present disclosure. In some embodiments, the method 400 is a sub-method of operation 302 of FIG. 3. In some embodiments, the method 400 can be implemented by, for example, the failure prediction model 124, the failure prediction system 102, a computer, a computational node, a processor, or another combination of hardware or software or both.
[0061] Operation 402 includes aggregating historical data 202 into a training subset 204-1 and a test subset 204-2. Operation 404 includes labeling the historical data 202 to generate labeled historical data 206 including data indicative of a normal condition 208-1 and data indicative of a degraded condition 208-2. Operation 406 includes applying a lookback window 212 and a prediction range 214 to each data point in the historical data 202 to generate data samples 216-1 associated with a normal condition and data samples 216-2 associated with a degraded condition.
[0062] Operation 408 includes clustering the data samples 216 into a number of clusters 220. In some embodiments, operation 408 utilizes K-means clustering. Operation 410 includes calculating a cluster score 222 associated with each of the clusters 220. The cluster scores 222 may include a respective normal score 224 and a respective anomaly score 226 for each cluster in the clusters 220.
[0063] Operation 412 includes adjusting the failure prediction model 124. Adjusting the failure prediction model 124 may include, for example, (i) modifying the labeled historical data 206 by changing the definition of the data indicative of normal conditions 208-1 and the data indicative of degraded conditions 208-2, (ii) modifying the size of the lookback window 212, (iii) modifying the size of the prediction horizon 214, (iv) modifying parameters associated with the clusters 220 (e.g., number of clusters, shape of clusters, size of clusters, etc.), or other model adjustment techniques or strategies, or both, or a combination thereof. In some embodiments, the failure prediction model 124 is adjusted based on results from testing the test subset 204-2.
[0064] Operation 414 includes outputting the trained failure prediction model 124. In some embodiments, outputting the trained failure prediction model 124 includes storing the trained failure prediction model 124 in a computer-readable storage medium, such as, for example, the virtually provisioned model container 114.
[0065] 5A, a flowchart of an example method 500 for generating results based on output from a failure prediction model 124 is shown, in accordance with some embodiments of the present disclosure. In some embodiments, the method 500 is a sub-method of operation 308 of FIG. 3. In some embodiments, the method 500 can be implemented by, for example, the failure prediction model 124, the failure prediction system 102, a computer, a computational node, a processor, or another combination of hardware or software or both.
[0066] The operation 502 includes associating a binary classification result 232 with each data sample (e.g., time-boxed data sample 230) previously input to the failure prediction model 124. In some embodiments, the classification 232 is either “normal” (e.g., 0), or “abnormal”, “fault”, or another non-normal indicator (e.g., 1). The classification 232 may be based on the normal score 224 and anomaly score 226 of the cluster 220 capturing each time-boxed data sample 230. More specifically, if the normal score 224 is greater than the anomaly score 226, the corresponding time-boxed data sample 230 is deemed normal. Conversely, if the anomaly score 226 is greater than the normal score 224, the corresponding time-boxed data sample 230 is deemed anomalous or otherwise indicative of a fault.
[0067] Operation 504 includes generating a fault signal 234 for each input data sample (e.g., time-boxed data sample 230). The fault signal 234 may be based on the normal score 224, the anomaly score 226, or the classification 232, or a combination thereof, associated with the time-boxed data sample 230. In some embodiments, the fault signal 234 represents a more reliable indicator of truly anomalous data (e.g., reduces false positives). The fault signal 234 is described in further detail below with respect to FIG. 5B.
[0068] 5B, a flowchart of an example method 510 for generating a fault signal 234 is shown, in accordance with some embodiments of the present disclosure. In some embodiments, the method 510 is a sub-method of operation 504 of FIG. 5A. In some embodiments, the method 510 can be implemented by, for example, the fault prediction model 124, the fault prediction system 102, a computer, a computational node, a processor, or another combination of hardware or software or both.
[0069] Operation 512 includes calculating an average anomaly score for the clusters 220 ranked as normal. Calculating the average anomaly score during normal operation may include averaging the anomaly scores 226 for each cluster in the clusters 220 where the normal score 224 is greater than the anomaly score 226. This may be beneficial insofar as it establishes a baseline level of anomalies that may function to reduce false positives.
[0070] Operation 514 includes assigning the time-windowed incoming data sample 230 to a cluster 220 associated with an anomaly score 226 and a normal score 224. If the anomaly score for the designated cluster is greater than the average anomaly score during normal operation (determined in operation 512), then the data sample is classified as being in failure mode (classified as 1); otherwise, the data sample is classified as being in normal mode (classified as 0).
[0071] Operation 516 includes calculating the failure signal 234 as the average value of the binary classification results of 0 or 1 for the asset over a predefined period (e.g., 10 days) or predefined number of time-framed data samples 230. In some embodiments, operation 516 may utilize Equation 3:
[0072] 1-5, one particular application of embodiments of the present disclosure relates to detecting performance degradation caused by rotor wear in a PCP. Because performance degradation caused by rotor wear in a PCP is a non-linear phenomenon, embodiments of the present disclosure are well suited to accurately predict rotor degradation caused by wear in a PCP.
[0073] Artificial lift systems utilizing PCPs enable a variety of non-thermal oil and gas recovery methods, such as cold heavy oil production with sand (CHOPS). PCPs can lift viscous mixtures of oil and sand from underground reservoirs to the surface, improving production costs, improving maintenance costs, improving application flexibility, and reducing environmental impacts compared to other artificial lift systems (e.g., electric submersible pumps (ESPs)).
