Deep Hybrid Convolutional Neural Networks for Wind Turbine Gearbox Fault Diagnosis

Abstract

To provide a fault diagnosis system for a wind turbine.SOLUTION: During operation, a system collects current signals associated with a physical object including a rotating machine. The system demodulates the collected signals to obtain current envelope signals, which eliminates fundamental frequencies and retains fault-related frequencies. The system resamples the current envelope signals, which converts the fault-related frequencies to constant frequency components. The system enlarges the resampled envelope signals to obtain fault information by a fault-amplifying convolution layer. The system provides the fault information as input to a deep convolutional neural network (CNN). The system generates an output including fault diagnosis for the physical object by the deep CNN.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present disclosure generally relates to machine learning and data classification. More specifically, the present disclosure relates to deep hybrid convolutional neural networks for fault diagnosis of rotating machinery such as wind turbines and associated gearboxes.

Background Art

[0002] Machine learning algorithms have been applied to a wide class of rotating systems. One exemplary dataset is the operation of wind turbines. In rotating machinery, certain characteristics such as manufacturing, energy, and transportation can be monitored. Faults in these fields, for example, faults associated with one or more components of rotating machinery, can lead to increased costs due to significant downtime, as well as an increase in accidents and related safety issues. Therefore, rotating machinery can benefit from systems that can predict problems when they begin to occur (e.g., "soft faults") and before "hard" faults.

[0003] One problem in detecting soft faults relates to how sensor data can be collected. For example, in a rotating machinery system, data can be obtained from a sensor at a first location, which may be at a close distance or a far distance from another sensor incorporated into the system at another location. Fault signals may be obscured by other operating characteristics or hidden among them, thereby making it difficult to detect the fault signals.

Summary of the Invention

Problems to be Solved by the Invention

[0004] Therefore, efficient detection of soft faults, including detecting and analyzing relevant fault signals from various sensors, leaves a problem in the field of mechanical tilt for fault diagnosis of rotating machinery systems, such as wind turbines.

[0005] One embodiment provides a system for facilitating fault diagnosis. During operation, the system collects a current signal associated with a physical object that includes a rotating machine. The system demodulates the collected signal to obtain a current envelope signal, thereby eliminating the fundamental frequency and retaining the fault-related frequencies. The system resamples the current envelope signal, thereby converting the fault-related frequencies into constant-frequency components. The system amplifies the resampled envelope signal by a fault amplification convolutional layer to obtain fault information. The system provides the fault information as an input to a deep convolutional neural network (CNN). The system generates an output including a fault diagnosis for the physical object by the deep CNN.

[0006] In some embodiments, the rotating machine includes one or more of a wind turbine, a wind turbine gearbox, a machine including a rotating shaft, a machine including one or more rotating components, and at least one component from which a current signal can be collected or obtained.

[0007] In some embodiments, demodulating the collected signal, resampling the current envelope signal, amplifying the resampled envelope signal, and providing the fault information as an input to the deep CNN are performed by a physics-based module.

[0008] In some embodiments, demodulating the collected signal is performed by an amplitude demodulation module of the physics-based module and is based on a Hilbert transform. The retained fault-related frequencies are non-stationary fault-related frequencies.

[0009] In some embodiments, resampling the current envelope signal is performed by an angular resampling module of a physics-based module and is based on an angular resampling algorithm. The angular resampling algorithm is based on an order tracking method, and the resampled envelope signal has equal phase increments in the angular domain, thereby eliminating spectral smearing.

[0010] In some embodiments, the physics-based module includes a fault amplification convolution layer. Expanding the resampled envelope signal further includes constructing a kernel by the fault amplification convolution layer based on the amplitude corresponding to the constant frequency component, and extracting features by measuring the similarity between the kernel and the local input signal.

[0011] In some embodiments, the system provides fault information as an input to a deep CNN by performing a fast Fourier Transform (FFT) analysis on the expanded and resampled envelope signal. The fault information provided to the deep CNN includes the magnitude of a predetermined frequency range. The predetermined frequency range is configured by a system or user associated with the rotating machine.

[0012] In some embodiments, the deep CNN processes the fault information based on zero-padding, batch normalization, and a plurality of pooling layers following a plurality of convolution layers.

[0013] In some embodiments, the deep CNN further processes the fault information based on two fully connected layers by using a softmax function to determine the conditional probability regarding the health state of the rotating machine. The fault diagnosis includes fault classification related to the health state of the rotating machine.

Brief Description of the Drawings

[0014]

Figure 1

[0015]

Figure 2

[0016]

Figure 3A

[0017]

Figure 3B

[0018]

Figure 4

[0019]

Figure 5A

[0020]

Figure 5B

[0021]

Figure 5C

[0022]

Figure 6

[0023]

Figure 7

[0024] In the drawings, the same reference numerals refer to the same graphical elements.

DETAILED DESCRIPTION

[0025] The following description is presented to enable any person skilled in the art to make and use the invention and is provided in the context of a particular application and its requirements. Various modifications to the disclosed embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other embodiments and applications without departing from the spirit and scope of the invention. Thus, the invention is not limited to the embodiments shown, but is to be accorded the widest scope consistent with the principles and features disclosed herein. General Overview

[0026] The embodiments described herein provide a system for improving the accuracy and robustness of fault diagnosis of a rotating machine system using a deep hybrid convolutional neural network.

[0027] As described above, faults associated with one or more components of a rotating machine can lead to increased costs due to significant downtime, as well as an increase in accidents and related safety issues. Thus, rotating machines can benefit from a system that can predict these problems when they begin to occur (e.g., "soft faults") and before "hard" failures. One problem in detecting soft faults relates to how sensor data can be collected. For example, in a rotating machine system, data can be obtained from a sensor at a first location, which may be in close proximity to or at a great distance from another sensor incorporated into the system at a different location. Fault signals can be obscured by or hidden among other operating characteristics, thereby making it difficult to detect the fault signals.

[0028] Accordingly, efficient detection of soft faults, including detecting and analyzing related fault signals from various sensors, remains an issue in the field of mechanical tilt for fault diagnosis of rotating machine systems, such as wind turbines.

