System and method for induction machine eccentricity severity estimation using sparsity-driven regression models

Abstract

A fault detection system for eccentricity severity of an induction machine including a rotor and a stator is provided. The fault detection system includes a sensor interface configured to acquire sensor signals from sensors disposed at predetermined locations on the induction machine, the sensor signals indicating a rotor eccentricity level of the induction machine. The fault detection system further includes a memory coupled to a processor. The memory stores a training data set and instructions implementing a learning-based fault detection method for the induction machine. The instructions include generating an eccentricity feature matrix from the sensor signals, the sensor signals including load torque, rotor speed, rotor vibration acceleration, rotor vibration velocity, and a stator or induction machine current spectrum. The instructions further include determining an eccentricity level of the induction machine based on the eccentricity feature matrix using the learning-based fault detection method, the learning-based fault detection method being configured to find the eccentricity level from the learning-based eccentricity feature matrix dataset.

Description

[Technical Field]

[0001] The present invention relates to a system for estimating eccentricity severity, and more particularly to machine learning and artificial intelligence adapted to estimating the level of eccentricity severity of induction motors. [Background technology]

[0002] Eccentricity is a common problem in induction machines or other types of electric motors, which can be caused by imperfections in the manufacturing process or long-term operation. When eccentricity exists, the air gap between the stator and rotor is not evenly distributed, causing fluctuating torque and undesirable vibrations. Worse, it can lead to insulation damage or sudden failure during operation. Therefore, detecting eccentricity and estimating the severity of eccentricity are very important for preventive maintenance.

[0003] Over the past few decades, electric motor eccentricity detection has attracted considerable attention in the field of electric motor fault detection. The most commonly used invasive method for eccentricity diagnosis is motor current signature analysis (MCSA), which aims to detect characteristic frequency components in the spectrum associated with a certain type of eccentricity. In addition to MCSA, instantaneous active and reactive power spectra have also been utilized for motor eccentricity detection, where the ratio between the amplitude of the rotational speed-dependent characteristic component and the DC component is defined as the fault signature. Other methods, such as magnetic field-based eccentricity detection, which aims to examine the magnitude of characteristic harmonics by analyzing the spectrum of stray magnetic flux, have also been explored.

[0004] These methods are primarily used to detect the presence of eccentricity and are inaccurate or impractical for estimating the severity of the eccentricity. However, knowing the severity of the eccentricity can be beneficial for cost-effective maintenance of induction motors and their control.

[0005] Therefore, there is a need to estimate the level of severity of the eccentricity of the induction motor. Summary of the Invention

[0006] An object of certain embodiments is to provide a system and method suitable for estimating not only the presence or absence of eccentricity in an induction motor, but also the level of severity of the eccentricity. Additionally or alternatively, some embodiments are intended to estimate different levels of eccentricity from motor current spectra and / or other measurements indicative of the state of operation of the induction motor. Additionally or alternatively, some embodiments are intended to estimate different levels of eccentricity in an induction motor under different load conditions.

[0007] Some embodiments are based on the understanding that estimating different levels of eccentricity in an induction motor requires a model that links measurements of the motor's operating conditions to the level of eccentricity. Such a model can be derived based on an analysis of the motor's dynamic characteristics. However, such a model would suffer from uncertainties in both the quality of the measurements of different operating parameters and the relationships between those parameters. For example, if the model relates measurements of different harmonics of a current to a particular load driven by the motor, the relationship between the measurements of different harmonics for that current and other loads is at least partially unknown. In addition, the measurements themselves are susceptible to noise. These may not be an issue when the goal is to determine the presence or absence of eccentricity. However, these uncertainties can make the severity level estimation inaccurate or even impractical.

[0008] Some embodiments are based on the recognition that these inaccuracy problems are due, at least in part, to unknown relationships of noise in the measurements of the operating parameters. Indeed, when the measurements of the parameters of the operation of the electric motor are not precisely known and their impact on the severity of eccentricity is uncertain, the noise of different measurements and their relationships are typically added together as one aggregate noise affecting the model. However, this approach leads to amplification of the individual noise of each measurement due to the uncertainty of the relationships between the measurements, leading to a decrease in the signal-to-noise ratio (SNR) in the estimation of the severity level of eccentricity.

