Alerts manager

Abstract

Systems and methods for improved asset health monitoring systems which takes in time series data from various equipment sensors as input. Machine learning models learn the behavior and relationships among these data and send alerts to the user when the equipment deviates from normal operation, potentially indicating an anomaly or impending failure. These detected events enable plant operators to take timely actions to avoid costly repairs or operation shutdowns. Novel methodologies for sensor rank calculation and data interpolation are incorporated into that asset health monitoring system offer improved explanations of machine learning model outputs that facilitate a deeper understanding of the underlying mechanisms of the monitored process driving the measured events in multi-dimensional datasets.

Description

1086.2106001ALERTS MANAGER RELATED APPLICATION(S)

[0001] This application claims the benefit of U.S. Provisional Application No.63 / 749,270, filed on January 24, 2025. The entire teachings of the above application(s) are incorporated herein by reference.BACKGROUND

[0002] A plant, industrial facility, factory, and the like can be comprised of a complex system of equipment. Asset health monitoring software, for example Mtell® by Aspen Technology, Inc. (of Applicant Assignee), receives as input periodic data from various sensors monitoring a chemical or the like process and, based on the learnt behavior of the monitored process and relationship amongst the sensors, sends alerts / alarms to a user when the monitored process or equipment involved deviates from its normal operation. The alerts / alarms can be generated by software or modules such as machine learning based agents. Deviation from normal or expected operation can be a sign of an anomaly in plant operation, indicate impending failure, or provide notice of required maintenance. However, to make these signs actionable, an asset health monitoring system needs to provide additional information explaining the output of models that predict plant behavior, determine normal / abnormal operation, and generate alerts.

[0003] As the complexity of monitored processes increases, so too does the complexity and frequency of alerts generated by asset health monitoring software and systems. Because generated alerts may be dependent on multiple sensors and the data they provide, it may be difficult for a user to understand which sensors, and to what extent, contributed to the generation of the alert. This can create a delay in evaluating and understanding the alert and what it represents or is warning about in the monitored process. This extended evaluation process delays crucial decisions on actionable responses. A need exists for a way to calculate how each sensor and its data contributes to a generated alert and to present that information in a clear and easy to understand manner to enable fast interpretation and reaction.

[0004] As used herein “interpretability” in machine learning models refers to the degree to which the inner workings and decisions of the model can be understood and explained by humans. It is the ability to explain why a model makes certain predictions or provides certain outputs based on its input data. Providing interpretability for machine learning models is- 1 - 5000876. vl1086.2106001crucial to their efficient and proper use, particularly in asset health monitoring systems where plant operators need to understand why an alarm has been generated and what actions they should take to address it. This is especially important in complex systems such as when multi-dimensional sensor analysis is conducted, as intricate relationships between input variables may exist.

[0005] Existing methodologies offer limited or no interpretability. A system may monitor various sensors and detect events, yet users are often left to manually inspect sensor behavior to understand the rationale behind each event. This approach fails to provide comprehensive insights, especially when dealing with complex interrelations across multiple dimensions. Human observation alone is insufficient to discern such intricate patterns, given the limitations of the naked eye. Therefore, there is a need for methodologies that offer transparent explanations that facilitate a deeper understanding of the underlying mechanisms driving the observed or predicted events in multi-dimensional datasets.SUMMARY

[0006] Applicant addresses the foregoing needs and shortcomings in the art.Embodiments of the present disclosure include improvements to asset health monitoring software and systems which take in time series data from various equipment sensors as input. Embodiments can use or interface with Machine Learning (ML) models to learn the behavior and relationships among received data, sending alerts to the user when the equipment deviates or is predicted to deviate from normal operation, potentially indicating an anomaly or impending failure. Novel methodologies for improving the interpretability of generated alerts are used to enable quick and decisive user understanding which subsequently allows operators to take timely actions to avoid costly repairs or operation shutdowns.

[0007] Embodiments include a computer-implemented method of plant process monitoring, comprising receiving an alert regarding a process at a given plant, the alert being based on measurements generated by a set of sensors of the given plant, said receiving being performed by a digital processor and for each sensor in the set of sensors, automatically calculating, by the digital processor, a contribution score, the contribution score based on each sensor’s contribution to the alert. The method also includes calculating, by the digital processor, a sensor ranking based on the calculated contribution scores, the sensor ranking relatively ordering the sensors in the set of sensors based on each sensor’s contribution to the alert and generating, based on the calculated sensor ranking, an output indicative of a- 2 - 5000876. vl1086.2106001contribution of at least one sensor in the set of sensors to the alert in a manner enabling increased accuracy in interpreting the alert and in monitoring the process at the given plants.

[0008] The received alert may be generated by a machine learning model. The received alert can be indicative of failure, damage, likelihood of failure, or malfunction of a process unit of the process at the given plant.

[0009] The calculated contribution score may be a Shapley value. In such embodiments, the method may also include post-processing, by the digital processor, the calculated contribution scores to eliminate negative Shapley values.

[0010] The method can further include transforming, by the digital processor, the calculated contribution scores into normalized contribution scores, the normalized contribution scores sum to a probability value of the alert.

[0011] The measurements generated by the set of sensors of the given plant may be timeseries data during a monitored time period, and the calculated contribution scores can be based on each sensor’s contribution to the alert at an at least one sampled time within the monitored time period. In such embodiments, the method can also include determining the at least one sampled time based on the time-series data during the monitored time period. In such embodiments, the method can also include determining the contribution scores at an at least one interpolated time based on interpolated measurements for each sensor of the set of sensors, the interpolated measurements determined using the time-series data during the monitored time period. Akima spline interpolation or linear interpolation may be used to generate the interpolated measurements for each sensor of the set of sensors.

[0012] Embodiments further include a system for plant process monitoring comprising a digital processor communicatively coupled to an asset management tool. The digital processor configured to receive an alert regarding a process at a given plant, the alert being based on measurements generated by a set of sensors of the given plant and for each sensor in the set of sensors, automatically calculate a contribution score, the contribution score based on each sensor’s contribution to the alert. The digital processor also configured to calculate a sensor ranking based on the calculated contribution scores, the sensor ranking relatively ordering the sensors in the set of sensors based on each sensor’s contribution to the alert and generate, based on the calculated sensor ranking, an output indicative of a contribution of at least one sensor in the set of sensors to the alert in a manner enabling increased accuracy in interpreting the alert and in monitoring the process at the given plants.- 3 - 5000876. vl1086.2106001

[0013] The received alert can be generated by a machine learning model. The received alert can be indicative of failure, damage, likelihood of failure, or malfunction of a process unit of the process at the given plant.

[0014] The calculated contribution score may be a Shapley value and the digital processor may be further configured to post-process the calculated contribution scores to eliminate negative Shapley values.