[0029] Embodiments described herein address these issues by providing a system comprising two modules: a physics-based module that provides amplitude demodulation, angular resampling, and fault amplification to obtain fault information, and a deep convolutional neural network (CNN) that incorporates the fault information from the physics-based module and captures more comprehensive features for fault diagnosis. Embodiments described herein refer to a wind turbine as an exemplary rotating machine (specifically, a gearbox associated with the wind turbine and the overall rotating machine system), but the embodiments can be applied to any rotating machine system and related components and are not limited to wind turbines, wind turbine gearboxes, or any component or system associated with a wind turbine.

[0030] Maintenance costs can contribute a significant portion of the levelized cost of electricity (LCOE) generated by wind turbines. For example, inspection and maintenance costs can account for 10 - 15% of the LCOE of onshore wind turbines and up to 40% of the LCOE of offshore wind turbines. One essential component in wind turbines is the gearbox. Failures associated with the gearbox can lead to significant downtime and financial losses for wind turbines. Therefore, effective and efficient fault diagnosis of wind turbine systems and components such as gearboxes can achieve improvements in system availability, safety, and reliability, and can also reduce downtime and maintenance costs.

[0031] One technique for fault diagnosis of wind turbine gearboxes is to use the current signal of the generator. These "current-based" techniques can offer several advantages over the widely used vibration-based techniques. First, since current signals are already used in wind turbine control systems, there is no need to install additional sensors or data collection devices. This can reduce the cost and complexity of implementing current-based methods. Second, current-based fault diagnosis methods can potentially be integrated into supervisory control and data acquisition systems or control systems to automatically trigger appropriate alarms when problems occur. This feature can be particularly essential for unmanned wind turbine operation in remote or inaccessible locations. Third, current signals are generally less affected by environmental noise and the location of sensors compared to vibration signals. Current signals can be recorded at the bottom of the tower, are easily accessible, and do not interfere with the wind turbine.

[0032] Current-based techniques can provide fault diagnosis for wind turbine gearboxes, but leave several challenges in extracting fault features useful for gearbox fault diagnosis from current signals. First, the current signal can include a fundamental component and a fault-related component. The fault-related component is modulated by the fundamental component, which is the dominant component and typically much larger than the fault-related component. When explaining the "signal-to-noise ratio" (SNR), the "signal" can correspond to the fault-related component, while the "noise" can correspond to other components including the fundamental component. Current signals generally have a very low signal-to-noise ratio (SNR), which can make it difficult to extract fault features, especially for pure data-driven methods.

[0033] Furthermore, due to the various shaft rotation speeds of the wind turbine, the signals collected from the gearbox condition monitoring system are usually non-stationary. In this way, the fault-related information (i.e., the fault characteristic frequency) often changes over time in the collected signals. Therefore, advanced signal processing algorithms may be required to identify and extract useful fault features.

[0034] After fault features are identified and extracted from the collected sensor signals, machine learning techniques such as support vector machines (SVMs) and artificial neural networks (ANNs) can be applied to detect and classify faults associated with the gearbox. Deep learning algorithms may be used for different fault diagnosis applications and can be trained to adaptively learn the high-order features of the input through multiple non-linear and special operations. As a result, some of the inherent drawbacks of traditional machine learning algorithms can be mitigated.

[0035] Convolutional neural network (CNN) is one of the most powerful deep learning algorithms and can be adopted for fault diagnosis purposes. Some conventional CNNs are designed as 1-D structures to directly process data and facilitate the analysis of one-dimensional (1-D) sensor data. For example, some conventional CNNs can utilize raw current data as input and use 1-D CNN to integrate feature extraction and classification for real-time motor fault detection. Other conventional CNNs can directly learn features from the frequency spectrum of vibration signals for the condition monitoring of gearboxes.

[0036] However, although conventional CNNs have made progress in fault diagnosis applications, existing CNNs and related methods still face certain challenges. One challenge is that most conventional CNNs use a pure data-driven framework that does not consider physical knowledge in the design of the CNN structure. This can lead to the loss of important fault information, especially in scenarios where the data has a very low SNR, such as current signals in wind turbines. Another challenge is that the CNN structure is essential to obtain desirable fault diagnosis results. Different from image recognition, the hyperparameters of CNN should be carefully adjusted based on the characteristics of the input signal and features. This can be important for obtaining more accurate and robust diagnosis results in fault diagnosis applications.

[0037] Embodiments described herein address these challenges by providing a system that includes a deep hybrid convolutional neural network (DHCNN) for fault diagnosis of a wind turbine gearbox using current signals. The current signals can include three-phase stator and rotor currents. The system includes two modules. The first module is a physics-based module that provides amplitude demodulation, angular resampling, and fault amplification (e.g., via Hilbert transform) to obtain fault information. The second module is a deep convolutional neural network (CNN) that incorporates the fault information from the physics-based module and captures more comprehensive features for fault diagnosis.

[0038] The physics-based module can bring about an improvement in the SNR of the current signal and can further provide more fault information from a physical perspective. In the amplitude demodulation part of the physics-based module, the system can demodulate the collected current signal using the Hilbert transform to obtain its "current envelope signal" or "envelope". Thereby, the dominant fundamental frequency can be excluded while retaining the non-steady fault-related frequencies.

[0039] The system can resample the current envelope signal using an angular resampling algorithm through the physics-based module to convert the non-steady fault-related components into constant-frequency components within the "resampled envelope signal". Subsequently, the system can expand or amplify the resampled envelope signal through the fault amplification part of the physics-based module. The fault amplification part can be a convolutional layer and can construct a kernel. The fault characteristic frequency can determine the kernel size and the number of filters, eliminating the need to train and optimize to determine relevant information. This can also reduce the computational cost in the learning process. In this way, the system can suppress the fundamental rotation frequency and subsequently amplify the higher harmonics where the characteristics of the fault may generally be buried.