[0009] Some embodiments recognize that the problem of noise accumulation can be addressed by determining in advance the structure of a data-driven model and learning certain coefficients of that structure from data. For example, some embodiments determine a model of eccentricity severity as a weighted combination of measurements of operating parameters of an induction motor, with the weights in the weighted combination learned from training data using machine learning. Such a model structure is advantageous because it handles the uncertainty of each measurement separately. Indeed, the weights learned using machine learning may incorporate not only the contribution of each measurement to the severity level estimation, but also the uncertainty of that measurement.

[0010] Additionally, fixing the model's structure as a weighted combination limits the nonlinearity of interrelationships between different measurements. However, the potential mutual effects of different measurements on each other remain uncertain. To address this issue, some embodiments manipulate the learning process toward identifying physically consistent solutions according to machine learning principles that draw information from physical properties. For example, some embodiments use machine learning subject to sparsity constraints, since the model's structure allows each measurement to be considered individually. By doing so, only measurements with dominant relationships to eccentricity are kept, thereby reducing the number of uncertain interrelationships between different measurements while preserving the statistical distribution of models trained using machine learning.

[0011] For example, assume that an eccentricity estimation system receives measurements of 10 different operating parameters of an induction motor. If the weights of a weighted combination of a model relating eccentricity severity to the measurements, according to some embodiments, are learned using machine learning without sparsity constraints, the level of eccentricity would be determined as a weighted combination of the 10 measurements. However, if a sparsity constraint is implemented during training, perhaps only three weights will have non-zero values. As a result, noisy contributions of different measurements can be reduced and / or avoided.

[0012] The benefits of using machine learning with a sparsity constraint can be further analyzed from the perspective of the machine learning itself. When faced with the task of learning weights for a proposed model, even with unlimited labeled training data (which is rarely the case), machine learning may learn weights that are not completely accurate enough to estimate severity levels with the desired accuracy. This is because the training data for machine learning differs from the input data measured during the operation of the induction motor. Therefore, machine learning relies on the expectation that the statistical distribution of measurements used to learn the weights will resemble the statistical distribution of measurements collected during the operation of the motor. Therefore, the measurements used for machine learning and severity estimation are important because this statistical relationship should be stable. While this statistical relationship is generally unknown in advance, implementing a sparsity constraint during machine learning makes it possible to eliminate the noisy contributions of different measurements. These noisy contributions typically disrupt the statistical relationship between the training data and the testing data. Therefore, removing them from the model improves the accuracy of severity estimation. As a result, in some embodiments, a single set of weights can be used to determine different eccentricity levels.

[0013] Additionally, evaluation of the model during operation of the electric motor is simplified because, in this example, only three measurements contribute to the model, and ten different measurements are considered to form weights for combining these three measurements. Furthermore, this principle can be used as a test to verify whether weights are learned in machine learning that are subject to sparsity constraints. This is because three weights learned with a sparsity constraint from training data containing measurements of ten parameters will differ from weights learned without a sparsity constraint from the corresponding measurements of three parameters.

[0014] According to some embodiments of the present invention, there is provided a fault detection system for estimating a severity of eccentricity of an induction machine including a rotor and a stator. The fault detection system may include a sensor interface configured to acquire sensor signals from sensors disposed at predetermined locations on the induction machine, the sensor signals indicating an eccentricity level of the rotor of the induction machine, and may further include a memory coupled to a processor, the memory storing instructions implementing a learning-based fault detection method for the induction machine, the instructions, when executed by the processor, performing the steps of generating an eccentricity feature matrix by extracting the sensor signals via the sensor interface, the sensor signals including at least one of a load torque, a rotor speed, a rotor vibration acceleration, a rotor vibration velocity, and a stator current spectrum, and the instructions, when executed by the processor, further performing the steps of determining an eccentricity level of the induction machine based on the eccentricity feature matrix using the learning-based fault detection method, the learning-based fault detection method being trained to find the eccentricity level from the learning-based eccentricity feature matrix dataset.

