Method for predicting downtime of production for an activity and predicting a cause of the downtime of production

Abstract

The present invention relates to a computer-implemented method (200) for predicting downtime of production for an activity and predicting a cause of the downtime of production. The method comprises the steps of: receiving (210) input data comprising a production description of the activity, the production description of the activity including a raw natural language text description made by an operator; predicting (220), by an artificial intelligence, Al, model, a downtime of the production using at least a part of the input data; predicting (230), by the Al model, a cause of the downtime of the production using at least a part of the input data; indicating (240) a predicted downtime of the production for the activity and a predicted cause of the downtime. In addition, the present invention relates to a corresponding Al model, a computer-implemented method for training the Al model as well as a data-processing device and computer program.

Description

[0001] METHOD FOR PREDICTING DOWNTIME OF PRODUCTION FORAN ACTIVITY AND PREDICTING A CAUSE OF THE DOWNTIME OF PRODUCTION

[0002] Field of the invention

[0003] The present disclosure relates to a computer-implemented method, a data processing device and a computer program for predicting downtime of production for an activity and predicting a cause of the downtime of production. In addition, the present disclosure relates to an artificial intelligence model for use in predicting downtime of production for an activity and predicting a cause of the downtime of production and a training method of an artificial intelligence model for use in predicting downtime of production for an activity and predicting a cause of the downtime.

[0004] Background

[0005] Unplanned downtime is a concern in some industries, in particular in the oil and gas industry. The offshore sector, for example, has been loaded with unnecessarily costs and time due to the unplanned downtime and this problem worsens as aging assets and the loss of industry experience take their toll.

[0006] These challenges make efficiency of operation monitoring and maintenance a choke point for many industries, in particular in oil and gas industries, bringing opportunities to become increasingly efficient.

[0007] The oil and gas industry has grappled with quantifying the cost of unplanned downtime and studies revealed that offshore oil and gas industries experience on a large operating loss to unplanned downtime. In addition, fewer than 24% of industries describe their maintenance approach as a predictive one based on data and analytics. Therefore, over three-quarters either take a reactive or time-based approach. In addition, industries using a predictive, data-based approach experience 36% less unplanned downtime than those with a reactive approach. In addition, the existing system and process to report production and downtime losses are based on some traditional methods, which include reporting data in spreadsheet formats and communicating information through e-mails while using several tools, repository systems, and databases which vary. In fact, the reports are generated by operators. Operators who are in charge of operational and production activities have to write textual reports of all observations on events and the production activities. These reports contain, for instance, data, findings, unfavorable events, state of the equipment, activity description and summaries of the situation. The report is a reference for studying the current situation and to contextualize any event or operation which had happened in the past. Every day, many reports are written to describe in detail the operational and production activities that took place during the day for each activity.

[0008] In operating industries which follow a reactive approach by reporting, on a daily-basis, through excel file, the production by field along with descriptions of the main events that have affected the production target, the monitoring part (i.e. , identifying losses and relevant downtime analyses from the reports as well as finding the root causes of the critical challenges) is mainly done manually by the stakeholders.

[0009] However, traditional methods carry several business limitations like inefficiency, long cycle time, inaccuracies in the data reporting, and sub-optimal planning. Moreover, the ability to respond quickly is not guaranteed as data collection from multiple sources is a critical bottleneck that further delays the decision-making process leading to efficiency losses.

[0010] Therefore, the production monitoring and maintaining process faces the different challenges and limitations, in particular, a lack of automatic allocation of losses (planned / unplanned and their activities), a lack of relevant downtime analyses, a lack of operating efficiency improvements recommendations, a lack of visualization dashboard to display, production, targets, and operating efficiency for a different combination of fields / assets, a lack of management of change and delay in decision making.

[0011] Against this background, there is a need for a method for predicting downtime of production for an activity and predicting a cause of the downtime of production. Summary of the invention

[0012] The above-mentioned problem is at least partly solved by a computer-implemented method according to aspect 1, by an Al model according to aspect 7, by a computer- implemented method according to aspect 13, by a data processing device according to aspect 18 and a computer program according to aspect 19.