[0040] The second module is a deep CNN module. The system can calculate the fast Fourier transform (FFT) of six convolutional signals respectively, and for fault diagnosis, the FFT spectra can be supplied in parallel as a deep 1-D CNN. The whole system can include both a physics-based module and a deep CNN module, and can be referred to as a deep hybrid CNN (DHCNN). The DHCNN can utilize the concept of feature-level sensor data fusion to capture richer data and a more robust health state of the wind turbine. For effective training of the DHCNN, hyperparameters can be fully adjusted and batch normalization can be adopted. The effectiveness and superiority of the described embodiments can be verified by various gearbox failures in a doubly-fed induction generator (DFIG)-based wind turbine drive train test bed. Fault characteristic frequencies, amplitude modulation, angular resampling, and the background of standard CNN -- Fault characteristic frequencies in current signals

[0041] Mechanical faults in the gearbox can be identified by current signals based on the electromechanical connection between the gearbox and the generator. The fault-related frequencies of the gears, i.e., the vibrations at the shaft rotation frequency, can modulate the current signal. In a doubly-fed induction generator (DFIG) system, the power electronic interface can control the rotor current to obtain the variable speed necessary to capture maximum energy from the changing wind. For DFIG current, in the single-phase stator / rotor current, the fault characteristic frequency components exist at f±f i (i = 1, 2, 3...), where f is the fundamental frequency of the current signal, and f i is the vibration characteristic frequency of the gear fault. In the gearbox of a DFIG-based wind turbine, f i is proportional to the shaft rotation frequency f r . Depending on the operating mode, f of the stator current is constant at 60 Hz, and f of the rotor current is equal to (60±f r ) Hz. During the operation of a DFIG-based wind turbine, f±f iThe amplitude observed may be different from the level or amplitude observed in a healthy state. This may indicate that there is a possibility of gear failure within the gearbox, inducing additional vibrations at frequency f i in some cases. In this way, these frequency amplitudes can be used as effective fault features for gear fault diagnosis.

[0042] However, the amplitudes of such fault characteristic frequency components are generally much smaller than the fundamental frequency in the current signal and can result in a low SNR. Thus, the problem of improving the SNR of the fault-related components remains. Furthermore, since wind turbines operate at a time-varying shaft rotation frequency due to fluctuations in wind speed and direction, the fault characteristic frequencies in the current are not proportional to a constant f r either. This is another problem, indicating that further signal processing may be required for better fault feature extraction. --Amplitude Demodulation

[0043] To eliminate the fundamental frequency and increase the SNR, the described embodiments may extract the envelope signal e(t) of the current signal using amplitude demodulation. Hilbert transform may be used for amplitude demodulation, and the Hilbert transform may correspond to a 90-degree phase shift in the time domain. The Hilbert transform of a single-phase current signal (e.g., i a (t)) denoted by H[i a (t)] can be defined by the following integral transform.

Equation

Equation

[0044] Angle resampling is a technique that can solve the problem of spectral smearing of signals from wind turbines operating under variable shaft speed conditions. The general idea of angle resampling is to resample a fixed sampling rate signal into a signal with fixed phase intervals in the phase domain. Angle resampling has relatively high resolution in the frequency domain compared to time-frequency domain analysis methods, and thus can be more effective in extracting fault features in the frequency domain. To obtain the angle resampling of the envelope signal e(t) acquired after amplitude demodulation, an order tracking-based method can be used. The acquired and resampled envelope signal e’(t) can have equal phase increments in the angle domain, and thus does not have the problem of spectral smearing. In this way, conventional spectral analysis can be performed on the resampled signal for feature extraction. -- Architecture of Standard CNN

[0045] A convolutional neural network (CNN) is a multi-stage feed-forward neural network, typically consisting of multiple convolutional layers, pooling layers, and fully connected layers. Using these layers, tasks of feature learning and classification can be performed. The embodiments described herein focus on 1-D CNN because the input to the CNN is a 1-D current signal.

[0046] The convolutional layer can convolve the input 1-D vector as a set of kernels w l ∈R JxHxI and then perform an activation operation to generate output features, where J is the number of kernels, H is the fixed length of each kernel, and I is the number of channels (depth) within the kernel. Using the kernels, local features within a local region of the input can be extracted.

[0047] The output feature vector

Equation

Mathematics

Mathematics

Mathematics

Mathematics

Mathematics

[0048] The pooling layer can usually be stacked after the convolutional layer within the CNN architecture. The pooling layer can function as a downsampling operation that reduces the size of the features and the parameters of the CNN, and thus can reduce the training time and memory requirements, and can further control overfitting. One commonly used pooling function is "max pooling" that extracts the maximum value of the local region of the input features.

[0049] The fully connected layer is the last few layers within the CNN structure. The fully connected layer can flatten the features learned from the previous layer and can be used for classification purposes. Overview of the DHCNN with a physics-based module

[0050] As described above, the embodiments described herein provide improvements over existing deep learning architectures for fault diagnosis. The described embodiments provide a system that includes a DHCNN having two modules: a physics-based module and a deep CNN module. This system collects three-phase stator and rotor current signals from a wind turbine generator terminal. The system processes the signals collected via the physics-based module to improve the SNR in order to obtain better diagnostic results. Specifically, the system may use a Hilbert transform-based amplitude modulation algorithm to eliminate the current fundamental frequency and extract the envelope signal. Next, the system may use an angular resampling method to convert the non-stationary envelope signal in the time domain into a stationary signal in the angular domain. Subsequently, the system may use a fault amplification convolutional layer based on the vibration characteristic frequency f i of the gearbox to increase the SNR for feature extraction. Finally, due to the potential time delay between input signals and different kernel sizes in the fault amplification convolutional layer, the system may perform FFT analysis to convert the signal from the time domain to the frequency domain. This may also significantly reduce the input size to the deep CNN module. The FFT spectrum may be provided to a deep CNN module that includes multiple convolutional layers, batch normalization, max pooling layers, and fully connected layers or used as input to the deep CNN module. --Detailed Overview of the Physics-Based Module

[0051] After the system performs amplitude modulation and angular resampling, the acquired and resampled envelope signal e’(t) is a stationary signal, and the characteristic frequency f i is a constant value in the frequency domain spectrum of e’(t).

Number

[0052] This convolutional layer aims to extract features by measuring the similarity between the kernel and the local input signal. In this way, the kernel for fault diagnosis purposes should be useful for identifying whether the magnitude of the input signal at the fault characteristic frequency is large. It is known that a well-trained kernel within the convolutional layer of a CNN can be a set of filters with single or multiple characteristic frequencies. Thus, in the described embodiments, a physics-based convolutional layer Conv0 is designed to include certain fault frequencies [Number] and should be able to increase their magnitudes when a fault occurs.

