Automatically locating and tracking a transient object in a hydrocarbon well conduit
A machine-learning model trained on pressure data from conduit monitoring systems accurately predicts the position and movement of transient objects within hydrocarbon well conduits, addressing noise interference and enhancing maintenance efficiency.
Patent Information
- Application Number
- SA1120260838
- Authority / Receiving Office
- SA · SA
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2026-01-28
- Publication Date
- 2026-10-05
AI Technical Summary
Existing systems struggle to accurately locate and track transient objects, such as scrapers or obstructions, within hydrocarbon well conduits using pressure data, due to noise interference and the complexity of fluid dynamics, which affects the precision of predictive models.
A machine-learning model is trained using pressure data from conduit monitoring systems to predict the position and movement of transient objects by analyzing key attributes like pressure pulse introduction and reflection, employing techniques like FFT and supervised/unsupervised learning, and filtering noise to enhance model accuracy.
The system provides precise location, speed, and direction of transient objects, enabling timely maintenance actions and improving operational efficiency in hydrocarbon well processes.
Smart Images

Figure 36A24C6A-FF47-45A7-9C76-9263AABD8324 
Figure 625E786C-7482-4227-AA58-647990FFD3CD 
Figure 687D66C1-4332-44EC-8167-F9DBB1BA0B50
Abstract
Description
Certain aspects and examples described in this document relate to a transient object locator and tracker (TOT) system in a hydrocarbon well operation conduit (also referred to in this document as the “conduit”) by training a machine-learning model and applying that trained machine-learning model to pressure measurements obtained from a conduit monitoring system of interest. Examples of conduit monitoring systems can be configured to record pressure within a conduit and can be installed, for example, in hydrocarbon well operation conduits such as wellhead casings, flow lines, and pipelines. Examples of conduit monitoring systems can use a combination of sensors and data acquisition devices to produce data on pressure conditions within a conduit. This pressure data may be used as training data and may be sent to a computing device, such as a TOT system (i.e., a transient object locator and tracker), to train a machine-learning model to produce a predictive model.The predictive model can then be applied to pressure measurements of a stream of interest to predict the position and movement of a passing object in the stream of interest. A flow monitoring system used to provide data to a system, such as a transit object locator and tracker (TOT), works by introducing a pressure wave into the flowing fluid and using one or more sensors to measure the amplitude of pressure wave reflections caused by transit object deflections, such as scrapers or obstructions within the flow. The pressure data generated by the sensor(s) can be recorded, stored, or otherwise aggregated and sent to the TOT for machine learning model training or predictive purposes when the pressure data is associated with a flow of interest. Pressure data collected by a conduit monitoring system can be used to train a machine learning model. For example, pressure data collected by the conduit monitoring system can be automatically sent to a computer for use in training a machine learning model. The computer can be part of a transient object locating and tracking system. The computer can access the pressure data in the conduit pressure dataset produced by the conduit monitoring system to generate a training dataset and use this training dataset to train the machine learning model and produce a trained (predictive) model. The trained model can then be applied to new pressure data associated with a conduit of interest to predict the position and movement of a transient object in that conduit. Ideally, a machine learning model should be trained on at least one key attribute, which may be a set of key attributes. For example, the machine learning model could be trained to recognize the point in the collected pressure data where a pressure pulse is introduced into a monitored stream (e.g., when the valve opens). This typically results in the highest point of acoustic energy generated in the stream during a test. Similarly, the machine learning model could be trained to recognize the point in the collected pressure data where the mechanism used to introduce the pressure pulse into the monitored stream is stopped (e.g., when the valve closes), or the point where the pressure is restored within the stream after the pressure pulse is introduced. The pressure data regression or the first derivative regression of the pressure data at one or more specific points during a test could also be analyzed in this regard.For the purposes of producing a first (and possibly second) derivative of pressure and training a machine learning model to recognize these key features, pressure data can be transformed using a technique such as the fast Fourier transformation (FFT). The machine learning model can also be trained on other key attributes related to the stream under observation, such as, for example, stream length. Training the machine learning model using stream length as a key attribute, for instance, may also allow the model to be trained on pressure data collected only during a period of interest, while pressure data collected before or after that period can be excluded. This can also help the machine learning model understand the shape of the pressure data of interest and fit the pressure data to the goal of minimizing the standard deviation or the mean average between calculations of the transient object's position. It is possible to train various examples of a machine learning model using supervised, semi-supervised, or unsupervised machine learning methods. In at least one example, the trained (predictive) model could be, for instance, a classification model. In another example, the predictive model could instead be a regression model. Artificial neural networks and deep learning can also be used to generate a predictive model. An example of a transient object locator and tracker (TAT) system could be reporting the movement (e.g., direction, speed) of a transient object in a stream, such as for personnel responsible for operating or maintaining the stream of an affected hydrocarbon well process, or a scraper placed in the stream. Another example of a TAT system, in response to transient object locator locator locator locator locator locator 1 2 2 3 3 4 ... In addition to using examples of Transient Object Locators and Trackers (TANDs) to locate a transient object in a specific stream, TANDs can utilize sensor data obtained from a stream monitoring system in various other ways. For example, TANDs can use sensor data collected from a flowline of an existing hydrocarbon well, pipeline, or drilling rig to apply predictive modeling for future hydrocarbon well operations, such as, but not limited to, drilling, flowline, or pipeline