[0015] Further, some embodiments may provide an apparatus for estimating a severity level of an eccentricity of an induction motor, the apparatus may include an input interface configured to accept, via a network, values of a set of parameters of an operation state of the induction motor at different time steps, a memory configured to store a set of weights learned for the set of parameters of the operation of the induction motor from training data using machine learning subject to a sparsity constraint, and a processor configured to determine a severity level as a weighted combination of the values of the set of parameters accepted at a time step and retrieved from the memory with corresponding weights, wherein the processor uses the same weights for the sets of parameters accepted at different time steps, and the apparatus may further include an output interface configured to render the determined severity level.

[0016] Some embodiments may further provide an artificial intelligence (AI) training system for learning the weights stored in the memory of the device described above. The AI training system may include a training processor and a training memory having stored thereon instructions to cause the training processor to collect training data indicating measurements of a plurality of parameters, including a set of parameters paired with labeled values of severity levels, train a set of weights for a weighted combination of the plurality of parameters that reduces a loss function comprising the difference between the severity level estimated using the current weights and the labeled value subject to a sparsity constraint, and submit the weights to the device described in claim 1 via one or a combination of wired and wireless communication channels.

[0017] The accompanying drawings are included to provide a further understanding of the invention, illustrate embodiments of the invention, and together with the detailed description, serve to explain the principles of the invention. The drawings shown are not necessarily to scale, emphasis instead generally being placed upon illustrating the principles of embodiments of the present disclosure. [Brief explanation of the drawings]

[0018] [Figure 1A] 1 is a schematic diagram of a fault severity detection system for controlling and monitoring an induction motor according to one embodiment of the present invention; [Figure 1B] 1 is a schematic diagram of a fault severity detection system operated by a user (operator) over a network for controlling and monitoring an induction motor according to one embodiment of the present invention; [Figure 2A] FIG. 10 is a diagram of eccentricity severity levels indicated by relative air gap, according to one embodiment of the present invention. [Figure 2B] FIG. 10 is a diagram of eccentricity severity levels indicated by relative air gap, according to one embodiment of the present invention. [Figure 2C] FIG. 10 is a diagram of eccentricity severity levels indicated by relative air gap, according to one embodiment of the present invention. [Figure 3] 1 is an exemplary diagram and illustration of an induction motor control and monitoring system. [Figure 4A] 1A-1C are exemplary plots of time domain stator current and corresponding frequency spectrum, respectively. [Figure 4B] 1A-1C are exemplary plots of time domain stator current and corresponding frequency spectrum, respectively. [Figure 5A] 1A-1C are exemplary plots of vibration acceleration and corresponding vibration velocity, respectively. [Figure 5B] 1A-1C are exemplary plots of vibration acceleration and corresponding vibration velocity, respectively. [Figure 6A] 1 is an exemplary plot of the 30 Hz frequency component at different loads. [Figure 6B] 1 is an exemplary plot of the 90 Hz frequency component at different loads. [Figure 7] 1 is an exemplary plot of a correlation matrix of features used in training a regression model. [Figure 8]FIG. 1 is a block diagram of a method for training a learning-based eccentricity detection method (model) in an induction motor, according to one embodiment of the present invention. [Figure 9] FIG. 6 illustrates an algorithm of a method for training a regression model of an induction motor according to another embodiment of the present invention. [Figure 10] 10 is an exemplary plot of feature weights after a trained regression model. [Figure 11] 1 is an exemplary plot of predicted eccentricity severity levels compared to true eccentricity levels. DETAILED DESCRIPTION OF THE INVENTION

[0019] Various embodiments of the present invention will now be described with reference to the drawings. It should be noted that the drawings are not drawn to scale, and that elements of similar structure or function are represented by similar reference numerals throughout the drawings. It should also be noted that the drawings are intended only to facilitate the description of particular embodiments of the present invention. They are not intended as an exhaustive description of the invention or as limitations on the scope of the invention. Furthermore, aspects described in connection with a particular embodiment of the present invention are not necessarily limited to that embodiment and may be implemented in any other embodiment of the present invention.

