Using Online Machine Learning to Detect and Predict Machine Failures

Abstract

This disclosure relates to a method and a machine monitoring system for predicting faults in industrial machines. The system is configured to receive sensor data related to machines such as large industrial machinery and select indicative data features of machine faults. The system then applies an unsupervised machine fault detection process and a supervised machine fault prediction process to the selected indicative data features. When new sensor data of the machine is received, the machine fault detection process is applied to at least one selected indicative data feature associated with the new sensor data. This allows the disclosed system to determine whether at least one machine fault indicator is detected, and if so, mark the machine fault. The system then updates the supervised machine fault prediction process with the newly marked machine fault indicator, such that the supervised machine fault prediction process is continuously updated and improved.

Description

[0001] Cross - Reference to Related Applications

[0002] This application claims the benefit of U.S. Provisional Application No. 62 / 832,467, filed on April 11, 2019, the content of which is incorporated herein by reference. Technical Field

[0003] The present disclosure generally relates to maintenance systems for machines, and more particularly to using online machine learning to automatically detect and predict machine failures for continuous improvement and adaptive prediction models. Background Art

[0004] Communication, processing, cloud computing, artificial intelligence, and other computerized technologies have made significant progress in recent years, heralding new areas of technology and production. In addition, many industrial technologies adopted since or before the 1970s are still in use today. Existing solutions related to these industrial technologies usually have minor improvements, thus only slightly increasing production and output.

[0005] In modern manufacturing practices, manufacturers typically require strict production schedules and provide perfect or near - perfect production quality. Therefore, whenever an unexpected machine failure occurs, these manufacturers are at risk of suffering significant losses. A machine failure is an event that occurs when a machine deviates from correct operation. An error, which is a deviation from the correct or expected state of the machine, is not necessarily a failure but may lead to and indicate a potential future failure. In addition, an error may cause abnormal machine behavior that may affect performance.

[0006] The average machine downtime due to failures for a typical manufacturer is 17 days per year (i.e., the average amount of time during which production is partially or fully stopped due to machine failures), which means 17 days of lost production and revenue. For example, in the case of a typical 450 - megawatt power turbine, one day of downtime may cost the manufacturer more than $3 million in lost revenue. Such downtime may have additional costs associated with repairs, safety precautions, etc.

[0007] In energy power plants, billions of dollars are spent each year to ensure reliability, especially for backup systems and redundancies used to minimize production downtime. In addition, monitoring systems can be used to quickly identify failures, thus accelerating the resumption of production when a downtime occurs. However, existing monitoring systems typically only identify failures after or before the downtime has started.

[0008] Some existing monitoring and maintenance solutions use detection functions to predict upcoming machine failures. Such solutions are based on data collected by sensors connected to such machines. The processing of sensor data is limited to the signals collected by the sensors and is limited to static prediction. However, these solutions have several deficiencies, such as becoming obsolete and irrelevant as machine data changes, requiring continuous maintenance of prediction mechanisms, static prediction, and detection models for processing dynamic data, etc.

[0009] Therefore, it would be advantageous to provide a solution that can overcome the above challenges. Summary of the Invention

[0010] The following is an overview of several example embodiments of the present disclosure. This overview is provided to facilitate the reader's basic understanding of these embodiments and does not fully define the scope of the present disclosure. This overview is not an extensive overview of all expected embodiments and is neither intended to identify the key or important elements of all embodiments nor to depict the scope of any or all aspects. Its sole purpose is to present some concepts of one or more embodiments in a simplified form as a preface to the more detailed description presented later. For convenience, the term "certain embodiments" may be used herein to refer to a single embodiment or multiple embodiments of the present disclosure.

[0011] Certain embodiments disclosed herein include an online machine learning-based method for detecting and predicting industrial machine failures. The method includes receiving sensor data related to at least one industrial machine; generating a plurality of data features based on at least a portion of the sensor data; selecting at least one indicative data feature for machine failure detection from the plurality of data features; applying an unsupervised machine failure detection process to the selected at least one indicative data feature, wherein the unsupervised machine failure detection process is configured to detect a machine failure indicator based on the selected at least one indicative data feature; receiving new sensor data related to the at least one industrial machine; determining whether at least one machine failure indicator is detected in the new sensor data by applying the unsupervised machine failure detection process to the selected at least one indicative data feature associated with the new sensor data; and marking the at least one machine failure indicator when it is determined that at least one machine failure indicator is detected, wherein when it is determined that no machine failure indicator is detected, the unsupervised machine failure detection process continuously searches for machine failure indicators.