design, maintenance scheduling, and so on. The following are illustrative examples, provided to familiarize the reader with the general topic discussed in this document and not intended to limit the scope of the concepts revealed. The following sections describe several additional features and examples, referencing diagrams where similar numbers indicate similar elements. Directional descriptive terms have been used to describe illustrative aspects, but, like the illustrative examples, they should not be used to limit the current disclosure. The Stream 100 monitoring system, which monitors and reports conditions within a stream of interest, is depicted in Figure 1, along with the Transient Object Locator and Tracker 102 system, which is connected to the Stream 100 monitoring system. The Stream 100 monitoring system works by introducing a pressure wave into the fluid within the stream, and when the pressure wave propagates, it uses one or more sensors to measure the amplitude of the reflected pressure waves resulting from sedimentation or leaks in the stream. Stream Monitoring System 100 records, stores, or otherwise aggregates pressure data generated by the sensor(s) in response to receiving reflections of pressure waves traveling through the stream. The pressure data collected by Stream Monitoring System 100 can be sent to a computer for analysis or for use in training a machine learning model. As shown, Computer 104 is part of Transient Object Locating and Tracking System 102 in this example. Pressure data can be sent to Computer 104 of Transient Object Locating and Tracking System 102 as defined by Stream Monitoring System 100, upon request from Computer 104 of Transient Object Locating and Tracking System 102, or otherwise. Computer 104 of Transient Object Locating and Tracking System 102 can be located locally within the Stream Monitoring System 100 components and can be coupled to them via a local interface.Alternatively, Computer 104 and Transient Object Locator and Tracker System 102 can be located remotely from the components of Stream Monitoring System 100, and Computer 104 can receive pressure data from Stream Monitoring System 100 via a network, such as, but not limited to, the internet. Communication between Stream Monitoring System 100 and Transient Object Locator and Tracker System 102 may be wired or wireless. Wireless communication between Stream Monitoring System 100 and Computer 104 of Transient Object Locator and Tracker System 102 is shown in Figure 1 for illustrative purposes. Figure 2 illustrates an example of a computing environment involving a Transient Object Locator and Tracker (TANDER) system 200 with a computer 202 configured to train a Machine Learning Model 204 and to execute a trained (predictive) model 206 for TANDER locating and tracking a hydrocarbon well stream. In other examples, the computer used to train the machine learning model may be separate from the TANDER system 200. The machine learning model may be trained using pressure data samples from a first stream pressure set 208, a second stream pressure set 210, and a third stream pressure set 212, as described in more detail below. In some examples, it is possible to train the machine learning model 204 on only one or the other of the first stream pressure set 208 and the second stream pressure set 210, or the first stream pressure set 208 and the third stream pressure set 212.The first stream pressure dataset 208, the second stream pressure dataset 210, and the third stream pressure dataset 212 can be stored in the memory (for example, on a local drive) of the computer 202, or on a network drive, cloud storage system, or any other applicable storage medium. The computer 202 Transient Object Locator and Tracker System 200 can use existing pressure data to train the machine learning model 204, specifically pressure data obtained through a previous flow monitoring system operation, such as, but not limited to, the flow monitoring system 100 shown in Figure 1 and described similarly above. In one example, the first flow pressure data set 208 could be recorded or otherwise obtained pressure data from a flow monitoring system sensor when monitoring a perfect hydrocarbon well flow—that is, a flow without transient objects or other diversions such as sedimentation or seepage. In another example, the second flow pressure data set 210 could be recorded or otherwise obtained pressure data from a flow monitoring system when monitoring a hydrocarbon well flow with a transient object in the form of a blockage caused by some type of material (for example, hydrates).In one example, the third stream pressure data set 212 could be recorded pressure data or otherwise obtained by a stream monitoring system when monitoring the stream of a hydrocarbon well containing a passing object in the form of a scraper. In one example, a machine learning model can be trained using only a portion of the pressure data available for the first stream pressure set 208, the second stream pressure set 210, or the third stream pressure set 212, while the remaining pressure data for the same stream pressure sets 208, 210, or 212 can be withheld for later use in validating the predictive model. In other words, once predictive model 206 is produced by training machine learning model 204, predictive model 206 can be validated against a portion of the first stream pressure set 208, the second stream pressure set 210, or the third stream pressure set 212 that is held in reserve. It is also possible to validate predictive model 206 against results produced for the same pressure data by a physics-based model known for its accuracy. Machine Learning Model 204 can be trained, as illustrated in one example, using a guided learning method. In this example, the pressure data samples in the first stream pressure dataset 208, the second stream pressure dataset 210, and the third stream pressure dataset 212 are labeled. As such, during training, Machine Learning Model 204 understands that the pressure data for the first stream pressure dataset 208 represents pressure measurements taken by a stream monitoring system relative to an optimal hydrocarbon well stream. Similarly, during training, Machine Learning Model 204 understands that the pressure data for the second stream pressure dataset 210 represents pressure signal measurements taken by a stream monitoring system relative to a hydrocarbon well stream with a blockage other than the scraper, and that the pressure data for the third stream pressure dataset 212 represents pressure signal measurements taken by a stream monitoring system relative to a hydrocarbon well stream with a scraper.It is also possible to train the 204 machine learning model using a single conduit pressure dataset that includes appropriately labeled pressure data representing the pressure inside each of an ideal conduit, a conduit with an obstruction other than a scraper, and a conduit with a scraper. In one example, a guided teaching method could be a trained classification method implemented with a classification algorithm, such that the predictive model 206 is the classification model. Various types of