Method for providing physically interpretable fault information of a bearing by a fault detection model
By mapping sensor data from the input domain to the physical meaning domain and performing feature attribution, the problem of uninterpretable bearing fault detection models in existing technologies is solved, achieving physical interpretability and accurate prediction of bearing faults.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SIEMENS AG
- Filing Date
- 2023-08-01
- Publication Date
- 2026-05-19
AI Technical Summary
Existing technologies struggle to effectively detect bearing faults using machine learning models, and traditional signal processing methods lack flexibility and physical interpretability, resulting in models that are uninterpretable and difficult to understand.
By mapping sensor data from the input data domain to a selected physical meaning data domain, an extended fault detection model is used and feature attribution is performed to provide physically interpretable information about bearing faults.
The physical interpretability of the machine learning model has been achieved, enabling it to identify the root causes of bearing failures and provide reliable predictions, thereby improving the interpretability and accuracy of fault detection.
Smart Images
Figure CN119816717B_ABST
Abstract
Description
[0001] This disclosure relates to a fault detection device and a computer-implemented method for providing physically interpretable fault information of bearings embedded in a machine through a fault detection model.
[0002] Sensors are ubiquitous in various heavy machinery, motors, and similar equipment. A particularly important application area for sensors is the detection of bearing failures in rotating machinery such as motors, turbines, and pumps. Bearing failures are one of the most common causes of failure in rotating equipment. These failures can be detected in vibration patterns. To obtain the necessary information, sensors such as position sensors, velocity sensors, accelerometers, and spectral emission energy sensors can be mounted directly on the bearings or on these machines. This allows for the extraction of vibration frequencies and amplitudes, yielding measurements that can be used for vibration analysis. If the bearing suffers from various forms of damage, geometric defects, or faults, the sensor values often indicate suspicious patterns and anomalies. While various tools and methods derived from physical theory exist in principle that would allow this information to be obtained from sensor measurements, this remains a very challenging task.
[0003] However, several challenges remain related to detecting bearing failures from sensor data. Because time-series sensor data are only indirect measurements of the real physical mechanisms and have highly complex data structures, the machine learning models trained to detect bearing failures are themselves complex and therefore often incomprehensible to humans, frequently referred to as black-box algorithms. This means it's impossible to understand how the model behaves, why it predicts bearing failure, and therefore whether the algorithm's results align with physical reasoning. This presents a barrier for data scientists / model developers to build better and more robust models, and for domain experts (such as engineers or operators) to understand and trust the model results. This also leads to problems in detecting the root causes of bearing failures (e.g., bearing or crankshaft failures), and consequently, a lack of acceptance from users.
[0004] Signal processing methods used for detecting faults in bearings are standard in application and well-founded in theory. Identified faults in bearings can occur in different parts of the bearing. In a bearing, a fault can occur in the outer ring, inner ring, cage, or balls of a ball bearing. Different methods and formulas are needed to describe the physical relationship between measured sensor data and individual fault types. Because these formulas are typically based on bearing properties and rotational speed to describe the damage frequencies of different types of bearing damage, different phenomena can be physically explained. To perform fault diagnosis on these different types of damage in acceleration data, it is important to know these physical properties of the bearing and the rotational speed at which the bearing was operating when the data was recorded. Because some of the fault effects are amplitude-modulated and masked by resonance effects in the vibration spectrum, different preprocessing steps are applied to the recorded raw data to reveal specific fault frequencies.
[0005] However, signal processing methods often encounter several problems. For example, the generated vibration spectrum is highly dependent on the installation location of the velocity sensor and the possible load on the machine. Furthermore, noisy sensor signals with multiple confounding factors hinder signal processing.
[0006] Furthermore, machine learning methods have been considered for fault detection in bearings. To train a machine learning algorithm, it is necessary to have access to a sufficient amount of ideally labeled training data containing real vibration signals measured during the actual operation of a machine of interest, with both healthy and defective parts. However, such algorithms are inherently highly complex black-box models. This means that the logic upon which such methods form their decisions is completely unknown.
[0007] Therefore, the purpose of this application is to provide an apparatus and method for fault detection of rolling objects using machine learning methods, which outputs physically interpretable information that may even indicate the root cause of the detected fault. A further specific objective is to improve the interpretability of bearing fault detection by utilizing machine learning algorithms trained on vibration or current data from rotating machinery.