[0013] A 1staspect of the present invention refers to a computer-implemented method for predicting downtime of production for an activity and predicting a cause of the downtime of production, the method comprising the steps of: receiving input data comprising a production description of the activity, the production description of the activity including a raw natural language text description made by an operator; predicting, by an artificial intelligence, Al, model, a downtime of the production using at least a part of the input data; predicting, by the Al model, a cause of the downtime of the production using at least a part of the input data; indicating a predicted downtime of the production for the activity and a predicted cause of the downtime.

[0014] Generating a predicted downtime of production for an activity and a predicted cause of the downtime of production based on a production description of the activity including a raw natural language text description made by an operator allow an automatic monitoring of textual production reports with an automatic detection of downtime and an automatic detection of root causes of downtime. The result of the predictive analysis allows a better production performance monitoring, an operational efficiency improvements recommendation, and proactive identification of areas of focus and optimization as well as a reduction in unplanned losses due to downtime of production for an activity.

[0015] According to a 2ndaspect in the 1staspect, the downtime predicted by the Al model comprises a probability of the downtime, and the cause of the downtime predicted by the Al model comprises a probability of the cause. The probability of the downtime of the production for the activity and the probability of the cause of the downtime give the user a measure of uncertainty involved in the prediction.

[0016] According to a 3rdaspect in the 2ndaspect, the method further comprises determining, by the Al model, a data drift report comprising an analysis of the input data compared to historical data.

[0017] Determining a data drift report provides a comprehensive analysis of the drifting patterns of data and Al model behavior with respect to the data.

[0018] According to a 4thaspect in any one of the 2ndto 3rdaspect, the method further comprises determining, by the Al model, a model drift report comprising an analysis of the prediction made by the Al model.

[0019] Determining a model drift report provides a description of the quality / robustness of model outputs.

[0020] According to a 5thaspect in the 3rdand 4thaspects, a part of the input data comprises information on at least one of: a target production value of the activity, a production value of the activity and a time-series operating data of the activity.

[0021] According to a 6thaspect in any one of the 1stto 5thaspect, the activity comprises a well production and the production description of the activity includes a description of the well production in a log file.

[0022] A 7thaspect refers to a Al model for use in predicting downtime of production for an activity and predicting a cause of the downtime of production, the model having a structure comprising a natural language processing to process a production description of the activity to predict a downtime and a cause of the downtime.

[0023] According to an 8thaspect in the preceding aspect, the natural language processing is implemented by a neural network, preferably a transformer network. According to a 9thaspect in any one of the 7thto 8thaspect, the Al model further comprises a data drift module to generate a data drift report comprising an analysis of input data compared to historical data.

[0024] According to a iothaspect in the preceding aspect, the data drift module is implemented by an adversarial machine learning model.

[0025] According to a 11thaspect in any one of the 9thto 10thaspect, the Al model further comprises a learning module to train the natural language processing based on a predicted downtime and a predicted cause of the downtime and on a data drift report generated.

[0026] According to a 12thaspect in the preceding aspect, the learning module further comprises a continual learning module and an Al model monitoring module to analyze the prediction of the Al model and to have a feedback loop with the natural language processing.

[0027] A 13thaspect refers to a computer-implemented method for training an intelligence, Al, model for use in predicting downtime of production for an activity and predicting a cause of the downtime, the method comprising a step of: receiving training data to train the artificial intelligence model, wherein the training data at least comprises: an production description of an activity, a predicted downtime and a predicted cause of the downtime.

[0028] According to a 14thaspect in the preceding aspect, the training data further comprises a data drift analysis.

[0029] According to a 15thaspect in any one of the 13thto 14thaspect, the training data further comprises a model monitoring analysis.

[0030] According to a 16thaspect in any one of the 1stto 6thaspect, the Al model has been trained according to the method of any one of the 13thto 15thaspect. According to a 17thaspect in any one of the 1stto 6thaspect and the 16thaspect, the method comprises determining that new training data is available and training the Al model according to the method of any one of the 13thto 15thaspect using the new training data.

[0031] An 18thaspect refers to a data processing device comprising means configured to perform the method according to the 1stto 6thand / or 12thto 17thaspect.

[0032] A 19thaspect refers to a computer program comprising instructions, which when executed by a computer, causes the computer to perform the method according to the 1stto 6thor 12thto 17thaspect.

