Diagnostic device, diagnostic system, and diagnostic method

Abstract

A control unit (3) of a diagnostic device (10) selects one or more first features (T1) from time series data obtained from operating data about components (42, 43, 44) constituting a rotating machinery system (40) and one or more second features (T2) from frequency data obtained through frequency analysis of the operating data, and diagnoses the state of the rotating machinery system (40) on the basis of the selected first feature(s) (T1) and the selected second feature(s) (F1).

Description

[Technical Field]

[0001] The present disclosure relates to a diagnostic device, a diagnostic system, and a diagnostic method. [Background Art]

[0002] Conventionally, there are abnormality diagnostic devices and abnormality diagnosis methods that diagnose abnormalities in a rotating mechanical system including rotating machinery, power transmission mechanisms, load equipment, etc. using acquired sensor signal information. Such abnormality diagnosis devices and abnormality diagnosis methods require processing of a huge amount of data such as vibration, temperature, pressure, sound, etc. obtained from sensors, which may lead to reduced efficiency and accuracy of diagnosis. Therefore, in order to realize efficient and accurate diagnosis of rotating mechanical systems, the following diagnostic devices using AI (Artificial Intelligence) have been disclosed.

[0003] That is, in a conventional diagnostic device, the current of the rotating machinery to be diagnosed is measured, noise is removed by a sequential cyclic statistical information filter, and the noise-removed frequency-domain current signal or time-domain current signal is used to extract features reflecting the state of the rotating mechanical system to be diagnosed from the current signal through deep learning, thereby identifying abnormalities in the rotating machinery (see, for example, Patent Document 1). [Prior Art Documents] [Patent Documents]

[0004] [Patent Document 1] Japanese Unexamined Patent Publication No. 2021-76564 [Summary of the Invention] [Problem to be Solved by the Invention]

[0005] In the conventional diagnostic device disclosed above, the anomaly diagnosis method for diagnosing abnormalities in rotating machinery removes noise from the measured current signal, extracts features that reflect the state of the rotating machinery from the data obtained from the sensor, and diagnoses the anomaly using deep learning. However, among the multiple features obtained from the data acquired by the sensor, the features that reflect the anomaly differ depending on the type and magnitude of the anomaly, which sometimes leads to a decrease in diagnostic accuracy.

[0006] This disclosure provides technologies to solve the above-mentioned problems, and aims to provide diagnostic devices, diagnostic systems, and diagnostic methods that offer high diagnostic accuracy and efficiency. [Means for solving the problem]

[0007] The diagnostic device disclosed herein is A diagnostic device comprising: an acquisition unit for acquiring operating data of components constituting a rotating machinery system; and a control unit for diagnosing the state of the components constituting the rotating machinery system based on characteristic quantities obtained from the acquired operating data, The control unit, The system includes a first database that shows the correlation between time-series data obtained from the aforementioned operating data and frequency data obtained by frequency analysis of the aforementioned operating data, regarding abnormalities occurring in the rotating machinery system. Based on the first database, the associations made by the first database, From time series data, the aforementioned features Multiple types First feature And, as stated above From frequency data as the aforementioned feature quantity Multiple types Second feature Choose broth, The selected plurality of first features include at least two first features that are distinct from each other, and the selected plurality of second features include at least two second features that are distinct from each other. The selected Multiple types First feature and the above Multiple types A two-dimensional array with the second feature arranged Constructing a feature space A trained model for diagnosing the state of the rotating machinery system is constructed using machine learning with array data. 、 before The recording control unit is, The state of the rotating machine system is diagnosed based on the trained model and the distribution characteristics of multiple types of features, namely the first and second features, in the array data constituting the feature space. It is. The diagnostic system of the present disclosure includes: the diagnostic device configured as described above, and a notification unit connected to the diagnostic device via a network and configured to notify an administrator of a diagnostic result obtained by the diagnostic device, . The diagnostic method of the present disclosure is a diagnostic method using the diagnostic device configured as described above, Based on the aforementioned first database, and associated by the aforementioned first database, selecting, as the feature amounts, a Multiple types first feature amount and from the time-series data of the operation data, and a second feature amount as the feature amount from frequency data obtained by frequency-analyzing the operation data, Multiple types second feature amount and , The selected plurality of first features include at least two first features that are distinct from each other, and the selected plurality of second features include at least two second features that are distinct from each other. constructing a trained model for diagnosing a state of the rotating machinery system by machine learning using two-dimensional array data in which the selected Multiple types first feature amount and the selected Multiple types second feature amount are arrayed, Constructing a feature space and diagnosing the state of the rotating machinery system 、 before based on the trained model and distribution characteristics of a plurality of types of feature amounts including the first feature amount and the second feature amount in the array data constituting the feature space, . Effects of the Invention

