DEVICE FOR FAULT DETECTION AND FAILURE PREDICTION BY MONITORING VIBRATIONS OF OBJECTS IN PARTICULAR INDUSTRIAL ASSETS

Abstract

A device configured to monitor a vibrating object, the device comprising a common housing holding an accelerometer for sampling vibration signatures of the vibrating object, resulting in vibration samples, and a computing device comprising a data processor and a memory having stored thereon a computer program product for monitoring the vibrating object, the computer program product comprising an input module to receive the vibration samples, an analysis module to analyze the vibration samples to derive asset health scores, a machine learning model to determine asset operating ranges, and an output module to output messages, wherein the computer program product when running on the data processor causes the computing device to receive during a time interval, having an end time t1, the vibration samples from the accelerometer, resulting in a time series vector array, and to analyze the vibration samples comprising deriving from the time series vector array a baseline asset health score and deriving from the time series vector array a time series asset health score, and to subject at least a part of the time series asset health score to the machine learning model for determining at least one asset operating range, and to receive a further vibration sample at a monitor time t2 wherein the monitor time t2 is subsequent to the end time t1, and to derive from the further vibration sample an asset health score, and to determine if the asset health score falls within an operating range determined by the machine learning model, resulting in a monitor result, and to output a message depending on the monitor result, and wherein the device consumes less current than 20 mA.

Claims

1. A device configured to monitor a vibrating object, the device comprising a common housing holding: an accelerometer for sampling vibration signatures of the vibrating object, resulting in vibration samples, and a computing device comprising a data processor and a memory having stored thereon a computer program product for monitoring the vibrating object, the computer program product comprising: an input module to receive the vibration samples; an analysis module to analyze the vibration samples to derive asset health scores; a machine learning model to determine asset operating ranges, and an output module to output messages, wherein the computer program product when running on the data processor causes the computing device to: receive during a time interval, having an end time t1, the vibration samples from the accelerometer, resulting in a time series vector array; analyze the vibration samples comprising: deriving from the time series vector array a baseline asset health score, and deriving from the time series vector array a time series asset health score; subject at least a part of the time series asset health score to the machine learning model for determining at least one asset operating range; receive a further vibration sample at a monitor time t2 wherein the monitor time t2 is subsequent to the end time t1; derive from the further vibration sample an asset health score; determine if the asset health score falls within an operating range determined by the machine learning model, resulting in a monitor result; output a message depending on the monitor result, and wherein the device consumes less current than 20 mA.

2. The device of claim 1, wherein furthermore the common housing holding a power source;

3. The device of claim 2, wherein the power source is selected from one or more batteries, PV cells with suitable capacitors, energy harvesting systems with suitable capacitor, and a combination thereof.

4. The device of claim 1, wherein the vibrating object is corresponding to one selected from a vibrating pump, a vibrating compressor, a vibrating turbine, a vibrating vehicle, a vibrating satellite, a vibrating motor, a vibrating fan, a vibrating pipe, a vibrating construction part, a vibrating steel sheet, a vibrating cover, a vibrating surface of the earth, and a combination thereof.

5. The device of claim 1, wherein furthermore the computer program product comprises a further machine learning model to predict future asset health scores, and wherein the computer program product when running on the data processor causes the computing device to: subject at least a part of the time series asset health score to the further machine learning model for predicting a future asset health score at a future time t4 subsequent to the end time t1; derive, at a predict time t3, the future asset health score from the further machine learning model wherein the predict time t3 is equal or subsequent to the end time t1 and wherein the predict time t3 is prior to the future time t4; output a message to the RF module depending on a deviation of the future asset health score from the baseline asset health score.

6. A system configured to monitor a vibrating object, comprising: a power source; a power management circuit to control flow and direction of electrical power from the power source; an accelerometer for sampling vibration signatures of the vibrating object, resulting in vibration samples; a RF module; an antenna, and a computing device comprising a data processor and a memory having stored thereon a computer program product for monitoring the vibrating object, the computer program product comprising: an input module to receive the vibration samples; an analysis module to analyze the vibration samples to derive asset health scores; a machine learning model to determine asset operating ranges, and an output module to output messages, wherein the computer program product when running on the data processor causes the computing device to: receive during a time interval, having an end time t1, the vibration samples from the accelerometer, resulting in a time series vector array; analyze the vibration samples comprising: deriving from the time series vector array a baseline asset health score, and deriving from the time series vector array a time series asset health score; subject at least a part of the time series asset health score to the machine learning model for determining at least one asset operating range; receive a further vibration sample at a monitor time t2 wherein the monitor time t2 is subsequent to the end time t1; derive from the further vibration sample an asset health score; determine if the asset health score falls within an operating range determined by the machine learning model, resulting in a monitor result; output a message to the RF module depending on the monitor result, and wherein the RF module uses the antenna to transmit the message, and wherein the system consumes less current than 1 A.