[0033] Brief description of the drawings

[0034] Various aspects of the present invention are described in more detail in the following by reference to the accompanying figures, without the present invention being limited to the embodiments of these figures.

[0035] Fig. 1 illustrates an overview of a method for predicting downtime of production for an activity and predicting a cause of the downtime of production according to aspects of the present invention.

[0036] Fig. 2 illustrates a flow chart of a method for predicting downtime of production for an activity and predicting a cause of the downtime of production according to aspects of the present invention.

[0037] Fig. 3 illustrates a deep learning model architecture combining different layers of encoder according to aspects of the present invention.

[0038] Fig. 4 illustrates an embodiment of the data drift module according to aspects of the present invention.

[0039] Fig. 5 illustrates an embodiment of the continual learning module according to aspects of the present invention. Detailed description

[0040] In the following, certain aspects of the present invention are described in more detail.

[0041] Fig. i illustrates an overview of an algorithm 100 according to aspects of the present invention used for predicting downtime of production for an activity and predicting a cause of the downtime of production. The algorithm 100 is based on a workflow, namely an artificial intelligence, Al, workflow 120 implementing an Al model. The input data on which the method operates is referred to as operation report 110.

[0042] An operation report 110 may be a daily operation report and may be generated by an operator. An operation report 110 on which downtime for an activity and a cause of the downtime may be predicted is associated with a production for an activity. The operation report may comprise at least a production description of an activity 110a. The production description of the activity 110a may comprise all the information about the production, targets and reasons for daily activity. The production description of the activity 110a maybe realized by an operator who is in charge of supervising the activity. The operator may write in the operation report a production description of the activity to report events and activities that occur, for example during a day. The production description of the activity may comprise, for instance, findings, unfavorable events and summaries of the operation and of the production. The production description of the activity may include a natural language text description.

[0043] The operation report 110 may also comprise information on at least one of: a target production value of the activity, a production value of the activity and a time-series operating data of the activity 110b. The operation report 110 may also comprise other data acquired or collected from sensors describing for instance, the production state or the equipment state of the activity.

[0044] The activity may comprise a well production and the production description of the activity may include a description of the well production for example in a log file. While the Al workflow 120 operates on the operation report 110, it is possible that it only operates on certain parts of the operation report. For example, the IA workflow 120 may only use the production description of the activity as input data.

[0045] The input data maybe inputted into an artificial intelligence, Al, model for predicting downtime of production for an activity and predicting a cause of the downtime of production using at least a part of the operation report, in particular the production description of the activity.

[0046] The Al workflow 120 may include a module 120a for standardizing and transforming the input data, i.e. the operation report 110, a natural language processing (NLP) engine 120b, a data drift module 120c and a learning module i2od. The learning model I2od may comprise a continual learning module I2odi and an Al model monitoring module I2od2.

[0047] The data standardization and transformation module 120a may preprocess the operation report 110, in particular the production description of the activity, e.g., by converting it to lower case or by converting it to upper case to standardize the input data. The output of the data standardization and transformation module 120a may be passed to the NLP engine 120b and the data drift module 120c.

[0048] The NLP engine 120b may be intertwined with the continual learning module I2odi and the Al model monitoring module I2od2 of the learning module I2od which forms into a feedback loop with continuous input from the data drift module 120c as illustrated in Fig. 1. This Al workflow 120 addresses the issues of model uncertainty and data drift, while laying the foundation for continuous learning in a semi-supervised approach.

[0049] The NLP engine 120b may comprise a fine-tuned transformers module I2obi that may include different fine-tuned transformers adapted to perform different tasks and an uncertainty module I2ob2 which may estimate the uncertainty of the Al model. The output 130 from the Al workflow 120 may comprise and may indicate a predicted downtime of the production for the activity 130a and a predicted cause of the downtime 130b. The predicted cause of the downtime may include one or more root causes.

[0050] The downtime predicted by the Al model may comprise a probability of the downtime, and the cause of the downtime predicted by the Al model may comprise a probability of the cause. The probabilities further give the user the measure of uncertainty involved in the prediction.