[0012] Certain embodiments disclosed herein also include a system based on an online machine learning method for detecting and predicting industrial machine failures. The system includes a processing circuit; and a memory containing instructions that, when executed by the processing circuit, configure the system to: receive sensor data related to at least one industrial machine; generate a plurality of data features based on at least a portion of the sensor data; select at least one indicative data feature for machine failure detection from the plurality of data features; apply an unsupervised machine failure detection process to the selected at least one indicative data feature, wherein the unsupervised machine failure detection process is configured to detect a machine failure indicator based on the selected at least one indicative data feature; receive new sensor data related to the at least one industrial machine; determine whether at least one machine failure indicator is detected in the new sensor data by applying the unsupervised machine failure detection process to the selected at least one indicative data feature associated with the new sensor data; and mark the at least one machine failure indicator when it is determined that the at least one machine failure indicator is detected, wherein when it is determined that no machine failure indicator is detected, the unsupervised machine failure detection process continuously searches for a machine failure indicator. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The subject matter disclosed herein is particularly pointed out and distinctly claimed at the end of the specification. The foregoing and other objects, features, and advantages of the disclosed embodiments will become apparent from the following detailed description taken in conjunction with the accompanying drawings.

[0014] Figure 1 is a network diagram for describing various disclosed embodiments.

[0015] Figure 2 is a schematic diagram of a machine management server according to an embodiment.

[0016] Figure 3 is a flowchart showing a method for automatic detection and prediction of machine failures according to an embodiment.

[0017] Figure 4A is an example graph showing a training process of a machine failure detection process according to an embodiment.

[0018] Figure 4B is an example graph showing the application of a machine failure detection or prediction process to new sensor data according to an embodiment. DETAILED DESCRIPTION

[0019] It is important to note that the embodiments disclosed herein are merely examples of many useful applications of the innovative teachings herein. Generally, statements made in the specification of this application do not necessarily limit any of the various claimed embodiments. Additionally, certain statements may apply to some inventive features but not to others. Generally, unless otherwise stated, a singular element may be plural and vice versa without loss of generality. In the figures, the same reference numerals represent the same components through multiple views.

[0020] The various disclosed embodiments include methods and machine monitoring systems for predicting machine failures using machine learning techniques. In one embodiment, the machine monitoring system is configured to receive sensor data associated with a machine (such as large industrial machinery) and select indicative data features of a machine failure. The system then applies an unsupervised machine failure detection process and a supervised machine failure prediction process to the selected indicative data features. When new sensor data of the machine is received, the machine failure detection process is applied to the selected at least one indicative data feature associated with the new sensor data. This allows the disclosed system to determine whether at least one machine failure indicator is detected, and if so, mark the machine failure. The system is then configured to automatically update the supervised machine failure prediction process using the newly marked machine failure indicator, such that the supervised machine failure prediction process is continuously updated and improved.

[0021] Figure 1 An example network diagram 100 for describing the various disclosed embodiments is shown. The example network diagram 100 includes a machine monitoring system (MMS) 130, a management server 140, a database 150, and a client device 160 communicatively connected via a network 110. The example network diagram 100 also includes a plurality of sensors 120-1 to 120-n (hereinafter referred to individually as sensor 120 and collectively as sensors 120 for simplicity only), where n is an integer equal to or greater than 1, connected to the machine monitoring system 130. The network 110 can be, but is not limited to, a wireless network, a cellular or wired network, a local area network (LAN), a wide area network (WAN), a metropolitan area network (MAN), the Internet, the World Wide Web (WWW), similar networks, and any combination thereof.

[0022] The client device 160 can be, but is not limited to, a personal computer, a laptop computer, a tablet computer, a smart phone, a wearable computing device, or any other device capable of receiving and displaying notifications indicating maintenance and failure timing predictions, results of supervised analysis, unsupervised analysis of machine operation data, etc.

[0023] Sensor 120 is located near industrial machine 170 (e.g., physical proximity within a predetermined threshold). Industrial machine 170 can be any machine whose performance can be represented by sensor data, such as but not limited to, turbines, engines, welding machines, three-dimensional (3D) printers, injection molding machines, combinations thereof, parts thereof, etc.