[0020] FIG. 1A is a schematic diagram of a fault severity detection system 100 for controlling and monitoring an induction motor, according to one embodiment of the present invention.

[0021] Induction motor (system) 200 includes a rotor assembly 102, a stator assembly 104, a main shaft 106, and two main bearings 108. In this example, induction motor 200 is a squirrel-cage induction motor.

[0022] Controller 220 is powered by power supply 230 and may be used to monitor and control the operation of induction motor 200 in response to various inputs in accordance with embodiments of the present invention. For example, controller 220 coupled to induction motor 200 may control the speed of the induction motor based on input received from fault detection system 100, which is configured to obtain data regarding the operating conditions of induction motor 200 from sensors 150. According to certain embodiments, the electrical signals of sensors 150 may be current and voltage sensors for obtaining current and voltage data related to induction motor 200. For example, the current sensors sense current data from one or more of the induction motor's phases. More specifically, if the induction motor is a three-phase induction motor, the current and voltage sensors sense current and voltage data from the three phases of the three-phase induction motor. While certain embodiments of the present invention are described with respect to a polyphase induction motor, other embodiments of the present invention may be applicable to other polyphase electromechanical machines.

[0023] Some embodiments of the present invention describe a system for fault detection in an electric machine, such as an induction motor 200. The system configured for detection includes a fault detection module 100 for detecting the presence and severity of a rotor 102 fault condition, including an eccentricity fault, within an induction motor assembly. In one embodiment, the fault detection module 100 is implemented as a subsystem of a controller 220. In an alternative embodiment, the fault detection module 100 is implemented using a separate processor. The fault detection module 100 may be a hardware circuit module operably connected to the controller 220. In some implementations, the fault detection module 100 and the controller 220 may share information. For example, the fault detection module 100 may reuse sensor data used by the controller to control the operation of the induction motor.

[0024] Fault detection module 100 further includes a processor 110, a memory 120, and a fault detection program stored in memory 120 when instructions thereof are executed by processor 110. Module 100 further includes a sensor interface 130 configured to acquire signals from sensors 150. Interface 130 includes an A / D (analog-to-digital) and A / D (analog-to-digital) converter in data communication with processor 110, memory 120, the fault detection program, user interface 140, and sensors 150. Processor 110 may be multiple processors, and memory 120 may be a memory module including multiple memories. User interface 140 is configured to connect to a keyboard and display unit configured to display normal / fault status information of induction motor 200 in response to an output of fault detection module 200.

[0025] If induction motor 200 has an eccentricity fault, the torque of the rotor fluctuates, causing induction motor 200 to vibrate.

[0026] In some cases, some embodiments of the present invention provide an apparatus for estimating a severity level of an eccentricity of an induction motor, the apparatus including an input interface configured to accept, via a network, values of a set of parameters of an operation state of the induction motor in the form of a feature matrix at different time steps, and a memory configured to store a set of weights learned for the set of parameters of the operation of the induction motor from training data using machine learning subject to sparsity constraints. The apparatus also includes a processor configured to determine the severity level as a weighted combination of the values of the set of parameters accepted at a time step with corresponding weights retrieved from the memory, the processor using the same weights for sets of parameters accepted at different time steps, and an output interface configured to render the determined severity level.

[0027] FIG. 1B is a schematic diagram of a fault severity detection system for controlling and monitoring an induction motor operated by a user (operator) via a network 250 according to one embodiment of the present invention. In this case, fault detection module 100 may be included in an operating system for the induction motor located at the user / operator's location. When fault detection module 100 determines that a fault has occurred based on sensor signals from sensors 150 via a network (communication network) 250 during operation of induction machine (induction motor) 200, fault detection module 100 may transmit a control signal via network 250 to slow down or stop driving of induction motor 200 based on a pre-programmed algorithm (not shown) stored in memory 120. In some cases, sensor 150 may be calibrated based on a sensor calibration program (not shown) stored in the memory of fault detection module 100. In some cases, network 250 may be an optical fiber network, a wireless network, an Internet network, or a data communication network consisting of a combination of at least two of an optical fiber network, a wireless network, and an Internet network.