[0074] Despite the suitability of PCPs to handle the higher sand content in heavy fuel oils, one issue resulting from continued sand entrapment is abrasive wear failure. Abrasive wear can refer to the gradual deterioration in pump performance as the hard chrome plating on the rotor wears away, and is the most common type of failure in PCPs. This wear can be limited to the surface of the chrome plating on the rotor or can extend to the base metal. In either case, the original rotor geometry changes. This change in geometry can affect the performance of the PCP insofar as the meshing between the rotor and stator changes. Severe abrasive wear, where the hard chrome plating wears away down to the base metal, can permanently damage the elastomer in the rotor, requiring the pump to be replaced.
[0075] Pump failures in oil wells result in significant costs in terms of lost production time. Therefore, the ability to predict pump wear-out failures reduces these costs by providing proactive scheduled maintenance of the PCP prior to failure. Additionally, improved awareness of pump performance degradation during the early stages of wear-out can help operators make appropriate adjustments in operational decision making to extend operational life.
[0076] However, predicting failures in PCPs is difficult for similar reasons as predicting performance degradation with nonlinear characteristics. For one, the failure mechanisms are gradual, causing problems in appropriately labeling historical data as "normal" or "abnormal" since the PCP may remain functioning at suboptimal performance for long periods while the rotor gradually wears out. Another challenge is related to the unbalanced set of historical data available for the PCP (e.g., significantly more normal data than abnormal data). Yet another challenge is related to variable failure mechanisms that vary with the operating environment (e.g., formations with high sand content compared to formations with low sand content). Thus, PCP failures are difficult to predict accurately. Nevertheless, aspects of the present disclosure, when implemented in the field of predictive monitoring of PCPs, can accurately predict PCP performance degradation due to wear-related mechanisms.
[0077] For example, returning again to the discussion of FIG. 4 as it relates to training a failure prediction model 124 for an asset 110, such as a PCP, operation 402 may aggregate sensor data, such as pump speed, pump torque, casing pressure, and production rate. The aggregated sensor data may be separated into a training subset 204-1 (e.g., approximately 80% of the data) and a testing subset 204-2 (e.g., approximately 20% of the data). Operation 404 may label the historical data as data indicative of a normal condition 208-1 (approximately 97% of the unbalanced data in training subset 204-1) or data indicative of a degraded condition 208-2 (approximately 3% of the unbalanced data in training subset 204-1). In some embodiments, data indicative of a degraded condition 208-2 may be all data from 25 days to 3 days prior to a known PCP replacement date (determined from maintenance records), while data prior to the 25 days prior to a known PCP replacement date may be considered data indicative of a normal condition 208-1.
[0078] Operation 406 may generate data samples 216 by applying the lookback window 212 and the prediction horizon 214 to each data point. Operation 408 may cluster the data samples 216 using K-means clustering, and operation 410 may calculate the normal score 224 and the anomaly score 226 according to Equation 1 and Equation 2, respectively. Operation 412 may train the failure prediction model 124, and operation 414 may output the trained failure prediction model 124.
[0079] Applying the test subset 204-2 to the trained failure prediction model 124 for a PCP application illustrates the utility of aspects of the present disclosure. FIG. 6A illustrates experimental results of a graph of cumulative anomaly score 226 (y-axis) as a function of days to known pump replacement date (x-axis). As shown, there is a steep increase in the trend line between 25 days prior to failure and the known pump replacement date. Thus, FIG. 6A illustrates that the anomaly score 226 can be used to successfully predict PCP failure.
[0080] 6B shows experimental results of a graph of the intensity of the failure-indicating classification 232 (y-axis) as a function of days to failure (x-axis). As shown, the intensity of the failure-indicating classification 232 increases significantly between about 25 days prior to failure and the known pump replacement date. Thus, FIG. 6B shows that classification 232 can be successfully used to predict PCP failure.
[0081] Notably, although no graphs are provided relating to the failure signal 234, it will be apparent to one skilled in the art that the failure signal 234 exhibits similar predictive power with reduced noise compared to the results shown in Figures 6A and 6B, since the failure signal 234 represents an average score, whereas Figure 6A shows cumulative scores and Figure 6B shows intensity scores.
[0082] FIG. 7 shows experimental results related to confusion matrices for various time periods prior to a known pump replacement date using the test subset 204-2 for the PCP failure prediction model 124, as described above. As shown in FIG. 7, for 30 days prior to failure 700-1, the normal (actual label 702-1)-normal (predicted label 704-1) box is 0.76, the failure-normal box is 0.38, the normal-failure box is 0.24, and the failure-failure box is 0.62. For 14 days prior to failure 700-2, the normal (actual label 702-2)-normal (predicted label 704-2) box is 0.76, the failure-normal box is 0.26, the normal-failure box is 0.24, and the failure-failure box is 0.74. For the 5 days before the fault 700-3, the normal (actual label 702-3)-normal (predicted label 704-3) box is 0.76, the fault-normal box is 0.12, the normal-fault box is 0.24, and the fault-fault box is 0.88.
[0083] In general, FIG. 7 shows improvement in prediction performance as a function of proximity to the actual pump replacement date. Moreover, FIG. 7 shows recall (e.g., true detections divided by the total number of true detections and false negatives) of about 88% 5 days before failure 700-3, about 75% 14 days before failure 700-2, and about 62% 30 days before failure 700-1. Meanwhile, aspects of the present disclosure achieved precision (e.g., true detections divided by the total number of true detections and false negatives) of about 78% (5 days before failure 700-3), 76% (14 days before failure 700-2), and 72% (30 days before failure 700-1). Thus, FIG. 7 shows that aspects of the present disclosure achieve a robust failure prediction model 124 for predicting PCP performance degradation as a result of rotor wear.