[0053] Assume that the fault of the gearbox has a fault characteristic frequency equal to n f . Then, the n f kernels are designed such that each kernel includes four consecutive periods of one sine wave having a certain fault characteristic frequency [Number] . The convolution operation of two signals in the time domain can correspond to multiplication in the frequency domain.

[0054] Let the Fourier transform of the signal e’(t) and the kernel C0 (t) in Conv0 be E’(jω) and C 0 (jω), respectively. [Number] where F is the Fourier transform operation. The system can use the Fourier transform by convolving with a kernel designed to amplify the fault frequency. That is, the Fourier transform domain can be a multiplier, as seen, for example, in the design of the kernel function for amplifying Y(jω). In this way, the magnitude of the fault characteristic frequency can be amplified after layer Conv0, and as a result, the system can improve the SNR of the signal.

[0055] For different faults [Number] Typically, it can vary from several Hz to several hundred Hz. Therefore, the length of the kernel in Conv0 can vary significantly. The system can perform FFT analysis after Conv0 and use only the magnitude of the selected frequency range in the FFT spectrum as the input to the CNN. --Overview of the deep CNN module

[0056] As shown in the overall structure of the DHCNN in Figure 2, the output of the physics-based module is the FFT spectrum of the preprocessed signal supplied to the deep CNN module. The CNN module can include four blocks of convolutional layers, batch normalization, and max-pooling layers that gradually reduce the dimensions of the input tensor while increasing the number of channels. Subsequently, the DHCNN can apply flattening and two fully connected layers for fault diagnosis.

[0057] The first block is described as an example for introducing the structure designed within the DHCNN module. In the convolutional layer ("Conv1"), the system can use zero-padding to keep the input size constant after the convolution operation. ReLU can be selected as the activation function omega because it can accelerate the convergence of the training process using backpropagation learning. The system can add batch normalization to Conv1 between the convolution function and the activation function to reduce the internal covariance shift of the CNN and accelerate the training process. The system can stack a max-pooling ("Pool1") layer after Conv1, thereby determining the maximum value of adjacent points and further reducing the output dimension.

[0058] In the DHCNN, the fault classification stage can be composed of two fully connected layers by taking the flat fault features learned from the previous layer. The system can use, for example, the softmax function to determine the conditional probability O j for the health state of the j-th gearbox.

Mathematics

Mathematics

[0059] In the training stage, the loss function can be defined as the categorical cross-entropy between the estimated softmax output probability distribution and the actual class. The system can apply the Adam stochastic optimization algorithm to minimize the loss function. - Detailed description of an exemplary environment for facilitating fault diagnosis

[0060] FIG. 1 shows an exemplary environment 100 for facilitating fault diagnosis according to one embodiment of the present invention. The environment 100 may include a device 102, an associated user 122, and an associated display 103, a device 104 and an associated user 124, and a device 106. The device 102 may be a client computing device, such as a laptop computer, a mobile phone, a smartphone, a tablet, a desktop computer, and a handheld device. Additionally, the devices 102, 104, and 106 may be, for example, computing devices, servers, network entities, and communication devices. The devices 102, 104, and 106 can communicate with each other via a network 120. The environment 100 may also include a physical object having components that rotate and generate an electric current or a signal. For example, the physical object may be a wind turbine 108 having a plurality of blades, such as a blade 110 having a plurality of sensors 110.1 to 110.n. Each sensor can record data and transmit it to another device. The wind turbine 108 may include a gearbox 112 and a current sensor 114, and the current sensor 114 can detect generated signals, such as three-phase stator and rotor currents generated by a doubly-fed induction generator (DFIG) or a wound rotor induction generator (WRIG) associated with the gearbox 112 (as will be described later with respect to FIGS. 3A and 3B).

[0061] During operation, device 104 may send training data 140 to device 106 via network 120. Device 106 may receive the training data 140 (as training data 142) and train a deep CNN based on the training data 142 (via the network training 144 operation). The training data 140 may be sent to device 106 from device 104 or another device (not shown) at periodic intervals or in response to a command from a device (e.g., 102, 106, or another device). Thereafter, device 102 may send a fault diagnosis command 150 to device 106 via user 122. Here, the command 150 is a request for fault characteristic information related to a physical object, physical component, or physical system such as a wind turbine 108 (or specifically, a component within the gearbox 112 of the wind turbine 108).

[0062] Device 106 may receive the fault diagnosis command 150 (as fault diagnosis command 152) and perform an operation 154 to obtain time - series data, generating a message 156 to obtain time - series data addressed to device 104. Device 104 may receive the message 156 to obtain time - series data (as message 158 to obtain time - series data). The current sensor 114 may send time - series data 160 to device 104 based on a request from device 104 (not shown). The current sensor 114 may also send time - series data 160 to device 104 at periodic intervals based on a first predetermined threshold or continuously based on a second predetermined threshold. Device 104 may return the time - series data 160 (as time - series data 162) to device 106.

[0063] Device 106 may receive the time - series data 162 (as time - series data 164), and then may perform the following operations. Device 106 may perform the operation of obtaining the envelope signal 166 to demodulate, for example, the collected time - series data 164 (the "signal"), eliminate the fundamental frequency, and retain the non - stationary fault - related frequencies. Device 106 may perform the operation of resampling the envelope signal 168 to convert, for example, the non - stationary fault - related components into constant - frequency components. Device 106 may perform the operation of expanding the SNR of the resampled envelope signal 170, where, for example, the fault - amplification convolutional layer amplifies the fault and constructs a kernel based on the amplitude at the constant frequency. Device 106 may perform the operation of calculating the FFT spectrum 172 and subsequently perform the operation of inputting the FFT spectrum to the DHCNN for obtaining the fault diagnosis 174. Device 106 may return the fault diagnosis 176 to device 102.

[0064] Device 102 may receive the fault diagnosis 176 (as fault diagnosis 178) and may cause the information related to the fault characteristics to be displayed on the display 103. Exemplary display information may include information related to a physical object (wind turbine 108), the specific design or component architecture of the gearbox 112, the time - series data 160, the FFT spectrum 172, the classification or fault type, other information indicating the fault type, and one or more components associated with the indicated fault type.