[0051] The output 130 from the Al workflow 120 may also include a data drift report 130c and a model drift report I3od. They provide a comprehensive analysis of the drifting patterns of data and Al model behavior with respect to the data. In particular, the data drift report may comprise an analysis of the input data compared to historical data and the model drift report may comprise an analysis of the prediction made by the Al model. This, the data drift report provides an analysis of the quality of the data whereas the model drift report provides an analysis of the quality / robustness of the Al model.

[0052] The algorithm has the advantage of identifying and / or detecting the downtime / shortfall from an unstructured textual operation report in near real-time to enhance the efficiency in achieving the production targets and helps the production industry to manage the risks associated with shortfall and mitigate it further. The algorithm may also use the target and production values for each asset / field across time.

[0053] The output 130 of the Al workflow 120 may go into an interactive visualization layer where an asset / filed profile is created across time which lets the user analyze historical production trends, measure the quantum of production impact across time, and other import key performance indicators. Moreover, the visualization layer may also illustrate the projected trend of downtime time along with some inferential statistics and forecasts.

[0054] The visualization layer of the method for predicting downtime of production for an activity and predicting a cause of the downtime of production may also help in monitoring key performance indicators which can be customized on the fly as per user convenience.

[0055] The algorithm pertains to the use of artificial intelligent technology to enable a predictive optimization approach for asset management, monitoring and planning the production for an activity for example in Oil and Gas operator industries. More specifically, the method conveys to a methodology to drive operational efficiency and profitability by reducing unplanned deferred production.

[0056] The Al workflow 120 may also identify and at the same time predict the probable cause for the production shortfall for the activity. In addition, it may help to identify losses, analyze relevant downtime and forecast deficits. It may also address the lack of comprehensive and specific strategies for asset repair and maintenance.

[0057] Further the Al engine 120 may provide asset performance management facilities where indicative strategies could be found for each field about asset maintenance and repair.

[0058] Fig. 2 illustrates a flow chart of a method 200 for predicting downtime of production for an activity and predicting a cause of the downtime of production according to aspects of the present invention. The method 200 maybe based on the algorithm too. In step 210, input data maybe received. Step 210 may refer to the operation report 110 of the algorithm too.

[0059] The operation report may comprise at least a production description of the activity 110a, the production description of the activity including a raw natural language text description made by an operator. The production description of the activity 110a may be realized by an operator and may include a natural language text description. The operation report may also include data collected from sensors present to manage the production of the activity or the equipment of this activity.

[0060] In step 220, a predicted downtime of the production maybe generated using a trained Al model based on at least a part of the input data, i.e. the production description of the activity made by the operator. In particular, the description used by the trained Al model may be a raw natural language text description written by the operator, on which no treatment on the description has been carried out, i.e., error correction, automated adding text, etc. The trained Al model may have a structure comprising a natural language processing to process a production description of the activity to predict a downtime of production and a cause of the downtime of production. The natural language processing maybe implemented by a deep learning model, and in particular a neural network, preferably a transformer network. The natural language processing may be adapted for example for oil and gas domain to help the model to capture syntactic and semantic features of the text production description which is specific to the industry. The deep learning model may extract the contextual embedding from the in-domain trained deep learning network which serves as a feature space in form of embedding further to be used in the transformer model. Step 220 may refer to the Al workflow 120 of the algorithm too. The predicted downtime of production for the activity may comprise a probability of the downtime of production.

[0061] In step 230, a predicted cause of the downtime of the production may be generated using the trained Al model based on at least a part of the input data, i.e. the production description of the activity made by the operator. The predicted cause of the downtime od production may comprise a probability of the cause of the downtime.

[0062] In step 240, a predicted downtime of the production for the activity and a predicted cause of the downtime may be indicated. The step of indicating the result may comprise a visualization which lets the user to analyze historical production trends, measure the quantum of production impact over time, and other key performance indicators. Moreover, the visualization layer may also illustrate the projected trend of downtime time along with some inferential statistics and forecasts.

[0063] According to the method, the prediction of a downtime of production from the production description which is reported, is based on an unstructured report format. These reports which contain production description of an activity are processed, transformed and standardized with the use of Natural Language Processing (NLP) to match production volumes and target volumes, which maybe indexed across date time. Further, the production descriptions of the activity are used to teach an Al model in a self-supervised method to detect the downtime and its root causes in isolation from the production description. In addition, certain data attributes may also be collected over time from the input data and may provide an insight in the form of a report, i.e. data drift report 130c from the data drift module 120c which captures the shifts in linguistic patterns of the textual production description reports over time by tracking the syntactic and semantic differences in the text.