[0084] FIG. 8 illustrates a block diagram of an exemplary computer 800 according to some embodiments of the present disclosure. In various embodiments, the computer 800 can perform any or all of the methods described in FIGS. 3-5 and / or implement the functions described in any one of FIGS. 1-2 and / or 6-7. In some embodiments, the computer 800 receives instructions related to the aforementioned methods and functions by downloading processor-executable instructions from a remote data processing system over the network 850. In other embodiments, the computer 800 provides instructions for the aforementioned methods and / or functions to a client machine such that the client machine performs the method or part of the method based on the instructions provided by the computer 800. In some embodiments, the computer 800 is incorporated into the failure prediction system 102 of FIG. 1, the failure prediction model 124 of FIG. 1, or another aspect of the present disclosure (or is virtually provisioned with functions similar to the computer 800).
[0085] The computer 800 includes memory 825, storage 830, an interconnect 820 (e.g., a bus), one or more CPUs 805 (also referred to herein as processors), an I / O device interface 810, I / O devices 812, and a network interface 815.
[0086] Each CPU 805 retrieves and executes programming instructions stored in memory 825 or storage 830. Interconnect 820 is used to move data, such as programming instructions, between CPU 805, I / O device interface 810, storage 830, network interface 815, and memory 825. Interconnect 820 may be implemented using one or more buses. In various embodiments, CPU 805 may be a single CPU, multiple CPUs, or a single CPU including multiple processing cores. In some embodiments, CPU 805 may be a digital signal processor (DSP). In some embodiments, CPU 805 includes one or more 3D integrated circuits (3DICs) (e.g., 3D wafer-level packaging (3DWLP), 3D interposer-based integration, 3D stacked ICs (3D-SICs), monolithic 3D ICs, 3D heterogeneous integration, 3D system in package (3DSiP), and / or package on package (PoP) CPU configurations). Memory 825 is included generally to represent random access memory (e.g., static random-access memory (SRAM), dynamic random access memory (DRAM), or flash). Storage 830 is included generally to represent non-volatile memory, such as a hard disk drive, solid state device (SSD), removable memory card, optical storage, or flash memory device.In alternative embodiments, storage 830 may be replaced by a storage area-network (SAN) device, cloud, or other device connected to computer 800 via I / O device interface 810 or to network 850 via network interface 815.
[0087] In some embodiments, memory 825 stores instructions 860. However, in various embodiments, instructions 860 are stored partially in memory 825 and storage 830, respectively, or stored entirely in memory 825 or storage 830, or accessed over network 850 via network interface 815.
[0088] The instructions 860 may be computer readable and computer executable instructions for performing some or all of any of the methods of Figures 3-5 or implementing the functions described in Figures 1-2 and / or 6-7. In some embodiments, the instructions 860 may be referred to as a non-linear performance degradation prediction protocol (or instructions, mechanism, etc.) or a failure prediction protocol (or instructions, mechanism, etc.). Although the instructions 860 are shown in memory 825, the instructions 860 may include program instructions executable by one or more CPUs 805 collectively stored across multiple computer readable storage media.
[0089] In various embodiments, I / O device(s) 812 include an interface that can present information and receive input. For example, I / O device(s) 812 can present information to and receive input from a user interacting with computer 800.
[0090] The computer 800 is connected to a network 850 via a network interface 815. The network 850 can include a physical network, a wireless network, a cellular network, or a different network.
[0091] Although this disclosure includes detailed descriptions of cloud computing, it should be understood that implementation of the subject matter presented herein is not limited to a cloud computing environment. Rather, embodiments of the invention may be implemented in conjunction with any other type of computing environment now known or later developed.
[0092] Cloud computing is a service delivery model for enabling convenient, on-demand network access to a shared pool of configurable computational resources (e.g., networks, network bandwidth, servers, processing, memory, storage, applications, virtual machines, and services) and for rapidly provisioning and releasing these resources with minimal administrative effort or interaction with a service provider. This cloud model may include at least five characteristics, at least three service models, and at least four deployment models.
[0093] The features are as follows: On-Demand Self-Service: Cloud customers can automatically provision server time, network storage, and other computing power as they need it, without any unilateral, human interaction with the service provider. Broad network access: Capabilities are available over the network and can be accessed using standard mechanisms, facilitating use by heterogeneous thin-client or thick-client platforms (e.g., cell phones, laptops, and PDAs). Resource Pool: The provider's computing resources are pooled and offered to multiple consumers using a multi-tenant model. Different physical and virtual resources are dynamically allocated and reallocated according to demand. There is a sense of location independence, consumers usually have no control or knowledge regarding the exact location of the resources offered, although at a higher level of abstraction they may be able to specify a location (e.g. country, state, or data center).
[0094] Rapid Elasticity: Capacity is provisioned quickly and elastically, in some cases automatically, so it can be quickly scaled out and quickly released to quickly scale in. Capacity available for provisioning often appears to the consumer as unlimited, available for purchase in any quantity at any time.
[0095] Metered Services: Cloud systems leverage metering to automatically control and optimize resource usage at a level of abstraction appropriate to the type of service (e.g., storage, processing, bandwidth, and active user accounts). Resource usage can be monitored, controlled, and reported, providing transparency to both providers and consumers of the services utilized.