[0008] This objective is achieved through the features of the independent claim. The dependent claims contain further developments of the invention.
[0009] The first aspect relates to a computer-implemented method for providing physically interpretable fault information of bearings embedded in a machine through a fault detection model, comprising the following steps:
[0010] - Obtain sensor data measured at the bearing as input data associated with the input data domain, and a fault detection model that has been trained on the sensor data associated with the input data domain to output a predicted fault value for the bearing by processing the obtained sensor data.
[0011] - The measured sensor data is mapped from the input data domain to a selected data domain, generating an augmented fault detection model. This model outputs augmented predicted fault values that are correlated with the selected data domain, rather than the input data domain, where the selected data domain has physical meaning for the bearing fault.
[0012] - Perform feature attribution on the augmented fault detection model based on the acquired sensor data to quantify the importance of at least one individual feature of the input data to the augmented fault value associated with the selected data domain, and
[0013] - Display individual features and their corresponding quantified importance in the selected data domain at the user interface.
[0014] This method is based on a "traditional" fault detection model that associates an input signal, such as sensor data measured over time, with a predicted fault value, e.g., whether the fault exists (value 1) or does not exist (value 0). Feature attribution applied to the predicted fault value provides data points in the input data domain (e.g., the time domain). These data features include at least one, but in most cases, several adjacent sensor data points. These data features in the input domain do not provide information about the underlying specific fault. The quantified data features in the input data domain neither provide clues to the root cause of the predicted fault value nor are they interpreted in a manner consistent with accepted physical theory. In contrast, feature attribution performed on an extended fault detection model associated with a selected data domain provides features from the selected data domain, not those from the input data domain. The extended predicted fault value generated from the extended fault detection model is equivalent to, and in particular, even identical to, the predicted fault value generated from the obtained fault detection model. The selected data domain has physical meaning for the bearing fault and is therefore interpretable, for example, comparable to the typical failure frequency of the bearing under consideration.
[0015] In an embodiment of this method, the domain mapping consists of multiple cascaded domain mappings.
[0016] Such cascaded domain mappings can simulate / reflect different domain transformations and thus provide greater flexibility for physically interpretable faults in another domain that serves as the input data domain, in which the fault detection model is trained.
[0017] In an embodiment of the method, at least one domain mapping is performed by applying a reversible bijective transformation function to the measured input data.
[0018] This ensures a unique and explicit mapping between the output of the fault detection model in the input data domain (i.e., the predicted fault value) and the output of the augmented fault model in the selected data domain (i.e., the augmented fault value). The predicted fault value and the augmented predicted fault value are equivalent, meaning they have the same value.
[0019] In embodiments of this method, feature attribution is performed using any model-agnostic feature attribution method applicable to the type of machine learning model used for the fault detection model.
[0020] This ensures that the data features most relevant to the output of the augmented fault detection model can be analyzed independently on the augmented learning model and depend only on the "original" fault detection model.
[0021] In embodiments of this method, the fault detection model is a deep neural network, particularly an autoencoder, a convolutional neural network, or a deep belief network.
[0022] Deep neural networks can learn extremely complex patterns, and they are flexible and capable of handling high-dimensional and complex sensor input data.
[0023] In embodiments of the method, the sensor data is vibration data or current data measured at or near the bearing.
[0024] Vibration data is particularly indicative of bearing failure, as defects in the bearing, such as defects in the balls, can cause periodic interruptions. Machines with defective bearings typically require more current than those without defects.
[0025] In an embodiment of the method, the sensor data measured at the bearing is measured in the time domain and mapped to the frequency domain.
[0026] Sensors primarily measure physical parameters of a machine that change over time, and therefore can be used in the time domain in most cases. Since bearing failure causes periodic interference in the sensor data measured over time, the frequency domain of the signal envelope appears to be the most probable domain for physical interpretation.
[0027] In an embodiment of the method, if the importance of quantification of the extended fault detection model is detected at a predefined frequency, an alert is automatically output to the user interface, the predefined frequency being related to the root cause of the bearing.
[0028] In such situations, the alarm draws the operator's attention, indicating a high probability of bearing failure. This allows for rapid action at the machine, such as changing settings, stopping the machine, or scheduling maintenance work.