[0024] Each sensor 120 is configured to collect sensor inputs based on the operation of machine 170, such as but not limited to sound signals, ultrasonic signals, light, motion tracking indicators, temperature, energy consumption indicators, etc. Sensor 120 can include but not limited to sound capture sensors, motion tracking sensors, energy consumption meters, thermometers, etc. Any one of sensors 120 can be connected to machine 170, but it is not necessary to be connected to machine 170 (such connection is not shown in Figure 1 for simplicity only and does not limit the disclosed embodiments).

[0025] Sensor 120 is connected to machine monitoring system 130. Machine monitoring system 130 can be configured to store and preprocess the raw sensor data received from sensor 120. Optionally or jointly, machine monitoring system 130 can be configured to periodically retrieve the collected sensor data stored in, for example, database 150.

[0026] Preprocessing can include but not limited to data cleaning, normalization, rescaling, detrending, reformatting, noise filtering, combinations thereof, etc. Preprocessing can also include data feature extraction. The result of data feature extraction can include data features used by management server 140 during machine learning to detect data features indicating machine failure when a machine failure occurs, or data features indicating an impending machine failure as further described below.

[0027] In one embodiment, management server 140 can be configured to identify multiple data features represented by at least one statistical feature in the timestamped sensor data. The multiple data features represent the behavior of at least one component of the machine. Data feature extraction can include but not limited to dimensionality reduction techniques, such as but not limited to singular value decomposition, discrete Fourier transform, discrete wavelet transform, segment method, or combinations thereof. When using such dimensionality reduction techniques, preprocessing can result in, for example, a lower-dimensional space of the sensory input. Machine monitoring system 130 is configured to send the preprocessed sensory input to management server 140.

[0028] In one embodiment, the management server 140 is configured to receive timestamp sensor data associated with at least one machine (e.g., machine 170) via the network 110. The timestamp sensor data can be received from the machine monitoring system 130. The timestamp sensor data can be received from one or more sensors, such as sensor 120. The sensor data can be received continuously and can be received in real time. Each type of sensor data can be related to at least one process executed by the machine associated with the machine, etc. That is, the first type of sensor data can be related to the temperature of the industrial machine 170, the second type of sensor data can be related to the speed of a certain gear of the machine 170, and so on. In another embodiment, the management server 140 is configured to receive preprocessed sensor data.

[0029] In one embodiment, the management server 140 can be configured to store the sensor data (raw data, preprocessed data, or both) received from the machine monitoring system 130. Alternatively or in combination, the sensor data can be stored in the database 150. The database 150 can further store sensory inputs (raw, preprocessed, or both) collected from a plurality of other sensors (not shown) associated with other machines (also not shown). The database 150 can also store indicators, anomaly patterns, behavior trends, fault predictions, machine learning models for analyzing sensory input data, or combinations thereof. In one embodiment, the management server 140 is configured to preprocess the raw sensory inputs as further described above.

[0030] In one embodiment, the management server 140 is configured to generate one or more data features based on the received sensor data. The data features can be represented by features calculated mathematically. In another embodiment, the generation can be performed by converting the preprocessed sensor data and / or each type of raw sensor data into one or more data features represented by features calculated mathematically. The data features can be a mathematical representation of the raw sensor data, allowing the sensor data to be represented in a clearer manner. The generation can be achieved using at least one statistical analysis technique. The statistical analysis techniques can include, but are not limited to, calculating the average of the raw sensory inputs, calculating the median of the raw sensory inputs, calculating the standard deviation of the raw sensory inputs, etc.

[0031] In one embodiment, performing statistical analysis techniques on raw or preprocessed sensor data allows the management server 140 to generate multiple data features. The data features allow facilitating the identification of associations between multiple anomalies associated with multiple processes involving the machine 170. That is, the data features are new information representations of the raw or preprocessed sensor data that allow identifying hidden structures in the raw sensor data. In another embodiment, the transformation includes reducing the size of the raw sensor data, e.g., by converting raw data at second resolution to indicate minute resolution. The transformation may include singular value decomposition, discrete Fourier transform, discrete wavelet transform, line segment method, etc.

[0032] In another embodiment, the transformation includes normalizing the raw sensor data and / or the preprocessed sensor data to a unified scale. That is, the raw sensor data may be presented in different scales, and thus the management server 140 may be configured to normalize the raw sensor data by generating a unified scale for all the raw sensor data. The sensor data may include a particular gear sensed by a first sensor, the oil temperature sensed by a second sensor, etc. The unified scale can be used to identify associations between different types of sensor data, correlations between abnormal behaviors of different types of sensor data, etc.