[0096] The service model is as follows: SaaS (Software as a Service): The capability provided to the consumer is the use of the provider's applications running on a cloud infrastructure. Those applications can be accessed from a variety of client devices through thin-client interfaces such as web browsers (e.g., web-based email). The consumer does not manage or control the underlying cloud infrastructure, including the network, servers, operating systems, storage, or individual application functionality, except for the possibility of limited user-specific application configuration settings. PaaS (Platform as a Service): The capability provided to a consumer is to deploy applications that the consumer creates or acquires, written using programming languages and tools supported by the provider, onto a cloud infrastructure. The consumer does not manage or control the underlying cloud infrastructure, including networks, servers, operating systems, or storage, but does have control over the deployed applications and, in some cases, the configuration of the application hosting environment. Infrastructure as a Service (IaaS): The capability provided to a consumer is the provisioning of processing, storage, network, and other basic computing resources over which the consumer can deploy and run any software, which may include operating systems and applications. The consumer does not manage or control the underlying cloud infrastructure, but has control over the operating systems, storage, and deployed applications, and in some cases has limited control over selected network components (e.g., host firewalls).
[0097] The deployment model is as follows: Private Cloud: This cloud infrastructure is operated solely for one organization. It can be managed by this organization or a third party and can exist on-premise or off-premise. Community Cloud: This cloud infrastructure is shared by multiple organizations to support a specific community with shared concerns (e.g., mission, security requirements, policy, and compliance considerations). It can be managed by these organizations or a third party and can reside on-premise or off-premise. Public Cloud: This cloud infrastructure is available for use by general users or large industry organizations and is owned by an organization that sells cloud services. Hybrid cloud: This cloud infrastructure is a combination of two or more clouds (private, community, or public) that remain unique but are bound together by standardized or proprietary technologies that allow for the portability of data and applications (e.g., cloud bursting to balance the load between clouds).
[0098] A cloud computing environment is a service-oriented environment with an emphasis on statelessness, loose coupling, modularity, and semantic interoperability. At the heart of cloud computing is an infrastructure that contains a network of interconnected nodes.
[0099] 9, an exemplary cloud computing environment 50 is shown. As shown, the cloud computing environment 50 includes one or more cloud computing nodes 10 with which local computing devices used by cloud consumers (e.g., a personal digital assistant (PDA) or mobile phone 54A, a desktop computer 54B, a laptop computer 54C, and / or an automobile computer system 54N) can communicate. The nodes 10 may communicate with each other. The nodes 10 may be physically or virtually grouped in one or more networks (not shown), such as in a private cloud, community cloud, public cloud, or hybrid cloud, or combination thereof, as previously described herein. This allows the cloud computing environment 50 to provide an infrastructure, platform, and / or SaaS that does not require the cloud consumer to maintain resources on a local computing device. The types of computing devices 54A-N illustrated in FIG. 9 are intended as examples only, and it will be understood that the computing node 10 and cloud computing environment 50 may communicate with any type of computer-controlled device via any type of network and / or network-addressable connection (e.g., a connection using a web browser).
[0100] Referring now to Figure 10, there is shown a set of functional abstraction layers provided by cloud computing environment 50 (Figure 9). It should be understood in advance that the components, layers, and functions shown in Figure 10 are intended to be illustrative only and that embodiments of the present invention are not limited thereto. As shown, the following layers and corresponding functions are provided:
[0101] The hardware and software layer 60 includes hardware and software components. Examples of hardware components include mainframes 61, Reduced Instruction Set Computer (RISC) architecture based servers 62, servers 63, blade servers 64, storage devices 65, and networks and network components 66. In some embodiments, the software components include network application server software 67 and database software 68.
[0102] The virtualization layer 70 provides an abstraction layer that can provide virtual entities such as virtual servers 71 , virtual storage 72 , virtual networks including virtual private networks 73 , virtual applications and operating systems 74 , and virtual clients 75 .
[0103] As an example, the management layer 80 may provide the following functions: Resource provisioning 81 provides dynamic procurement of computational and other resources utilized to execute tasks within the cloud computing environment. Metering and pricing 82 provides cost tracking and billing or invoicing for the utilization of resources within the cloud computing environment. As an example, the resources may include application software licenses. Security provides identity verification of cloud users and tasks, as well as protection of data and other resources. User portal 83 provides access to the cloud computing environment to users and system administrators. Service level management 84 provides allocation and management of cloud computational resources to meet required service levels. Service level agreement (SLA) planning and execution 85 provides pre-provisioning and procurement of cloud computational resources in anticipation of upcoming demands and in accordance with SLAs.
[0104] The Workload Layer 90 illustrates examples of functionality available in a cloud computing environment. Examples of workloads and functionality that may be provided from this layer include mapping and navigation 91, software development and lifecycle management 92, virtual classroom instruction delivery 93, data analytics processing 94, transaction processing 95, and non-linear performance degradation prediction 96.
[0105] The embodiments of the invention may be a system, a method, or a computer program product, or a combination thereof, at any possible level of technical detail of integration. The computer program product may include a computer-readable storage medium including computer-readable program instructions for causing a processor to perform aspects of the invention.
[0106] A computer readable storage medium may be a tangible device that can hold and store instructions for use by an instruction execution device. The computer readable storage medium may be, for example, but is not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof. A non-exhaustive list of more specific examples of computer-readable storage media includes portable computer diskettes, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital versatile disk (DVD), memory sticks, floppy disks, mechanically encoded devices such as punch cards or ridges in grooves on which instructions are recorded, and any suitable combinations of these. As used herein, computer-readable storage media should not itself be construed as being ephemeral signals such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through a waveguide or other transmission medium (e.g., light pulses passing through a fiber optic cable), or electrical signals transmitted over wires.