[0064] A continual learning module I2odi may also allow the Al model to supervise its learning without the availability of an annotator or a subject matter expert.

[0065] Additional information may be indicated on the uncertainty of the Al model predictions which helps the end user to estimate the risks involved for incorrect predictions, in particular in a report, i.e. the model drift report I3od.

[0066] The invention has the advantage to increase production efficiency on activities. The monitoring needs to be conducted at a high resolution in order to understand the performances of the current processes in the field and highlight indicators of deficiencies.

[0067] The invention has also the advantage to provide an efficient monitoring, asset performance management, predictive optimization, and comprehensive strategies for asset repair and maintenance. This can be done in real time with a high level of accuracy (for example of more than 95%) which will increase over time with the learning module i2od. The twining architecture of the data drift module 120c and the learning model i2od serves as a solid basis to continual learning which makes it selfsupervised and thus will help to improve the robustness of predictions.

[0068] An Al model for use in predicting downtime of production for an activity and predicting a cause of the downtime of production may have a structure comprising a natural language processing to process a production description of the activity to predict a downtime and a cause of the downtime. The natural language processing may be implemented by a deep learning model and in particular a neural network, preferably a transformer network. According to an embodiment, the deep learning model implementing the natural language processing is a customized transformer architecture which may combine different layers 310, 320, 330 of encoder structurally along with some modifications in training mechanism as illustrated in Fig. 3. The modifications may include intelligent masking of embedding feature space in each pass of the dataset to the model flow as illustrated in Fig. 3.

[0069] The uncertainty module I2ob2 shown in Fig. 1 may be, according to an exemplary embodiment, a pipeline for estimating deep learning model uncertainty. The module may use isotonic regression concepts coupled with the calibration error metric and predictive entropy to combat epistemic and random uncertainty. With this module, the quality of uncertainty is determined and addresses it using deep ensembles.

[0070] The Al model further may comprise a data drift module to generate a data drift report comprising an analysis of input data compared to historical data. The data drift module maybe implemented by an adversarial machine learning model.

[0071] The Al model may further comprise a learning module to train the natural language processing based on a predicted downtime and a predicted cause of the downtime and on a data drift report generated. The learning module may further comprise a continual learning module and an Al model monitoring module to analyze the prediction of the Al model.

[0072] Fig. 4 illustrates an embodiment of the data drift module 120c illustrated in Fig 1.

[0073] The data drift module 120c may generate a data drift report 130c comprising an analysis of input data compared to historical data. The data drift module 120c maybe implemented by an adversarial machine learning model.

[0074] The data drift module 120c may encapsulate three different components namely variance shift detector 410, evaluate feature space 420 and adversarial machine learning model 430. Variance shift detector 410 is a module that may determine the variance shift in the new data 440 as compared to the historical data 450. The determination may be based on the principles of covariance matrix which may give a score of dispersion.

[0075] Similarly, the feature space of the text maybe extracted 460, 470 from the in-domain trained deep learning network for the new as well as historical data and evaluated by the component evaluate feature space 420 using a combination of dissimilarity metrics. These metrics may include cosine similarity, Jaccard distance and Minkowski distance. The result of this operation may be a dissimilarity score.

[0076] The third component of the data drift module 120c may concern an adversarial machine learning model module 430 which may output area under the receiver operating characteristic curve (ROC). The output score 480 allows to know how indistinguishable the new data with respect to historical data is.

[0077] According to a particular embodiment, an adversarial machine learning model adaption module 490 may retrain the adversarial machine learning model 430 in production if the score 480 is below a prescribed range and update the weights of the model. It also allows to track the decay in drift patterns over time and provides a comprehensive report on the patterns over time, i.e. a data drift report 130c.

[0078] Altogether the scores from the three components 410, 420 and 430 may be combined using weighted average to attain one score which may give some insights about the distribution of new data from the existing data.

[0079] The continual learning module I2odi illustrated in Fig. 1 is detailed in Fig. 5 according to an embodiment.