[0008] According to the diagnostic device, diagnostic system, and diagnostic method of the present disclosure, a highly accurate and efficient diagnostic device, diagnostic system, and diagnostic method can be obtained. Brief Description of the Drawings

[0009] [Figure 1] FIG. 1 is a block diagram showing a schematic configuration of a diagnostic system including the diagnostic device according to Embodiment 1. [Figure 2] FIG. 2 is a diagram showing a schematic configuration of a rotating machinery system that is a diagnostic target of the diagnostic device according to Embodiment 1. [Figure 3] It is a flow diagram showing the flow of state diagnosis processing performed by the diagnosis device according to the first embodiment. [Figure 4] It is a conceptual diagram showing operation data stored by the feature calculation unit according to the first embodiment. [Figure 5] It is a conceptual diagram showing feature amounts extracted by the feature calculation unit according to the first embodiment. [Figure 6] It is a conceptual diagram showing feature amount conversion processing according to the first embodiment. [Figure 7] It is a diagram showing a detailed configuration of map data according to the first embodiment. [Figure 8] It is a current frequency spectrum obtained by performing FFT analysis on current of a rotating machinery system according to the first embodiment. [Figure 9] It is a conceptual diagram for explaining the AI function in the abnormality diagnosis unit according to the first embodiment. [Figure 10] It is a conceptual diagram for explaining the AI function in the abnormality diagnosis unit according to the first embodiment. [Figure 11] FIG. 11A is a diagram showing an example of accuracy of abnormality diagnosis when no threshold is set. FIG. 11B is a diagram showing an example of accuracy of abnormality diagnosis when a threshold is set. [Figure 12] It is a flow diagram showing another example of the flow of state diagnosis processing performed by the diagnosis device according to the first embodiment. [Figure 13] It is a diagram showing the hardware configuration of a control device of the diagnosis device according to the first embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0010] First Embodiment. FIG. 1 is a block diagram showing a schematic configuration of a diagnosis system 30 including a diagnosis device 10 according to the first embodiment. FIG. 2 is a diagram showing a schematic configuration of a rotating machinery system 40 that is a diagnosis target of the diagnosis device 10 according to the first embodiment. The diagnosis system 30 of the present embodiment diagnoses a state of the rotating machinery system 40 such as an abnormality and notifies an administrator performing maintenance of the diagnosis result.

[0011] First, let me explain the rotating machinery system 40 that is the subject of the diagnosis. As shown in Figure 2, the rotating machinery system 40 is composed of a power supply 41, a rotating machine 42, a power transmission mechanism 43, and load equipment 44.

[0012] Power supply 41 supplies power from the power grid. Alternatively, power supply 41 may convert the power supplied from the power grid to any frequency, any current, or any voltage using an inverter before supplying it. The rotating machinery 42 refers to machines that perform rotational movements, including electric motors, single-phase / three-phase induction machines, synchronous machines, etc., and has the function of converting electrical energy into mechanical energy.

[0013] The power transmission mechanism 43 consists of belts, gears, couplings, etc., and is a mechanism that connects the rotating machine 42 and the load equipment 44. The load equipment 44 includes pumps, fans, compressors, etc., and is powered by mechanical energy converted by the rotating machinery 42. Alternatively, the power transmission mechanism 43 may be omitted, and the rotating machine 42 and the load equipment 44 may be connected directly. The components constituting the rotating machinery system 40 are not limited to those listed above, but may be any equipment related to the operation of the rotating machinery 42.

[0014] Multiple sensors 1 are provided near each component of the rotating machinery system 40 to detect operating data related to the operating state of each component, such as current, voltage, vibration (vibration velocity, vibration acceleration), sound, strain, temperature, pressure, AE (Acoustic Emission), and power. These sensors 1 enable the detection of appropriate operating data for any type of rotating machinery system 40. This detected operating data is input to the diagnostic device 10.

[0015] It should be noted that the operating data acquired by sensor 1 is not limited to the operating data described above, as long as it indicates the state of each component constituting the rotating machinery system 40. The operating data acquired by sensor 1 is affected by the environmental conditions of the rotating machinery system 40. For example, if the load equipment 44 is a pump, the pump may be submerged, in which case diagnosis based on vibration, sound, etc., is difficult. Therefore, in such cases, acquiring operating data such as current and voltage is prioritized over vibration, sound, etc.

[0016] Next, the configuration of the diagnostic system 30 will be described. As shown in Figure 1, the diagnostic system 30 includes a diagnostic device 10 for diagnosing abnormalities in the rotating machinery system 40, and a display unit 20 as a notification unit for informing the administrator of the diagnostic results.