[0029] In an embodiment of this method, the quantitative importance of the output is displayed using color codes related to the quantity.
[0030] This helps identify key or highly relevant values in a continuous spectrum, such as importance values. Color codes can indicate different numbers of values using different colors of a predefined color scale or by different intensities of a single color.
[0031] In an embodiment of the method, the bearing fault detection model is trained on the signal envelope of sensor data measured in the time domain and analyzed in the frequency domain for a specific fault frequency.
[0032] The signal envelope is highlighted as the time distance between two peaks in the sensor data measured in the time domain. Bearing characteristic failures depend on parameters such as the diameter of the balls or the pitch circle diameter of the bearing, and generate interference at different time intervals. This interference is directly related to the frequency, which depends on the values of these parameters.
[0033] In an embodiment of this method, domain mapping is performed using a Fourier transform function.
[0034] The Fourier transform or fast Fourier transform function is well-known and requires very little processing power.
[0035] In embodiments of the method, the machine is a rotating machine, particularly an electric motor, turbine, pump, or press.
[0036] The second aspect relates to a fault detection device for providing physically interpretable fault information of a bearing embedded in a machine through a fault detection model, the fault detection device comprising at least one processor configured to perform the following steps:
[0037] - Obtain sensor data measured at the bearing as input data associated with an input data domain, and a fault detection model trained on the sensor data associated with the input data domain to output a predicted fault value for the bearing by processing the obtained sensor data.
[0038] - The measured sensor data is mapped from the input data domain to a selected data domain, thereby generating an extended fault detection model. This extended fault detection model outputs an extended predicted fault value related to the selected data domain, where the selected data domain has physical meaning for the bearing fault.
[0039] - Perform feature attribution on the augmented fault detection model based on the acquired sensor data to quantify the importance of at least one individual feature of the input data to the augmented fault value associated with the selected data domain, and
[0040] - Display individual features and their corresponding quantified importance in the selected data domain at the user interface.
[0041] The third aspect relates to a computer program product that can be directly loaded into the internal memory of a digital computer, including a software code portion for performing the steps as described above when the product is run on the digital computer.
[0042] The invention will be explained in more detail with reference to the accompanying drawings. Similar objects will be labeled using the same reference numerals.
[0043] Figure 1 An embodiment of the computer-implemented method of the present invention is illustrated by a flowchart.
[0044] Figure 2 The interactions between the various processing steps are illustrated schematically in more detail.
[0045] Figure 2A The diagram schematically illustrates the domain mapping and the resulting extended fault detection model.
[0046] Figure 2B The characteristics of the transformation function applied to a domain mapping are illustrated schematically.
[0047] Figure 3A The output of the fault detection model in the input domain is illustrated schematically.
[0048] Figure 3B The output of an embodiment of the extended fault detection model in the selected domain is illustrated schematically.
[0049] Figure 4 The diagram schematically illustrates a comparison between the output of feature attribution from the fault detection model in the input domain and the output of feature attribution from the extended fault detection model in the selected domain.
[0050] Figure 5 An embodiment of the fault detection device of the present invention is illustrated as a block diagram.
[0051] Note that in the detailed description of the embodiments below, the drawings are merely schematic, and the elements illustrated are not necessarily shown to scale. Rather, the drawings are intended to illustrate the function and cooperation of components. It is to be understood here that any connection or coupling of functional units, devices, components, or other physical or functional elements may also be achieved by directly or indirectly connecting coupling elements (e.g., via one or more intermediate elements). Connections or couplings of entities or components may be achieved, for example, based on wired, wireless, and / or a combination of wired and wireless connections. Functional units may be implemented by dedicated hardware (e.g., processor, firmware) or by software, and / or by a combination of dedicated hardware, firmware, and software. Further note that each functional unit described for the apparatus may perform functional steps of the associated method, and vice versa.
[0052] First, a description of standard methods for detecting bearing failures is provided. These methods are standard in application and well-founded in theory. However, they often suffer from a number of problems. Examples include the (too) noisy raw signal of sensor data and multiple confounding factors. These standard methods lack flexibility and cannot handle high-dimensional, complex sensor data. Since this method is based on a physical derivation of the bearing behavior during failure conditions, it will be described in detail below.
