Method and monitoring system for detecting fault in machine

Abstract

A method for detecting a fault in a machine (700) comprises the steps of: a) receiving (S1) a vibration signal (100) recorded by one or more sensors (706) sensing vibrations of the machine (700); b) determining (S2) a real fault signal (102) by applying a neural network (104) to the vibration signal (100); and c) applying (S3) an envelope spectrum analysis (106) to the real fault signal (102) to detect a fault. According to the solution, the advantages of a mature envelope spectrum analysis method and machine learning are combined.

Description

Technical Field

[0001] The present invention relates to a method, a monitoring system and a computer program product for detecting faults in a machine. Background Art

[0002] Sensors are ubiquitous in all kinds of heavy machinery, electric motors, etc. and are therefore highly relevant to many different business areas. One particularly important application area for sensors is the detection of faults in rotating parts in machinery such as motors, turbines and pumps.

[0003] Faults in rotating mechanical components like gears or bearings are one of the most common causes of failure of rotating equipment. This failure can be detected in vibration patterns. To obtain the required information, sensors such as position transducers, velocity sensors, accelerometers and spectral emission energy sensors can be mounted directly on rotating parts or on these machines. This allows to obtain measurements that can be used for vibration analysis by extracting vibration frequency and amplitude. If the rotating elements suffer from different damages, geometric defects or failures, the sensor values ​​often show suspicious patterns and anomalies.

[0004] Although there are various tools and methods derived from physical theory that would, in principle, allow obtaining this information from sensor measurements, this remains a very challenging task. Furthermore, there are several issues related to detecting rotating element faults from sensor data that still need to be addressed. The sensor time series (a) is only an indirect measurement of the real physical mechanism and (b) the measured sensor data is overlaid with a large number of other effects (e.g. caused by the load on the motor) that have to be filtered out at great expense. (c) Furthermore, in order to accurately detect and classify damage in the spectrum, the exact geometry of the mounted bearing must be known, which is usually not the case.

[0005] One approach is to use a composite fault diagnosis method that utilizes intrinsic component filtering (Zongzhen Zhang et al. (2019). A novel composite fault diagnosis method that utilizes intrinsic component filtering). In this paper, a method for composite fault diagnosis of gearboxes is described that can prevent damage accidents and minimize production losses.

[0006] In addition, there has been a push to apply more powerful machine learning algorithms (Liu, R., Yang, B., Zio, E., & Chen, X. (2018). Artificial intelligence for rotating machinery fault diagnosis: an overview. Mechanical Systems and Signal Processing, 108, 33-47.). However, such methods are often not human-understandable and are often referred to as black-box algorithms. This means that it is impossible to understand how the model behaves, why the model predicts a fault, and therefore whether the results of the algorithm are consistent with the existing domain knowledge about the problem. This creates obstacles for developers to build better and more powerful models and for domain experts (engineers and physicists) to understand and trust these results. This also creates problems in detecting the root cause of the fault. Summary of the invention

[0007] It is therefore an object of the present invention to provide an improved method of machine fault detection.

[0008] According to a first aspect, there is provided a method for detecting a fault in a machine, comprising the following steps:

[0009] a) receiving vibration signals recorded by one or more sensors sensing vibrations of the machine;

[0010] b) determining the true fault signal by applying a neural network to the vibration signal; and

[0011] c) Apply envelope spectrum analysis to real fault signals to detect faults.

[0012] Envelope spectrum analysis consists of a set of signal processing steps and can be considered as one of the most popular techniques for identifying faults in vibration signals of rotating elements like bearings or gears (Randall, RB, & Antoni, J. (2011). Rolling element bearing diagnostics - a tutorial. Mechanical Systems and Signal Processing, 25(2), 485-520). This is due to the fact that all analysis steps are well justified by physical domain knowledge.

[0013] It can be difficult to obtain information about the impulse frequency from the raw vibration signal, since it is an amplitude modulated version of the signal of interest. Therefore, it is advantageous to analyze a demodulated version of the signal such as the signal envelope (as done in envelope spectrum analysis). The signal envelope can be viewed as a modified version of the vibration signal, where the presence of the characteristic impulse caused by the fault is more obvious. Therefore, defects can be detected by checking whether the characteristic frequency is present in the envelope spectrum of the measured vibration signal.