[0033] In one embodiment, the management server 140 is configured to select at least one indicative data feature from the multiple data features for machine fault detection and / or machine fault prediction. The indicative data feature is a representation of the sensor data that, when analyzed, allows more accurately indicating a machine fault and / or an upcoming machine fault relative to other data features that contribute less to the machine fault prediction process or the machine fault detection process.

[0034] In one embodiment, the selection of the indicative data feature is performed by scanning a large and comprehensive feature database to obtain two subsets of information features for detection and prediction. The indicative data feature may include descriptive statistical features. The indicative feature for event detection indicates a machine fault once such a fault has occurred. For example, the indicative data feature for machine fault detection may be related to the water temperature, the revolutions per minute (RPM) of a particular industrial machine component, etc. The indicative feature for fault prediction is an indicative feature that shows a gradual degradation before the fault occurs. For example, the indicative data feature for event prediction may be related to the vibration sound level of an industrial machine, the oil pressure of a particular component of an industrial machine, etc. It should be noted that the management server 140 may use sensory inputs sensed by the same sensor to detect and predict events in the industrial machine 170. For example, the oil pressure may be an indicative data feature for machine fault detection, and it may also be an indicative data feature for machine fault prediction. Feature selection is iteratively performed in each retraining iteration of the supervised model.

[0035] In another embodiment, a plurality of indicative data features may be selected based on at least one distribution of a plurality of data features. The distribution may indicate a development association among the plurality of data features towards a machine failure. In one embodiment, the distribution may indicate an association among the plurality of data features during a machine failure.

[0036] In another embodiment, at least one indicative data feature is selected from a plurality of data features based on a probability of predicting and / or detecting a machine failure. For example, an industrial machine including five components (e.g., machine 170) is being monitored, and during a specific time period, parameters of three indicative data features associated with three components of machine 170 indicate abnormal parameters for each component.

[0037] According to the same example, the management server 140 may determine a distribution of the indicative data features, i.e., abnormal parameters for each of them, indicating an association among the three indicative data features, which may indicate an upcoming machine failure. In another embodiment, the selection of indicative data features having a better probability of contributing more to predicting a machine failure relative to other data features may be achieved by identifying an increased change in the data feature distribution relative to the normal state of the machine prior to the machine failure.

[0038] In one embodiment, the management server 140 is configured to apply an unsupervised machine failure detection process and a supervised machine failure prediction process to the selected indicative data features. The unsupervised machine failure detection process is configured to detect a machine failure indicator based on the selected indicative data features. The machine failure indicator may be, for example, a value indicating a machine failure associated with a specific parameter of at least one component of the machine (e.g., machine 170). For example, an oil temperature of 90 degrees Celsius for a certain industrial machine may be classified as an indicator of a machine failure.

[0039] The supervised machine failure prediction process is configured to predict a machine failure based on the selected indicative data features. In one embodiment, the unsupervised machine failure detection process and the supervised machine failure prediction process may be applied to at least a portion of timestamp sensor data that has been previously labeled or annotated with respect to one or more machine failure indicators. Thus, a training phase is achieved by applying the process to the labeled or annotated sensor data. The training phase may include recording features associated with each machine failure indicator, such as an average value of the machine failure, continuous time, etc.

[0040] In one embodiment, the management server 140 is configured to receive new sensor data associated with at least one machine (e.g., machine 170). The new sensor data may include at least a portion of information received from at least one sensor (e.g., sensor 120) that the management server 140 has never processed before. That is, the new sensor data may include, for example, a machine failure that has not been previously recorded or flagged. In one embodiment, the new sensor data may be associated with one or more components of at least one machine 170. For example, when the same sensor data is received regarding a first component of machine 170, at least one new set of data is received regarding three other components of machine 170.

[0041] In one embodiment, the management server 140 is configured to determine whether one or more machine fault indicators are detected in the new sensor data by applying an unsupervised machine fault detection process to the selected indicative data feature or indicative data features. The unsupervised machine fault detection process is designed to detect machine fault indicators and new machine fault indicators.

[0042] In one example, ten machine fault indicators are detected in the sensor data. When new sensor data is received at the management server 140, two additional new machine fault indicators are detected. In one embodiment, the new machine fault indicators may be associated with machine components that have never previously indicated a machine fault. In a further embodiment, the new machine fault indicators may be associated with a known machine fault type of the machine component, but at a different new scale. For example, the new machine fault indicator may indicate abnormal behavior represented by the revolutions per minute (RPM) of the machine engine, which is a parameter that has never been indicated on a machine fault before. According to another example, the new machine fault indicator may indicate abnormal behavior of the machine represented by the oil temperature of the machine, which is a parameter that has been indicated on machine faults many times before but at different scales.