[0080] In a particular embodiment, a method to adapt the NLP engine 120b to data shifts and uncertainties in the model prediction by continuously teaching itself in a self-supervised manner may be implemented. In particular, the combination of a small amount of labeled data with a large amount of unlabeled data during training may be used. The continual learning module i2odi allows the Al model to be continuously and semi-automatically adapted to changes and evolutions in the data by re-training the deep learning model in a semi-supervised manner.

[0081] An Al model for use in predicting downtime of production for an activity and predicting a cause of the downtime of production may be trained according to a computer-implemented method for training an Al model for use in predicting downtime of production for an activity and predicting a cause of the downtime according to aspects of the present invention. Such a method may comprise a step of receiving training data to train the artificial intelligence model. The input training data may at least comprise: a production description of an activity, a predicted downtime and a predicted cause of the downtime. The training method makes the Al model more robust and accurate.

[0082] While during application of the trained Al model, the downtime of production for an activity and the cause of the downtime to be predicted are unknown, the downtime of production for an activity and the cause of the downtime during training is known to be able to optimize the Al model.

[0083] In some embodiments, the training data may continuously be updated once new input data or training data is available. Thus, the new report with predicted downtime of production for an activity and predicted cause of the downtime of production maybe added to the existing training data.

[0084] The training data may further comprise a data drift analysis that can be memorized for example in a data drift report and / or a model monitoring analysis that can be memorized for example in a model drift report.

[0085] The training data may be used to train the Al model.

[0086] In particular, the continual learning workflow of the continual learning module I2odi illustrated in Fig. 5 may take input from the NLP engine 120b and the model monitoring module I2od2 and may have a feedback loop attached to it, which guides the training procedure in real time. Also, it may take input from the data drift module 120c and combine it with input from the Al model monitoring module I2od2, to help the continual learning decisions.

[0087] According to the embodiment described in Fig. 5, a self-supervised module 510 may be linked to three in-domain deep learning networks 520, 530 and 540 to handle different tasks at once.

[0088] Each deep learning network maybe implemented by a neural network, preferably a transformer network which may be adapted. Indeed, each fine-tuned transformer may be adapted to perform different tasks namely binary classification 520, Next sentence prediction (NSP) 530 and Named Entity Recognition (NER) 540 on the same set of text. The method for continual learning based on different deep learning networks transforms the input into different tasks which trains and evaluates to capture different inherent characteristic of the inputs via different representations.

[0089] Each transformer 520, 530 and 540 implemented by means of a model container transforms the text to embedding suited for the task, which trains the deep learning model and flows the output in conjunction to ensemble module 550. The outputs from the three different deep learning models are assimilated and a strong learner is created with the help of a meta model as these models capture different nuances of the text.

[0090] The output from ensemble module 550 flows to self-supervised module 510 where prior information on the data drift and model drift is combined with new reports directs the adaptive model update module 560 to retrain the deep learning networks.

[0091] The method for continual learning may contain a training pipeline to teach the model in isolation.

[0092] In addition, the self-supervised module may trigger adaptive model update module 560 to save and store the model weights for further re training.

[0093] The aspects according to the present invention may be implemented in terms of a computer program which may be executed on any suitable data processing device comprising means (e.g., a memory and one or more processors operatively coupled to the memory) being configured accordingly. The computer program may be stored as computer-executable instructions on a non-transitory computer-readable medium. Embodiments of the present disclosure maybe realized in any of various forms. For example, in some embodiments, the present invention maybe realized as a computer-implemented method, a computer-readable memory medium or a computer system. The present invention may also be realized in an integrated computer environment comprising a user computer system, a network and a server, wherein the user computer system implements a web browser, and the user computer system displays web pages provided to the web browser by the server.

[0094] In some embodiments, a non-transitory computer- readable memory medium may be configured so that it stores program instructions and / or data, where the program instructions, if executed by a computer system, cause the computer system to perform a method, e.g. any of the method embodiments described herein, or any combination of the method embodiments described herein, or any subset of any of the method embodiments described herein, or any combination of such subsets.

[0095] In some embodiments, a computing device maybe configured to include a processor (or a set of processors) and a memory medium, where the memory medium stores program instructions, where the processor is configured to read and execute the program instructions from the memory medium, where the program instructions are executable to implement any of the various method embodiments described herein (or any combination of the method embodiments described herein, or any subset of any of the method embodiments described herein, or any combination of such subsets). The device maybe realized in any of various forms.