[0014] However, the success of envelope spectrum analysis may be limited due to the presence of various additional vibrations that may mask the fault signal. For example, such interference may result from deterministic vibrations caused by other machine components like shafts or simply from other background noise.

[0015] The current solution advantageously combines the proven envelope spectrum analysis method with the power of machine learning. Machine learning is here used to provide the envelope spectrum analysis with improved data known as the "real fault signal". In particular, by applying a neural network to the vibration signal, the noise content in the real fault signal is reduced when compared to the vibration signal. In other words, those frequencies that are relevant to fault detection are more prominent in the real fault signal.

[0016] The vibration signal may be a raw vibration signal or a pre-processed signal (eg, using conventional processing techniques such as filtering).

[0017] For example, the sensor for sensing vibrations may be an accelerometer.

[0018] Faults can be detected by identifying one or more (apparent) fault frequencies in the envelope spectrum.

[0019] According to an embodiment, the neural network in step b) is an unsupervised neural network.

[0020] Therefore, no labeled data is required, which significantly simplifies training.

[0021] According to another embodiment, the unsupervised neural network is designed to determine the real fault signal such that an impulse component in the real fault signal is maximized.

[0022] By increasing the impulse component, the portion of the signal relevant to fault detection is increased. The reason is that fault signals are usually impulsive. In one embodiment, the unsupervised neural network is designed to determine the real fault signal so that the impulse component in the real fault signal is increased compared to the vibration signal.

[0023] According to another embodiment, the unsupervised neural network is designed to determine the true fault signal such that the difference between the true fault signal and the vibration signal is minimized.

[0024] This ensures that the true fault signal still adequately represents the vibration signal.

[0025] According to another embodiment, the unsupervised neural network is designed such that:

[0026]

[0027] in: and is a hyperparameter,

[0028] SEN θ is an unsupervised neural network,

[0029] x is the vibration signal,

[0030] θ is a set of parameters, including SEN θ The weights and biases of , where θ is adjusted during unsupervised learning,

[0031] d() is a distance function configured to determine the difference between a true fault signal and a vibration signal, and

[0032] imp() is a function configured to determine the impulse component in the true fault signal.

[0033] This function represents the cost function used in the learning (optimization) of the neural network (SEN). Its advantage is that it ensures the increase of impulse while maintaining the authenticity of the original vibration signal.

[0034] According to another embodiment, the function d() is designed to determine the L1 norm, the Euclidean norm, the distance in the spectrum and / or the kernel-based distance.

[0035] Experiments have shown that with this function d(), a suitable representation of the (original) vibration signal is found.

[0036] According to another embodiment, the function d() is a further neural network.

[0037] The function d() can be determined using a supervised or unsupervised neural network.

[0038] According to another embodiment, the function imp() is designed to determine the kurtosis, entropy or non-Gaussianity.

[0039] The inventors have found that using this function imp(), fault frequencies in real fault signals are particularly well identified.An example of a non-Gaussian function is log(cosh()).

[0040] According to another embodiment, the unsupervised neural network is a deep, convolutional or recurrent neural network.

[0041] In experiments, the inventors found that convolutional neural networks (especially combinations of convolutional neural networks) provide particularly good results. According to an embodiment, the unsupervised neural network includes a convolutional layer, followed by a transposed convolutional layer network. A linear rectifier function can be used between the convolutional layer and the transposed convolutional layer. The convolutional layer, the linear rectifier function and the transposed convolutional layer can form a block. Two or more such blocks can form an unsupervised neural network. The filter size of the neural network or layer can range from 1 to 100, preferably 5 to 25. In one embodiment, for example, the number of filters used ranges from 5 to 50 or 10 to 20.

[0042] According to another embodiment, the envelope spectrum analysis comprises:

[0043] Determine the vibration envelope of the real fault signal,

[0044] Determine the frequency spectrum of the vibration envelope,

[0045] One or more significant frequencies in the frequency spectrum are determined and correlated with known fault frequencies.

[0046] The rich domain knowledge about known faults available in envelope spectrum analysis can be used for fault identification.

[0047] According to another embodiment, the machine comprises a rotating machine element.

[0048] For example, the machine element is a bearing, in particular a roller element bearing. For example, the vibrations may result from defects in the rolling elements, races or cages.