[0015] The measurements generated by the set of sensors of the given plant can be timeseries data during a monitored time period, and the calculated contribution scores based on each sensor’s contribution to the alert at an at least one sampled time within the monitored time period. In such embodiments, the digital processor may be further configured to determine the at least one sampled time based on the time-series data during the monitored time period. In such embodiments, the digital processor may be further configured to determine determining the contribution scores at an at least one interpolated time based on interpolated measurements for each sensor of the set of sensors, the interpolated measurements determined using the time-series data during the monitored time period.

[0016] Embodiments also include a non-transitory computer-readable data storage medium comprising instructions to cause a computer to receive an alert regarding a process at a given plant, the alert being based on measurements generated by a set of sensors of the given plant and for each sensor in the set of sensors, automatically calculate a contribution score, the contribution score based on each sensor’s contribution to the alert. The instructions further cause a computer to calculate a sensor ranking based on the calculated contribution scores, the sensor ranking relatively ordering the sensors in the set of sensors based on each sensor’s contribution to the alert and generate, based on the calculated sensor ranking, an output indicative of a contribution of at least one sensor in the set of sensors to the alert in a manner enabling increased accuracy in interpreting the alert and in monitoring the process at the given plants.

[0017] The instructions may also cause a computer toto determine the contribution scores at an at least one interpolated time based on interpolated measurements for each sensor of the set of sensors, the interpolated measurements determined using the measurements generated by a set of sensors of the given plant.- 4 - 5000876. vl1086.2106001BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The patent or application file contains at least one drawing executed in color. Copies of this patent or patent application publication with color drawing(s) will be provided by the Office upon request and payment of the necessary fee.

[0019] The foregoing will be apparent from the following more particular description of example embodiments, as illustrated in the accompanying drawings in which like reference characters refer to the same parts throughout the different views. The drawings are not necessarily to scale, emphasis instead being placed upon illustrating embodiments.

[0020] FIG. l is a block diagram illustrating an example network environment for alert analysis and sensor ranking of example embodiments disclosed herein.

[0021] FIG. 2 displays a table of sensor ranks created according to existing methodologies.

[0022] FIG. 3 A displays a plot of point-wise sensor ranks of a system created according to an example embodiment.

[0023] FIG. 3B displays a plot of point-wise sensor ranks of a system created according to an example embodiment.

[0024] FIG. 4 is a bar plot of event-based sensor ranks for a select anomaly from FIG. 3 A created according to an example embodiment.

[0025] FIG. 5 is an example plot that displays the pointwise sensor ranks plot across the whole available dataset created according to an example embodiment.

[0026] FIGs. 6-8 are histogram plots for select sensors included in plot of FIG. 5 created according to an example embodiment.

[0027] FIG. 9 is a flow diagram illustrating the workflow of an example method of alert management and maintenance response according to an example embodiment.

[0028] FIG. 10 is a flow diagram illustrating an additional workflow of an example method of alert management and maintenance response according to an example embodiment.

[0029] FIG. 11 is a schematic view of a computer network in which embodiments can be implemented.

[0030] FIG. 12 is a block diagram of a computer node or device in the computer network ofFIG. 12.- 5 - 5000876. vl1086.2106001DETAILED DESCRIPTION

[0031] A description of exemplary embodiments follows.

[0032] Monitoring and controlling large plant processes, such as but not limited to chemical and refinery processing, requires the aggregation and analysis of large datasets created and provided by sensors monitoring the process. These datasets may be created by monitoring the process over a select time period and include live or historical data. Asset health monitoring tools and programs, for example Aspen Technology Inc.’s Mtell®, are used to analyze this large amount of data and provide indications, alerts, or warnings if normal operation is interrupted or may be interrupted in the foreseeable future. Such indications, alerts, and warnings allow for timely corrective action to be taken.

[0033] Asset health monitoring tools may include the function of creating agents, for example a machine learning module. Alternatively, agents may be created, trained, or generated separately using other tools, systems or software. A user can create or select an agent that targets and monitors for various failures in a certain sub assembly of a piece of equipment, such as a lubrication system of a rotating machine, or could target a specific failure state such as seal failure, for non-limiting example. Other sub-assemblies, assemblies, pieces of equipment, sub-systems, and systems are similarly suitable. Other target or failure modes, states, or phases of operation are similarly suitable. Multiple or combinations of such agents can be created or chosen based on various failure modes (states) or equipment subassemblies they are trying to monitor. Agents can be user-selectively enabled by providing users with a choice of including or excluding a sensor into an agent’s model based on the relationships they want to capture. The agent can use data from the included sensors to provide predictions of current or future sub assembly failure and to provide alerts based on those predictions. In other words, a user can select a set of sensors to include as inputs for a monitoring agent; and the monitoring agent is defined or otherwise configured to use a model, which may be a machine learning module or an equivalent, to determine or predict the monitored sub assembly or process’ behavior including failure, upcoming failure, or abnormal operation. If the model predicts or determines failure, the corresponding monitoring agent provides an alert to a user notifying them (in specific context and detail) and enabling corrective action.

[0034] During plant process monitoring, data from the equipment sensors can be used as an input to machine learning models, including created agents. When an anomaly or failure is- 6 - 5000876. vl1086.2106001predicted, an alert is sent to the plant operators. Alerts can be accompanied with or comprised of the sensor ranks. Sensor ranks are a ranked list, in some applications in the form of percentage values, that quantify the contribution of each sensor to the predicted event. Without sensor ranks, plant operators would have difficulty identifying what sensors, and therefore their monitored processes elements, may be the cause of or contributing to the alert.

[0035] As the complexity of the monitored process increases, the number of agents and therefore possible number of alerts can grow exponentially threatening to overwhelm a user, incurring the drawbacks discussed above, and possibly preventing or delaying corrective action. This necessitates improving interpretability of the machine learning model generating the alerts. Embodiments of the present disclosure use a game theory approach to calculate the contribution of each sensor to the prediction of an event. This in turn, informs a user of which monitored elements or subprocesses are the likely source of predicted or detected abnormal operation or failure. In game theory, Shapley values are used to fairly distribute the payoff of a cooperative game among its players. They represent the average marginal contribution of each player to all possible coalitions in the game. Embodiments of the present disclosure improve machine learning interpretability by applying calculated Shapley values to determine and explain the contribution of each feature (e.g., sensors, their data, and the process / element monitored by each sensor) to the prediction made by a model, including but not limited to a machine learning agent. Embodiments further help in explaining the model's output by attributing the prediction outcome to the individual features. Shapley values quantify the impact of each feature on the model's prediction across all possible combinations of features. Embodiments then use the calculated Shapley values to create sensor ranks, which provide a user with the knowledge of both absolute and relative sensor contribution to the alert.

[0036] By calculating Shapley values, embodiments provide insights to users on which features are most influential in the model's decision-making process, thus enhancing the interpretability of the model's predictions. This is particularly useful in complex models where understanding the role of each feature is needed for trust, further analysis, and corrective or preventative action. Embodiments of the present disclosure further provide a methodology where Shapley values are post-processed to be transformed into percentages and reported in an intuitive way as sensor ranks, which can assist users (e.g., field engineers) in making decisions and taking action. Alternative embodiments may use additional or other- 7 - 5000876. vl1086.2106001means to calculate or determine sensor contribution and sensor ranks in a manner to improve user’s interpretation and understanding of alerts.