[0065] In the embodiments described herein, the output of the DHCNN is a fault diagnosis, which may include not only the identification of the presence or absence of a fault but also, if a fault exists, the type of the fault. That is, the system may provide a classification of the type of the detected fault as shown above with respect to fault type 248 in FIG. 2 and fault type 370 in FIG. 3B. The fault diagnosis may be returned to the requesting user and displayed in various forms of information on a display screen associated with the requesting user, as described above with respect to fault diagnosis 178 in FIG. 1 and display 103. In an exemplary rotating machine system of a wind turbine, the fault diagnosis may be used by any other person or group of persons who may be interested in or in need of information regarding a physical asset or a rotating machine system (e.g., a wind turbine), which may also be referred to as an "entity of interest" (singular or plural).

[0066] For example, a maintenance technician may use a fault diagnosis that classifies a specific gear having a specific fault (e.g., gear 2 having a two-tooth missing (TTM) fault) to replace the specific gear or a set of gears associated with the specific gear. Another entity of interest may include a plant owner who may use the fault diagnosis at a high level to more efficiently address issues related to the overall plant and its components. Another entity of interest may include a power grid operator who may use a fault diagnosis or a set of fault diagnoses to plan for other requirements for obtaining energy (e.g., if a specific fault diagnosis or a set of fault diagnoses affects the capacity of a wind turbine to provide a predicted amount of electrical power). Another entity of interest may include a manufacturing lead or a user associated with a system including a wind turbine. This entity may use the fault diagnosis to plan maintenance during off-peak production cycles, for example, to plan the timing of production such that production is not or is less dependent on the rotating machine.

[0067] Fault diagnosis output by the system can be provided to an operations dashboard or other graphical user interface (GUI) for any of the exemplary entities of interest listed above. Each entity of interest can be a user (such as user 122 in FIG. 1) that generates and transmits a command (e.g., command 150 in FIG. 1) to diagnose a fault. The system can display the identification, classification, and other relevant information of the fault diagnosis (the "detected fault") on a display screen associated with the user (e.g., display 103 in FIG. 1). Based on the fault diagnosis output of the detected fault, the user can perform corrective actions to address the diagnosed or detected fault. The user can then use the operations dashboard or GUI (e.g., by pressing a widget or other executable button or widget) to generate another command to diagnose the fault to determine whether the corrective actions performed adequately addressed the detected fault. In this way, the described embodiments provide an improvement in the classification of faults in rotating machinery systems (such as within a wind turbine and associated gearbox) by using the DHCNN together with a physics-based module that includes a fault amplification layer.

[0068] FIG. 2 shows an exemplary architecture 200 of a DHCNN and a physics-based module according to an embodiment of the present application. The architecture 200 can include a data collection module 210, a physics-based module 220, and a deep CNN module 240. During operation, the data collection module 210 can obtain time-series data by monitoring, observing, and detecting current signals 214 such as those generated from a physical object (e.g., a wind turbine 212 or a signal generation component associated with the wind turbine 212). The current signal 214 can be represented as time-series data. The data acquisition module 210 can transmit the current signal 214 as time-series data to the physics-based module 220 (via communication 250).

[0069] The physics-based module 220 may receive the current signal 214 as time-series data via the amplitude demodulation module 222. The amplitude demodulation module 222 may demodulate the collected current signal to obtain a current envelope signal, thereby eliminating the fundamental frequency and retaining the non-stationary fault-related frequencies. Next, the angular resampling module 224 may resample the current envelope signal based on an angular resampling algorithm to convert the non-stationary fault-related components into constant-frequency components. Subsequently, the fault amplification convolution layer 226 (denoted as "Conv0" in the present disclosure) may expand the resampled envelope signal and construct a kernel based on the constant-frequency amplitude. The FFT spectrum 228 module may perform an FFT analysis of the expanded or amplified envelope signal and provide only the FFT spectrum having the magnitude of the selected frequency range as an input to the deep CNN module 240.

[0070] The output of the FFT spectrum 228 module may be transmitted to the deep CNN module 240 via communication 252. This FFT spectrum is further provided as an input to the convolution / batch normalization / pooling layer 244 of the deep CNN module 240 via communication 254. For example, following four of the convolution and pooling layers, the deep CNN module 240 may process the data via, for example, two fully-connected layers 246. Finally, the deep CNN module 240 may provide the fault type 248 as its output that may be returned to the requesting device or related user for display on the display screen of the requesting device or user and further analysis as described above with respect to FIG. 1. The further analysis may include user measures for repairing or otherwise addressing the diagnosed faults related to any associated physical components of the rotating shaft system. Exemplary Datasets and Results Using Physics-Based Modules --Exemplary Environment: Wind Turbine Emulator

[0071] Figure 3A shows an exemplary environment 300 for using a physics-based module according to an embodiment of the present application. Environment 300 may depict a system with a wind turbine emulator to demonstrate a DHCNN for fault diagnosis. Environment 300 may be driven by wind 310 and rotor 312. For example, an induction motor driven by a variable frequency alternating current (AC) drive can be used as a prime mover together with a step-down type gearbox 314 that reduces the shaft rotation frequency of the induction motor (e.g., the doubly-fed induction generator (DFIG) 316 or wound-rotor induction generator (WRIG) 350 of FIG. 3B). This can emulate the prime power of the wind turbine rotor. Another two-stage helical gearbox (e.g., the two-stage helical gearbox 360 of FIG. 3B) can be used to emulate the gearbox in the power transmission device with some artificially generated faults and is connected to the DFIG 316 with two pole pairs.

[0072] The stator 318 of the DFIG 316 can be connected to a programmable AC power supply (shown as grid 322) that can be used to emulate a power grid. The rotor 320 of the DFIG 316 can be connected to the same AC power supply (e.g., grid 322) via two back-to-back connected three-phase insulated-gate bipolar transistor (IGBT) power converters, which are a rotor side converter (RSC) 326 and a grid side converter (GSC) 324, respectively. The system can record signals used by a DFIG control scheme, including three-phase rotor currents, using a dSPACE 1005 board (not shown) at a sampling frequency of, for example, 5 kHz.