[0053] Localized failures in bearings (especially rolling element bearings) can occur in different parts of the bearing, either in the outer ring, inner ring, cage, or rolling elements. Different approaches are required depending on the type of failure. For the formulas below, we will use the following symbols: d is the bearing ball diameter, D is the pitch circle diameter, f... r Φ is the shaft speed, n is the number of rolling elements, and Φ is the bearing contact angle.
[0054] Inner ring damage is caused by irregularities on the inner ring of the bearing. When the rolling elements impact this fault, an shock is introduced, resulting in high-frequency resonance. The envelope spectrum shows this fault at the BFPI frequency:
[0055]
[0056] The outer ring damage was caused by irregularities on the outer ring of the bearing. The envelope spectrum shows this fault at the BFPI frequency:
[0057]
[0058] Wear or deformation will cause the cage to move from its center position. This creates an unbalanced force, which results in impact pulses.
[0059]
[0060] The damaged rolling element periodically contacts the inner and outer rings of the bearing, generating impact signals. Sidebands are expected because the rolling element rotates about itself and simultaneously experiences relative movement through the cage. The envelope spectrum shows peaks at the following locations:
[0061]
[0062] Because these formulas describe the failure frequencies of different types of bearing damage based on bearing physical properties and rotational speed, they can physically explain different phenomena. To perform fault diagnosis on these different types of bearing damage from acceleration data, it is important to know these physical properties of the bearing and the rotational speed at which the bearing was operating when the data was recorded.
[0063] Because some of the fault effects are amplitude-modulated and masked by resonance in the vibration spectrum, different preprocessing steps are applied to the recorded raw data to reveal specific fault frequencies. It should be noted that the resulting vibration spectrum is highly dependent on the installation location of the velocity sensor and the possible load on the machine.
[0064] The signal processing-based method applies a bandpass filter to the acceleration signal, which includes bearing operating noise. The envelope of the pre-filtered signal is then calculated and transferred to the frequency domain.
[0065] By calculating the envelope, the amplitude-modulated damage fault signal can be demodulated, and the resulting envelope spectrum shows different bearing faults in the form of peak values at the characteristic frequencies BPFI, BPFO, FTF and BSF as defined above.
[0066] The common method for obtaining the envelope spectrum is to calculate and analyze the signal, and then transfer it to the frequency domain. Since strong background noise (such as pulsed electromagnetic noise and periodic harmonic noise generated by shaft rotation) has a significant impact on the selection of the resonant frequency band, it is important to correctly select the upper and lower limits of the bandpass filter and an appropriate cutoff frequency for the center frequency.
[0067] To achieve this, a kurtogram is calculated to determine the frequency band with the highest signal-to-noise ratio. The kurtogram shows the spectral kurtosis at different window widths and center frequencies. Kurtosis is a measure of the "tail" of the probability distribution of a real-valued random variable.
[0068]
[0069] Because of the fourth power, the pulse deviates from the mean, resulting in a large kurtosis value.
[0070] Selecting the center frequency and bandwidth from the kurtosis plot with the highest peak value is a promising approach for the bandpass filter being applied. As mentioned above, this method requires detailed knowledge of the bearings being installed and a significant amount of manual work by a domain expert.
[0071] In the following text about Figure 1 Embodiments of the method of the present invention are described, and regarding Figure 2 and Figure 3A / B provides a more detailed explanation of embodiments of the method of the present invention.
[0072] First step S1 (see...) Figure 1 This is used to provide physically interpretable fault information about bearings built into the machine, i.e., to obtain sensor data measured at the bearing (see [reference]). Figure 2 The input data 10, associated with the input data domain, and the fault detection model 11, are trained on sensor data associated with the input data domain. Preferably, the sensor data 10 is vibration data or current data measured at or near the bearing. The sensor data 10 measured at the bearing is measured in the time domain. The output of the fault detection model 11 is a predicted fault value 12 of the bearing obtained by processing the obtained sensor data 10.
[0073] The measured sensor data 10 is mapped from the input data domain to a selected data domain, thereby generating an extended fault detection model that outputs an extended predicted fault value related to the selected data domain, see step S2. The selected data domain has physical meaning for the bearing fault and is therefore a semantic representation of the sensor data. For bearing monitoring, the mapping is preferably performed from time, which is the input data domain, to frequency, which is the selected data domain. Figure 2 In the figure, the extended fault detection model is described by reference symbol 13.