[0037] The novel Shapley-values-based approach used by embodiments of the present disclosure improve assets health monitoring systems by providing local explainability and opening up a “black-box” machine learning model to bring localized and accurate explanations for detected events. Local explainability is highly beneficial as it can explain system behavior during select time periods, for example periods when an alert is generated or periods of high volatility. Embodiments of the present disclosure bring accurate and easy-to-understand explainability of models used by asset health monitoring systems, allowing users to take targeted actions to better address the equipment failure or anomaly. In some embodiments this is done by using a state-of-the-art SHAP algorithm to measure the sensor contributions, post-processing to bring generated numbers to a scale that is understandable to plant operators and generating easy-to-follow plots or other visualization elements.

[0038] There are multiple advantages embodiments of the present disclosure provide over existing solutions.

[0039] First, existing solutions lack explainability in general. While the system monitors several sensors and detects events, users must manually check sensor data to understand why an event occurred. However, existing solutions don’t assist users in recognizing complex patterns across multiple dimensions as these patterns are not easily visible to the human eye.

[0040] Second, existing solutions can only provide global explainability, while the proposed solution can provide local and global explainability. There are two approaches for model interpretation - local and global. Global interpretation pertains to how a model makes decisions according to its overall structure, explains the complete behavior of the model, and understanding the suitability of the model for deployment. One example of global interpretation is predicting the risk of disease in patients. Local interpretation pertains to how the model makes decisions for a single instance, explains individual predictions; and understanding the behavior of the model in the local neighborhood or during a specific time. One example of local interpretation is understanding why a specific person has a high risk of a disease.

[0041] Examples of global explainability are coefficients of the linear model or feature importance in tree-based methods. However, asset health monitoring systems customers need actionable alerts - they need to know what caused a particular anomaly or failure to be able- 8 - 5000876. vl1086.2106001to act on it. Local explainability provides better actionable insights into the cause of anomaly than global.

[0042] Third, using a hybrid sampling strategy, embodiments of the present disclosure can dramatically reduce computational requirements while ensuring that sensor rank information is available at every time point, thus supporting high-resolution interpretability, diagnostics, and root-cause analysis. This approach leverages the physical property that, in most engineered systems, sensor readings change smoothly over time, allowing for intelligent sampling and reconstruction.

[0043] By combining uniform, low-resolution sampling with adaptive, denser sampling in event intervals and advanced, time-aware interpolation, embodiments of the present disclosure achieve a unique balance between computational efficiency and interpretability. They can provide high-resolution, continuous sensor rank information across the entire time series, making it especially valuable for monitoring, diagnostics, and root-cause analysis in equipment and industrial applications. The method’s design leverages the physical properties of engineered systems, ensuring that the interpolated results are both reliable and meaningful for engineering and operational decision-making.

[0044] Finally, embodiments the present disclosure provide a more intuitive visualization of received alerts, sensor data, and sensor contribution to those alerts. It would be understood by a person skilled in the art that the visualizations disclosed herein are illustrative examples and alternative displays, graphs, user interfaces, may also be utilized.

[0045] Example Network Environment for Plant Processes

[0046] FIG. 1 is a block diagram of an example network environment 100 providing alert aggregation and analysis in embodiments. System computers 101, 102 may operate as controllers, host asset health monitoring tools, and execute the disclosed alert manger and its functions. In some embodiments, each one of the system computers 101, 102 may operate alone, or the system computers 101, 102 may operate together as distributed processors contributing to real-time operations, monitoring, and alert generation. In some embodiments, additional system computers 112 may also operate as distributed processors contributing to the real-time operation as a controller or provide or assist with the other functions of system computers 101, 102. System computers 101, 102 may include a user interface to provide (display or otherwise render) to a user alerts, generated by asset health monitoring tools, and outputs of the disclosed alert manager including those improving the interpretability of the alerts. System computers 101, 102 may also enable a user to update or alter plant processes- 9 - 5000876. vl1086.2106001using their functions as controllers based on the provided alerts generated by asset health monitoring tools and outputs of the disclosed alert manager. In some embodiments, asset health monitoring tools and the disclosed alert manager may be hosted on and run by computers external to example network environment 100, which can be dedicated to process control applications.

[0047] The system computers 101 and 102 may communicate with the data server 103 to access collected data for measurable process variables from a historian database 111.Historian database 111 (the data therein) may be accessed and utilized by the asset monitoring tools and alert manager. The data server 103 may be further communicatively coupled to a distributed control system (DCS) 104, or any other plant control system, which may be configured with instruments 109A-109I, 106, 107 that collect data at a regular sampling period (e.g., one sample per minute) for the measurable process variables.Instruments 109A-109I, 106, 107 can be or include sensors that are utilized by monitoring agents or used to train their component machine learning modules. Instruments 106, 107 are online analyzers (e.g., gas chromatographs) that collect data at a longer sampling period. The instruments 109 A - 1091, 106, 107 may communicate the collected data to an instrumentation computer 105, also configured in the DCS 104, and the instrumentation computer 105 may in turn communicate the collected data to the data server 103 over communications network 108. The data server 103 may then archive the collected data in the historian database 111 for model calibration and inferential model training purposes. The data collected varies according to the type of target plant process (such as crude processing in a refinery plant, or chemical processing in a pharmaceutical industrial or similar processing plant, for nonlimiting example).

[0048] The collected data and achieved data may include measurements for various measurable process variables used for process control, agent training and creation, alert management, and / or asset monitoring. These measurements may include, for example, a feed stream flow rate as measured by a flow meter 109B, a feed stream temperature as measured by a temperature sensor 109C, component feed concentrations as determined by an analyzer 109A, and reflux stream temperature in a pipe as measured by a temperature sensor 109D. The collected data may also include measurements for process output stream variables, such as, for example, the concentration of produced materials, as measured by analyzers 106 and 107. The collected data may further include measurements for manipulated input variables, such as, for example, reflux flow rate as set by valve 109F and determined by flow meter- 10 - 5000876. vl1086.2106001109H, a re-boiler steam flow rate as set by valve 109E and measured by flow meter 1091, and pressure in a column as controlled by a valve 109G. The collected data reflects the operation conditions of the representative plant during a particular sampling period. The collected data is archived in the historian database 111 for model calibration, inferential model training purposes, agent training and creation, alert management, and / or asset monitoring.