7. The system of claim 6, wherein the power source is selected from AC power using a DC adapter, one or more batteries, PV cells with suitable capacitors, energy harvesting systems with suitable capacitor, and a combination thereof.

8. The system of claim 6, wherein the system is fitted in a common housing.

9. The system of claim 8, wherein the system is operationally coupled to the vibrating object.

10. The system of claim 6, wherein furthermore the computer program product comprises a further machine learning model to predict future asset health scores, and wherein the computer program product when running on the data processor causes the computing device to: subject at least a part of the time series asset health score to the further machine learning model for predicting a future asset health score at a future time t4 subsequent to the end time t1; derive, at a predict time t3, the future asset health score from the further machine learning model wherein the predict time t3 is equal or subsequent to the end time t1 and wherein the predict time t3 is prior to the future time t4; output a message to the RF module depending on a deviation of the future asset health score from the baseline asset health score.

11. A method for monitoring a vibrating object with a system consuming less current than 150 μA in standby mode and consuming less current than 20 mA in active mode, comprising: waking up the system at a pre-determined event from the standby mode to the active mode, and while the system is in the active mode: receiving during a time interval, having an end time t1, vibration samples corresponding to the vibrating object, resulting in a time series vector array; analyzing the vibration samples comprising: deriving from the time series vector array a baseline asset health score, and deriving from the time series vector array a time series asset health score; subjecting at least a part of the time series asset health score to a machine learning model for predicting a future asset health score at a future time t4 subsequent to the end time t1; deriving, at a predict time t3, the future asset health score from the machine learning model wherein the predict time t3 is equal or subsequent to the end time t1, and wherein the predict time t3 is prior to the future time t4, and outputting a message depending on a deviation of the future asset health score from the baseline asset health score.

12. The method of claim 11, wherein the pre-determined event is a pre-determined interval.

13. The method of claim 11, wherein the duration of the time interval is longer than the duration of a period between the predict time t3 and the future time t4.

14. The method of claim 11, comprising furthermore while the system is in the active mode: subjecting at least a part of the time series asset health score to a further machine learning model for determining at least one asset operating range; receiving a further vibration sample at a monitor time t2 wherein the monitor time t2 is subsequent to the end time t1; deriving from the further vibration sample an asset health score; determining if the asset health score falls within an operating range determined by the further machine learning model, resulting in a monitor result, and outputting a message depending on the monitor result.

15. A non-transitory computer readable medium having stored thereon computer program instructions for monitoring a vibrating object that, when executed by a processor in a computing device consuming less current than 10 mA, configure the computing device to perform: receiving during a time interval, having an end time t1, vibration samples corresponding to the vibrating object, resulting in a time series vector array; analyzing the vibration samples comprising: deriving from the time series vector array a baseline asset health score, and deriving from the time series vector array a time series asset health score; subjecting at least a part of the time series asset health score to a machine learning model for predicting a future asset health score at a future time t4 subsequent to the end time t1; deriving, at a predict time t3, the future asset health score from the machine learning model wherein the predict time t3 is equal or subsequent to the end time t1 and wherein the predict time t3 is prior to the future time t4, and outputting a message depending on a deviation of the future asset health score from the baseline asset health score.

16. The non-transitory computer readable medium having stored thereon computer program instructions for monitoring a vibrating object of claim 15 that furthermore configure the computing device to perform: subjecting at least a part of the time series asset health score to a further machine learning model for determining at least one asset operating range; receiving a further vibration sample at a monitor time t2 wherein the monitor time t2 is subsequent to the end time t1; deriving from the further vibration sample an asset health score; determining if the asset health score falls within an operating range determined by the further machine learning model, resulting in a monitor result, and outputting a message depending on the monitor result.