[0043] In one embodiment, the management server 140 is configured to flag the one or more machine fault indicators when it is determined that one or more machine fault indicators are detected. In one embodiment, an electronic flag may be generated and associated with each new machine fault indicator. The electronic flag may include descriptive information related to the machine fault indicator, such as a title indicating the fault type, fault level, etc.

[0044] In one example, a new vibration level of a particular machine is detected through a machine fault detection process and classified as a machine fault. According to the same example, the new vibration level, i.e., the value associated with the new vibration level, is marked by the management server 140. The marking may include, for example, the value of the new vibration level, the time when the new vibration level was detected, the sensor used to sense the new vibration level, the machine components affected by the new vibration level, etc. It should be noted that when no machine fault indicator is detected, the unsupervised machine fault detection process continuously searches for machine fault indicators.

[0045] In one embodiment, the management server 140 is configured to update the supervised machine fault prediction process with one or more marked machine fault indicators such that the supervised machine fault prediction process is continuously updated. By updating the machine fault prediction process with newly marked machine fault indicators, the prediction ability of the machine fault prediction process remains high over time. That is, the machine fault prediction process is trained using the methods disclosed herein to predict machine faults even in the absence of human intervention for long periods of time. Predicting machine faults may include identifying patterns, trends, etc. as shown by the machine sensor data discussed further above.

[0046] In one example, the embodiments disclosed herein allow for the detection and prediction of the downtime of industrial machines. For this purpose, the management server 140 requires an initial marking period during which faults are known and marked. The time range of this marking can be provided by the customer in the form of a fault log or generated internally in the absence of such a log. When log data is not available, the disclosed management server 140 is configured to continue with the initial training and generate two machine learning models: one for fault detection and another for fault prediction (each based on a relevant subset of indicative features).

[0047] These initially trained models are continuously updated over time (to adapt to changes and new types of faults). Each new fault detected by the management server 140 is also fed back into the management server 140 and used to retrain the detection and prediction models on-the-fly. Once the updated models are trained and ready, they replace the previous models and are applied to new streaming data, and the process continues iteratively.

[0048] In one embodiment, the disclosed method is based on online machine learning techniques. Online machine learning is a machine learning method in which data becomes available in a sequential order and is used to update the most indicative predictors of future data at each step, as opposed to batch learning techniques that generate the most indicative predictors by learning the entire training data set at once. That is, by using the disclosed method, the process of determining whether a new machine fault indicator is detected occurs continuously, and the labeling process and the updating process of the supervised machine fault prediction process occur continuously.

[0049] In one embodiment, the disclosed method may be implemented using semi-supervised learning. Semi-supervised learning is a class of machine learning tasks and techniques that typically use a small amount of labeled data and a large amount of unlabeled data. When using semi-supervised learning, the process requires learning from a data set that includes both labeled and unlabeled data. Semi-supervised techniques are a combination of supervised and unsupervised machine learning methods. One such technique involves completing the unlabeled samples with an unsupervised machine learning method and then allowing a supervised method to be applied on the complete labeled data set.

[0050] Figure 2 An example block diagram of a management server 140 implemented according to one embodiment is shown. The management server 140 includes a processing circuit 210 coupled to a memory 220, a storage device 230, a network interface 240, and a machine learning (ML) processor 250. In one embodiment, the components of the management server 140 may be communicatively connected via a bus 260.

[0051] The processing circuit 210 may be implemented as one or more hardware logic components and circuits. For example, but not limited to, illustrative types of hardware logic components that may be used include field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), system on a chip systems (SOCs), graphics processing units (GPUs), tensor processing units (TPUs), general purpose microprocessors, microcontrollers, digital signal processors (DSPs), etc., or any other hardware logic components that may perform the computation or other manipulation of information.

[0052] The memory 220 may be volatile (e.g., RAM), non-volatile (e.g., ROM, flash memory, etc.), or a combination thereof. In one configuration, computer-readable instructions for implementing one or more embodiments disclosed herein may be stored in the storage device 230.

[0053] In another embodiment, the memory 220 is configured to store software. The software should be interpreted broadly as any type of instruction, whether referring to software, firmware, middleware, microcode, hardware description language, or others. The instructions may include code (e.g., source code format, binary code format, executable code format, or any other suitable code format). When executed by one or more processors, the instructions cause the processing circuit 210 to perform the various processes described herein.