[0049] The system computers 101 or 102 may execute various types of process controllers for online or offline deployment purposes to control the monitored system or plant process. The output values generated by the controller(s) on the system computers 101 or 102 may be provided to the instrumentation computer 105 over the network 108 for an operator to view or may be provided to automatically program any other component of the DCS 104, or any other plant control system or processing system coupled to the DCS system 104. In some embodiments, the instrumentation computer 105 can store the historian database 111 in the data server 103 and execute the plant process controlled s) in a standalone mode. Collectively, the instrumentation computer 105, the data server 103, and various sensors and output drivers (e.g., 109A-109I, 106, 107) form the DCS 104 and can work together to implement and run embodiments of the present disclosure. Process controllers, executed by system computers 101 or 102, may control or be used to control the monitored system or plant (chemical, industrial, etc.) process based on outputs provided by the asset monitoring tool and the alert manager.

[0050] The example architecture 100 of the computer system supports the process operation of a representative plant (factory, refinery, and the like). In some embodiments, the representative plant may be, for non-limiting example, a refinery or a chemical processing plant having a number of measurable process variables, such as, for example, temperature, pressure, and flow rate variables. It should be understood that the present disclosure may use a wide variety of other types of technological processes and / or equipment in the useful arts in addition to or as alternatives to those described herein.

[0051] FIG. 2 displays a table 200 of sensor ranks in accordance with existing solutions. Table 200 may be an output of an alert management system or software. Table 200 is presented to a user and ranks sensors, identified by name in column 202, by order of their contribution to an alert, shown in column 201. Table 200 presents the sensors in the order of importance assigning the numerical values 201a-c, as well as marking the top 3 contributors with the shades of red in column 203. This approach has several downsides, first sensor contributions 201 are calculated as an aggregated average across all events in a time range- 11 - 5000876. vl1086.2106001and sensors are ranked in column 202 according to that universal contribution calculation. If a time range has multiple events or sensor contribution changes over time, table 200 is unable to provide the level of detail needed to provide key information to a user. Furthermore table 200 presents a user with out of context number 201a-e and unspecified colors 203 which is difficult to clearly comprehend.

[0052] In contrast, the present disclosure proposes and enables improved options for sensor contribution calculation and visualization.

[0053] FIG. 3 A displays a plot 300a of point- wise sensor ranks of a system created according to an example embodiment. Plot 300a displays calculated sensor contribution to the predicted probability of having an event for every sensor at every time interval of a select time period. Plot 300a allows a user to better pinpoint the contribution of sensors to the anomaly and see how it evolves over time. Also, a user can see how different anomalies may have different contributing sensors. Plot 300a transforms and displays the information used to create table 200 to improve interpretability and provide additional information. Plot 300a tracks sensor contribution 301a over time 301b for five sensors 302a-303e. Plot 300a shows how sensor contributions, on y-axis 301a, to the model's prediction evolve over time, on x-axis 301b. Periods of alerts are highlighted using indicators 303. From October 1 to November 8, when the system is healthy and the probability of an event is low, none of the sensors 302a-303e has a major contribution. However, when the system starts to operate abnormally and the probability of an event increases due to the correlation breaking between the 4_HP DE temperature 302e and 3 HP NDE 302d temperature sensors, a high contribution from these two sensors during the event regime is observed. Plot 300a allows a user to see when and where sensors contribute to an alert and how that changes over time, something not possible with the “snapshot” information provided in table 200.

[0054] FIG. 3B displays a plot 300b of point-wise sensor ranks of a system created according to an example embodiment. For some anomalies, such as alert 303, as shown in FIG. 3B, it is possible to have a change in top contributing sensor 302s: when one sensor played a significant role in the beginning, but later another one became more influential. Again, this information would be lost if sensor contribution information was presented as shown in table 200

[0055] Plot 300b discloses point-wise sensor ranks of a system based on sensor contribution 301a over time 301b. Plot 300b shows a prolonged detected anomaly 303 that starts in November 2018 and lasts till May 2019. At the very start of anomaly, in the first- 12 - 5000876. vl1086.2106001weeks of November, the main contributing sensor is 8 COJ RDL ACCNTO CMP 08. In midNovember, sensor 11 VIBRACIONES COMPR LOA joins as the important for anomaly detection. After the first week of February, the role of 8 COJ RDL ACCNTO CMP 08 and 11 VIBRACIONES COMPR LOA sensors in prediction anomaly diminishes, and the top contribution sensor becomes 10 VIBRACIONES COMPR LA. In the traditional approach, shown in table 200, of presenting only aggregated values over an entire period, it is extremely hard to comprehend such scenarios.

[0056] In addition to point-wise plots (e.g., plots 300a, 300b), the embodiments of the present disclosure may also include the use of bar or other types of plots to give users a quick glance at the sensor ranks for the anomaly or failure.

[0057] FIG. 4 is a bar plot 400 of event-based sensor ranks 401 for a select anomaly (#4) from FIG. 3 A. Plot 400 includes the sensor contribution 401 for the five sensors 402a-e and shows that sensors 4_HP DE temperature 402a and 3 HP NDE temperature 402b contribute the most to the predicted anomaly. Embodiments of the present disclosure can generate bar plot, such as plot 400, for alerts / anomalies 303 identified in or other timepoints included in the larger timescale tables 200 and 300. This enables users to view both general and local sensor contributions.

[0058] In addition to local explanations, such as plot 400, embodiments of the present disclosure may generate the to better understand the individual sensors' contributions across the whole dataset.

[0059] FIG. 5 is an example plot 500 that displays the pointwise sensor ranks 502 plot across the whole available dataset. Plot 500 include a point-wise sensor ranks of a system that shows how sensor contributions 501a to the model's prediction evolve over time 501b.Sensors 1_LO SUP TEMP AFTER OIL COOLER TEMP and 2 OIL FILTER DIFF.PRESSURE are the main contributors during anomalies that were predicted in Jan 2017- Jan 2018. Sensor 2 OIL FILTER DIFF. PRESSURE is mostly the main contributor during July 2018-Jan 2019. For the anomalies that happened in Feb 2019, Apr 2019, Mar-Jun 2020, and Mar- Apr 2021, the O OIL RESERVOIR LEVEL is the main contributor.

[0060] FIGs. 6-8 are histogram plots 600, 700, 800 for select sensors included in plot 500 of FIG. 5. A user may select sensors 502 shown in table 500 to include in histogram plots 600, 700, 900. FIG. 6 displays a histogram 600 of sensor O OIL RESERVOIR LEVEL: sensor weights during anomalies vs. normal operation. Red color 602 indicates sensor weights during anomalies. Green color 601 indicates sensor weights during normal operation.- 13 - 5000876. vl1086.2106001As shown in plot 600, Sensor O OIL RESERVOIR LEVEL usually downplays the decision on whether we have an anomaly during normal operation and strongly emphasizes it during actual anomalies.

[0061] FIG. 7 displays a histogram 700 of sensor 1_LO SUP TEMP AFTER OIL COOLER TEMP: sensor weights during anomalies vs. normal operation. Red color 702 indicates sensor weights during anomalies. Green color 701 indicates sensor weights during normal operation. As shown in plot 700, sensor 1_LO SUP TEMP AFTER OIL COOLER TEMP usually downplays the decision on whether there is an anomaly both during normal operation and, especially, during actual anomalies. In other words, even when other tags say to the model that something unusual is happening, a user would be able to quickly interpret data from tag 1_LO SUP TEMP AFTER OIL COOLER TEMP as indicating, “Be calm, there is no anomaly”.