classification algorithms can be used for this purpose, such as, but not limited to, logistic regression and random forest classification algorithms. Other suitable classification algorithms known to experts in the field can also be used. The output of the predictive model 206 in this example could be a prediction (predictive output) 216 of the position and movement (if any) of a transient object in a stream of interest based on new pressure (sensor) data 214 obtained from a related stream monitoring system, to which the predictive model 206 is subsequently applied. The 204 machine learning model can also be trained, according to one example, using a semi-guided learning method. In the example of the semi-guided learning method, only some of the pressure data in the first stream pressure dataset 208, the second stream pressure dataset 210, and the third stream pressure dataset 212 are labeled. The 204 machine learning model can then be trained only on the partially labeled pressure data in the second stream pressure dataset 210 or the third stream pressure dataset 212, or on the partially labeled pressure data in both the first stream pressure dataset 208 and the second stream pressure dataset 210 or the third stream pressure dataset 212. As such, the 204 machine learning model is unaware of whether some of the training data it is trained on includes pressure data representing an ideal stream or a stream with sedimentation or seepage. To overcome this problem, the learning that occurs during the training of the 204 machine learning model on the labeled pressure data can be used to predict a sign for the unlabeled pressure data.The pressure data containing the predicted markers can then be used to retrain machine learning model 204 or to train a new machine learning model. The resulting predictive model 206, produced by training machine learning model 204 using a semi-directed learning method, can again be a classification model. Machine learning model 204 can also be trained, as illustrated in one example, using an unguided teaching method. In this example, model 204 can be trained using pressure data obtained from a stream monitoring system and known to be associated with an ideal stream to produce a predictive model 206 that reflects normal stream conditions. Once the normal predictive model 206 is produced, a predictive algorithm can use it to identify and track a transient object within a stream of interest based on the deviations in the new pressure (sensor) data 214 to which the predictive model 206 is subsequently applied, based on the degree of deviation from the normal model. The predictive algorithm can use, among other things, clustering, deviation detection, or other techniques to achieve this. For example, a trained or untrained artificial neural network (ANN) can be used to locate and track the movement of a passing object. A deep learning network can also be trained and used. Cluster learning techniques, such as primitives or clustering, can be used regardless of the modeling method employed. Cluster learning techniques can also be used to reduce model variance. Feature engineering techniques can also be used during machine learning model training, for example, to extract features (e.g., key features) from the initial training data and use these features to improve the machine learning model's training process. Similarly, feature selection techniques can be used to remove unnecessary elements from the initial training data for the machine learning model. Figure 3 shows a schematic diagram of an example of the Transient Object Locator and Tracker 200 (TOT 202) computer, which can be used to train a machine learning model 204. TOT 202 can also be used to implement a predictive model 206 for TOT locating and tracking a transient object in a hydrocarbon well stream. While Figure 3 depicts TOT 202 with certain components, other examples may include more, fewer, or different components than those shown in Figure 3. In one example, TOT 202 can be implemented as TOT 200, as described above with reference to Figure 2. As shown, the computer system 202 includes a processor 218 connected to memory 220 via bus 222. A processor 218 can consist of a single processor or multiple processors. Non-exclusive examples of a processor 218 include a field-programmable gate array (FPGA), an application-specific integrated circuit (ASIC), a microprocessor, or a combination thereof. Instructions 224 can be stored in memory 220. These instructions are executable by the processor to cause it to perform various operations. In some examples, instructions 224 may include processor-specific instructions generated by a compiler or interpreter from code written in any suitable computer programming language, such as C++, C#, or Java. Memory 220 can include a single memory device or multiple memory devices. Memory 220 can be persistent and may include any type of memory device that retains stored information when powered off. Examples of memory 220 include electrically erasable and programmable read-only memory (EEPROM), flash memory, or any other type of persistent memory. At least some memory devices include a non-transitory computer-readable medium from which the 218 processor can read the 224 instruction. A non-transitory computer-readable medium may include electronic, optical, magnetic, or other storage devices capable of supplying the 218 processor with the 224 instruction or other program code.Non-exclusive examples of a non-temporary medium that can be read by a computer include a magnetic disk(s), a memory chip(s), ROM, random-access memory (RAM), ASIC, a configured processor, an optical storage device, or any other medium from which a computer processor can read Instruction 224. Computer 202 can include machine learning model 204, which can be trained using pressure data samples from the first stream pressure dataset 208, the second stream pressure dataset 210, or the third stream pressure dataset 212. Machine learning model 204 can be trained using the pressure data samples from the first stream pressure dataset 208, the second stream pressure dataset 210, or the third stream pressure dataset 212 by computer 202 to produce predictive model 206. Computer 202 can then execute predictive model 206 on new pressure data 214 received from or generated by a stream monitoring system to produce predictive output 216 regarding the position and motion of a transient object in a stream of interest, related to the pressure data on which predictive model 206 was applied.When training machine learning model 204 using a semi-directed learning method, the outputs of the predictive model 206 may include predicted markers for the unlabeled data found in the first drain pressure dataset 208, the second drain pressure dataset 210, or the third drain pressure dataset 212, which can serve as additional training data 226 that can be used to further train machine learning model 204 or to train a new machine learning model. Output 216 can provide useful information to the user of the Transient Object Locator and Tracker System 200. For example, the output can produce a notification that at least