[0107] The computer readable program instructions described herein can be downloaded from a computer readable storage medium to each computing device / processing device or to an external computer or storage device over a network (e.g., the Internet, a local area network, a wide area network, or a wireless network, or a combination thereof) that can include copper transmission cables, optical transmission fiber, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface within each computing device / processing device receives the computer readable program instructions from the network and forwards the computer readable program instructions for storage on a computer readable storage medium within each computing device / processing device.
[0108] The computer readable program instructions for carrying out the operations of the present invention may be assembler instructions, instruction-set-architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, configuration data for an integrated circuit, or source or object code written in any combination of one or more programming languages, including object-oriented programming languages such as Smalltalk®, C++, and procedural programming languages such as the “C” programming language or similar programming languages. The computer readable program instructions may execute entirely on the user's computer, partially on the user's computer as a standalone software package, partially on the user's computer and on a remote computer, or entirely on a remote computer or server. In the latter scenario, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection may be made to an external computer (e.g., via the Internet using an Internet Service Provider). In some embodiments, to carry out aspects of the invention, electronic circuitry including, for example, a programmable logic circuit, a field-programmable gate array (FPGA), or a programmable logic array (PLA), can execute computer readable program instructions to customize the electronic circuitry by utilizing state information of the computer readable program instructions.
[0109] Aspects of the present invention are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks included in the flowchart illustrations and / or block diagrams, can be implemented by computer readable program instructions.
[0110] These computer readable program instructions may be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, where the instructions, executed by the processor of the computer or other programmable data processing apparatus, produce means for performing the functions / operations specified in the blocks of the flowcharts and / or block diagrams. These computer readable program instructions may also be stored on a computer readable storage medium and capable of directing a computer, programmable data processing apparatus, or other device, or combination thereof, to function in a particular manner, such that the computer readable storage medium on which the instructions are stored comprises an article of manufacture containing instructions for performing aspects of the functions / operations specified in the blocks of the flowcharts and / or block diagrams.
[0111] Computer-readable program instructions may also be loaded into a computer, other programmable data processing apparatus, or other device such that the instructions, which execute on the computer, other programmable apparatus, or other device, perform the functions / acts specified in the blocks of the flowcharts and / or block diagrams, thereby causing a series of operable steps to be performed on the computer, other programmable apparatus, or other device to generate a computer-implemented process.
[0112] The flowcharts and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or subset of instructions, including one or more executable instructions for implementing the specified logical function. In some alternative implementations, the functions shown in the blocks may occur in a different order than that shown in the figures. For example, two blocks shown in succession may in fact be executed substantially simultaneously or in reverse order, depending on the functionality involved. It is also noted that each block in the block diagrams and / or flowchart illustrations, as well as combinations of blocks included in the block diagrams and / or flowchart illustrations, may be implemented by a dedicated hardware-based system that performs the specified functions or operations or executes a combination of dedicated hardware and computer instructions.
[0113] Although it will be appreciated that the process software (e.g., any of the instructions stored in instructions 860 of FIG. 8 configured to perform any portion of the method described with respect to FIGS. 3-5 or implement any portion of the functionality described in FIGS. 1-2 and / or 6-7) can be deployed by manually loading it directly into the client, server, and proxy computers via loading a storage medium such as a CD, DVD, or the like, the process software can also be automatically or semi-automatically deployed to the computer systems by sending the process software to a central server or group of central servers. The process software is then downloaded to the client computers that will execute the process software. As an alternative, the process software is sent directly to the client systems via email. The process software is then isolated into a directory or loaded into the directory by executing a set of program instructions that isolates the process software into a directory. Another alternative is to send the process software directly into a directory on the hard drive of the client computer. If a proxy server is present, the process selects the proxy server code, determines the computer on which to place the proxy server code, transmits the proxy server code, and then installs the proxy server code on the proxy computer. The process software is transmitted to the proxy server and then stored on the proxy server.
[0114] Embodiments of the invention may also be delivered as part of a service agreement with a client company, non-profit organization, government agency, internal organizational structure, etc. These embodiments may include configuring a computer system to execute software, hardware, and web services that implement some or all of the methods described herein, and deploying such software, hardware, and web services. These embodiments may also include analyzing client behavior, making recommendations in response to the analysis, building a system that implements a subset of the recommendations, integrating the system into existing processes and infrastructure, metering usage of the system, allocating costs to users of the system, and billing, invoicing (e.g., generating invoices), or otherwise receiving payment for use of the system.
[0115] The terms used herein are for the purpose of describing particular embodiments only and are not intended to limit the various embodiments. As used herein, the singular forms "a", "an" and "the" are intended to include the plural unless the context expressly indicates otherwise. It will be further understood that the terms "comprise" and / or "comprising", when used herein, indicate the presence of stated features, integers, steps, operations, elements, or components, or combinations thereof, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, or groups thereof, or combinations thereof. In the preceding detailed description of exemplary embodiments of the various embodiments, reference is made to the accompanying drawings, which form a part hereof (where like numbers represent like elements), in which specific example embodiments are shown by way of example and in which the various embodiments may be practiced. These embodiments have been described in sufficient detail to enable one skilled in the art to practice the embodiments, but other embodiments may be used and logical, mechanical, electrical and other changes may be made without departing from the scope of the various embodiments. In the preceding description, many specific details have been set forth to enable a thorough understanding of the various embodiments. However, various embodiments may be practiced without these specific details. In other instances, well-known circuits, structures and techniques have not been shown in detail so as not to obscure the embodiments.