[0054] The storage device 230 can be a magnetic storage device, an optical storage device, etc., and can be implemented as, for example, flash memory or other storage technologies, CD-ROM, digital versatile disc (DVD), SSD, or any other medium that can be used to store the desired information.

[0055] The network interface 240 allows the management server 140 to communicate with the machine monitoring system 130, for example, to receive raw and / or preprocessed sensory input. Additionally, the network interface 240 allows the management server 140 to communicate with the client device 160 to send notifications related to, for example, machine abnormal activities, machine fault prediction, etc.

[0056] The machine learning process 250 is configured to perform machine learning processing based on the sensor data received via the network interface 240, as further described herein. In one embodiment, the machine learning unit 250 is also configured to predict machine failures, update one or more machine fault prediction processes, etc. The machine learning process 250 can be implemented as, for example, a GPU, TPU, general microprocessor, DSP, etc.

[0057] It should be understood that the embodiments described herein are not limited to Figure 2 the specific architecture shown, and other architectures can be equally used without departing from the scope of the disclosed embodiments.

[0058] Figure 3 is an example flowchart 300 showing a method for detecting and predicting machine failures according to an embodiment. In one embodiment, the method can be executed by the machine fault predictor 140 (see Figure 1 and 2 ).

[0059] At S310, timestamp sensor data related to at least an industrial machine (e.g., machine 170) is received. The timestamp sensor data can be received from one or more sensors of the machine 170. Each type of sensor data can be related to at least one process associated with the machine, performed by the machine, etc.

[0060] At S320, multiple data features are generated based on at least a portion of the timestamp sensor data. The data features are extracted using data extraction techniques. Data feature extraction can include, but is not limited to, dimensionality reduction techniques such as, but not limited to, singular value decomposition, discrete Fourier transform, discrete wavelet transform, the segment method, or a combination thereof. In one embodiment, the data features can be represented by at least one statistical feature and / or can represent the behavior of at least one component of an industrial machine.

[0061] At S330, at least one indicative data feature is selected from the multiple data features for at least one of machine fault detection and machine fault prediction.

[0062] An indicative data feature is a representation of the sensor data that, when analyzed, allows for a more accurate indication of a machine fault and / or an impending machine fault relative to other data features that contribute less to the machine fault prediction process or the machine fault detection process. As described above, the indicative data feature can be selected based at least on the distribution of the multiple data features.

[0063] At S340, an unsupervised machine fault detection process and a supervised machine fault prediction process are performed on the at least one selected indicative data feature. The unsupervised machine fault detection process is configured to detect a machine fault indicator based on the at least one selected indicative data feature, and the supervised machine fault prediction process is configured to predict a machine fault based on the at least one selected indicative data feature, as further discussed Figure 1 below.

[0064] At S350, new sensor data related to the machine is received. The new sensor data can include at least a portion of information received from at least one sensor (e.g., sensor 120) that the management server 140 has never processed before, such as the value of a certain component that has reached a new level.

[0065] At S360, based on the at least one selected indicative data feature, it is determined whether one or more machine fault indicators are detected in the received new sensor data. If so, execution continues to S370; otherwise, execution continues to S350. This determination is achieved by applying the unsupervised machine fault detection process to the new sensor data or the at least one selected indicative data feature associated with the new sensor data. It should be noted that when it is determined that no machine fault indicator is detected, the unsupervised machine fault detection process continuously searches for machine fault indicators.

[0066] At S370, when one or more machine fault indicators are determined to be detected, mark one or more machine fault indicators. In one embodiment, an electronic mark may be generated and associated with each newly detected machine fault indicator. The electronic mark may include descriptive information related to the machine fault indicator, as discussed further above.

[0067] At S380, update the supervised machine fault prediction process (previously applied) with the marked one or more machine fault indicators. In some embodiments, semi-supervised or self-supervised methods may also be used. It should be noted that the supervised machine fault prediction process is continuously updated. It should be further noted that the updated and marked one or more machine fault indicators may be merged with known or old sensor data associated with a machine (such as machine 170) previously detected and stored in, for example, database 150.

[0068] Figure 4A FIG. 400A is an example graph showing a representation of a training process of a machine fault detection process according to an embodiment. Figure 4A The graph shown includes graph 400A, in which curve 410A is shown and represents sensor data of a certain parameter of the machine, such as revolutions per minute (RPM) of a machine engine. Curve 420A represents the marked machine faults provided to the machine fault detection process for training the machine fault detection process to detect machine faults. The point where the machine fault starts is represented by 430A, and the point where the machine fault ends is represented by 440A.