[0062] FIG. 8 displays a histogram 800 of sensor 2 OIL FILTER DIFF. PRESSURE: sensor weights during anomalies vs. normal operation. Red color 802 indicates sensor weights during anomalies. Green color 801 indicates sensor weights during normal operation. As shown in plot 800, sensor 2 OIL FILTER DIFF. PRESSURE usually is neutral regarding the decision on whether there is an anomaly both during normal operation and during actual anomalies. In other words, a user would be able to quickly interpret data tag 2 OIL FILTER DIFF. PRESSURE contributions to the model as indicating “I don’t know if there is an anomaly or not, I’ll keep my silence.”

[0063] The histograms 600, 700, 800 of FIGS. 6-8 can provide users better intuition about the overall sensor behavior, as well as whether to keep the sensor in the model.

[0064] Previous methodologies are not generalizable for different machine learning models. Asset management systems can use different kinds of machine learning models for failure and anomaly prediction. If model-specific approaches are used for white-box models like feature importance for tree-based algorithms or model coefficients or linear models, it would require different explanation approaches shown to users, which, apart from lacking precision from being global model explanators, can also be inconsistent and may point to different sensors for different models created on the same equipment.

[0065] For example, in a linear regression model y = bO + b\ * x, when x increases by 1% then will increase by M% keeping other factors constant. However, that approach is not generalizable for other models such as black-box models, including neural networks often used by assert management systems.- 14 - 5000876. vl1086.2106001

[0066] A need exists for an approach to quantify sensor contributions that can work with any model, including machine learning modes such as neural networks. The SHAP -based algorithm used by embodiments of the present disclosure provides exactly that. Apart from the benefits of being consistent in explanations across different models, it can also be easier and less burdensome to maintain code for a unified approach than for each model separately.

[0067] Previous approaches at providing sensor contribution across different models and time frames are not accurate in many cases. The SHAP -based algorithm employed by embodiments of the present disclosure provides sensor rank values for both specific data instances and aggregated over individual events. This offers a better and more precise understanding of the anomaly's cause, sensor contribution, and the monitored process overall.

[0068] Moreover, the sensor ranks calculated by previous existing solutions have a problem for cases with highly correlated tags. For example, if they are based on calculating the Mahalanobis distance from the particular point to the center of the cluster. The inverted covariance matrix, used for calculation of the Mahalanobis distance, creates an ill-posed problem for highly correlated sensors. Because of this, the sensor rank determinations by previous existing solutions may produce sensor ranks that correspond to the user’s intuition only for the cases in which covariance is the dominating factor in anomaly prediction.Basically, the implementation of previous existing solutions will only give correct sensor ranks if the top contributors happen to be the ones with the highest correlation in the training range.

[0069] In contrast, the SHAP -based algorithm utilized by embodiments of the present disclosure produces sensor ranks that match users' expectations even for cases with highly correlated tags, as verified on numerous customer defects.

[0070] Users (e.g., manufacturing facilities operators) would like to know about potential anomalies or failures in advance so they can plan maintenance, or other corresponding actions to avoid more costly repairs or operation shutdowns. Asset health monitoring systems can use data from sensors to build and train predictive models that detect potential failure, or deviations from a normal regime of operation, called anomalies. After detection, certain actions may need to be taken. Operators would like to know which sensor led to model predictions and outputs and plan corresponding actions based on this information. This creates a need to quantify the impact of each sensor on the model’s prediction.

[0071] Embodiments of the present disclosure include the use of Shapley values for machine learning interpretability. Using sensor data, Shapley values can be calculated that- 15 - 5000876. vl1086.2106001determine sensor contribution to a model’s output such as alert or warning, both at a specific time and over a select time period. This sensor data can be in the form of time series data comprising sensor measurements. However, in some cases, the calculation of Shapley values can be slow due to computational complexity. This is because Shapley values involve calculating the marginal contributions of each feature across all possible subsets of features, which can become prohibitively time-consuming, especially for models with a large number of features.

[0072] Therefore, embodiments of the present disclosure include the use of the SHAP (SHapley Additive exPlanations) algorithm, that speeds up this process by approximating Shapley values efficiently and accurately. It does this by leveraging a game-theoretic framework to compute Shapley values in a more tractable manner. Instead of exhaustively computing Shapley values for all possible feature subsets, the SHAP algorithm uses an approximation technique based on sampling subsets of features. This sampling approach significantly reduces the computational burden while still providing accurate estimates of feature contributions. The SHAP algorithm offers an approach to deciphering the output of any predictive algorithm used for asset management systems and process control.Additionally, the SHAP algorithm employs a specific property of Shapley values known as linearity. This property allows the SHAP algorithm to decompose a model's prediction into a sum of individual feature contributions, making the calculation more efficient.

[0073] Overall, by combining sampling techniques and leveraging the linearity property of Shapley values, the SHAP algorithm accelerates the computation of Shapley values, making it feasible to apply them to larger and more complex models.

[0074] While predictive models provide answers to "how much," the SHAP algorithm illuminates the "why" behind those answers. In other words, the SHAP algorithm allows users to demystify a black-box model. As mentioned earlier, values of the SHAP algorithm draw upon the concept of Shapley values, which originates from game theory. However, game theory relies on two fundamental components: a game and players.

[0075] Embodiments use a novel application of game theory to explainability and interpretability in machine learning and predictive models. Consider the following scenario of a predictive model: In this context, the "game" involves reproducing the model's outcome. The "players" represent the features (e.g., sensors and their provided data) incorporated into the model. The role of Shapley's values is to quantify the contribution that each player (sensor) brings to the game (alert output or other prediction). Similarly, values of the SHAP- 16 - 5000876. vl1086.2106001algorithm quantify the contribution that each feature (sensors) makes to the model's predictions (alert or process behavior prediction.) The term "game" refers to a single observation at single point in time. Each observation or entry in a time-series dataset can be treated as a distinct game.

[0076] The SHAP algorithm focuses on and improves the local interpretability of a predictive model. By using the SHAP algorithm, embodiments of the present disclosure obtain accurate localized explanations of the model prediction. In other words, which and to what extent sensors contributed to a model’s prediction of the monitored plant behavior or event. The outputted values of the SHAP algorithm can be both positive and negative. High positive values indicate the sensor that drove the model decision toward the “it’s likely an event” decision, while negative values indicate driving to “it’s likely normal regime” prediction.

[0077] The combination of values from the SHAP algorithm for every time point equals the model prediction of the probability of an event. However, often plant operators are interested in knowing the sensors that lead to an anomaly or failure and outputting negative contributors may be confusing. Additionally, when explaining an event, intuitively the sum of the contributors should be equal to one. For pointwise sensor ranks plots or other outputs provided to users, embodiments enable the sum of sensor ranks to be equal to the prediction of the probability of the event without any negative values.