identifies the location of the transient object in the stream of interest and may also include motion information (e.g., speed, direction) associated with the transient object. The output may also predict the time required for the transient object to reach a specific location in the stream of interest, such as the time it takes for the scraper to reach a pre-existing removal point. (Predictive) output 216, which includes a notification as part of the output, may be textual and / or graphical in nature. In one example, output 216, which is a notification or otherwise, could be displayed on a 234 display connected to processor 218 via bus 222. Pressure data received by a Transient Identification and Tracking (TIFT) system is typically affected by noise, such as pump noise or other external noise, and may occur across a wide frequency range. Consequently, examples of a TIFT computer (202) might include 200 filtering functions. For example, as shown in Figure 3, a 202 computer could perform a first filtering operation (228) that can be applied to pressure data from the first stream pressure dataset (208), the second stream pressure dataset (210), or the third stream pressure dataset (212) to produce a training dataset (230) containing the filtered pressure data samples. Alternatively, a first filtering operation (228) could be applied to a single stream pressure dataset containing the combined pressure data samples from the first stream pressure dataset (208), the second stream pressure dataset (210), or the third stream pressure dataset (212) to produce a training dataset (230) containing the filtered pressure data samples. The training dataset 230 can then be used for the candidate pressure data samples to train the machine learning model 204.The computer can also perform a second filtering operation (202) that can be applied to new pressure data for a stream monitoring system associated with a stream of interest before the predictive model (206) is applied to it. In one example, the first filtering process 228 can be used to remove noise from collected pressure data samples before using those samples to generate the training dataset 230 for training machine learning model 204. The first filtering process 228 can also focus the training dataset 230 on the frequency range most likely to be of interest to the Transient Object Locator and Tracker (TOT) system 200. In another example, the second filtering process 232 can be applied for similar purposes to new pressure data associated with a stream of interest on which the predictive model 206 is applied. Either the first filtering process 228 and the second filtering process 232, or both, can be two-step processes. For example, a low-pass filter, such as, but not limited to, the Butterworth filter, can be initially applied to collected pressure data samples to remove or attenuate portions of them associated with frequencies higher than the frequency range of interest to TOT 200.A second filter, such as a Gaussian filter or a notch filter, can then be applied to the one-time filtered pressure data samples to isolate pressure data points that fall within a frequency range of interest (e.g., 6 Hz to 7 Hz). These filtered pressure data samples can then be used to produce the training dataset 230, which can be used to train the machine learning model 204. The filtering process(s) can make the machine learning model training process more efficient and may lead to a more accurate predictive model. During testing with the 100 Gutter Monitoring System, the total time period during which the pressure wave propagates through the duct and continues to cause measured pressure wave reflections may significantly exceed a time period of interest, which may be the time period associated with the largest acoustic energy points generated in the duct during the test. This typically occurs, for example, during a time period beginning at the first point where a dimensional relief valve opens to produce a pressure wave in the duct monitored by the 100 Gutter Monitoring System, and a subsequent second point where the valve closes. Because the machine learning model can be trained on pressure data collected over many (e.g., hundreds or thousands) of monitorings, it may be beneficial to focus on the pressure curve in the duct only during the time period of interest. Figure 4 graphically shows a set of 300 overlapping stream pressure curves, each limited to a specific time period of interest and calculated based on pressure data samples collected by a stream monitoring system during a unique stream test on a given stream. A training dataset can include a large number of these candidate pressure data samples. The corresponding time period of interest for each pressure curve, in this example, is the period beginning at the first pressure vent point 302, which results from the opening of a vent valve that also generates a pressure wave in a stream monitored by the stream monitoring system, and a subsequent second point 304, where the reflection of a pressure wave from a transient object in the stream is received by the stream monitoring system sensor. The pressure recovery portion 306 of the 300 pressure curve occurs between the pressure vent point 302 and the point 304, where the reflection of the pressure wave from the transient object is received by the stream monitoring system sensor. The pressure values observed in the stream during the time period of interest for each corresponding stream monitoring system test are represented by a set of 300 overlapping individual stream pressure curves.Figure 4 also shows the average pressure curve 308, which represents the average of the pressure values associated with the overlapping individual pressure curves 300, as is the case with the first derivative 310 of the average stream pressure curve 308. The pressure rates and pressure curves, the position of a passing object, the velocity and direction of its movement, and the distance it travels can all be calculated using predictive model 206 based on established techniques. For example, by knowing, measuring, or otherwise determining other characteristics of the stream of interest (e.g., stream length and diameter) and the fluid within it, the predictive model can use mathematical equations to calculate the pressure rates and pressure curves for that stream. Similarly, by knowing, measuring, or otherwise determining the velocity of the pressure wave through the fluid in the stream of interest, the position, velocity, and direction of a passing object can be predicted using the predictive model. For example, by knowing the velocity of the pressure wave through the fluid in the stream of interest, the position of a passing object can be predicted based on the time interval between the stream monitoring system introducing a pressure wave into the fluid and the arrival of a reflected pressure wave from the passing object at the system's sensor.Similarly, by measuring the time difference between successive reflected pressure waves (as determined from time-repeated data sets), the speed and direction of motion of a passing object can be determined. Other