[0017] The diagnostic device 10 comprises a data acquisition unit 2 as an acquisition unit and a control unit 3. The data acquisition unit 2 acquires the operating data detected by the sensor 1. The data acquisition unit 2 also includes a filter function to remove noise from the detected operating data, an AD conversion function, and a memory function to store the data.

[0018] Furthermore, the operating data saved using this memory function may be saved to another external computer via the communication function. Additionally, the memory function may be configured to include both a function to delete saved operating data and a communication function. For example, past operating data, simulation data, etc., may be acquired using the communication function and stored in the memory function.

[0019] The feature calculation unit 4 calculates feature quantities for abnormality detection from each operating data acquired by the data acquisition unit 2, and comprises an extraction unit 4A, a time-series data storage unit 4B, a frequency data storage unit 4C, a conversion unit 4D, a first database DB1, and a second database DB2.

[0020] The anomaly diagnosis unit 5 performs anomaly diagnosis using AI based on the feature quantities calculated by the feature calculation unit 4, and comprises a model generation unit 5A, an AI diagnosis unit 5B, and a prediction unit 5C. Details of each component constituting the above-mentioned feature calculation unit 4 and anomaly diagnosis unit 5 will be described later.

[0021] The display unit 20 is a device that includes a display device such as a display capable of displaying the diagnostic results from the diagnostic device 10 and the set information, and controls the image displayed on the display. The display unit 20 may be integrated with the diagnostic device 10, or it may be a separate unit equipped with a communication function. In this case, the control of the display may be performed by a separate computer connected via the communication function, or by the display unit 20.

[0022] Furthermore, the display unit 20 may have a function to store the diagnostic results information from the diagnostic device 10 and the set information, and may also have a function to organize and search this information in chronological order. For example, the diagnostic results of abnormalities may be displayed as a graph showing a time-series trend.

[0023] In the above example, a display unit 20 with a display was used as the notification unit, but it is not limited to this. The notification unit may use a printer or the like to print on paper to notify the administrator of the diagnostic results and set information on paper, or it may use indicator lights or the like to notify the administrator only of normal and abnormal conditions. Alternatively, the notification unit may be a speaker or the like with an audio output function, as long as it has the function to notify the administrator of the abnormal diagnosis results of the diagnostic device 10.

[0024] Here, an abnormality in the rotating machinery system 40 includes an abnormality in the rotating machinery 42, an abnormality in the power transmission mechanism 43, and an abnormality in the load equipment 44. Examples of abnormalities in the rotating machine 42 include bearing abnormalities, rotor bar abnormalities, eccentricity abnormalities, unbalance abnormalities, layer shorting abnormalities, torque abnormalities, etc. Abnormalities in the power transmission mechanism 43 include, for example, belt loosening and breakage, belt cracking and wear, gear tooth wear, tooth chipping and breakage, etc. Abnormalities in the load equipment 44 include, for example, in the case of a pump, cavitation abnormalities, air entrapment abnormalities, seal wear and damage, foreign matter contamination, bearing abnormalities, etc. Furthermore, abnormalities in the rotating machinery system 40 include not only the abnormalities described above, but also, for example, abnormalities in fans, compressors, etc.

[0025] The following describes the diagnostic process performed by the diagnostic device 10 to diagnose abnormalities and other conditions in the rotating machinery system 40. Figure 3 is a flowchart showing the flow of the condition diagnosis process performed by the diagnostic device 10 of this embodiment. Figure 4 is a conceptual diagram showing the operating data stored in the time-series data storage unit 4B and frequency data storage unit 4C of the feature calculation unit 4 of this embodiment. Figure 5 is a conceptual diagram showing the feature quantities extracted by the extraction unit 4A of the feature calculation unit 4 in this embodiment. Figure 6 is a conceptual diagram showing the feature quantity conversion process in the conversion unit 4D of the feature calculation unit 4 in this embodiment.

[0026] First, when the condition diagnosis process is started, the diagnostic device 10 acquires the operating data of each component constituting the rotating machinery system 40, which has been detected by each sensor 1, using the data acquisition unit 2 (step S001).

[0027] Next, the feature calculation unit 4 extracts one or more time-series data from the acquired operating data by the extraction unit 4A (step S002), and stores the extracted time-series data in the time-series data storage unit 4B as shown in Figure 4. Figure 4 shows an example in which the time-series data stored includes the current flowing through the rotating machinery system 40, vibration acceleration, AE (Acoustic Emission), which is sound waves when the material deforms and breaks, and sound.