[0028] This configuration may be a maintenance system managed by a user / client that operates induction machine system 20 remotely from the induction machine site. For example, this system configuration may be used in a user-operated power generation system or a train system that controls an induction motor that drives a train. In some cases, when fault detection module 100 detects that the induction machine location is remote from the fault detection system location, data communication between the induction machine location and the fault detection system location occurs over a network. Network 250 may be a data communication network consisting of an optical fiber network, a wireless network, an Internet network, or a combination of at least two of an optical fiber network, a wireless network, and an Internet network.

[0029] Additionally, a fault detection system is included as part of the user's maintenance system, and when the determined induction machine eccentricity level is equal to or exceeds a critical threshold level, the fault detection system sends a control signal to the induction machine controller over the network using the sensor interface / control interface 130 to stop the induction machine from operating.

[0030] In one embodiment of the present invention, the current and voltage sensors respectively detect stator current data from the stator assembly 104 of the induction motor 200. The current and voltage data obtained from the sensors is communicated to the inverter and / or the fault detection module for further processing and analysis. The analysis includes performing a motor current signature analysis (MCSA) to detect a fault in the induction motor 200. In some embodiments, upon detecting a fault by using the fault detection module 100, the controller 220 receives a fault detection signal via the interface 130 of the fault detection module 100, which stops operation of the induction motor 200 by sending a signal to the controller 110 to interrupt the stator current of the induction motor 200 for further inspection or repair. In some cases, the sensor 150 may include a controller interface (not shown) configured to receive the fault detection signal from the interface 130 and send a fault status signal to the controller 220 so that the controller 220 interrupts the stator current of the induction motor 200 to stop operation of the induction motor 200. If the sensor 150 does not include a controller interface, the interface 130 may be configured to connect to the controller 220, and the controller, in response to a fault detection signal from the fault detection module 100 via the interface 130, interrupts the stator current of the induction motor 200 to stop operation of the induction motor 200.

[0031] The system also includes a memory for storing signal measurements and various parameters and coefficients for implementing the fault severity detection method.

[0032] Some embodiments of the present invention are based on the recognition that motor eccentricity detection is one of the important techniques for motor fault detection. The most used invasive method for eccentricity diagnosis is Motor Current Signature Analysis (MCSA), which aims to detect characteristic frequency components for a particular type of eccentricity in the frequency spectrum.

[0033]

number

[0034]

number

[0035] Other methods, such as magnetic field-based eccentricity detection, which aims to determine the magnitude of characteristic harmonics by analyzing the spectrum of stray magnetic flux, have also been investigated. However, these have not been widely accepted due to the high cost of installing sensors. In contrast to the binary eccentricity detection problem, eccentricity severity estimation is more challenging due to its complexity and the influence of operating conditions. Current spectrum-based indicators have been proposed to qualitatively assess eccentricity levels, but there are no clear, standardized standards for quantitative estimation, especially under fluctuating operating conditions.

[0036] 2A, 2B, and 2C are diagrams of eccentricity severity levels indicated by relative air gaps, according to one embodiment of the present invention.

[0037] Based on our physical model of induction machines and fault detection methods using different features, we aim to estimate the eccentricity severity of induction machines by a learning-based method that incorporates different eccentricity-related features.

[0038] We assume that N experiments were conducted under different eccentricity levels and various load conditions. For each experiment, we obtained multiple measurement time series, including eccentricity levels, torque load conditions, and rotation speed, vibration acceleration, and three-phase current.

number

[0039] These characteristic values in X can be called a set of parameters of the state of operation of the induction motor.

[0040]

number

[0041]

number

[0042]

number

[0043] We then iteratively update w, z, and μ using the Alternating Direction of Multipliers Method (ADMM). The detailed update process is summarized in Figure 8.