[0078] The following post-processing for pointwise sensor ranks may be implemented to eliminate negative Shapley values and to bring the sum of sensor ranks to the probability output:1

[0079] For cases where event Jhreshold * c <= expected_value\additive = expected value I m Eq.2shap values- = shap values- + additive; V i 6 (1, n),Vj G (1, m) Eq.3>q.4- 17 - 5000876. vl1086.21060015

[0080] For cased where eventjhreshold * c > expected_value><8

[0082] c is constant, 0<=c<=l, and indicates how small the event threshold should be compared to the expected value of the model before a user wants to add additive. In some examples, c=0.8 is recommended empirically.

[0083] The following post-processing for event-based sensor ranks may be implemented to aggregate the pointwise sensor ranks within the event and bring the sum of the values to one:

[0084] For every event:1

[0085] Embodiments of the present disclosure also utilize a method for efficiently and comprehensively computing sensor importance scores (“sensor contribution” or “sensor ranks”) from high-frequency, multivariate time-series data, as encountered in equipment monitoring, industrial analytics, and predictive maintenance. The method is designed to dramatically reduce computational requirements while ensuring that sensor rank information is available at every time point, thus supporting high-resolution interpretability, diagnostics, and root-cause analysis including but not limited to the creation of plots 300a, 300b, 400, 500, 600, 700, and 800. The approach leverages the physical property that, in most5000876. vl1086.2106001engineered systems, sensor readings change smoothly over time, allowing for intelligent sampling and reconstruction.

[0086] The method used by embodiments of the present disclosure begins by establishing a hybrid sampling strategy that combines two complementary approaches:

[0087] Uniform Sparse Sampling: Across the entire time series, data points are sampled at a fixed interval. The internal may be selected by a user or determined based on data volatility or other properties. This ensures that the general behavior of the system during normal operation is represented, even when no events or anomalies are present. In some embodiments, sampling interval is determined based on the total length of the dataset and the number of features (sensors), balancing computational efficiency with adequate coverage. The sparser the interval, the lower the computational burden. Referring to plot 300a shown in FIG. 3 A as an example, this would include samples from time periods outside of alerts 303.

[0088] Adaptive Dense Sampling in Event Intervals: The method also identifies intervals corresponding to events of interest (such as failures, anomalies, or transitions). Referring to plot 300a shown in FIG. 3A as an example, this would include time periods inside of alerts 303. Within each event interval, a denser, adaptive sampling is applied. The sampling rate within events is determined by the event’s duration and the number of features, ensuring that both short and long events are adequately covered. For short events, every point may be sampled; for longer events, a coarser but still representative sampling is used. Critical endpoints, such as the first and the last points of each event, are included to capture transitions and boundaries. Critical endpoints may be determined by machine learning models, user input, or other know methods for determining event boundaries.

[0089] The final set of indices for sensor rank computation is constructed by combining the uniformly sampled points and the adaptively sampled event points. Any indices corresponding to excluded or invalid data (e.g., due to data quality issues) are removed, ensuring that only valid and relevant data are used.

[0090] Direct Computation of Sensor Ranks at Sampled Points: At each selected sample point, the method computes sensor ranks using advanced model explainability techniques, such as SHAP (SHapley Additive exPlanations) values. These techniques quantify the contribution of each sensor to the model’s prediction at that point, providing a detailed and interpretable measure of sensor importance. Because this computation is resource-intensive some embodiments may restrict it to a strategically chosen subset of points- 19 - 5000876. vl1086.2106001yields significant computational savings without sacrificing interpretability. This subset of points may be determined by machine learning models, user input, or other know methods for data subset creation and determination.

[0091] Time-Aware Interpolation for Full Temporal Coverage: To provide sensor rank information at all time points, including those not directly sampled, the method can employ a sophisticated interpolation strategy. For each sensor (feature), the time series of computed sensor ranks is examined. If at least five valid (non-missing) points are available, Akima spline interpolation is used to fill in the gaps. This method is robust to outliers and preserves the smoothness and continuity of the underlying signal, which is especially important for physical sensor data. If fewer than five points are available, linear interpolation is used as a fallback, ensuring that all missing values are estimated. Alternative embodiments may utilize alternative point number threshold values or interpolation techniques.

[0092] In some embodiments, interpolation is performed independently for each sensor, using the actual time difference (in seconds) as the interpolation axis. This ensures that the temporal structure of the data is respected, and that interpolated values are physically meaningful.

[0093] The interpolated values are combined with the original computed values, filling in only the missing entries, so that no original information is lost.

[0094] Preservation of Physical Realism and Suitability for Equipment Analysis: Embodiments applying this approach are particularly effective for equipment and industrial process analysis, where sensor readings typically evolve gradually and do not exhibit abrupt, random changes. By leveraging the inherent smoothness of sensor signals, the method ensures that interpolated sensor ranks closely approximate the values that would have been obtained by direct measurement and computation, maintaining high fidelity and interpretability.

[0095] Embodiments utilizing this hybrid sampling strategy and data interpolation are robust to data gaps, irregular sampling, and periods of missing or excluded data, ensuring that sensor rank information is always available for downstream analysis, visualization, and decision-making. By combining uniform, low-resolution sampling with adaptive, denser sampling in event intervals and advanced, time-aware interpolation, embodiments of the present disclosure achieve a unique balance between computational efficiency and interpretability. They provide high-resolution, continuous sensor rank information across the entire time series, making it especially valuable for monitoring, diagnostics, and root-cause- 20 - 5000876. vl1086.2106001analysis in equipment and industrial applications. The novel methodology leverages the physical properties of engineered systems, ensuring that the interpolated results are both reliable and meaningful for engineering and operational decision-making.

[0096] FIG. 9 is a flow diagram illustrating the workflow 900 of an example method of alert management and maintenance response according to an example embodiment.Workflow 900 is executed at both a manufacturing facility 910 and an asset health monitoring system 920. The asset health monitoring system 920 may be local or remote to manufacturing facility 910. Environment 100, shown in FIG. 1, may include one or both of manufacturing facility 910 and an asset health monitoring system 920 and its components can be used to execute the steps of workflow 900.

[0097] At manufacturing facility 910, equipment performs 901 the monitored processes. Sensors record and stream 902 readings reflective of the monitored process being performed by equipment 901. As used herein “stream” includes the transfer of recorded data to downstream elements. Sensor readings may be streamed 902 live or recorded in a database as historical data for offline or later use. Machine learning models use the provided sensor readings 902 to monitor 903 for failures of anomalies in processing being performed by equipment 901. Embodiments can use any manner of available machine learning models deployed in both an offline and online environment. Machine learning models then output 904 a prediction of process behavior. This can take the form of a probability of the monitored process or manufacturing facility 910 having a failure or anomaly. This output may be reflective of past, present, or future process behavior.