[0049] According to another aspect, there is provided a monitoring system for detecting a fault in a machine, the monitoring system comprising:

[0050] a receiving unit for receiving a vibration signal recorded by a sensor for sensing vibration of the machine;

[0051] a determination unit for determining a true fault signal by applying a neural network to the vibration signal;

[0052] as well as

[0053] An application unit for applying envelope spectrum analysis to a real fault signal to detect the fault.

[0054] Each unit (e.g., receiving unit) can be implemented in hardware and / or software. If the unit is implemented in hardware, it can be implemented as a device, for example, as a part of a computer or processor or system, for example, a computer system. If the unit is implemented in software, it can be implemented as a computer program product, a function, a routine, a program code or an executable object.

[0055] According to another aspect, there is provided a computer program product comprising instructions which, when the program is executed by a computer, cause the computer to perform the above method.

[0056] The computer program product, such as a computer program device, can be embodied as a memory card, a USB stick, a CD-ROM, a DVD or as a file which can be downloaded from a server in a network.

[0057] For example, such a file may be provided by transmitting a file comprising the computer program product from a wireless communication network.

[0058] Any implementation of the first aspect may be combined with any implementation of the other aspect to obtain other implementations of the first aspect, and vice versa.

[0059] The use of "a" or "an" herein does not exclude the presence of more than one of an element.

[0060] Other possible implementations or alternative solutions of the present invention also include combinations of features described above or below with respect to the embodiments (not explicitly mentioned herein). Those skilled in the art may also add independent or isolated aspects and features to the most basic form of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Further embodiments, features and advantages of the present invention will become more apparent from the following description and the dependent claims taken in conjunction with the accompanying drawings, in which:

[0062] Figure 1 shows different signals employed in a method according to one embodiment;

[0063] Figure 2 Shows the corresponding Figure 1 A flowchart of the method;

[0064] Figure 3 shows details of an unsupervised neural network used in the present method according to one embodiment;

[0065] Figure 4 shows a comparison between the original vibration signal and the real fault signal according to an experiment;

[0066] Figure 5 Shows that from Figure 4 The noise removed from the real fault signal;

[0067] Figure 6 Shows Figure 4 The envelope spectrum of the original vibration signal and the real fault signal; and

[0068] Figure 7 A monitoring system according to an embodiment is shown.

[0069] In the drawings, like reference numbers designate similar or functionally equivalent elements unless otherwise indicated. DETAILED DESCRIPTION

[0070] Figure 1 A raw vibration signal 100 that has been recorded by a sensor is shown. The sensor (e.g., an accelerometer) may be attached to a machine or otherwise record its vibrations. For example, the machine may be a rotating machine including one or more bearings, in particular a motor, a turbine, etc.

[0071] Figure 7A bearing housing 700 is shown housing a ball bearing 702. The ball bearing 702 holds a rotating shaft 704. An accelerometer 706 is attached to the bearing housing 700 and detects vibrations, including vibrations of individual balls, one of which is indicated by reference numeral 708.

[0072] Figure 7 Also illustrated is a monitoring system 750, which may include a processing unit, such as a microprocessor and a memory. The monitoring system 750 includes a receiving unit 752, a determining unit 754 and an applying unit 756. The monitoring system 750 may also include an accelerometer 706 or any other vibration sensor.

[0073] The receiving unit 752 receives the detected original vibration signal 100. This corresponds to the flowchart of FIG. Figure 2 Step S1 in .

[0074] In step S2 , the determination unit 754 determines the true fault signal 102 by applying the neural network 104 to the vibration signal 100 .

[0075] Next( Figure 2 In step S3 ), the application unit 756 applies the envelope spectrum analysis 106 to the real fault signal 102 .

[0076] The envelope spectrum analysis 106 (e.g., an algorithm) outputs, for example, a spectrum 108. In particular, the envelope spectrum analysis 106 includes: determining a vibration envelope of the real fault signal 102; determining a spectrum 108 of the vibration envelope; determining one or more significant frequencies in the spectrum 108 and correlating the one or more significant frequencies with known fault frequencies. For example, based on domain knowledge, it is well known that roller bearing faults are usually identified at the ball passing frequency. This knowledge can then be used to determine whether a fault exists in the spectrum 108. This identification can be done automatically by the monitoring system 750.

[0077] Figure 3 Details of one embodiment of the neural network 104 used in step S2 above are illustrated.