techniques can also be used to determine location and motion. Figure 5 shows a hybrid bar chart that graphically represents a predicted range of transient body locations within the stream associated with the pressure data samples shown in Figure 4. More specifically, each bar of the hybrid bar chart represents a predicted transient body location selected from one of the 300 corresponding pressure curves shown in Figure 4. As shown, the set of 314 bars together defines the predicted location range of the transient body, which appears to be approximately 1200–1250 m from the location of the stream monitoring system sensor in this example. The average location, based on a set of predicted locations, is also represented by bar 316. Figure 5 also indicates a deviation of approximately 50 meters between the predicted positions of a transient body based on the 300 pressure data samples shown in Figure 4. As described above, the machine learning model can be trained in a way that seeks to minimize the standard deviation between the calculated transient body positions. A notification issued by a Transient Object Location and Tracking (TOT) system regarding a transient object in a stream of interest can provide various types of information. For example, when the notification is related to the location and tracking of a scraper in a stream of interest, it might include information such as the scraper's position, speed, and remaining distance (i.e., the remaining planned distance). The information provided by the notification may be alphanumeric, graphical, both, and so on. Figure 6 shows different information that can be provided in one example of Notice 318 issued by a Transient Object Locator and Tracker (TOT) system regarding a transient object in a stream of interest. As illustrated in this particular example, Notice 318 is related to the location and tracking of a scraper in a stream of interest and includes information such as the scraper's position (320), speed (322), and remaining distance (i.e., the remaining planned distance the scraper still needs to travel) (324). The information provided by Notice 318 may be alphanumeric, graphical, or both. Other information, in addition to that provided in Notice 318 shown in Figure 6, may be provided in other examples of notices. Figure 7 shows a flowchart of the training method for a machine learning model to locate and track a transient object in a hydrocarbon well stream. According to the method example shown in Figure 7, a stream pressure dataset containing a large number of pressure data samples from the hydrocarbon well process is accessed at frame 400 by a processor, such as, but not limited to, the transient object location and tracking system processor described above. At frame 402, the processor filters the pressure data samples from the stream pressure dataset. Filtering can be performed to remove noise and to focus the training dataset. For example, filtering can be done by applying a low-pass filter to the pressure data samples of the stream pressure dataset and then by applying a second filter to the pressure data samples after applying the low-pass filter. At frame 404, the processor, by filtering the pressure data samples from the stream pressure dataset, produces a training dataset with many filtered pressure data samples.At frame 406, the processor can assign key features to the candidate pressure data samples of the training dataset. In at least some examples, these key features can be assigned based on calculated first derivatives of the candidate pressure data samples for the stream pressure dataset. At frame 408, the processor can then train a machine learning model using the training dataset and the key features to produce a predictive model. The machine learning model can be trained to minimize the standard deviation and the arithmetic mean between the distance-to-transient predictions calculated from at least some of the candidate pressure data samples in the training dataset. In at least some examples, the predictive model can then be used to predict the position of a transient object in a stream of interest, as well as the speed and direction of the transient object's movement, by applying the predictive model to pressure measurements associated with the stream of interest. In at least some examples, the processor can also generate a prediction notification.The notification may include, for example, the location, speed, and direction of movement of the object passing through the waterway of interest, or it may cause other action to be taken. For illustrative purposes, various examples have been presented above concerning hydrocarbon well streams, fluids, and processes. However, it should be noted that these examples can also be used to track transient bodies in other types of streams. For instance, a system can be used to track transient bodies in streams carrying water, hydrogen, carbon dioxide, or other fluids. In one specific example, a system and method can be used to track transient bodies in a stream related to a carbon capture, utilization, and storage (CCUS) process, where carbon dioxide is captured from one source and transported to another location for use or geologic sequestration in an underground formation. Based on the current disclosure aspects, a system, method, and computer-readable non-temporal media are provided, according to one or more of the following examples. As used below, it should be understood that any reference to a set of examples is a reference to each of those examples separately (for example, it should be understood that "Examples 1-4" means "Examples 1, 2, 3, or 4"). Example 1 is a system, comprising: a processor; and memory containing instructions that can be executed by the processor to cause the processor to do the following: access a stream pressure dataset containing a large number of measured pressure data samples for the stream of a hydrocarbon well process; filter the pressure data samples of the stream pressure dataset by applying a low-pass filter to the pressure data samples and by applying a second filter to the pressure data samples after applying the low-pass filter to the pressure data samples, to produce a training dataset with a large number of filtered pressure data samples; assign a set of key features to each of the filtered pressure data samples of the training dataset, at least one of the key features set being a point of the largest measured acoustic energy in the stream; train a machine learning model using the training dataset and the key feature set to reduce the standard deviation and arithmetic mean between the distance-to-transient predictions calculated from at least some of the filtered pressure data samples in the training dataset,To produce a predictive model. Example 2 is the system according to Example 1, where: Pressure data samples are produced by a stream monitoring system configured to introduce a pressure wave into the stream and to measure the amplitude of the reflected pressure waves using at least one sensor; the stream monitoring system is coupled to a system computer, which includes the processor, memory, and