[0028] Furthermore, the extraction unit 4A performs frequency analysis such as FFT (Fast Fourier Transform) on the operating data stored in the time-series data storage unit 4B (step S003). The extraction unit 4A extracts one or more frequency data from the frequency data converted to a frequency scale in this way (step S004), and stores the extracted frequency data in the frequency data storage unit 4C as shown in Figure 4. Figure 4 shows an example in which frequency spectral data obtained by performing FFT on the current flowing through the rotating machine 42, vibration acceleration, AE, and sound are stored as frequency data.

[0029] In Figure 4, current, vibration acceleration, AE, and sound are shown as examples of time-series data and frequency data extracted from operating data, but voltage, strain, pressure, power, etc., may also be used. Furthermore, while an example using FFT (Fast Fourier Transform) for frequency analysis was shown, STFT (Short Time Fourier Transform), wavelet transform, etc., can also be used.

[0030] Next, the extraction unit 4A extracts one or more time series data from the time series data stored in the time series data storage unit 4B as the first feature, which is the time series feature T1, as shown in Figure 5, to be used as feature quantities for the anomaly determination processing by the anomaly diagnosis unit 5 described later (step S005). Furthermore, the extraction unit 4A extracts one or more frequency data points from the frequency data stored in the frequency data storage unit 4C as the second feature, which is the frequency feature F1, as shown in Figure 5, to be used as feature quantities for the anomaly determination processing by the anomaly diagnosis unit 5 described later (step S006).

[0031] Here, the extraction unit 4A extracts time-series features T1 on a time-series scale and frequency features F1 on a frequency scale based on the first database DB1, which will be described below. The characteristic quantities obtained from each component constituting the rotating machinery system 40 reflect the type and degree of the abnormality occurring in the rotating machinery system 40 differently. The first database DB1 records frequency features F1 that are correlated with each time-series feature T1. For example, for a first feature that is affected by a certain anomaly, a second feature that is similarly affected is recorded and associated with it.

[0032] The correspondence between time-series features T1 and frequency features F1 in this first database DB1 is not limited to a one-to-one relationship. For example, for a single current time-series data as time-series feature T1, three frequency features F1—current frequency spectrum, vibration frequency spectrum, and sound wave frequency spectrum—may be recorded in association. This mapping of frequency features F1 to time series features T1 may be based on verification through experiments, simulations, etc., or it may be based on estimation using AI. Furthermore, the first database DB1 may also record time-series features T1 and frequency features F1 that are not correlated with each other. For example, time-series features T1 and frequency features F1 relating to different components among the components that make up the rotating machine system 40 may be recorded. In this way, the first database DB1 records correlations, including whether or not there is a correlation between the time-series features T1 and frequency features F1.

[0033] The extraction unit 4A may also extract all time-series data and frequency data stored in the time-series data storage unit 4B and the frequency data storage unit 4C as features.

[0034] Next, the feature calculation unit 4 generates map data M as two-dimensional array data by arranging the time-series feature quantity T1 and the frequency feature quantity F1 in a two-dimensional space as shown in Figure 6, using the conversion unit 4D (step S007). Thus, the conversion unit 4D converts the time-series feature T1 and frequency feature F1 into two-dimensional map data M by plotting them as a heatmap. While a heatmap is used as an example of the conversion method to two-dimensional array data, it is not intended to be limited to heatmaps. For example, it may be converted to two-dimensional array data such as a bubble chart or area graph. When using map data M, if an anomaly is detected in the machine learning using AI described later, the result of the anomaly determination is visualized using a heatmap.

[0035] The following describes the detailed structure of map data M. Figure 7 shows the detailed configuration of the map data M in this embodiment. Figure 8 shows the current frequency spectrum obtained by performing an FFT analysis on the current flowing through the rotating machine 42 as time-series data, and the individual frequency data used in the map data M are shown.

[0036] Map data M is divided into multiple cells, each having a set number of rows and examples; in this example, it is configured with 6 rows and 6 columns. Each cell in each column of row 1 and row 2 of the map data M contains a time-series feature T1. Here, the time series feature T1 is calculated based on statistics for each time series data. For example, the statistics include at least one of the following for each time series data: root mean square, maximum value, mean, variance, standard deviation, crest ratio, sum, clearance coefficient, skewness, kurtosis, total harmonic distortion (THD), variance of the RMS value at any given period, and variance of the maximum value at any given period.

[0037] Each cell in each column from row 3 to row 6 of the map data M contains the frequency feature F1. The input frequency feature F1 is a spectral value that reflects the anomaly in the rotating machinery system 40, as described below, including the sideband of the maximum spectral peak, sideband frequency, floor level, and sum of sidebands.