Description

DRAWINGS

[0122] Embodiments of the current invention will now be described, by way of example only, with reference to the accompanying schematic drawings (which are not necessarily drawn to scale) in which corresponding reference symbols indicate corresponding parts, and in which:

[0123] FIG. 1A and FIG. 1B schematically depict a simplified diagram of an embodiment of a device configured to monitor a vibrating object.

[0124] FIG. 2 schematically depicts an embodiment of a device monitoring a wind turbine.

[0125] FIG. 3 schematically depicts a simplified diagram of an embodiment of a system configured to monitor a vibrating object.

[0126] FIG. 4 schematically depicts an embodiment of a system monitoring an engine with a connecting to a cloud server.

[0127] FIG. 5 schematically depicts a vibration spectrum analysis of a complex vibration over time in three frequency bands.

[0128] FIG. 6 depict a simplified flow chart of an example firmware stack.

[0129] FIG. 7A and FIG. 7B depict a simplified block diagram of a computer program product.

DESCRIPTION

[0130] FIG. 1A and FIG. 1B schematically depict a simplified diagram of an embodiment of a device 1 configured to monitor a vibrating object, comprising a common housing holding an accelerometer 2 for sampling vibration signatures (11 and 11′) of the vibrating object, resulting in vibration samples, and a computing device 3 comprising a data processor 4 and a memory 6 having stored thereon a computer program product 5 for monitoring the vibrating object. Computer program product 5 comprises an input module to receive vibration samples, and an analysis module to analyze vibration samples to derive asset health scores, and a machine learning model to determine asset operating ranges, and an output module to output messages 10.

[0131] In this embodiment, device 1 consumes less current than 20 mA.

[0132] In this embodiment device 1 receives vibration signatures (11 and 11′) from a vibrating object. Accelerometer 2 samples the vibration signatures (11 and 11′) of the vibrating object. Computer program product 5 runs on data processor 4 and causes the computing device 3 to receive during a time interval, having an end time t1, the vibration samples from accelerometer 2, resulting in a time series vector array, and to analyze the vibration samples, the analyzing comprises deriving from the time series vector array a baseline asset health score and deriving from the time series vector array a time series asset health score, and to subject at least a part of the time series asset health score to the machine learning model for determining at least one asset operating range, and to receive a further vibration sample at a monitor time t2 wherein the monitor time t2 is subsequent to the end time t1, and to derive from the further vibration sample an asset health score, and to determine if the asset health score falls within an operating range determined by the machine learning model, resulting in a monitor result, and to output a message 10 depending on the monitor result.

[0133] As illustrated in FIG. 1A, the device 1 receives vibration signatures 11 from a vibrating object, resulting in a vibration sample that has a health score falling within an operating range determined by the machine learning model and as a result device 1 does not send a message 10.

[0134] As illustrated in FIG. 1B, the device 1 receives vibration signatures 11′ from a vibrating object, resulting in a vibration sample that has a health score not falling within an operating range determined by the machine learning model and as a result device 1 does send a message 10.

[0135] In a further embodiment when device 1 receives vibration signatures 11 from a vibrating object, resulting in a vibration sample that has a health score falling within an operating range determined by the machine learning model and as a result device 1 does send a message 10, for example a message 10 indicating “everything is fine”.

[0136] In a further embodiment, device 1 comprises a power source.

[0137] In a further embodiment, the power source of device 1 is selected from one or more batteries, PV cells with suitable capacitors, energy harvesting systems with suitable capacitor, and a combination thereof. Battery types to be used may be rechargeable lithium polymer (LiPo), lithium ion or alkaline.

[0138] In a further embodiment, the power source of device 1 has a minimum surge/peak current support of 750 mA.

[0139] In a further embodiment, device 1 optimizes power consumption by actively partitioning processing between analog, digital and RF domains. As a result, the device consumes very little current (typically less than 100 μA in standby mode and 5 mA in active mode). This embodiment of device 1 therefore can support a wide variety of power sources.

[0140] In a further embodiment, device 1 consumes less current than 150 μA in standby mode and less current 20 mA in active mode.