[0044]

number

[0045] Figure 3 illustrates an exemplary setup for estimating the eccentricity severity level of an induction machine. To generate different eccentricity levels, the two original bearings 108 between the rotor 102 and stator 104 are removed. Instead, two larger external bearings are used to support the rotor so that the static eccentricity level of the motor can be manually adjusted with high precision within a certain range. A magnetic particle brake with adjustable torque by varying the input operating current is used as the load 210. The entire motor drive system is enclosed in a transparent cage for safety purposes. During operation, multiple sensors 150 are used to record synchronized time-series data: i) one tachometer 150 measures the rotational speed; ii) two accelerometers 150 record motor vibrations along the horizontal and vertical directions, respectively; and iii) three current probes 150 connected to the terminals of the stator 104 record the corresponding three-phase stator currents.

[0046] Experiments are conducted under various conditions of eccentricity level and load by adjusting the input operating current of the external bearing and the magnetic powder brake. For each experiment, given a pair of eccentricity level and load, we follow three steps: i) shift the bearing supporting the rotor to the eccentricity level under stationary conditions, ii) set the input current of the magnetic powder brake to provide the desired load torque, and iii) start the motor and record data while the motor is operating at steady state.

[0047]

number

[0048] To investigate the relationship between the motor operating characteristics and the eccentricity level, we preprocess the raw measurements to fit them as inputs to our regression model. For each experiment, we recorded time series of torque, rotational speed, horizontal acceleration, vertical acceleration, and three-phase current for 60 seconds each, with a sampling rate of 104 The training and testing datasets are collected at Sa / s. To enhance the training and testing datasets, we first divided each time series of the original 60-second measurements into 12 non-overlapping segments of 5 seconds each, resulting in a total of N=480 datasets for all 40 experiments. Each dataset includes load torque, rotational speed, acceleration time series, and three-phase stator current sequences. Then, we randomly select half of the 480 datasets for training and the other half for testing. Data features for each dataset are extracted as described in detail below.

[0049] The average vibration velocity is calculated as a feature according to the following three steps: Vibration acceleration time series A x and A y are independently integrated to obtain the raw vibration velocity time series. · The speed trend due to the accumulated error is calculated by the moving average method, where the window size is set as 10 samples. Calculate the mean absolute value of the net vibration velocity based on the horizontal and vertical detrended vibration velocities.

[0050] 4A and 4B show exemplary plots of vibration acceleration and vibration velocity, respectively.

[0051]

number

[0052]

number

[0053]

number

[0054] 5A and 5B show exemplary plots of collected time-domain stator current data and its corresponding frequency spectrum, illustrating the range of rotor frequency harmonics with various magnitudes.

[0055] To further investigate the magnitude of harmonics, Figures 6A and 6B plot the magnitude of 30 Hz and 90 Hz components in the current spectrum against the 60 Hz operating frequency component for different load conditions. We can observe that the magnitude can vary significantly with load, especially when the eccentricity level is relatively low. Therefore, estimating the eccentricity severity level based solely on the magnitude of 30 Hz or 90 Hz is unreliable. In order to develop an eccentricity severity estimation method, it is essential to properly incorporate multiple features.

[0056] As a result, the load torque, rotor speed, vibration acceleration, vibration velocity, and current spectrum features are provided for further model training and testing.

number

[0057] All features are normalized to have mean 0 and variance 1 to ensure that all features are weighted equally without prior knowledge. An example plot of the feature correlation matrix is shown in Figure 7.

[0058] FIG. 8 illustrates a block diagram of a method for training a learning-based eccentricity detection method (fault severity detection method) in an induction motor, according to one embodiment of the present invention.

[0059]

number

[0060]

number

[0061] A plot of the sparse weights learned from our training data is shown in Figure 10. We found that the 90 Hz frequency component plays a dominant role in severity level estimation. In addition to the 90 Hz frequency component, other features such as vibration and some higher harmonics also contribute to the final estimation. This is consistent with the literature that uses vibration and higher harmonics for eccentricity detection.