[0069] Figure 4B FIG. 400B is an example graph showing a representation of applying a machine fault detection and / or prediction process to new sensor data according to an embodiment. Figure 4B The graph shown therein includes graph 400B, in which curve 410B is shown and represents sensor data of a certain parameter of the machine, such as revolutions per minute (RPM) of a machine engine. Curve 420B represents the new machine faults detected by the machine fault detection process.

[0070] It should be noted that the information related to each newly detected machine fault is used as an input to the machine fault detection process as well as the machine fault prediction process. The trained model, i.e., the process, is continuously updated over time to adapt to changes and new types of faults. Each new fault detected by the system is also fed back into the system and used to retrain the detection and prediction models in real-time or near real-time. Once the updated process is trained above a certain level of certainty, the updated process replaces the previous process and is applied to new streaming sensor data, and the method iterates continuously.

[0071] The various embodiments disclosed herein can be implemented as hardware, firmware, software, or any combination thereof. Additionally, the software is preferably implemented as an application tangibly embodied on a program storage unit or computer-readable medium, which is composed of components, certain devices, and / or combinations of devices. The application can be uploaded to and executed by a machine including any suitable architecture. Preferably, the machine is implemented on a computer platform having hardware such as one or more central processing units ("CPUs"), memory, and input / output interfaces. The computer platform may also include an operating system and microinstruction code. The various processes and functions described herein can be part of the microinstruction code, part of the application, or any combination thereof, and can be executed by the CPU, whether or not such a computer or processor is explicitly shown. Additionally, various other peripheral units can be connected to the computer platform, such as additional data storage units and printing units. Further, a non-transitory computer-readable medium is any computer-readable medium other than a transitory propagated signal.

[0072] It should be understood that any reference to an element using names such as "first," "second," etc. in this document generally does not limit the number or order of those elements. Instead, these names are generally used herein as a convenient method of distinguishing between two or more elements or instances of an element. Thus, a reference to a first and a second element does not mean that only two elements can be used there or that the first element must somehow precede the second element. Additionally, unless otherwise stated, a group of elements includes one or more elements.

[0073] As used herein, the phrase "at least one" followed by a list of items means that any one of the listed items can be used alone or any combination of two or more of the listed items can be used. For example, if a system is described as including "at least one of A, B, and C," the system can include A alone; B alone; C alone; A and B in combination; B and C in combination; A and C in combination; or A, B, and C in combination.

[0074] All of the examples and conditional language recited herein are intended for pedagogical purposes to aid the reader in understanding the principles of the disclosed embodiments and the concepts contributed by the inventor to further the art, and should not be construed as being limited to these specifically recited examples and conditions. Additionally, all statements herein reciting principles, aspects, and embodiments of the disclosed embodiments and their specific examples are intended to cover their structural and functional equivalents. Moreover, such equivalents are intended to include both currently known equivalents and equivalents developed in the future, i.e., any elements developed that perform the same function regardless of structure.

Claims

1. An online machine learning-based method for detecting and predicting industrial machine failures, comprising: Receiving sensor data related to at least one industrial machine; Generating a plurality of data features based on at least a portion of the sensor data; Select at least one indicative data feature for machine fault detection from multiple data features; Apply an unsupervised machine fault detection process to the selected at least one indicative data feature, wherein the unsupervised machine fault detection process is configured to detect a machine fault indicator based on the selected at least one indicative data feature; Receive new sensor data related to the at least one industrial machine; Determine whether at least one machine fault indicator is detected in the new sensor data by applying the unsupervised machine fault detection process to the selected at least one indicative data feature associated with the new sensor data; and Mark the at least one machine fault indicator when it is determined that the at least one machine fault indicator is detected, wherein when it is determined that no machine fault indicator is detected, the unsupervised machine fault detection process continuously searches for machine fault indicators; Select at least one indicative data feature for machine fault prediction from multiple data features; Apply a supervised machine fault prediction process to the selected at least one indicative data feature, wherein the supervised machine fault prediction process is configured to predict a machine fault based on the selected at least one indicative data feature; and Update the supervised machine fault prediction process with the marked at least one machine fault indicator, such that the supervised machine fault prediction process is continuously and automatically updated and improved.

2. The method according to claim 1, wherein, The multiple data features represent the behavior of at least one component of the at least one industrial machine.

3. The method according to claim 1, wherein, The multiple data features are generated based on at least one statistical method.