[0141] FIG. 2 schematically depicts an embodiment of monitoring example 200 wherein device 1 is monitoring a wind turbine 21. Device 1 receives vibrations samples of wind turbine 21 and is able to monitor the health or condition of wind turbine 21 by inferring the results from the data derived from the vibration samples. As a result, device 1 can detect faults in the mechanism of the wind turbine, predict when failures are likely to occur, and when maintenance should be scheduled.

[0142] In an embodiment device 1 has a connection to a remote messaging system to update a dashboard displaying the health or condition of the wind turbine, and/or to notify when a fault is detected and repair or maintenance is needed.

[0143] In a further embodiment of device 1 inference results of device 1 are stored on a non-transitory computer readable medium (such as a SSD or HDD).

[0144] In further embodiments device 1 is monitoring one selected from a vibrating pump, a vibrating compressor, a vibrating turbine, a vibrating vehicle, a vibrating satellite, a vibrating motor, a vibrating fan, a vibrating pipe, a vibrating construction part, a vibrating steel sheet, a vibrating cover, a vibrating surface of the earth, and a combination thereof

[0145] FIG. 3 schematically depicts a simplified diagram of an embodiment of a system 31 configured to monitor a vibrating object comprising a power source 32, a power management circuit 33 to control flow and direction of electrical power from power source 32, a computing device 3 comprising a data processor 4 and a memory 6 having stored thereon a computer program product 5 for monitoring the vibrating object, an accelerometer 2 for sampling vibration signatures of the vibrating object resulting in vibration samples, a RF module 34, and an antenna 35. Computer program product 5 comprises an input module to receive vibration samples, and an analysis module to analyze vibration samples to derive asset health scores, and a machine learning model to determine asset operating ranges, and an output module to output messages 10.

[0146] In this embodiment, system 31 consumes less current than 20 mA. When system 31 receives vibration signatures from a vibrating object, accelerometer 2 samples vibration signatures of the vibrating object, resulting in vibration samples. Computer program product 5 runs on data processor 4 and causes computing device 3 to receive during a time interval, having an end time t1, the vibration samples from the accelerometer 2, resulting in a time series vector array, and to analyze the vibration samples, the analyzing comprises deriving from the time series vector array a baseline asset health score and deriving from the time series vector array a time series asset health score, and to subject at least a part of the time series asset health score to the machine learning model for determining at least one asset operating range, and to receive a further vibration sample at a monitor time t2 wherein the monitor time t2 is subsequent to the end time t1, and to derive from the further vibration sample an asset health score, and to determine if the asset health score falls within an operating range determined by the machine learning model, resulting in a monitor result, and to output a message 10 to RF module 34 depending on the monitor result, and wherein RF module 34 uses antenna 35 to transmit message 10.

[0147] In a further embodiment, the power source of system 31 is selected from AC power using a DC adapter, one or more batteries, PV cells with suitable capacitors, energy harvesting systems with suitable capacitor, and a combination thereof. Battery types to be used may be rechargeable lithium polymer (LiPo), lithium ion or alkaline.

[0148] In a further embodiment, the power source of system 31 has a minimum surge/peak current support of 750 mA.

[0149] In a further embodiment, system 31 optimizes power consumption by actively partitioning processing between analog, digital and RF domains. As a result, the system consumes very little current (typically less than 100 μA in standby mode and 5 mA in active mode). This embodiment of system 31 therefore can support a wide variety of power sources.

[0150] In a further embodiment, system 31 consumes less current than 150 μA in standby mode and less current than 20 mA in active mode.

[0151] In a further embodiment, system 31 is fitted in a common housing.

[0152] In a further embodiment, system 31 is operationally coupled to a vibrating object.

[0153] FIG. 4 schematically depicts an embodiment of a monitoring example 400 wherein system 31, with a battery 32, is monitoring an engine 41. System 31 has an uplink 48 to a cloud server 43, supported by a radio tower 42. Cloud server 43 processes messages 10 from system 31 received via uplink 48. Through various end-user computer devices (such as laptops, PCs and tablets), connected to cloud server 43, the messages 10 are distributed and presented as plain messages, graphs, statistics, KPIs, performance metrics, health metrics or a combinations thereof.

[0154] In a further embodiment system 31 can be configured and receive an update of computer program product 5 through downlink 49.