[0074] Feature attribution 14 is performed on the expanded fault detection model 13 to quantify the importance of at least one individual feature of the input data in the selected data domain to the expanded fault value, wherein the expanded fault value is equivalent to or even the same as the fault value of the obtained fault detection model, see step S3. Feature attribution is performed using any model-independent feature attribution method applicable to the fault detection model.
[0075] In the final step S4, the individual characteristics of the input data and the corresponding quantified importance in the selected data domain are displayed at the user interface. Figure 2An example of the output is shown in Figure 15. The color of the vertical line 16 indicates the importance of the frequency feature to the predicted fault value output by the fault detection model. The frequency feature is a dedicated frequency or a frequency band consisting of several subsequent frequencies that collectively contribute to the indicated importance. Based on predefined critical values related to the root cause of the bearing in the selected data domain, if the quantified importance of the extended fault detection model is detected at the predefined critical values, an alert is automatically output to the user interface. Preferably, the quantified importance of the output is displayed using color codes related to the quantity.
[0076] Figure 2A and Figure 2B The domain mapping and the resulting extended fault detection model 13 are shown in more detail. This is achieved through the transformation function... The data is applied to the sensor data 10 being measured and by using the inverse form of the transformation function. Domain mapping is performed on fault detection model 11. Therefore, the obtained sensor data 10 associated with the input data domain t is transferred to sensor data 101 in the selected data domain f. The inverse form of the transformation function applied to fault detection model 11... An extended fault detection function 13 is generated. Therefore, the obtained sensor data 10 mapped to the selected data domain is input into the extended fault detection model 13, which outputs an extended fault value 12. The input data in the selected data domain, the extended fault detection model 13, and the extended fault value 12 are input into feature attribution 14. Feature attribution 14 outputs a quantification of the importance of features in the obtained sensor data 101 in the selected data domain.
[0077] Figure 2B The transformation function is shown in more detail. Transformation function It is a reversible bijective transform function. This means that the transform function applied to the sensor data 100 obtained in the input data domain x is... The output is the sensor data 101 in the selected data domain z. On the other hand, the inverse form of the transformation function applied to the sensor data 101 in the selected data domain z is also given. The output is sensor data 100 in the input data domain x. The input data domain x can be any parameter, for example, such as... Figure 2A The time t shown is used. The selected data domain z can be any parameter, for example, such as... Figure 2A The frequencies shown.
[0078] In many operating conditions, sensor data is measured over time. Typically, one or more sensors detect bearing vibration based on the acceleration of the entire bearing or a portion of it. Another parameter measured to deduce bearing defects is the machine's current data. Sensors are usually located in parts of the machine close to the bearing. Figure 2 The obtained sensor data 10 shown provides the acceleration value a measured over time t. Time is the input data domain.
[0079] The machine learning model used for fault detection, i.e., the fault detection model 11, can be viewed as a function f that maps the input sensor data 10 to an output that provides a predicted fault value (e.g., a decision on whether a fault exists (value 1) or does not exist (value 0)). θ :R N →{0, 1}. The fault detection model is trained with a sufficient amount of ideally labeled training data, which includes actual sensor data, such as vibration signals measured over time in the input data domain during the actual operation of a machine of interest with healthy and defective parts.
[0080] The fault detection model 11 is preferably a deep neural network, especially an autoencoder, a convolutional neural network, or a deep belief network capable of learning extremely complex patterns. Most methods utilizing deep neural networks have shown good performance on test data, are flexible and capable of handling high-dimensional and complex sensor data, and their predictions are reliable. However, such models are inherently very complex black-box models. This means that it is completely unclear what logic such models are based on to form their decisions.
[0081] A feature attribution approach is presented to explain black-box machine learning models. This method quantifies the extent to which individual input data features contribute to the model's final predicted fault value. Input data features include one or more neighboring data points. The machine learning model's decision, i.e., predicting the fault value, can be viewed as applying d-dimensional sensor data x∈R... d Mapped to a function f: R that represents the real number f(x) of the predictive model decision. d →R. In the case of bearing fault detection, f(x) can, for example, indicate the logistic value or corresponding probability of the presence of a defect estimated by the fault detection model. The goal of feature attribution is to identify the importance vector φ∈R. d , making Φ i Quantize each input feature x i The importance of model prediction f(x) for a fixed input x.