instructions; and the instructions are also executable by the processor to cause the computer to receive pressure data samples from the stream monitoring system. Example 3 is the system according to Example 1, where the stream pressure data set includes a first set of pressure data associated with a stream known to contain a transient object in the form of a scraper, a second set of pressure data associated with a stream known to contain an obstruction other than the scraper, and a third set of pressure data associated with an ideal stream. Example 4 is the system according to Example 1, where the stream pressure dataset includes only pressure data samples recorded during a period of time of interest. Example 5 is the system according to Example 1, where: a low-pass filter and a second filter are used together to remove noise from compression data samples and to focus the training dataset on a frequency range of interest from 6 Hz to 7 Hz; the low-pass filter is a Butterworth filter and the second filter is a Gaussian or notch filter. Example 6 is the system according to Example 1, where the instructions are also executable by the processor to cause the processor to do the following: After filtering pressure data samples for the downstream pressure dataset, calculate at least the first derivative of each of the pressure data samples for the downstream pressure dataset; and assign the set of key features in each of the pressure data samples for the downstream pressure dataset from the first derivative of each of the pressure data samples for the downstream pressure dataset. Example 7 is the system according to Example 1, where the instructions are also executable by the processor to cause the processor to do the following: predict the position and movement of a transiting object in a stream of interest by applying the predictive model to datasets that include measured pressure data associated with the stream of interest and iterating over time, analyzing pressure profiles defined using measured pressure data; and outputting an order to perform an action chosen from the set consisting of issuing a notification indicating the position of at least the transiting object, scheduling the removal of the transiting object when the transiting object is a scraper, initiating a repair action with respect to the transiting object when the transiting object is a blockage, and combinations thereof. Example 8 is a computer-implemented method that involves: accessing, with a processor, a stream pressure dataset that includes a large number of measured stream pressure data samples for a hydrocarbon well process; filtering, with the processor, the stream pressure data samples by applying a low-pass filter to the data samples and by applying a second filter to the data samples after applying the low-pass filter to the data samples, to produce a training dataset with a large number of filtered data samples; assigning a set of key features to each of the filtered data samples for the training dataset, with at least one feature of the key feature set representing the point of the largest measured acoustic energy in the stream; training, with the processor, a machine learning model using the training dataset and the key feature set to reduce the standard deviation and arithmetic mean between the distance-to-transient predictions calculated from at least some of the filtered data samples in the training dataset, to produce a predictive model. Example 9 is the computer-implemented method according to Example 8, where the stream pressure dataset includes a first set of pressure data associated with a stream known to contain a passing object in the form of a scraper, a second set of pressure data associated with a stream known to contain an obstruction other than the scraper, and a third set of pressure data associated with an ideal stream. Example 10 is the computer-implemented method according to Example 8, where the pressure data samples in the stream pressure dataset include pressure data recorded only during a period of time of interest. Example 11 is the computer-implemented method according to Example 8, where: a low-pass filter and a second filter are used together to remove noise from compression data samples and to focus the training dataset on a frequency range of interest from 6 Hz to 7 Hz; the low-pass filter is a Butterworth filter and the second filter is a Gaussian or notch filter. Example 12 is the computer-implemented method according to Example 8, which also includes: after filtering the pressure data samples for the drain pressure dataset, calculating, by processor, the first slash for each of the pressure data samples for the drain pressure dataset; and assigning, by processor, the set of key features in each of the pressure data samples for the drain pressure dataset from the first derivative of each of the pressure data samples for the drain pressure dataset. Example 13 is the computer-implemented method according to Example 8, which also includes: predicting, by processor, the position and movement of a transient object in a stream of interest by applying the predictive model to datasets that include measured pressure data related to the stream of interest and its iteration over time, and analyzing specific pressure curves using the measured pressure data; and in response to the prediction of the position and movement of a transient object in the stream of interest, a command is issued, by processor, to perform an action chosen from the set consisting of issuing a notification indicating the position of at least the transient object, scheduling the removal of the transient object when the transient object is a scraper, initiating a repair action with respect to the transient object when the transient object is a blockage, and combinations thereof. Example 14 is the computer-implemented method according to Example 13, where predicting the position and motion of the transiting object in the stream of interest involves providing information selected from a set consisting of the transiting object's speed, the transiting object's direction of motion, the transiting object's travel distance, and combinations thereof. Example 15 is a non-temporary, computer-readable medium containing processor-executable instructions to cause the processor to: access a downstream pressure dataset containing a large number of measured pressure data samples from a hydrocarbon well process; filter the pressure data samples of the downstream pressure dataset by applying a low-pass filter to the pressure data samples and by applying a second filter to the pressure data samples after applying the low-pass filter, to produce a training dataset with a large number of filtered pressure data samples; assign a set of key features to each of the filtered pressure data samples of the training dataset, with at least one of the key features representing the point of largest measured acoustic energy in the downstream; and train a machine learning model using the training dataset and the key feature set to minimize the standard deviation and arithmetic mean between distance-to-transient predictions calculated from