[0155] FIG. 5 schematically depicts a vibration spectrum analysis 500 of a complex vibration 51 over time 52 in three frequency bands (51′, 51″ and 51″′) and arrow 53 is referring to frequency. Frequency spectrum data is derived from the complex vibration 51 by using a Fast Fourier Transform (FFT) algorithm, resulting in a FFT array. The FFT algorithm is implemented as a software program which runs on the data processor 4 or 4′. The spectrum data is divided into three spectral energy bands which contain low, medium and high frequency components respectively. The bands (51′, 51″ and 51″′) are individually analyzed for changes and trends to determine anomalous behavior. The trend of the three vibration spectrum bands (51′, 51″ and 51″′) may indicate the possible root cause of the developing faults.

[0156] FIG. 6 depicts a simplified flow chart of an example firmware stack 600, wherein vibration signatures, collected from an object to be monitored (hereafter referred to as “asset”), are input 61 for Data Collection action 62. A microcontroller wakes up at pre-determined events (such as intervals triggered by a timer within the microcontroller or by an external timer/clock, or a sensor), an accelerometer 2 samples at least one vibration signature of an asset (such as embodiments 21 and 41). The at least one vibration signature for 3-axis (x, y and z) is collected. Timestamps are assigned to each sample. Samples are transferred to a microcontroller subsystem and stored in RAM memory 6. The first sample collected is stored in RAM memory 6 as the baseline value.

[0157] The 3-axis data of a sample is input for a Data Shaping and Filtering action 63 and measured against a pre-determined threshold to detect asset activity level. If the asset is on, then the data is filtered to remove noise floor and the filtered data is combined into a time series vector array resulting from vibration samples that correspond with a complex vibration 51. If the asset is off, the sample could be discarded.

[0158] The time series vector array is input for a Signal Analysis action 64 and a baseline asset health score (for example from 1 to 100) is derived from the time series vector array.

[0159] In addition, a time series asset health score (for examples with values from 1 to 100) is derived from the time series vector array using a Fast Fourier Transform (FFT) algorithm.

[0160] A time series vector array is converted into a spectrum array by extracting the individual frequency components of the time series data by using the FFT algorithm.

[0161] The spectrum array is further optimized by dividing the spectral energy into three frequency bands: low 51′, medium 51″ and high 51″′. The result of dividing the spectral energy are three compressed spectrum arrays, corresponding to the three bands (51′, 51″ and 51″′).

[0162] Three compressed spectrum arrays are individually analyzed for changes and trends to determine anomalous behavior.

[0163] A trend of the time series vector array indicates the overall health of the asset and the trend of the three frequency bands indicate the possible root cause of developing faults.

[0164] A time series asset health score is input for a Machine Learning and Al action 65 and: [0165] I. analyzed for trends and patterns using previous baseline values, and [0166] II. a FFT array is analyzed for spectral energy change by measuring frequency bin of highest magnitude and comparing with baseline values

[0167] In both analyses, baseline values are initially used to set the nominal operating level of the asset. During subsequent sampling periods, the measured values are used to determine the normal asset operating range of the asset and a ML model is trained dynamically using unsupervised learning, by: [0168] 1. Each data point after the baseline is set is used to further refine the range. If the data point is close to the baseline (for example +/−25%), then the moving average is calculated. Subsequent values are compared with the moving average. [0169] 2. When a sample falls outside the range, an anomaly alert is generated. A new asset operating range with the anomaly value as the baseline of the new range is created. Steps I. and II. are repeated until all normal asset operating ranges or operating modes are classified. [0170] 3. After every sample, the overall moving average is calculated using all observed values and is used for dynamic regression analysis. [0171] 4. A regression analysis output is projected 3 weeks into the future using slope extrapolation and is used for failure probability calculation.

[0172] A time series vector array is input for a Failure Analysis 66 and is used to detect overall vibration. Overall vibration values correspond to general condition of the asset. As the asset deteriorates, the overall vibration will increase.

[0173] In addition, a FFT array is input for a Failure Analysis action 66 and is used to detect root-cause factors. Change in spectral energy of low frequency spectrum corresponds to mechanical/structural issues with the asset and change in high frequency spectrum corresponds to bearing-related issues.

[0174] Time series trends and FFT trends are input for a Predictive Analytics action 67 and are combined to determine failure probability by comparing the present values, change trajectory and projected moving average values resulting from signal analysis.