[0116] Different instances of the term "embodiment" as used within this specification do not necessarily refer to the same embodiment, but may refer to the same embodiment. All data and data structures shown or described herein are merely examples, and in other embodiments, different amounts of data, types of data, fields, numbers and types of fields, field names, numbers and types of rows, records, entries, or organization of data may be used. In addition, any data may be combined with logic such that separate data structures may not be necessary. Thus, the preceding detailed description should not be taken in a limiting sense.
[0117] The description of various embodiments of the present disclosure is presented for illustrative purposes, but is not intended to be exhaustive and is not limited to the disclosed embodiments. Many changes and modifications that do not depart from the scope and spirit of the described embodiments will be apparent to those skilled in the art. The terms used in this specification are selected to best explain the principles of the embodiments, practical applications, or technical improvements beyond the technology found in the market, or to allow other skilled in the art to understand the embodiments disclosed herein.
[0118] While the present disclosure has been described with respect to specific embodiments, it is anticipated that alterations and modifications thereto will become apparent to those skilled in the art, and it is therefore intended that the following claims be interpreted to cover all such alterations and modifications as fall within the true spirit and scope of the present disclosure.
[0119] All advantages described in this disclosure are exemplary advantages, and there may be embodiments of the disclosure that achieve all, some, or none of the described advantages while remaining within the spirit and scope of the disclosure.
[0120] In the following, a non-limiting list of examples is provided to illustrate some aspects of the present disclosure. Example 1 is a computer-implemented method. The method includes inputting a new data sample into a failure prediction model, where the failure prediction model is trained using a labeled historical data set, where each data point is associated with a lookback window and a prediction range to create each training sample, where each training sample is clustered into a plurality of clusters, where each of the plurality of clusters is associated with a normal score and an abnormality score, respectively, and outputting a classification associated with the new data sample based on comparing a first anomaly score of a first cluster of the plurality of clusters including the new data sample to an average anomaly score of clusters of the plurality of clusters having a normal score greater than the abnormality score.
[0121] Example 2 includes the method of example 1, including or excluding the optional feature, where the classification indicates a likelihood of performance degradation associated with wear on the asset associated with the new data sample.
[0122] Example 3 includes the method of any one of examples 1-2, including or excluding an optional feature, in which the lookback window defines a number of consecutive previous data points to include in each training sample.
[0123] Example 4 includes the method of any one of examples 1-3, including or excluding optional features, in which the prediction horizon defines a predefined amount of time into the future, and each label of each data point the predefined amount of time into the future is associated with each training sample.
[0124] Example 5 includes the method of any one of examples 1-4, including or excluding optional features. In this example, each training sample is clustered using K-means clustering.
[0125] Example 6 includes the method of any one of examples 1-5, including or excluding optional features. In this example, the method is performed by the failure prediction system in accordance with software downloaded to the failure prediction system from a remote data processing system. Optionally, the method further includes metering usage of the software and generating an invoice based on the metering of usage.
[0126] Example 7 is a system, including one or more processors and one or more computer-readable storage media having program instructions stored thereon, the program instructions, when executed by the one or more processors, configured to cause the one or more processors to perform a method including: inputting a new data sample into a failure prediction model, the failure prediction model being trained using a labeled historical data set, each data point being associated with a lookback window and a prediction range to create each training sample, each training sample being clustered into a plurality of clusters, each of the plurality of clusters being associated with a normal score and an abnormality score, and outputting a classification associated with the new data sample based on comparing a first anomaly score of a first cluster of the plurality of clusters including the new data sample to an average anomaly score of clusters of the plurality of clusters having a normal score greater than the anomaly score.
[0127] Example 8 includes the system of example 7, including or excluding the optional feature, where the classification indicates a likelihood of performance degradation associated with wear on the asset associated with the new data sample.
[0128] Example 9 includes the system of any one of examples 7-8, including or excluding the optional feature, in which the lookback window defines the number of consecutive previous data points to include in each training sample.
[0129] Example 10 includes the system of any one of examples 7-9, including or excluding optional features, in which the prediction horizon defines a predefined amount of time into the future, and each label of each data point the predefined amount of time into the future is associated with each training sample.
[0130] Example 11 includes the system of any one of examples 7-10, including or excluding the optional feature, in which each training sample is clustered using K-means clustering.
[0131] Example 12 is a computer program product. The computer program product includes one or more computer readable storage media and program instructions collectively stored on the one or more computer readable storage media, the program instructions including instructions configured to cause one or more processors to execute a method including inputting a new data sample to a failure prediction model, the failure prediction model being trained using a labeled historical data set, each data point being associated with a lookback window and a prediction range to create each training sample, each training sample being clustered into a plurality of clusters, each of the plurality of clusters being associated with a normal score and an anomaly score, and outputting a classification associated with the new data sample based on comparing a first anomaly score of a first cluster of the plurality of clusters including the new data sample to an average anomaly score of clusters in the plurality of clusters having a normal score greater than the anomaly score.
[0132] Example 13 includes the computer program product of example 12, including or excluding optional features, in which the classification indicates a likelihood of performance degradation associated with wear on the asset associated with the new data sample.
[0133] Example 14 includes the computer program product of any one of Examples 12-13, including or excluding an optional feature, in which the lookback window defines the number of consecutive previous data points to include in each training sample.