[0096] Although specific embodiments have been described above, these embodiments are not intended to limit the scope of the present disclosure, even where only a single embodiment is described with respect to a particular feature. Examples of features provided in the disclosure are intended to be illustrative rather than restrictive, unless stated otherwise. The above description is intended to cover such alternatives, modifications, and equivalents as would be apparent to a person skilled in the art having the benefit of this disclosure.

[0097] The scope of the present disclosure includes any feature or combination of features disclosed herein (either explicitly or implicitly), or any generalization thereof, whether or not it mitigates any or all of the problems addressed herein. In particular, with reference to the appended claims, features from dependent claims may be combined with those of the independent claims and features from respective independent claims may be combined in any appropriate manner and not merely in the specific combinations enumerated in the appended claims.

Claims

CLAIMS1. Computer-implemented method (200) for predicting downtime of production for an activity and predicting a cause of the downtime of production, the method comprising the steps of: receiving (210) input data comprising a production description of the activity, the production description of the activity including a raw natural language text description made by an operator; predicting (220), by an artificial intelligence, Al, model, a downtime of the production using at least a part of the input data; predicting (230), by the Al model, a cause of the downtime of the production using at least a part of the input data; indicating (240) a predicted downtime of the production for the activity and a predicted cause of the downtime.

2. The method of the preceding claim, wherein the downtime predicted by the Al model comprises a probability of the downtime and the cause of the downtime predicted by the Al model comprises a probability of the cause.

3. The method of any one of the preceding claims, wherein the method further comprises determining, by the Al model, a data drift report comprising an analysis of the input data compared to historical data.

4. The method of any one of the preceding claims, wherein the method further comprises determining, by the Al model, a model drift report comprising an analysis of the prediction made by the Al model.

5. The method of any one of the preceding claims, wherein a part of the input data comprises information on at least one of: a target production value ofthe activity, a production value of the activity and a time-series operating data of the activity.

6. The method of any one of the preceding claims, wherein the activity comprises a well production and the production description of the activity includes a description of the well production in a log file.

7. An Al model for use in predicting downtime of production for an activity and predicting a cause of the downtime of production, the model having a structure comprising a natural language processing (120b) to process a production description of the activity to predict a downtime and a cause of the downtime.

8. The Al model of the preceding claim, wherein the natural language processing is implemented by a neural network, preferably a transformer network.

9. The Al model of any one of claims 7 to 8, wherein the Al model further comprises a data drift module (120c) to generate a data drift report comprising an analysis of input data compared to historical data.

10. The Al model of the preceding claim, wherein the data drift module (120c) is implemented by an adversarial machine learning model.

11. The Al model of any one of claims 9 to 10, wherein the Al model further comprises a learning module (i2od) to train the natural language processing based on a predicted downtime and a predicted cause of the downtime and on a data drift report generated.

12. The Al model of the preceding claim, wherein the learning module further comprises a continual learning module (i2odi) and an Al model monitoring module (i2od2) to analyze the prediction of the Al model and to have a feedback loop with the natural language processing.13- A computer-implemented method for training an artificial intelligence, Al, model for use in predicting downtime of production for an activity and predicting a cause of the downtime, the method comprising a step of: receiving training data to train the artificial intelligence model, wherein the training data at least comprises: an production description of an activity, a predicted downtime and a predicted cause of the downtime.

14. The method of the preceding claim, wherein the training data further comprises a data drift analysis.

15. The method of any one of claims 13 to 14, wherein the training data further comprises a model monitoring analysis.

16. The computer-implemented method according to any one of claims 1 to 6, wherein the Al model has been trained according to the method of any one of claims 13 to 15.

17. The computer-implemented method according to any one of claims 1 to 6 and 16, wherein the method comprises determining that new training data is available and training the Al model according to the method of any one of claims 13 to 15 using the new training data.

18. Data processing device comprising means configured to perform the method of any one of the claims 1 to 6 and / or 12 to 17.

19. Computer program comprising instructions, which when executed by a computer, causing the computer to perform the method of any one of the claims 1 to 6 and / or 12 to 17.

Images