[0073] A position encoder with a resolution of 4096 cycles per revolution (e.g., encoder 356 of FIG. 3B) is mounted on the input shaft of the DFIG 316 to measure the shaft rotation frequency f r(t) can be measured. The shaft rotation frequency can vary randomly and generally can be within ±20% of the synchronous rotation frequency to meet the operating requirements of the DFIG316.

[0074] FIG. 3B shows an exemplary environment 340 for using a physics-based module according to an embodiment of the present application. The environment 340 may include a wound rotor induction generator (WRIG) 350 and a two-stage helical gearbox 360 joined by a coupler 358. The WRIG 350 may include a rotor winding 352, slip rings 354, and an encoder 356. The two-stage helical gearbox 360 may include four gears (labeled "gear 1", "gear 2", "gear 3", and "gear 4"), an input shaft 362, a pinion shaft 364, and an output shaft 366. Each labeled gear corresponds to a specific number of teeth z x corresponding thereto. For example, gear 1 has 52 teeth (z 1 = 52), gear 2 has 11 teeth (z 2 = 11), gear 3 has 38 teeth (z 3 = 38), and gear 4 has 17 teeth (z 4 = 17). Each gear may correspond to one of four failure types 370 of an exemplary faulty test gear, namely, one tooth missing (OTM) 372, two teeth missing (TTM) 374, chipping 376, and cracking 378.

[0075] The gear fault characteristic frequencies, i.e., the three shaft rotation frequencies, can be expressed by the following equations.

Equation

[0076] f r (t) = f 3 (t), so the constant in e'(t)

Equation

[0077] Using the following settings, exemplary results can be generated. The exemplary results are based on experiments conducted continuously for 100 minutes under each of the five fault types. The system records three-phase stator and rotor current signals as one raw data record for 100 seconds at an interval of 20 seconds between two consecutive data records, and obtains 50 raw data samples for each fault type. To increase the size of the training / test dataset, a simple data augmentation technique of slicing the raw data samples with a stride can be used to increase the number of data samples. For example, a 100-second raw data sample can be sliced into 36 data samples with a length of 30 seconds using a stride time of 2 seconds. In this way, there can be 1800 data samples for each fault type, and thus a total of 9000 data samples can be obtained. These data samples can be randomly shuffled and divided into training, validation, and test datasets respectively, which can contain 70%, 20%, and 10% of the augmented data samples.

[0078] To save as much fault information as possible, the frequency range of the FFT spectrum with the size supplied to the CNN model is

Number

[0079] FIG. 5A shows Table 500 with a summary of the parameters used in an exemplary DHCNN, including the results using a fault amplification convolutional layer within a physics-based module according to one embodiment of the present application. Table 500 may show details of the architecture of the exemplary DHCNN. Table 500 may include multiple entries for each layer, and each entry may include information related to the corresponding layer 502, kernel size / stride 504, number of kernels 506, output size (H * I) 508, and parameter size 510 (in columns). Entry 520 may correspond to the "Conv0" layer (e.g., the fault amplification convolutional layer 226 of the physics-based module 220 in FIG. 2). Entry 522 may correspond to the FFT (e.g., the FFT spectrum 228 shown as the output of the physics-based module 220 in FIG. 2). Subsequent entries 524-538 may correspond to multiple convolutional and pooling layers (e.g., 244 of the deep CNN module 240 in FIG. 2). Entries 540-542 may correspond to fully connected (FC) layers (e.g., 246 of the deep CNN module 240 in FIG. 2).

[0080] Table 500 shows that the length of the output size (column 508, "H") gradually decreases through the Conv layer and Pool layer, and the depth of the output size (column 508 "I") continues to increase. The exemplary DHCNN may include a total of 231,693 parameters, including 230,957 trainable parameters and 736 non-trainable parameters. The training process may be implemented using Keras with TensorFlow as the backend. --Exemplary Results Using a Physics-Based Module and a Fault Amplification Convolutional Layer

[0081] Figure 4 shows a diagram 400 of an extended data sample of 1 second with a TTM fault and its FFT spectrum, according to one embodiment of the present application. Diagram 400 shows a portion (upper part) of one data sample I ra (t) expanded from a single-phase rotor current collected under a TTM fault and its FFT spectrum. In Figure 4, the fundamental component is clearly identified (in the central part of box 430), and its frequency can vary in the range of 8 to 10 Hertz (Hz) due to various shaft speeds. At the same time, visual indicators of fault-related information are nearly impossible to detect. In fact, fault-related information (e.g., a fault-related range 432 that may include fault-related information 434 and 436) is hardly seen in the blurred range that appears in the expanded FFT spectrum. This demonstrates that the raw data sample has a very low SNR.

[0082] In the embodiments described herein, the system can eliminate the fundamental frequency and retain the non-stationary fault-related frequencies, for example, by collecting a current signal, demodulating the collected current signal to obtain a current envelope signal, resampling the current envelope signal to convert non-stationary fault-related frequencies into constant-frequency components, and expanding the resampled envelope signal to construct a kernel based on the amplitude at a constant frequency. An exemplary method for performing this signal processing procedure was described above in connection with Figure 2 and will be described later in connection with Figure 6.

[0083] FIG. 5B shows Table 550 with a comparison of the accuracy and standard deviation of four different methods according to an embodiment of the present application. Table 550 may include entries 560 - 566, and each entry may indicate a structure 552, an average accuracy 554, and a standard deviation 556. For example, entry 560 may correspond to the DHCNN of the described embodiment and shows the highest average accuracy (99.54%) and the lowest standard deviation (0.25%) among the four described methods. Entry 562 may correspond to "DCNN1", a conventional CNN structure that takes raw data samples as input without a physics-based module. Entry 564 may correspond to "DCNN2", a DHCNN without a fault amplification convolutional layer (i.e., Conv0). Entry 566 may correspond to a conventional feedforward artificial neural network (ANN). In this way, Table 550 shows that the described embodiment of the DHCNN has the highest accuracy and robust performance for fault diagnosis. Additionally, the average accuracy of DCNN1 is significantly lower than other methods, while the standard deviation of DCNN1 is significantly higher than other methods. This indicates that the physics-based module may be decisive or important for achieving high accuracy and robustness in fault diagnosis.