[0134] Example 15 includes the computer program product of any one of examples 12-14, including or excluding optional features, in which the prediction horizon defines a predefined amount of time into the future, and each label of each data point the predefined amount of time into the future is associated with each training sample.
[0135] Example 16 includes the computer program product of any one of Examples 12-15, including or excluding optional features, in which each training sample is clustered using K-means clustering.
[0136] Example 17 is a computer-implemented method for predicting wear-related deterioration of a PCP, the method including: inputting new data samples of the PCP into a model configured to predict wear-related deterioration of the PCP, the model being trained using a labeled historical PCP dataset, each data point being associated with a lookback window and a prediction range to create each training sample, each training sample being clustered into a plurality of clusters, each of the plurality of clusters being associated with a normal score and an abnormality score; and outputting a classification associated with the new data sample based on comparing a first anomaly score of a first cluster of the plurality of clusters including the new data sample to an average anomaly score of clusters in the plurality of clusters having a normal score greater than the anomaly score, the classification being indicative of wear-related deterioration of the PCP.
[0137] Example 18 is a computer-implemented method for predicting wear-related deterioration of a PCP, the method including: generating labeled historical data by performing binary labeling of historical data associated with one or more PCPs; generating a plurality of training data samples by applying a lookback window and a prediction range to each data point of the labeled historical data; clustering the plurality of training data samples into a plurality of clusters; calculating a cluster score for each cluster of the plurality of clusters; assigning a new data sample of the PCP to a first cluster of the plurality of clusters; and assigning a classification to the new data sample based on the cluster score associated with the first cluster, the classification being indicative of a likelihood of future wear-related deterioration of the PCP.
[0138] Example 19 includes the method of example 18, including or excluding optional features. In this example, the labeled historical data is labeled as faulty for a predefined period of time prior to a known pump replacement date.
[0139] Example 20 includes the method of any one of examples 18-19, including or excluding optional features. In this example, the labeled historical data includes pump speed data, pump torque data, casing pressure data, production rate data, and maintenance records.
[0140] Example 21 includes the method of any one of examples 18-20, including or excluding an optional feature. In this example, calculating a cluster score for each cluster further includes: calculating a normal score for the first cluster, the normal score being a first proportion of training data samples associated with a normal state in the first cluster divided by a second proportion of training data samples associated with a normal state in the plurality of training data samples; and calculating an anomaly score for the first cluster, the anomaly score being a third proportion of training data samples associated with a degraded state in the first cluster divided by a fourth proportion of training data samples associated with a degraded state in the plurality of training data samples. Optionally, the classification is based on a greater value of the normal score or the anomaly score for the first cluster.
[0141] Example 22 includes the method of any one of examples 18-21, including or excluding optional features. In this example, the method includes generating a fault signal for the new data sample, the fault signal including an average anomaly score for the new data sample over a predetermined number of previous data points. Optionally, generating the fault signal further includes calculating an average anomaly score for clusters of the plurality of clusters having a normal score greater than the anomaly score, associating, for each of the predetermined number of previous data points, a value of 1 to data points having an anomaly score of the first cluster greater than the average anomaly score and associating a value of 0 to data points having an anomaly score of the first cluster less than the average anomaly score, and calculating the fault signal as an average of the values of 1 and 0 associated with each of the predetermined number of previous data points.
Claims
1. inputting new data samples into a failure prediction model, the failure prediction model being trained using a labeled historical dataset, each data point being associated with a lookback window and a prediction range to create each training sample, each training sample being clustered into a number of clusters, and each of the clusters being associated with a normal score and an abnormal score; and outputting a classification associated with the new data sample based on comparing a first anomaly score of a first cluster of the plurality of clusters that includes the new data sample to an average anomaly score of clusters of the plurality of clusters having the normal score greater than the anomaly score.
2. The method of claim 1 , wherein the classification indicates a likelihood of performance degradation associated with wear of an asset associated with the new data sample.
3. The method of claim 1 , wherein the lookback window defines a number of consecutive previous data points to include in each training sample.
4. 2. The method of claim 1, wherein the prediction horizon defines a predefined amount of time into the future, and a respective label for each data point that is in the future by the predefined amount of time is associated with each of the training samples.
5. The method of claim 1 , wherein the training samples are clustered using K-means clustering.
6. The method of claim 1 , wherein the method is performed by a failure prediction system in accordance with software downloaded to the failure prediction system from a remote data processing system.
7. The method further comprising: Measuring usage of said software; and generating an invoice based on the measurement of usage.
8. 2. The method of claim 1, wherein the method is for predicting wear-related deterioration of a PCP, the input step includes inputting new data samples of a PCP into a model configured to predict wear-related deterioration of the PCP, the model being trained using a labeled historical PCP dataset, and the classification output during the output step being indicative of the wear-related deterioration of the PCP.
9. generating the labeled historical data by performing a binary labeling of historical data associated with one or more PCPs, wherein the plurality of training data samples are created by applying a look-back window and a prediction range to each data point of the labeled historical data, the method comprising: calculating a cluster score for each cluster of the plurality of clusters, the cluster scores being the normal score and the abnormal score; assigning the new data sample of a PCP to a first cluster of the plurality of clusters; and assigning a classification to the new data sample based on a cluster score associated with the first cluster.
10. 10. The method of claim 9, wherein the labeled historical data is labeled as faulty for a predefined period of time prior to a known pump replacement date.
11. 10. The method of claim 9, wherein the labeled historical data includes pump speed data, pump torque data, casing pressure data, production rate data, and maintenance records.