[0078] Neural network 104 is an unsupervised neural network. In other embodiments, for example, in addition to neural network 104, a supervised neural network is used.

[0079] The neural network 104 includes two blocks 300 and 302. In other embodiments, more than two blocks (e.g., 5 or 10) may be used. Each block 300, 302 may be configured in the same manner, except for some differences that will be described in detail below. Therefore, by way of example, the description focuses on block 300, but applies equally to block 302.

[0080] As shown in the dashed box, block 300 includes a convolutional layer including 64 filters, one of which is Figure 3 denoted by reference numeral 304. In different embodiments, the number of filters may vary. The size of the filters may also vary. In this experiment, a filter size of 10 was chosen, which means that the filter reads 10 values ​​306 from the vibration signal 100. The vibration signal 100 may be provided as a one-dimensional array comprising amplitude values ​​sampled at a fixed sampling rate.

[0081] Each filter includes weights, which are then used to calculate feature map 308. Initially, the weights are set using random numbers or statistical methods. One value of feature map 308 calculated using filter 304 is represented by reference numeral 310. Filter 304 is then moved to the next position on the array, and another value of feature map 308 is calculated. Once the entire array has been read, the process is repeated for the next filter (having, for example, different weights than the previous filter).

[0082] A rectified feature map 312 is calculated using a linear rectification function applied to each value 310 in the feature map 308. The rectified value calculated for the value 310 is indicated by 314. A linear rectification function (also called a rectifier) ​​is an activation function defined as the positive part of its parameter.

[0083] Using a transposed convolutional layer, an enhanced fault signal is determined 316. For example, the transposed convolutional layer again uses 64 filters of size 10 values. Figure 3 318 of the transposed convolutional layer. Using the filter 318, 10 values ​​320 are calculated for each value 314 in the rectified feature map 312. This step is repeated for each value in the rectified feature map 312. The filter 304 of the transposed convolutional layer preferably uses the same weights as the filter 318 of the convolutional layer. For this purpose, a process called weight sharing is used, which is indicated by the reference numeral 322.

[0084] The enhanced fault signal 316 is then passed to block 302 , which repeats the steps described above, however, it may use a smaller number of filters (e.g., 16 in this example). Thus, block 302 takes the enhanced fault signal 316 as input and outputs the true fault signal 102 .

[0085] Based on the real fault signal 102 , the cost function 324 is calculated as:

[0086]

[0087] in:

[0088] λ1 and λ2 are hyperparameters,

[0089] SEN θ is an unsupervised neural network 104,

[0090] x is the vibration signal 100,

[0091] θ is a set of parameters, including SEN θ The weights and biases of , where θ is adjusted during unsupervised learning,

[0092] d() is a distance function configured to determine the difference between the true fault signal 102 and the vibration signal 100, and

[0093] imp() is a function configured to determine an impulse component in the real fault signal 102 .

[0094] Here, the function d() describes a suitable distance metric (e.g., the Euclidean norm or a kernel-based distance) such that minimizing it ensures that the reconstructed signal remains correlated with the original signal and does not become arbitrary. d() may also include more complex transformations or additional neural networks in cases where a simple distance metric does not seem to be adequate. The second term imp() may be replaced by any impulsiveness measure, and maximizing it is equivalent to finding the minimum of its negative value. The log(cosh) function provides a more stable approximation of kurtosis.

[0095] In this implementation, the following parameters / functions are used:

[0096] λ1=λ2=1

[0097] d(SEN θ (x),x)=‖SEN θ (x)-x‖2

[0098] imp(SEN θ (x)) = log(cosh(SEN θ (x)

[0099] Now, as indicated by arrow 326, the parameters (such as weights and biases) of the neural network 104 (such as the weights of the filters of the convolutional layers in block 302) are adjusted. In particular, the weights in the filters of the convolutional layers in block 300 are adjusted based on the (new) weights used in block 302 using back-propagation, as indicated by reference numeral 328.

[0100] Using the adjusted parameters, a new true fault signal 102 is calculated and the cost function 324 is evaluated again. This process is repeated for, for example, 100 iterations. Then, in the spectrum analysis 106, the true fault signal 102 for which the cost function is found to be minimum is used. This may be the parameter set corresponding to the last iteration, but may also be a previous iteration.