[0062] The estimation results for the test dataset using the trained model and the true eccentricity severity levels are shown in Figure 11. We observe that our estimates of eccentricity levels fit well with the true eccentricity settings across all different loading conditions.

[0063]

number

[0064] Based on the above-described training procedure, some embodiments may provide an artificial intelligence (AI) training system for learning weights stored in a memory of the above-described apparatus, wherein the apparatus is provided for estimating a severity level of an eccentricity of an induction motor, and includes: an input interface configured to accept, via a network, values of a set of parameters of an operation state of the induction motor at different time steps; a memory configured to store a set of weights learned for the set of parameters of the operation of the induction motor from training data using machine learning subject to a sparsity constraint; and a processor configured to determine a severity level as a weighted combination of the values of the set of parameters accepted at a time step and retrieved from the memory, the weighted combination using corresponding weights, wherein the processor uses the same weights for the sets of parameters accepted at different time steps; and an output interface configured to render the determined severity level.

[0065] The AI training system is configured to include a training processor and a training memory, wherein the training memory stores instructions that cause the training processor to collect training data indicating measurements of a plurality of parameters including a set of parameters paired with labeled values of severity levels, train a set of weights for a weighted combination of the plurality of parameters that reduces a loss function that includes the difference between the severity level estimated using the current weights and the labeled value subject to a sparsity constraint, and submit the weights to a device via one or a combination of wired and wireless communication channels.

[0066] The above-described embodiments of the present invention can be implemented in any of numerous ways. For example, embodiments may be implemented using hardware, software, or a combination thereof. When implemented in software, the software code may be executed on any suitable processor or collection of processors, whether provided on a single computer or distributed among multiple computers. Such a processor may be implemented as an integrated circuit, with one or more processors being in an integrated circuit component. However, a processor may be implemented using circuitry in any suitable format.

[0067] Also, embodiments of the present invention may be embodied as methods for which examples have been provided. The acts performed as part of the method may be ordered in any suitable manner. Thus, embodiments may be constructed in which acts are performed in an order different from that illustrated, which may include performing some acts shown as sequential acts in the exemplary embodiment simultaneously.

[0068] The use of ordinal language such as "first," "second," etc. to modify a claim element in a claim does not, by itself, imply a priority, precedence, or ordering of one claim element over another claim element, or a chronological order in which method actions are performed, but is merely used as a label to distinguish one claim element having a certain name from another element having the same name (but for which ordinal language is used) to distinguish the claim elements.

[0069] Although the invention has been described by way of examples of preferred embodiments, it is to be understood that various other adaptations and modifications can be made within the spirit and scope of the invention.

[0070] Therefore, it is the object of the appended claims to cover all such variations and modifications as come within the true spirit and scope of the invention.

Claims

1. 1. An eccentricity severity fault detection system for an induction machine including a rotor and a stator, comprising: a sensor interface configured to acquire a sensor signal from a sensor disposed at a predetermined position of the induction machine, the sensor signal indicating an eccentricity level of a rotor of the induction machine, the fault detection system further comprising: a memory coupled to a processor, the memory storing instructions that, when executed by the processor, implement a learning-based fault detection method for the induction machine, the instructions comprising: performing a step of generating an eccentricity feature matrix by extracting the sensor signals via the sensor interface, the sensor signals including at least one of a load torque, a rotor speed, a vibration acceleration of the rotor, a vibration speed of the rotor, and a current spectrum of the stator; and the instructions, when executed by the processor, further comprising: a fault detection system performing the step of determining the eccentricity level of the induction machine based on the eccentricity feature matrix using the learning-based fault detection method, wherein the learning-based fault detection method is trained to find the eccentricity level from a learning-based eccentricity feature matrix dataset.

2. 2. The fault detection system of claim 1, wherein the eccentricity level is defined as an operating condition of the induction machine, the operating conditions including a load applied to the induction machine, a rotational speed of the rotor, vibration, and a current spectrum characteristic of the stator.