12. Calculating the cluster score for each of the clusters, calculating a normality score for the first cluster, the normality score being a first proportion of training data samples associated with a normal state in the first cluster divided by a second proportion of training data samples associated with the normal state in the plurality of training data samples; 10. The method of claim 9, further comprising: calculating an anomaly score for the first cluster, the anomaly score being a third proportion of training data samples associated with a deterioration state in the first cluster divided by a fourth proportion of training data samples associated with the deterioration state in the plurality of training data samples.
13. The method of claim 12 , wherein the classification is based on a greater value of the normal score or the abnormal score of the first cluster.
14. The method further comprising:
10. The method of claim 9, further comprising generating a fault signal for the new data sample, the fault signal comprising an average anomaly score for the new data sample over a predetermined number of previous data points.
15. generating the fault signal calculating an average cluster anomaly score for the plurality of clusters having a normal score greater than the anomaly score; for each of the predetermined number of previous data points, associating a value of 1 with data points having an anomaly score of the first cluster greater than the average anomaly score, and associating a value of 0 with data points having an anomaly score of the first cluster less than the average anomaly score; 15. The method of claim 14, further comprising: calculating the fault signal as an average value of the one and zero values associated with each of the predetermined number of previous data points.
16. one or more processors; and one or more computer-readable storage media storing program instructions, the program instructions, when executed by the one or more processors, causing the one or more processors to: inputting new data samples into a failure prediction model, the failure prediction model being trained using a labeled historical dataset, each data point being associated with a lookback window and a prediction range to create each training sample, each training sample being clustered into a number of clusters, and each of the clusters being associated with a normal score and an abnormal score; and outputting a classification associated with the new data sample based on comparing a first anomaly score of a first cluster of the plurality of clusters that includes the new data sample to an average anomaly score of clusters of the plurality of clusters having the normal score greater than the anomaly score.
17. The system of claim 16 , wherein the classification indicates a likelihood of performance degradation associated with wear of an asset associated with the new data sample.
18. 17. The system of claim 16, wherein the lookback window defines a number of consecutive previous data points to include in each training sample.
19. 17. The system of claim 16, wherein the prediction horizon defines a predefined amount of time into the future, and a respective label for each data point that is the predefined amount of time into the future is associated with each of the training samples.
20. 17. The system of claim 16, wherein the training samples are clustered using K-means clustering.
21. A computer-readable recording medium having recorded thereon a computer program for carrying out the method according to any one of claims 1 to 15.
22. 1. A computer-implemented method for predicting wear-related degradation in a PCP, the method comprising: inputting new data samples of a PCP into a model configured to predict wear-related deterioration of the PCP, the model being trained using a labeled historical PCP dataset, each data point being associated with a look-back window and a prediction range to create each training sample, each training sample being clustered into a number of clusters, each of the multiple clusters being associated with a normal score and an abnormal score, respectively; and outputting a classification associated with the new data sample based on comparing a first anomaly score of a first cluster of the plurality of clusters that includes the new data sample to an average anomaly score of clusters of the plurality of clusters having the normal score greater than the anomaly score, the classification being indicative of deterioration associated with the wear of the PCP.
23. 1. A computer-implemented method for predicting wear-related degradation in a PCP, the method comprising: generating labeled historical data by performing a binary labeling of historical data associated with one or more PCPs; generating a plurality of training data samples by applying a lookback window and a prediction horizon to each data point of the labeled historical data; clustering the plurality of training data samples into a plurality of clusters; calculating a cluster score for each cluster in the plurality of clusters; assigning new data samples of the PCP to a first cluster of the plurality of clusters; and assigning a classification to the new data sample based on a cluster score associated with the first cluster, the classification indicating a likelihood of deterioration associated with future wear of the PCP.
24. 24. The method of claim 23, wherein the labeled historical data is labeled as faulty for a predetermined period of time prior to a known pump replacement date.
25. 24. The method of claim 23, wherein the labeled historical data includes pump speed data, pump torque data, casing pressure data, production rate data, and maintenance records.
26. Calculating the cluster score for each of the clusters, calculating a normality score for the first cluster, the normality score being a first proportion of training data samples associated with a normal state in the first cluster divided by a second proportion of training data samples associated with the normal state in the plurality of training data samples; 24. The method of claim 23, further comprising: calculating an anomaly score for the first cluster, the anomaly score being a third proportion of training data samples associated with a deterioration state in the first cluster divided by a fourth proportion of training data samples associated with the deterioration state in the plurality of training data samples.
27. 27. The method of claim 26, wherein the classification is based on a greater value of the normal score or the abnormal score of the first cluster.
28. The method further comprising:
24. The method of claim 23, further comprising generating a fault signal for the new data sample, the fault signal comprising an average anomaly score for the new data sample over a predetermined number of previous data points.
29. generating the fault signal calculating an average cluster anomaly score for the plurality of clusters having a normal score greater than the anomaly score; for each of the predetermined number of previous data points, associating a value of 1 with data points having an anomaly score of the first cluster greater than the average anomaly score, and associating a value of 0 with data points having an anomaly score of the first cluster less than the average anomaly score; 30. The method of claim 28, further comprising: calculating the fault signal as an average value of the one and zero values associated with each of the predetermined number of previous data points.
30. A computer program causing a computer to execute the method according to any one of claims 1 to 15 and claims 22 to 29.
Citation Information
Patent Citations
Abnormality sign detection system and abnormality sign detection method
JP2015088078A
Numerical control device and machine learning device
JP2018097616A
Tool state estimation device and machine tool
JP2018138327A
Pump system
JP2019199874A