[0101] In one embodiment, the cost function 324 is defined as follows:

[0102]

[0103] Here, the difference with respect to the previous cost function is that the (total raw) vibration signal 100 corresponds to D. More generally, D specifies the set of training data used to train the network 104. Furthermore, D comprises a plurality of signal portions x of fixed length from the same machine 700 recorded during operation.

[0104] Figure 4 An original vibration signal 100 recorded from a bearing with an outer race fault and a true fault signal 102 after application of a neural network 104 are shown, with amplitude plotted on the y-axis and frequency (with respect to Figure 5 and Figure 6 The real fault signal 102 is obviously more impulsive than the original vibration signal 100, which indicates that the fault characteristics are more obvious.

[0105] Figure 5 The difference 500 between the original vibration signal 100 and the true fault signal 102 is depicted, the difference representing the amount of noise filtered by the neural network 104 .

[0106] Figure 6 The envelope spectrum 604 of the original vibration signal 100 and the envelope spectrum 602 of the real fault signal 102 are shown. In order to diagnose the outer race fault, clear peaks at the characteristic frequency BPFO (ball passing frequency outer race - indicated at 600) and its harmonics (2xBPFO, 3xBPFO, ... - indicated at 600-1, 600-2) need to be visible. As can be clearly seen, for the real fault envelope spectrum 602, the corresponding peaks 606, 606-1, 606-2 (also referred to as "significant frequencies" in the present invention) are significantly stronger, while the envelope spectrum 604 of the original vibration signal 100 is quite noisy. The experiment and data confirm that the proposed method significantly facilitates the detection of, for example, outer race faults.

[0107] Although the invention has been described in terms of preferred embodiments, it will be apparent to those skilled in the art that modifications are possible in all embodiments.

Claims

1. A method for detecting a fault in a machine (700), comprising the following steps: a) receiving (S1) a vibration signal (100) recorded by one or more sensors (706) sensing vibrations of the machine (700); b) determining (S2) a true fault signal (102) by applying a neural network (104) to said vibration signal (100); and c) applying (S3) envelope spectrum analysis (106) to said real fault signal (102) to detect said fault, Wherein, the neural network (104) in step b) is an unsupervised neural network, The unsupervised neural network (104) is designed to determine the real fault signal (102) so that an impulse component in the real fault signal (102) is maximized. The unsupervised neural network (104) is designed to determine the true fault signal (102) such that a difference between the true fault signal (102) and the vibration signal (100) is minimized.

2. The method according to claim 1, wherein: The unsupervised neural network (104) is designed such that: in, λ1 and λ2 are hyperparameters, SEN θ is the unsupervised neural network (104), x is the vibration signal (100), θ is a set of parameters, including SEN θ The weights and biases of , where θ is adjusted during unsupervised learning, d() is a distance function configured to determine the difference between the true fault signal (102) and the vibration signal (100), and imp() is a function configured to determine the impulse component in the real fault signal (102).

3. The method according to claim 2, wherein: The function d() is designed to determine the L1 norm, the Euclidean norm, the distance in the spectrum and / or the kernel-based distance.

4. The method according to claim 2 or 3, wherein: Function d() is another neural network.

5. The method according to any one of claims 2 to 4, wherein: The function imp() is designed to determine kurtosis, entropy, or non-Gaussianity.

6. The method according to any one of claims 1 to 5, wherein: The unsupervised neural network (104) is a deep, convolutional or recurrent neural network.

7. The method according to any one of claims 1 to 6, wherein: The envelope spectrum analysis (106) includes: A vibration envelope spectrum (602) of the real fault signal (102) is determined, and one or more significant frequencies (606, 606-1, 606-2) in the vibration envelope spectrum (602) are determined, and the one or more significant frequencies (606, 606-1, 606-2) are associated with a known fault frequency.

8. The method according to any one of claims 1 to 7, wherein: The machine (700) includes a rotating machine element (708).

9. A monitoring system (750) for detecting a fault in a machine (700), comprising: A receiving unit (752) for receiving a vibration signal (100) recorded by one or more sensors (706) for sensing vibration of the machine (700); A determination unit (754) for determining a true fault signal (102) by applying a neural network (104) to the vibration signal (100); as well as An application unit (756) is used for applying envelope spectrum analysis (106) to the real fault signal (102) to detect the fault.

10. A computer program product comprising instructions which, when said program is executed by a computer, cause said computer to perform the method according to any one of claims 1 to 8.

Images