at least some of the filtered pressure data samples in the training dataset, to producePredictive model. Example 16 is a non-temporary, computer-readable medium according to Example 15, where: Pressure data samples are produced by a stream monitoring system configured to introduce a pressure wave into the stream and to measure the amplitude of the reflected pressure waves using at least one sensor; the stream monitoring system is coupled to a system computer, which includes the processor, memory, and instructions; and the instructions are also executable by the processor to cause the computer to receive pressure data samples from the stream monitoring system. Example 17 is a non-temporary, computer-readable medium according to Example 15, where: The drain pressure dataset includes a first set of pressure data associated with a drain known to contain a transient object in the form of a scraper, a second set of pressure data associated with a drain known to contain an obstruction other than a scraper, and a third set of pressure data associated with an ideal drain; and the pressure data in the drain pressure dataset are pressure data recorded only during a period of time of interest. Example 18 is a non-temporary, computer-readable medium according to Example 15, where a low-pass filter and a second filter are used together to remove noise from compression data samples and to focus the training dataset on a frequency range of interest from 6 Hz to 7 Hz; the low-pass filter is a Butterworth filter and the second filter is a Gaussian or notch filter. Example 19 is a computer-readable non-temporary mediator according to Example 15, where the instructions are also processor-executable to cause the processor to do the following: After filtering pressure data samples for the downstream pressure dataset, calculate at least the first derivative of each of the pressure data samples for the downstream pressure dataset; and assign the set of key features in each of the pressure data samples for the downstream pressure dataset from the first derivative of each of the pressure data samples for the downstream pressure dataset. Example 20 is a non-temporary, computer-readable medium according to Example 15, where instructions are also executable by the processor to cause the processor to do the following: predict the position and motion of a transiting object in a stream of interest by applying the predictive model to datasets that include measured pressure data associated with the stream of interest and iterating over time, analyzing specific pressure curves using the measured pressure data; and outputting an order to perform an action chosen from the set consisting of issuing a notification indicating the position of at least the transiting object, scheduling the removal of the transiting object when the transiting object is a scraper, initiating a repair action with respect to the transiting object when the transiting object is a blockage, and combinations thereof. The preceding descriptions of specific examples, including those illustrated, are provided for illustrative and descriptive purposes only and are not intended to be exhaustive or to limit the detection to the exact images revealed. Those skilled in the field will find that many modifications, configurations, and applications are possible without straying from the scope of detection.
Claims
A system comprising a processor and memory containing executable instructions to cause the processor to: access a conduit pressure dataset containing a large number of measured pressure data samples from a hydrocarbon well operation; filter the pressure data samples of the conduit pressure dataset by applying a low-pass filter and then applying a second filter to produce a training dataset with a large number of filtered pressure data samples; and assign a set of key attributes to each of the pressure data samples selected for the training dataset, representing at least one attribute of the set.The key attributes include the point of largest measured acoustic energy in the conduit; a machine learning model is trained using the training dataset and the key attributes to minimize standard deviation and the mean average between distance predictions to the transient object calculated from at least some of the pressure data samples in the training dataset, to produce a predictive model. The system is according to protection element 1, where: pressure data samples are generated by a conduit monitoring system configured to produce a pressure wave in the conduit and to measure the amplitude of reflected pressure waves using at least one sensor; the conduit monitoring system is coupled to a computing device of the system, which includes the processor, memory, and instructions; and theInstructions can also be executed by the processor to cause the computing device to receive pressure data samples from the conduit monitoring system. The system, according to protection element 1, includes a conduit pressure dataset consisting of three sets: a first set of pressure data associated with a conduit known to contain a transient object (a pig), a second set of pressure data associated with a conduit known to contain an obstruction other than a pig, and a third set of pressure data associated with an ideal conduit. The system, according to protection element 1, includes pressure data samples recorded only during a specific time period. The system, according to protection element 1, uses a low-pass filter and a second filter together to remove noise from the pressure data samples and to focus the dataset.The training dataset operates within a frequency range of 6 Hz to 7 Hz. The low-pass filter is a Butterworth filter, and the second filter is a Gaussian filter or a notch filter. The system, according to protection element 1, where instructions are also executable by the processor, causes the processor to perform the following: after filtering the pressure data samples of the conduit pressure dataset, calculate at least the first derivative of each pressure data sample; and assign the key attributes to each pressure data sample of the conduit pressure dataset from the first derivative of each pressure data sample. The system, according to protection element 1, where instructions are also executable by the processorThe processor is used to: predict the position and movement of a transient object in a stream of interest by applying a predictive model to datasets that include measured pressure data associated with the stream of interest and its repetition over time; analyze pressure profiles defined using measured pressure data; and output a command to perform an action chosen from a set consisting of: issuing a notification indicating the position of the transient object; scheduling the removal of the transient object when the transient object is a pig scraper; initiating a repair action related to the transient object when the transient object is a blockage; and combinations thereof. A computer-implemented method involves: accessing, via a processor, a stream pressure dataset containing a large number of measured pressure data samples for a