[0084] Furthermore, by using the physics-based module in the described embodiment of the DHCNN herein, the system can use the operating conditions of the system (e.g., identifying the gearbox of a wind turbine as a common fault point in the components of a rotating machinery system) rather than relying solely on pure machine learning (like conventional CNNs), thereby providing an improvement in the analysis and diagnosis of fault-related information. Specifically, by using a hybrid approach of a physics-based module (including a fault amplification module) and a deep CNN, the described embodiment can result in improved fault diagnosis for a wide class of rotating machinery systems.

[0085] FIG. 5C shows a plot 570 having exemplary accuracy curves for some of the methods listed in FIG. 5B, according to one embodiment of the present application. Plot 570 shows that the training accuracy of all three listed methods (DHCNN, DCNN1, and DCNN2) can reach stable values as the number of epochs increases, but ultimately DHCNN has the highest accuracy. Further, the described embodiment of DHCNN has much higher accuracy at the start and can converge more quickly than DCNN1 and DCNN2 due to the use of Conv0 in the physics-based module. In this way, the utilization of Conv0 can provide more fault information to DHCNN without training, thereby resulting in more rapid and accurate diagnostic results, especially in time-sensitive systems. These improvements demonstrate that the described DHCNN can be better for faster and more efficient learning and can be implemented in real time for online adaptation. Exemplary Method for Facilitating Fault Diagnosis

[0086] FIG. 6 shows a flowchart 600 illustrating a method for facilitating fault diagnosis, according to one embodiment of the present application. During operation, the system collects a current signal associated with a physical object including a rotating machine (operation 602). The system demodulates the collected signal to obtain a current envelope signal, thereby eliminating the fundamental frequency and retaining the fault-related frequencies (operation 604). The system resamples the current envelope signal, thereby converting the fault-related frequencies into constant-frequency components (operation 606). The system expands the resampled envelope signal by a fault amplification convolutional layer to obtain fault information (operation 608). The system provides the fault information as an input to a deep convolutional neural network (CNN) (operation 610). The system generates an output including fault diagnosis of the physical object by the deep CNN (operation 612). This deep CNN may include a deep hybrid CNN (DHCNN) based on a physics-based module including a fault amplification convolutional layer. Exemplary Computer and Communication System

[0087] Figure 7 shows an exemplary computer and communication system that facilitates fault diagnosis, according to one embodiment of the present invention. The computer system 702 includes a processor 704, a memory 706, and a storage device 708. The memory 706 can include volatile memory (e.g., RAM) that functions as a management memory and can be used to store one or more memory pools. Further, the computer system 702 can be coupled to a display device 710, a keyboard 712, and a pointing device 714. The storage device 708 can store an operating system 716, a content processing system 718, and data 734.

[0088] When executed by the computer system 702, the content processing system 718 can include instructions that cause the computer system 702 to execute the methods and / or processes described in this disclosure. Specifically, the content processing system 718 can include instructions for transmitting and / or receiving data packets to / from other network nodes via a computer network (communication module 720). The data packets can include data, requests, commands, time series data, training data, and fault diagnosis or fault classification.

[0089] The content processing system 718 may further include instructions for collecting current signals associated with physical objects equipped with rotating machines (communication module 720 and data acquisition module 722). The content processing system 718 may include instructions for demodulating the collected signals to obtain a current envelope signal, thereby eliminating the fundamental frequency and retaining the fault-related frequencies (amplitude demodulation module 724). The content processing system 718 may include instructions for resampling the current envelope signal, thereby converting the fault-related frequencies into constant-frequency components (angular resampling module 726). The content processing system 718 may include instructions for expanding the resampled envelope signal by a fault amplification convolutional layer to obtain fault information (fault amplification module 728). The content processing system 718 may include instructions for providing the fault information as an input to a deep convolutional neural network (CNN) (communication module 720 and information providing module 730). The content processing system 718 may include instructions for generating an output including fault diagnosis of a physical object by a deep CNN (fault diagnosis module 732).

[0090] Data 734 may include any data that is required as an input or generated as an output by the methods and / or processes described in the present disclosure. Specifically, data 734 may store at least data, a set of data, data representing a current signal, an indicator or identifier of a physical object or a rotating machine, a demodulated signal, a current envelope signal, a fundamental frequency, a fault-related frequency, a resampled signal, a constant-frequency component, an expanded or amplified signal, fault information, information associated with or related to a CNN, DCNN, or DHCNN, an output, a fault diagnosis, a fault type or fault classification, an FFT spectrum, indicators of physics-based modules, an amplitude demodulation module, an angular resampling module, and a fault amplification module, and indicators or identifiers of convolutional layers, batch normalization, pooling layers, or fully connected layers, and a fault type.

[0091] The data structures and codes described in the "Best Mode for Carrying Out the Invention" section are typically stored in a computer-readable storage medium, which can be any device or medium capable of storing codes and / or data for use by a computer system. Examples of computer-readable storage media include, but are not limited to, volatile memory, non-volatile memory, magnetic and optical storage devices such as disk drives, magnetic tapes, CDs (compact discs), DVDs (digital versatile discs or digital video discs), or other media capable of storing computer-readable media known or developed in the future.

[0092] The methods and processes described in the "Best Mode for Carrying Out the Invention" section can be embodied as codes and / or data stored in a computer-readable storage medium as discussed above. When a computer system reads and executes the codes and / or data stored on the computer-readable storage medium, the computer system executes the methods and processes embodied as data structures and codes and stored in the computer-readable storage medium.

[0093] Furthermore, the methods and processes described above can be included in hardware modules or devices. Examples of hardware modules or devices include, but are not limited to, application-specific integrated circuit (ASIC) chips, field-programmable gate arrays (FPGAs), dedicated or shared processors that execute specific software modules or codes at specific times, and other programmable logic devices known or developed later. When the hardware modules or devices are activated, the methods and processes contained therein are executed.

[0094] The foregoing embodiments described in this specification are presented for purposes of illustration and description only. They are not intended to be exhaustive or to limit the invention to the forms disclosed. Accordingly, many modifications and variations will be apparent to those skilled in the art. In addition, the foregoing disclosure is not intended to limit the invention. The scope of the invention is defined by the appended claims.