[0082] To date, any such feature attribution method will retrieve feature attribution on a per-sensor data feature x, i.e., on a per-input data domain of sensor data 10. Therefore, for example, if fault detection model 11 is trained to classify sensor data in the time domain, then Φ will specify the importance of sensor data features in the time domain, such as... Figure 3A As depicted in the diagram, sensor data features with high importance for predicted bearing fault values are marked by dashed lines 21, while those with low importance are marked by dotted lines 22. Importance can also be indicated by color, where the importance score is encoded by the color scale and / or intensity of lines 21 and 22. In this case, where the fault detection model primarily focuses on frequency information, displaying feature importance in the time domain would lead to uninterpretable results and may be completely meaningless to domain experts.
[0083] If possible Figure 3A What we see here makes it difficult to infer any specific pattern or find concrete reasons to explain the decisions of the fault detection model. On the other hand, see... Figure 3B If we calculate the value of feature importance in the frequency domain, representing the frequency of the sensor data provided for measurement, it becomes immediately clear that the presence of a single frequency peak (see line 31) has a significant impact on the predicted fault value. Similar to... Figure 3A In this context, the importance values in the selected data domain (here, the frequency domain) are encoded by different structures or lines of different colors at the corresponding frequencies. See [link to relevant documentation]. Figure 3B .
[0084] To achieve feature attribution in a selected domain that differs from the input data domain of the measured sensor data 10, feature attribution is transformed into the selected domain by applying a bijective mapping that captures a reversible one-to-one correspondence between the input data domain of the fault detection model 11 and the selected domain. For example, the Fourier transform can be viewed as a mapping from the time domain to the frequency representation, which is also reversible.
[0085] In mathematics, a field mapping to D is defined by an invertible function. This specifies that there is another function. Make Domain mappings can also consist of multiple cascaded mappings, where all of these mappings are invertible. Because The goal is to transform the sensor data features in the input domain into a more meaningful chosen domain; we will... This is called an interpretable or semantic representation of x. If an appropriate domain mapping function is specified... This allows it to be combined with feature attribution methods to compute feature importance values based on a chosen domain (i.e., semantic representation z) rather than on sensor data features x. More specifically, It can be used to create extended models
[0086] in
[0087] Any model-independent feature attribution method can now be used in Instead of evaluating on f to produce significance values in the chosen domain, this is correct because model-agnostic methods are designed to work with any machine learning model. If a model-specific attribution method is to be applied, then it needs to be examined... Does the method still satisfy the necessary assumptions (e.g., differentiability), or does the method need further adjustments to suit such models (e.g., new LRP-rules)?
[0088] Semantic interpretation for bearing fault detection
[0089] Defects in bearings cause fault signals, i.e., measured sensor data, which in turn have an amplitude modulation effect on a specific carrier signal. Domain experts can detect such effects by analyzing the envelope spectrum and checking for the presence of specific fault frequencies. This logic is strictly based on a physical understanding of bearing fault defects. If a machine learning model is trained to identify bearing faults from raw or pre-processed signals (i.e., sensor data in the input data domain), the connection to existing domain knowledge about the physics of the problem may be overlooked or at least unknown. This is especially true for machine learning models, particularly deep neural networks, which have proven successful in bearing fault detection tasks.
[0090] Such models may exploit any potential characteristics of sensor data to make decisions based on it, and some of these characteristics may be spurious. This can lead to overfitting and potentially poor model performance in deployment. To prevent this and ensure a reliable model with high predictive quality, it is necessary to verify the extent to which the model follows the physical underlying routines of domain experts. More precisely, in the case of fault detection of bearings installed in a machine, if one wants to examine which features are important to the fault detection model 11 via a feature attribution method, it would be ideal to obtain values for feature importance based on the frequency components of the envelope spectrum of sensor data measuring vibrations of the machine near the bearing. This information would be readily available to domain experts and would make it easy to check whether the fault detection model conforms to the physical understanding of bearing faults.