4. The method according to claim 1, wherein, At least one indicative data feature for machine fault detection is selected from the multiple data features based on the probability of detecting a machine fault.

5. The method according to claim 1, wherein, At least one indicative data feature for machine fault prediction is selected from the multiple data features based on the probability of predicting a machine fault.

6. The method according to claim 1, further comprising: Selecting a plurality of indicative data features from the plurality of data features based on at least one distribution of the plurality of data features, wherein the at least one distribution indicates at least one development association towards machine failure among the plurality of data features.

7. The method according to claim 1, wherein, At least a portion of the sensor data is pre-marked with at least one machine fault indicator.

8. The method according to claim 1, wherein, Determining whether at least one machine fault indicator is detected in the new sensor data is based on semi-supervised machine learning.

9. A non-transitory computer-readable medium having instructions stored thereon for causing a processing circuit to execute a process, the process comprising: Receiving sensor data related to at least one industrial machine; Generating a plurality of data features based on at least a portion of the sensor data; Select at least one indicative data feature for machine fault detection from multiple data features; Apply an unsupervised machine fault detection process to the selected at least one indicative data feature, wherein the unsupervised machine fault detection process is configured to detect a machine fault indicator based on the selected at least one indicative data feature; Receive new sensor data related to the at least one industrial machine; Determine whether at least one machine fault indicator is detected in the new sensor data by applying the unsupervised machine fault detection process to the selected at least one indicative data feature associated with the new sensor data; Mark the at least one machine fault indicator when it is determined that the at least one machine fault indicator is detected, wherein when it is determined that no machine fault indicator is detected, the unsupervised machine fault detection process continuously searches for machine fault indicators; and Select at least one indicative data feature for machine fault prediction from multiple data features; Apply a supervised machine fault prediction process to at least one selected indicative data feature, wherein the supervised machine fault prediction process is configured to predict a machine fault based on the at least one selected indicative data feature; and Update the supervised machine fault prediction process with at least one marked machine fault indicator such that the supervised machine fault prediction process is continuously and automatically updated and improved.

10. A system for an online machine learning-based method for detecting and predicting industrial machine failures, comprising: A processing circuit; and A memory that contains instructions which, when executed by the processing circuit, configure the system to: Receive sensor data related to at least one industrial machine; Generate a plurality of data features based on at least a portion of the sensor data; Select at least one indicative data feature for machine fault detection from the plurality of data features; Apply an unsupervised machine fault detection process to the at least one selected indicative data feature, wherein the unsupervised machine fault detection process is configured to detect a machine fault indicator based on the at least one selected indicative data feature; Receive new sensor data related to the at least one industrial machine; Determine whether at least one machine fault indicator is detected in the new sensor data by applying the unsupervised machine fault detection process to the at least one selected indicative data feature associated with the new sensor data; and When it is determined that the at least one machine fault indicator is detected, mark the at least one machine fault indicator, wherein when it is determined that no machine fault indicator is detected, the unsupervised machine fault detection process continuously searches for a machine fault indicator, wherein the system is further configured to: Select at least one indicative data feature for machine fault prediction from the plurality of data features; Apply the at least one selected indicative data feature to the supervised machine fault prediction process, wherein the supervised machine fault prediction process is configured to predict a machine fault based on the at least one selected indicative data feature; and Update the supervised machine fault prediction process with at least one marked machine fault indicator such that the supervised machine fault prediction process is continuously and automatically updated and improved.

11. The system according to claim 10, wherein, The plurality of data features represent the behavior of at least one component of the at least one industrial machine.

12. The system according to claim 10, wherein, The plurality of data features are generated based on at least one statistical method.

13. The system according to claim 10, wherein, At least one indicative data feature for machine fault detection is selected from the plurality of data features based on the probability of detecting a machine fault.

14. The system according to claim 10, wherein, At least one indicative data feature for machine fault prediction is selected from the plurality of data features based on the probability of predicting a machine fault.

15. The system according to claim 10, wherein, The system is further configured to: Select a plurality of indicative data features from the plurality of data features based on at least one distribution of the plurality of data features, wherein the at least one distribution indicates at least one development association towards a machine fault among the plurality of data features.

16. The system according to claim 15, wherein, At least a portion of the sensor data is pre - marked with at least one machine fault indicator.

17. The system according to claim 15, wherein, The system is further configured to: Determine whether at least one machine fault indicator is detected in the new sensor data based on semi - supervised machine learning.

Images