3. The fault detection system of claim 1 , wherein the sensor interface repeatedly acquires the operational signal over a predetermined period of time.

4. The fault detection system of claim 1 , wherein the frequency region of the current spectrum characteristic associated with eccentricity is determined at about half an operating frequency of the induction machine and about one-and-a-half of the operating frequency.

5. The fault detection system of claim 4 , wherein the current spectrum corresponds to a frequency component of 90 Hz.

6. The obstacle detection system of claim 1 , wherein the sensors include a gap sensor and an acceleration sensor.

7. The fault detection system of claim 1 , wherein all elements of the eccentricity feature matrix are normalized to have a mean and unit variance.

8. 2. The fault detection system of claim 1, wherein a location of the induction machine is separated from a location of the fault detection system, and data communication between the location of the induction machine and the location of the fault detection system is via a network, the network being a data communication network consisting of an optical fiber network, a wireless network, an Internet network, or a combination of at least two of the optical fiber network, the wireless network, and the Internet network.

9. The fault detection system of claim 8 , wherein the fault detection system is included as part of a user's maintenance system.

10. 2. The fault detection system of claim 1, wherein when the determined eccentricity level of the induction machine is equal to or greater than a critical threshold level, the fault detection system sends a control signal over a network to a controller of the induction machine to stop operating the induction machine.

11. 1. An apparatus for estimating a severity level of eccentricity of an induction motor, comprising: an input interface configured to accept, via a network, values of a set of parameters of the operation state of the induction motor at different time steps in the form of a feature matrix; a memory configured to store a set of weights learned for the set of parameters of the operation of the induction motor from training data using machine learning subject to sparsity constraints; a processor configured to determine the severity level as a weighted combination of the values of the set of parameters received at a time step and weighted with corresponding weights retrieved from the memory, wherein the processor uses the same weights for the sets of parameters received at the different time steps, and the apparatus further comprises: An apparatus comprising an output interface configured to render the determined severity level.

12. The apparatus of claim 11 , wherein the weights for the parameters are determined from the training data that includes an additional parameter with a learned weight having a zero value.

13. 12. The apparatus of claim 11 , wherein the set of parameters of operation of the induction motor includes one or a combination of: torque generated by the induction motor to move a load, a rotor speed of the induction motor, a rotor vibration acceleration, a rotor vibration velocity, and a stator current spectrum of the induction motor.

14. 14. The apparatus of claim 13, wherein the set of parameters for the operation of the induction motor includes over 100 different parameters.

15. 12. The apparatus of claim 11, wherein the set of parameters for the operation of the induction motor includes a parameter indicative of a load carried by the induction motor.

16. 16. The apparatus of claim 15, wherein the parameter indicative of the load is a measure of torque of the induction motor.

17. 12. An artificial intelligence (AI) training system for learning the weights stored in the memory of the device of claim 11, comprising a training processor and a training memory having instructions stored therein, the instructions causing the training processor to: collecting training data indicative of measurements of a plurality of parameters including the set of parameters paired with labeled values of the severity levels; training the set of weights for the weighted combination of the plurality of parameters to reduce a loss function comprising the difference between a severity level estimated using current weights and the labeled value subject to the sparsity constraint; 10. An artificial intelligence (AI) training system that submits the weights to the device of claim 1 via one or a combination of wired and wireless communication channels.

18. 20. The AI of claim 17, wherein to train the weights, the training processor is configured to solve a regularization problem formulated as an optimization problem including a regularizer that enforces the sparsity constraint.

19. The training processor further comprises: configured to form an eccentricity feature matrix from the training data; 20. The AI of claim 18, configured to find a single set of weights that matches a weighted combination of different elements of the eccentricity feature matrix to corresponding labeled values of the severity levels.

20. The training processor further comprises: configured to receive an indicator of an unmeasured parameter from the plurality of parameters; 18. The AI of claim 17, configured to train the set of weights for the weighted combinations of the plurality of parameters while enforcing weights for the non-measured parameters to zero.

Images