[0098] Next, the SHAP algorithm approach is used to calculate 905 Shapley values that measure sensors’ contributions to the prediction of anomaly or failure outputted in step 904. These raw results can be post-processed 906 make it more intuitive to understand for plant operators and industrial equipment maintenance staff who are not data scientists.Improvements in model explainability can be utilized by all aspects of asset health monitoring systems 920, including those external to workflow 900. Some embodiments may exclude post-processing step 906 or use alternative or additional post-processing methodologies depending on the needs of asset health monitoring system 920 and its users.

[0099] By leveraging the SHAP algorithm to calculate 905 the contributions of individual sensors to the prediction of anomalies or failures, and subsequently refining and optionally transforming the raw results through post-processing 906 as outlined herein, workflow 900 enhances the interpretability of the data for plant operators and industrial maintenance- 21 - 5000876. vl1086.2106001personnel who lack expertise in data science. The transformed results can provide a clearer understanding by, as a non-limited example, highlighting only the positive sensor contributions that are most indicative of the anomaly or failure within a given time frame. Workflow 900 effectively eliminates confusion caused by negative values, which represent normal sensor activity, thereby improving the comprehensibility of the model and facilitating its integration into asset health monitoring systems. Using the clear results, a user can detect failures or anomalies based on the generated alerts and sensor rankings 907 and take any needed action 908 based on that understanding to protect, optimize, maintain or otherwise control the process being performed at manufacturing facility 910 by equipment 901. Sensor rankings generated in step 907 can be in any desired format herein including plots 300a, 300b, 400, 500, 600, 700, and 800. The output generated in step 907 may be presented in an interactive user interface allowing a user to determine both local and global process behavior.

[0100] FIG. 10 is a flow diagram illustrating an additional workflow 1000 of an example method of alert management and maintenance response according to an example embodiment. Workflow 1000 is executed at both a manufacturing facility 1010 and an asset health monitoring system 1020. The asset health monitoring system 1020 may be local or remote to manufacturing facility 1010. At the manufacturing facility 1010, sensor data is created that is reflective of physical properties of a monitored process. This sensor data may be in the format of a multivariate time-series data taken over a set or selected time period. Selective sampling and data interpolation may be used to either expand or narrow the streamed sensor readings 1001. Sensor data may also be live data of a current monitored process or historical data of a past monitored process. Sensor data is then sent or streamed 1001 to the asset health monitoring system 1020.

[0101] At the asset health monitoring system 1020, the sensor data is received by a machine learning model that uses the sensor data to predict process behavior including, in some embodiments, to generate the probability of a failure of anomaly 1002. Asset health monitoring system 1020 may also use the received sensor data to further update and improve the machine learning model, train additional machine learning models, use interpolation to determine unsampled information, or additional processes related to process monitoring or control.

[0102] As some in workflow 1000, some embodiments may include parallel processing paths for sensor rankings and anomaly detection. In such embodiments, the output of the machine learning model is used to both calculate the Shapley values that provide sensor- 22 - 5000876. vl1086.2106001contribution 1003 and detect predicted failures or anomalies 1004. The calculation of Shapley values 1003 may further include post-processing of those values to improve interpretability as well as apply the SHAP algorithm to increase calculation efficiently. To detect predicted failures or anomalies 1004 embodiments may use any know methods to intemperate, categorize, process, post-process, transform, or analyze the outputs of machine learning models 1002.

[0103] Next, asset health monitoring system 1020 generates sensor rankings 1004 using the calculated Shapley values and alerts 1006 based on the detection of failures or anomalies. Both the senor ranks and alerts are combined into an easy to interpret output 1005. This, as a non-limited example, may include sensor ranking vs. time plots such as plots 300a, 300b, and 500 bar graphs such as plot 400, and histograms such as histograms 600, 700, 800. The generated output 1005 may be presented to a user in an interactive user interface allowing the user to examine both local and global process behavior. Using the clear output shown sensor ranks and alerts 1005, a user take any needed action 1007 based on that understanding to protect, optimize, maintain or otherwise control the monitored process being performed at manufacturing facility 1010.

[0104] Example Digital Processing Environment

[0105] FIG. 11 illustrates a computer network or similar digital processing environment in which the disclosed embodiments 100, 900, 1000 may be implemented. Client computer(s) / devices 50 and server computer(s) 60 provide processing, storage, and input / output devices executing application programs and the like. Client computer(s) / devices 50 can also be linked through communications network 70 to other computing devices, including other client devices / processes 50 and server computer(s) 60. Communications network 70 can be part of a remote access network, a global network (e.g., the Internet), cloud computing servers or service, a worldwide collection of computers, Local area or Wide area networks, and gateways that currently use respective protocols (TCP / IP, Bluetooth, etc. to communicate with one another. Other electronic device / computer network architectures are suitable.

[0106] FIG. 12 is a block diagram of the internal structure of a computer (e.g., client processor / device 50 or server computers 60) in the computer system of FIG. 11. Each computer 50, 60 contains system bus 79, where a bus is a set of hardware lines used for data transfer among the components of a computer or digital processing system. Bus 79 is essentially a shared conduit that connects different elements of a computer system (e.g.,- 23 - 5000876. vl1086.2106001processor, disk storage, memory, input / output ports, network ports) that enables the transfer of information between the elements. Attached to system bus 79 is I / O device interface 82 for connecting various input and output devices (e.g., keyboard, mouse, displays, printers, speakers) to the computer 50, 60. Network interface 86 allows the computer to connect to various other devices attached to a network (e.g., network 70 of FIG. 11). Memory 90 provides volatile storage for computer software instructions 92 and data 94 used to implement an embodiment (e.g., workfl ow / method 900, 1000 of FIGs. 9 and 10, generation of plots 300a, 300b, 400, 500, 600, 700, 800, and the sensor ranking and calculation functions disclosed here). Disk storage 95 provides non-volatile storage for computer software instructions 92 and data 94 used to implement an embodiment. Data 94 may include plant operating plans, plant scheduling plans, datasets of plant sensor data, machine learning models, instructions for sensor rank calculation, data interpolation, post processing methods, and so forth as previously discussed. Central processor unit 84 is also attached to system bus 79 and provides for the execution of computer instructions.

[0107] In some embodiments, the processor routines 92 and data 94 are a computer program product (generally referenced 92), including a computer readable medium (e.g., a removable storage medium such as one or more DVD-ROM’s, CD-ROM’s, diskettes, tapes; or a non-removable storage medium, one or more cloud servers, etc.) that provides at least a portion of the software instructions for the disclosed system. Computer program product 92 can be installed by any suitable software installation procedure, as is well known in the art. In another embodiment, at least a portion of the software instructions may also be downloaded over a cable, communication, and / or wireless connection. In other embodiments, the programs are a computer program propagated signal product 75 (FIG. 11) embodied on a propagated signal on a propagation medium (e.g., a radio wave, an infrared wave, a laser wave, a sound wave, or an electrical wave propagated over a global network such as the Internet, or other network(s)). Such carrier medium or signals provide at least a portion of the software instructions for the routines / program 92.