[0091] In the following text, domain mappings are provided for three common scenarios of the bearing fault detection model (FD model) based on the type of sensor data measured in the input domain of the input data 10 used to train the obtained fault detection model 11, and also in Table 1 below.
[0092] First, the fault detection model 11 is trained on the signal envelope in the time domain, which is given by the amplitude of the analyzed signal (e.g., measured sensor data). In that case, the applied domain mapping... It is a Fourier transform, therefore
[0093]
[0094] Secondly, a fault detection model 11 is trained on sensor data in the time domain. To derive the signal envelope of the sensor data from the time-domain signal x, this signal envelope is calculated by analyzing the amplitude of the signal. This means that the signal envelope x... env In mathematics, x env =|x+iHT(x)| is given, where HT is similar to the Hilbert transform. It can now be obtained via FT(x) env The desired envelope spectrum is calculated. The goal is to find a domain mapping that derives the envelope spectrum from the time-domain signal and is invertible. However, absolute values will violate this requirement. To avoid this problem, it is also necessary to preserve the phase information of the analyzed signal by calculating its independent variable arg(x+iHT(x)). This yields the following domain mapping:
[0095] in
[0096] Its reverse This ultimately allows us to obtain an effective semantic representation of the time-domain signal based on its envelope spectrum.
[0097] Finally, the fault detection model 11 is trained on the sensor data in the frequency domain. In this case, the sensor data is first transformed to the time domain using the inverse Fourier transform, and then the above-specified... It is applied to mapping from the time domain to the frequency domain.
[0098] The table below summarizes the different domain mappings from different input data domains to the frequency domain as the chosen data domain, and their inverse transformations.
[0099]
[0100] Table 1
[0101] Thus, the proposed method provides a tool to assess the degree of consistency between a machine learning model trained to detect bearing failures and existing prior knowledge about the underlying physics. The domain mapping φ can be combined with existing feature attribution methods to estimate the extent to which the model has already utilized the presence of characteristic failure frequencies. Such information is readily available to domain experts compared to uninterpretable importance values in the input domain (i.e., also known as raw data, which are generated solely by feature attribution methods). This is in... Figure 4 Visualization in Chinese.
[0102] This is Figure 4 Visualized in the middle. The fault detection model was trained on sensor data signals measured in the time domain and has already detected bearing faults in the presented signals. On the left, the importance values 41 and 42 of the feature attribution evaluated in the time domain are depicted. Again, it is difficult to infer any useful information about the potential causes of bearing faults. In particular, it is unclear whether the fault detection model is consistent with existing domain knowledge, i.e., that a specific peak in the envelope spectrum should be present in the case of a fault. However, if feature attribution is computed based on semantic representation, i.e., the chosen data domain related to the physical interpretation, see [reference needed]. Figure 4 On the right-hand side, it can be verified that the fault detection model places great emphasis on the relevant fault frequencies indicated by the vertical line 46. At least one importance value 45 indicating high importance coincides with the relevant fault frequency 46, while the importance values 44 indicating low importance are clearly separated. This information is extremely useful to domain experts and can be used to validate or improve the fault detection model accordingly.
[0103] Figure 5 An embodiment of a fault detection device 50 is shown. The fault detection device 50 includes a data interface 51 configured to receive sensor data measured at a bearing mounted on machine 40 as input data. The input data is associated with an input data domain. Furthermore, a fault detection model is obtained via the data interface 51. The fault detection model is trained on the sensor data associated with the input data domain to output a predicted fault value for the bearing by processing the obtained sensor data. Machine 10 is a rotating machine, particularly an electric motor, turbine, pump, and press.
[0104] The fault detection device 50 includes a data mapping unit 52 configured to map measured sensor data from an input data domain to a selected data domain, thereby generating an extended fault detection model that outputs an extended predicted fault value related to the selected data domain. The selected data domain is configured to have physical meaning for bearing faults.
[0105] The fault detection device 50 includes a feature attribution unit 53 configured to perform feature attribution on an augmented fault detection model, thereby quantifying the importance of at least one individual feature of the input data to the augmented fault value associated with a selected data domain.
[0106] The fault detection device 50 includes a user interface 54, which is configured to display individual characteristics of the input data and the corresponding quantitative importance in the selected data domain.