Claims

1. 1. A computer-executable method for facilitating fault diagnosis, the method comprising: acquiring a current signal associated with a physical object including a rotating machine; demodulating the acquired signal to obtain a current envelope signal, thereby rejecting fundamental frequencies and retaining fault related frequencies; resampling the current envelope signal, thereby converting the fault related frequencies to constant frequency components; a fault amplifying convolutional layer amplifies the resampled envelope signal to obtain fault information; providing the fault information as an input to a deep convolutional neural network (CNN); generating, by the deep CNN, an output comprising the fault diagnosis for the physical object.

2. The rotating machine includes: Wind turbines, A wind turbine gearbox; a machine including a rotating shaft; 10. The method of claim 1, comprising one or more of: one or more rotating components; and a machine including at least one component from which a current signal can be collected or obtained.

3. 2. The method of claim 1 , wherein demodulating the collected signals, resampling the current envelope signal, upscaling the resampled envelope signal, and providing the fault information as input to the deep CNN are performed by a physics-based module.

4. demodulating the acquired signals is performed by an amplitude demodulation module of the physics-based module and is based on a Hilbert transform; The method of claim 3 , wherein the retained fault-related frequencies are non-stationary fault-related frequencies.

5. resampling the current envelope signal is performed by an angular resampling module of the physics-based module and is based on an angular resampling algorithm; The corner resampling algorithm is based on an order tracking method; The method of claim 3 , wherein the resampled envelope signal has equal phase increments in the angular domain, thereby eliminating spectral smearing.

6. the physics-based module includes the fault amplifying convolutional layer; Expanding the resampled envelope signal comprises: constructing a kernel based on the amplitude corresponding to the constant frequency component by the fault amplifying convolution layer; The method of claim 3 , further comprising: extracting features by measuring similarity between the kernel and a local input signal.

7. Providing the fault information as an input to the deep CNN includes: performing a Fast Fourier Transform (FFT) analysis on the resampled and expanded envelope signal; The fault information provided to the deep CNN includes a magnitude of a predetermined frequency range, The method of claim 1 , wherein the predetermined frequency range is configured by a system or a user associated with the rotating machine.

8. The method of claim 1 , wherein the deep CNN processes the fault information based on zero padding, batch normalization, and multiple pooling layers followed by multiple convolutional layers.

9. The deep CNN processes the fault information further based on two fully connected layers by determining a conditional probability for a health state of the rotating machine using a softmax function; The method of claim 8 , wherein the fault diagnosis includes a fault classification associated with the health condition of the rotating machine.

10. 1. A computer system for facilitating fault diagnosis, the computer system comprising: A processor; a storage device storing instructions that, when executed by the processor, cause the processor to perform a method, the method comprising: acquiring a current signal associated with a physical object including a rotating machine; demodulating the acquired signal to obtain a current envelope signal, thereby rejecting fundamental frequencies and retaining fault related frequencies; resampling the current envelope signal, thereby converting the fault related frequencies to constant frequency components; a fault amplifying convolutional layer amplifies the resampled envelope signal to obtain fault information; providing the fault information as an input to a deep convolutional neural network (CNN); generating, by the deep CNN, an output comprising the fault diagnosis for the physical object.

11. The rotating machine includes: Wind turbines, A wind turbine gearbox; a machine including a rotating shaft; 11. The computer system of claim 10, comprising one or more of: one or more rotating components; and a machine comprising at least one component from which a current signal can be collected or obtained.

12. 11. The computer system of claim 10, wherein demodulating the collected signals, resampling the current envelope signals, upscaling the resampled envelope signals, and providing the fault information as input to the deep CNN are performed by a physics-based module.

13. demodulating the acquired signals is performed by an amplitude demodulation module of the physics-based module and is based on a Hilbert transform; The computer system of claim 12 , wherein the retained fault-related frequencies are non-stationary fault-related frequencies.

14. resampling the current envelope signal is performed by an angular resampling module of the physics-based module and is based on an angular resampling algorithm; The corner resampling algorithm is based on an order tracking method; 13. The computer system of claim 12, wherein the resampled envelope signal has equal phase increments in the angular domain, thereby eliminating spectral smearing.

15. the physics-based module includes the fault amplifying convolutional layer; Expanding the resampled envelope signal comprises: constructing a kernel based on the amplitude corresponding to the constant frequency component by the fault amplifying convolution layer; 13. The computer system of claim 12, further comprising: extracting features by measuring a similarity between the kernel and a local input signal.

16. Providing the fault information as an input to the deep CNN includes: performing a Fast Fourier Transform (FFT) analysis on the resampled and expanded envelope signal; The fault information provided to the deep CNN includes a magnitude of a predetermined frequency range, The computer system of claim 10 , wherein the predetermined frequency range is configured by a system or a user associated with the rotating machine.

17. 11. The computer system of claim 10, wherein the deep CNN processes the fault information based on zero padding, batch normalization, and multiple pooling layers followed by multiple convolutional layers.

18. The deep CNN processes the fault information further based on two fully connected layers by determining a conditional probability for a health state of the rotating machine using a softmax function; The computer system of claim 17 , wherein the fault diagnosis includes a fault classification associated with the health condition of the rotating machine.

19. A system using a deep hybrid convolutional neural network (DHCNN), comprising: a data collection module configured to collect current signals associated with a physical object including a rotating machine; A fault amplifying convolutional layer of a physics-based module, comprising: A fault amplifying convolutional layer, the fault amplifying convolutional layer being configured to expand a resampled envelope signal based on the collected current signal to obtain fault information; an information providing module configured to provide the fault information as an input to a deep convolutional neural network (CNN); The DHCNN is configured to generate an output comprising a fault diagnosis of the physical object.

20. The physics-based module further includes an amplitude demodulation module, an angle resampling module, the impairment amplification convolution layer, and the information providing module; the amplitude demodulation module is configured to demodulate the acquired signal to obtain a current envelope signal, thereby rejecting fundamental frequencies and retaining fault related frequencies; 20. The system of claim 19, wherein the angle resampling module is configured to resample the current envelope signal, thereby converting the fault-related frequency to a constant frequency component to obtain the resampled envelope signal.

Images