hydrocarbon well process.The process involves filtering pressure data samples from a conduit pressure dataset by applying a low-pass filter and then applying a second filter to the pressure data samples after the initial low-pass filter. This results in a training dataset with a large number of filtered pressure data samples. Key attributes are assigned to each of the filtered pressure data samples in the training dataset, with at least one of these key attributes representing the largest measured acoustic energy point in the conduit. A machine learning model is then trained using the training dataset and key attributes to minimize the standard deviation.The deviation and mean average between the distance predictions to the transient object, calculated from at least some of the pressure data samples in the training dataset, are used to produce a predictive model. The computer-implemented method, according to safeguard element 8, uses a conduit pressure dataset that includes a first set of pressure data associated with a conduit known to contain a transient object (a pig), a second set of pressure data associated with a conduit known to contain an obstruction other than a pig, and a third set of pressure data associated with an ideal conduit. The computer-implemented method, according to safeguard element 8, uses pressure data samples in the conduit pressure dataset that are recorded only during a specific time period.For protection element 8, a low-pass filter and a second filter are used together to remove noise from the pressure data samples and focus the training dataset on a frequency range of interest from 6 Hz to 7 Hz. The low-pass filter is a Butterworth filter, and the second filter is a Gaussian filter or a notch filter. The computer-implemented method, according to protection element 8, also includes the following: After filtering the pressure data samples of the conduit pressure dataset, the processor calculates the first slash for each pressure data sample and assigns the key attributes to each pressure data sample from the first derivative of the first slash.The computer-implemented method, as per security element 8, also includes: predicting, via the processor, the position and movement of a transient object in a conduit of interest by applying a predictive model to datasets that include measured pressure data associated with the conduit of interest and its iteration over time, and analyzing pressure profiles defined using measured pressure data; and in response to the prediction of the position and movement of a transient object in the conduit of interest, the processor issues a command to perform an action chosen from a set consisting of: issuing a notification indicating the position of the transient object; scheduling the removal of the transient object when the transient object is a pig scraper; initiating a repair action related to the transient object when the transient object is a blockage; and combinations thereof.The computer-implemented method, according to safeguard element 13, involves predicting the position and motion of a transient object in the stream of interest. This requires information selected from a set consisting of the transient object's velocity, direction of motion, distance traveled, and combinations thereof. A non-transitory, computer-readable medium contains processor-executable instructions that cause the processor to: access a conduit pressure dataset containing a large number of measured pressure data samples from a hydrocarbon well operation; and filter the pressure data samples by applying a low-pass filter and a second filter.After applying a low-pass filter to the pressure data samples, a training dataset is produced containing a large number of filtered pressure data samples. A set of key attributes is assigned to each of the filtered pressure data samples in the training dataset. At least one of the key attributes represents the point of largest measured acoustic energy in the channel. A machine learning model is then trained using the training dataset and the key attributes to minimize the standard deviation and mean average between the distance-to-transient object predictions calculated from at least some of the filtered pressure data samples in the training dataset, thus producing a predictive model. The non-transitory medium is computer-readable.According to Security Element 15, a computer-readable medium is defined as follows: Pressure data samples are produced by a conduit monitoring system configured to generate a pressure wave in the conduit and to measure the amplitude of reflected pressure waves using at least one sensor. The conduit monitoring system is coupled to a computing device of the system. The computing device includes the processor, memory, and instructions. The instructions are also executable by the processor to cause the computing device to receive pressure data samples from the conduit monitoring system. The non-transitory computer-readable medium is defined according to Security Element 15 as follows: A conduit pressure dataset includes a first set of pressure data associated with a conduit known to contain a transient object, such as a pig scraper, and a second set of pressure data...Pressure data are associated with a conduit known to have an obstruction other than a scraper pig, and a third set of pressure data is associated with an ideal conduit. The pressure data in the conduit pressure dataset consists of pressure data recorded only during a specific time period. The non-transitory computer-readable medium, as per safeguards element 15, uses a low-pass filter and a second filter to remove noise from the pressure data samples and to focus the training dataset on a frequency range of interest between 6 Hz and 7 Hz. The low-pass filter is a Butterworth filter, and the second filter is a Gaussian or notch filter. The non-transitory computer-readable medium, as per safeguards element 15, also allows for the execution of instructions.The processor is to perform the following actions: after filtering pressure data samples from the conduit pressure dataset, calculate at least the first derivative of each pressure data sample from the conduit pressure dataset; and assign the key attributes in each pressure data sample from the first derivative of each pressure data sample. The non-transitory computer-readable medium, according to security element 15, where instructions are also executable, is to perform the following actions: predict the position and motion of a transient object in a stream of interest by applying a predictive model to datasets that include measured pressure data associated with the stream of interest and its iteration over time, and analyzeAnalyzing pressure profiles is defined using measured pressure data; and an order is issued to perform an action selected from a set consisting of: issuing a notification indicating the location of the transient object, scheduling the removal of the transient object when the transient object is a pig scraper, initiating a repair action with respect to the transient object when the transient object is a blockage, and combinations thereof.