[0175] The result of the failure analysis and predictive analyses is input for a Security action 68 and is converted into an encrypted message as output 69.

[0176] Although listed in a sequential order, these actions for monitoring an asset may in some instances be performed in parallel. Also, the various actions may be combined into fewer actions, divided into additional actions, and/or removed based upon the desired implementation.

[0177] For example, a machine learning is also described by these steps: [0178] 1. An asset health score (for example 0-100) derived from the first time series vector array forms a baseline asset health score. [0179] 2. A time series asset health score derived from a baseline asset health score is used to train a first ML model. [0180] 3. The first ML model is used to predict the asset health score into the future. [0181] 4. If the predicted asset health score is above a pre-determined threshold (>75), then an anomaly is generated and the probability of failure is calculated. [0182] 5. A second ML model for determining asset operating ranges, resulting from step 6 to 8. [0183] 6. The baseline asset health score (from step 1) is also used to create an asset operating range (for example: mode 0). [0184] 7. The time series asset health score (from step 2) is compared with the baseline asset health score and if it is close to the baseline (within 25%), then the asset is considered to still be in mode 0. [0185] 8. If the asset health score deviates from the baseline (>25%), then an additional asset operating range is created (for example: mode 1, 2, 3, . . . ) and this step is repeated for subsequent asset health scores. [0186] 9. Anomalies are generated when a new operating mode is created and when any outliers are detected.

[0187] In further examples a machine learning model uses a trend indicator, in particularly moving average, for trend analysis.

[0188] FIG. 7A and FIG. 7B depict a simplified block diagram of a computer program product 5 and 5′.

[0189] As illustrated in FIG. 7A an embodiment of computer program product 5 comprises: [0190] an input module 71 to receive the vibration samples; [0191] an analysis module 72 to analyze the vibration samples to derive asset health scores; [0192] a machine learning model 73 to determine asset operating ranges, and [0193] an output module 74 to output messages,
wherein the computer program product 5 when running on the data processor 4 causes the computing device to: [0194] receive during a time interval, having an end time t1, the vibration samples from the accelerometer 2, resulting in a time series vector array; [0195] analyze the vibration samples comprising: [0196] deriving from the time series vector array a baseline asset health score, and [0197] deriving from the time series vector array a time series asset health score; [0198] subject at least a part of the time series asset health score to the machine learning model 73 for determining at least one asset operating range; [0199] receive a further vibration sample at a monitor time t2 wherein the monitor time t2 is subsequent to the end time t1; [0200] derive from the further vibration sample an asset health score; [0201] determine if the asset health score falls within an operating range determined by the machine learning model 73, resulting in a monitor result, and [0202] output a message 10 depending on the monitor result.

[0203] As illustrated in FIG. 7B a further embodiment of computer program product 5′ furthermore comprises a further machine learning model 73′ to predict future asset health scores, and wherein the computer program product 5′ when running on the data processor 4 causes the computing device to: [0204] subject at least a part of the time series asset health score to the further machine learning model 73′ for predicting a future asset health score at a future time t4 subsequent to the end time t1; [0205] derive, at a predict time t3, the future asset health score from the further machine learning model 73′ wherein the predict time t3 is equal or subsequent to the end time t1 and wherein the predict time t3 is prior to the future time t4; [0206] output a message 10 depending on a deviation of the future asset health score from the baseline asset health score.

[0207] It will also be clear that the above description and drawings are included to illustrate some embodiments of the current invention, and not to limit the scope of protection. Starting from this disclosure, many more embodiments will be evident to a skilled person. These embodiments are within the scope of protection and the essence of this invention and are obvious combinations of prior art techniques and the disclosure of this patent. The terms “coupled,” “attached,” or “connected” may be used herein to refer to any type of relationship, direct or indirect, between the components in question, and may apply to electrical, mechanical, fluid, optical, electromagnetic, electromechanical or other connections. In addition, the terms “first,” “second,” etc. are used herein only to facilitate discussion, and carry no particular temporal or chronological significance unless otherwise indicated.

[0208] Those skilled in the art will appreciate from the foregoing description that the broad techniques of the embodiments can be implemented in a variety of forms. Therefore, while the embodiments have been described in connection with particular examples thereof, the true scope of the embodiments should not be so limited since other modifications will become apparent to the skilled practitioner upon a study of the drawings, specification, and following claims.