[0108] In some embodiments, the propagated signal is an analog carrier wave or digital signal carried on the propagated medium. For example, the propagated signal may be a digitized signal propagated over a global network (e.g., the Internet), a telecommunications network, or other network. In some embodiments, the propagated signal is a signal that is transmitted over the propagation medium over a period of time, such as the instructions for a software application sent in packets over a network over a period of milliseconds, seconds,- 24 - 5000876. vl1086.2106001minutes, or longer. In another embodiment, the computer readable medium of computer program product 92 is a propagation medium that the computer system 50 may receive and read, such as by receiving the propagation medium and identifying a propagated signal embodied in the propagation medium, as described above for computer program propagated signal product. Generally speaking, the term “carrier medium” or transient carrier encompasses the foregoing transient signals, propagated signals, propagated medium, storage medium and the like. In other embodiments, the program product 92 may be implemented as a so-called Software as a Service (SaaS), or other installation or communication supporting end-users.

[0109] It should be understood that the flow diagrams, block diagrams, and network diagrams may include more or fewer elements, be arranged differently, or be represented differently. Some non-limiting exemplary workflows 900, 1000 are shown in FIG. 9 and FIG. 10. But further it should be understood that certain implementations may dictate the block and network diagrams and the number of block and network diagrams illustrating the execution of the embodiments be implemented in a particular way. Accordingly, alternative embodiments may also be implemented in a variety of computer architectures, physical, virtual, cloud computers, and / or some combination thereof, and, thus, the data processors described herein are intended for purposes of illustration only and not as limitations of the embodiments.

[0110] The teachings of all patents, published applications and references cited herein are incorporated by reference in their entirety.

[0111] While example embodiments have been particularly shown and described, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the scope of the embodiments.- 25 - 5000876. vl

Claims

1086.2106001CLAIMS1. A computer-implemented method of plant process monitoring, comprising:receiving an alert regarding a process at a given plant, the alert being based on measurements generated by a set of sensors of the given plant, said receiving being performed by a digital processor;for each sensor in the set of sensors, automatically calculating, by the digital processor, a contribution score, the contribution score based on each sensor’s contribution to the alert;calculating, by the digital processor, a sensor ranking based on the calculated contribution scores, the sensor ranking relatively ordering the sensors in the set of sensors based on each sensor’s contribution to the alert; andgenerating, based on the calculated sensor ranking, an output indicative of a contribution of at least one sensor in the set of sensors to the alert in a manner enabling increased accuracy in interpreting the alert and in monitoring the process at the given plants.

2. The computer-implemented method of process monitoring of claim 1 wherein the received alert is generated by a machine learning model.

3. The computer-implemented method of process monitoring of claim 1 wherein the received alert is indicative of failure, damage, likelihood of failure, or malfunction of a process unit of the process at the given plant.

4. The computer-implemented method of process monitoring of claim 1 wherein for each sensor, the calculated contribution score is a Shapley value.

5. The computer-implemented method of process monitoring of claim 4 further comprising post-processing, by the digital processor, the calculated contribution scores to eliminate negative Shapley values.

6. The computer-implemented method of process monitoring of claim 1 further comprising transforming, by the digital processor, the calculated contribution scores- 26 - 5000876. vl1086.2106001into normalized contribution scores, the normalized contribution scores sum to a probability value of the alert.

7. The computer-implemented method of process monitoring of claim 1 wherein the measurements generated by the set of sensors of the given plant are time-series data during a monitored time period, and the calculated contribution scores are based on each sensor’s contribution to the alert at an at least one sampled time within the monitored time period.

8. The computer-implemented method of process monitoring of claim 7 further comprising determining the at least one sampled time based on the time-series data during the monitored time period.

9. The computer-implemented method of process monitoring of claim 7 further comprising determining the contribution scores at an at least one interpolated time based on interpolated measurements for each sensor of the set of sensors, the interpolated measurements determined using the time-series data during the monitored time period.

10. The computer-implemented method of process monitoring of claim 9 wherein Akima spline interpolation or linear interpolation is used to generate the interpolated measurements for each sensor of the set of sensors.

11. A system for plant process monitoring, the system comprising:a digital processor communicatively coupled to an asset management tool, the digital processor configured to:receive an alert regarding a process at a given plant, the alert being based on measurements generated by a set of sensors of the given plant;for each sensor in the set of sensors, automatically calculate a contribution score, the contribution score based on each sensor’s contribution to the alert;calculate a sensor ranking based on the calculated contribution scores, the sensor ranking relatively ordering the sensors in the set of sensors based on each sensor’s contribution to the alert; and- 27 - 5000876. vl1086.2106001generate, based on the calculated sensor ranking, an output indicative of a contribution of at least one sensor in the set of sensors to the alert in a manner enabling increased accuracy in interpreting the alert and in monitoring the process at the given plants.

12. The system of claim 11 wherein the received alert is generated by a machine learning model.

13. The system of claim 11 wherein the received alert is indicative of failure, damage, likelihood of failure, or malfunction of a process unit of the process at the given plant.

14. The system of claim 11 wherein for each sensor, the calculated contribution score is a Shapley value.

15. The system of claim 14 wherein the digital processor is further configured to postprocess the calculated contribution scores to eliminate negative Shapley values.

16. The system of claim 11 wherein the measurements generated by the set of sensors of the given plant are time-series data during a monitored time period, and the calculated contribution scores are based on each sensor’s contribution to the alert at an at least one sampled time within the monitored time period.

17. The system of claim 16 wherein the digital processor is further configured to determine the at least one sampled time based on the time-series data during the monitored time period.

18. The system of claim 16 wherein the digital processor is further configured to determine determining the contribution scores at an at least one interpolated time based on interpolated measurements for each sensor of the set of sensors, the interpolated measurements determined using the time-series data during the monitored time period.

19. A non-transitory computer-readable data storage medium comprising instructions to cause a computer to:receive an alert regarding a process at a given plant, the alert being based on measurements generated by a set of sensors of the given plant;- 28 - 5000876. vl1086.2106001for each sensor in the set of sensors, automatically calculate a contribution score, the contribution score based on each sensor’s contribution to the alert;calculate a sensor ranking based on the calculated contribution scores, the sensor ranking relatively ordering the sensors in the set of sensors based on each sensor’s contribution to the alert; andgenerate, based on the calculated sensor ranking, an output indicative of a contribution of at least one sensor in the set of sensors to the alert in a manner enabling increased accuracy in interpreting the alert and in monitoring the process at the given plants.

20. The non-transitory computer-readable data storage medium of claim 20 further comprising instructions to cause a computer to determine the contribution scores at an at least one interpolated time based on interpolated measurements for each sensor of the set of sensors, the interpolated measurements determined using the measurements generated by a set of sensors of the given plant.- 29 - 5000876. vl

Images