[0107] It should be understood that the descriptions of the above examples are intended to be illustrative, and the illustrated components are readily adaptable to various modifications. For example, the illustrated concepts can be applied to different technical systems, and in particular to different subtypes of the corresponding technical systems with only minor adjustments.
Claims
1. A computer-implemented method for fault detection, used to provide physically interpretable fault information of bearings embedded in a machine via a fault detection model (11), the computer-implemented method comprising the following steps: - Obtain (S1) sensor data (10) measured at the bearing as input data associated with the input data domain, and a fault detection model (11) trained on the sensor data associated with the input data domain to output a predicted fault value (12) of the bearing by processing the obtained sensor data (10). - The measured sensor data (10) is mapped (S2) from the input data domain (t) to the selected data domain (f), and an extended fault detection model (13) is generated, which outputs an extended predicted fault value associated with the selected data domain, wherein the selected data domain has physical meaning for the bearing fault. - Feature attribution (14) is performed (S3) on the extended fault detection model (13) of the obtained sensor data (10), thereby quantifying the importance of at least one individual feature of the input data to the extended predicted fault value associated with the selected data domain, and - Display (S4) the individual features and their corresponding quantification importance in the selected data domain at the user interface.
2. The computer-implemented method according to claim 1, wherein the mapping (S2) consists of a plurality of cascaded domain mappings.
3. The computer-implemented method according to claim 1 or 2, wherein at least one mapping (S2) is performed by applying a reversible bijective transformation function to the measured sensor data.
4. The computer-implemented method according to claim 1 or 2, wherein the features Attribution (14) is performed using any model-independent feature attribution method applicable to the fault detection model (11).
5. The computer-implemented method according to claim 1 or 2, wherein the fault detection model (11) is a deep neural network.
6. The computer-implemented method according to claim 1 or 2, wherein the fault detection model (11) is an autoencoder, a convolutional neural network or a deep belief network.
7. The computer-implemented method according to claim 1 or 2, wherein the sensor data (10) is vibration data or current data measured at or near the bearing.
8. The computer-implemented method according to claim 1 or 2, wherein the sensor data (10) measured at the bearing is measured in the time domain and a mapping to the frequency domain is performed (S2).
9. The computer-implemented method according to claim 1 or 2, wherein if the quantitative importance of the extended fault detection model is detected at a predefined frequency, an alarm is automatically output to the user interface, the predefined frequency being related to the root cause of the fault in the bearing.
10. The computer-implemented method according to claim 1 or 2, wherein the quantitative importance of the output (16) is displayed in color codes relating to the quantity.
11. The computer-implemented method according to claim 1 or 2, wherein the fault detection model (11) is trained on the signal envelope of sensor data (10) measured in the time domain and analyzed for a specific fault frequency in the frequency domain.
12. The computer-implemented method of claim 10, wherein the mapping is performed via a Fourier transform function (S2).
13. The computer-implemented method according to claim 1 or 2, wherein the machine is a rotating machine.
14. The computer-implemented method according to claim 1 or 2, wherein the machine is an electric motor, turbine, pump, or press.
15. The computer-implemented method according to claim 3, wherein mapping (S2) is performed by applying a transformation function to the measured sensor data (10) and by applying an inverse transformation function to the fault detection model (11).
16. A fault detection device (50) for providing physically interpretable fault information of a bearing embedded in a machine (40) via a fault detection model (11), the fault detection device (50) comprising at least one processor configured to perform the following steps: - Obtain sensor data (10) measured at the bearing as input data associated with the input data domain, and a fault detection model (11) trained on the sensor data associated with the input data domain to output a predicted fault value (12) of the bearing by processing the obtained sensor data (10). - The measured sensor data (10) is mapped from the input data domain (t) to the selected data domain (f), and an extended fault detection model is generated, which outputs an extended predicted fault value related to the selected data domain, wherein the selected data domain has physical meaning for the bearing fault. - Perform feature attribution (14) on the augmented fault detection model (13) based on the obtained sensor data, thereby quantifying the importance of at least one individual feature of the input data to the augmented predicted fault value associated with the selected data domain, and - Display individual features and their corresponding quantified importance in the selected data domain at the user interface.
17. A computer program product that can be directly loaded into the internal memory of a digital computer, comprising a software code portion for performing the steps of claims 1-15 when the product is run on the digital computer.