[0038] For example, if the acquired operating data is current or voltage, the maximum spectrum will be the power supply frequency fs of the rotating machine 42, as shown in Figure 8. In other words, when the acquired data is current or voltage, sidebands appear on both the low-frequency and high-frequency sides, centered around the spectral peak of the power supply frequency fs, and these sidebands are spectra caused by anomalies. As shown in Figure 8, sidebands f1(f1(C)·f1(D)), f2(f2(B)·f2(E)), and f3(f3(A)·f3(F)) are generated on both sides of the power supply frequency fs of the rotating machine 42.

[0039] For example, if the bearing of the rotating machine 42 is faulty, and the power supply frequency is fs, the actual rotational frequency of the rotating machine 42 is fr, and the order of the harmonic is a natural number n, then the sideband fbearing caused by the bearing fault is given by the following formula.

[0040]

number

[0041] Furthermore, for example, if the rotor bar of the rotating machine 42 is abnormal, and the slip is s, the sideband wave f(rotor_bar) caused by the rotor bar abnormality is given by the following equation.

[0042]

number

[0043] Furthermore, for example, if the belt of the power transmission mechanism 43 becomes loose or breaks, and the rotation frequency of the belt is fb, the sideband fbelt caused by the loosening or breakage of the belt is given by the following formula.

[0044]

number

[0045] In this way, the extraction unit 4A derives the rotation frequency of the rotating machine 42, the rotation frequency of the belt, etc. The extraction unit 4A then uses the frequencies or spectral peak values ​​of the sidebands f1, f2, and f3 that occur around the power supply frequency, which are determined by the difference or sum of the power supply frequency fs and the rotation frequencies fr and fb that drive the rotating machine 42, as second features. As a result, the features that reflect each of the above-mentioned abnormalities, such as bearing abnormalities, rotor bar abnormalities, and belt abnormalities, are arranged in the map data M.

[0046] Here, the frequency features F1 arranged in each column of the 5th row of the map data M are the average floor levels in the current frequency spectrum shown in Figure 8. Specifically, the extraction unit 4A derives the average of the floor-level spectral values ​​for each set frequency band. In this example, the extraction unit 4A derives the average floor-level values ​​for each of the set frequency bands: 0Hz~30Hz, 10Hz~40Hz, 20Hz~50Hz, 30Hz~60Hz, 40Hz~70Hz, and 50Hz~80Hz.

[0047] In deriving the average value, the extraction unit 4A sorts the spectral values ​​of the frequency components in each set frequency band in ascending order according to their magnitude. Then, to exclude spectral values ​​that are significantly outside the normal range, the extraction unit 4A sets a threshold range for the spectral values ​​for each set frequency band. After adjusting the spectral values ​​of each frequency component in each frequency band to values ​​within the threshold range, the extraction unit 4A derives the average value of those spectral values.

[0048] In the above example, we showed the use of the frequency or spectral peak value of the sidebands generated on both sides of the power supply frequency fs of the rotating machine 42, but we are not limited to this. We may also use the frequency or spectral peak value of the nth harmonic sideband, which is n times the power supply frequency fs (where n is a natural number), for example, the sideband (nfs-fb). Furthermore, the feature calculation unit 4 may be equipped with a memory function to store the extracted first and second feature quantities, and may also be equipped with a function to organize and search these feature quantities in chronological order. Additionally, the feature calculation unit 4 may be equipped with a memory function to store the converted two-dimensional array data, and may also be equipped with a function to organize and search this data in chronological order.

[0049] Furthermore, data saved using the above-mentioned memory function may be saved to another computer via the communication function. The memory function may also include a function to delete saved data and a function to retrieve past data via the communication function and store it in the memory function. Furthermore, the frequency feature F1 may be not only the spectral value resulting from the anomaly, but also a value calculated based on statistics. For example, the statistics may include at least one of the following: variance, standard deviation, kurtosis, and skewness.

[0050] Returning to Figure 3, the anomaly diagnosis unit 5 then, if the state determination process is in the learning phase, constructs a trained model using machine learning via the model generation unit 5A (step S008). In other words, the anomaly diagnosis unit 5 has AI functionality and performs machine learning and deep learning. Furthermore, if the state determination process in step S007 is not in the learning phase, that is, if a trained model has already been constructed, the learning process in step S008 will not be performed.

[0051] Next, the abnormality diagnosis unit 5 uses the trained model configured as described above by the AI ​​diagnosis unit 5B to diagnose abnormalities in the rotating machinery system 40 (step S009). Next, the diagnostic results from the diagnostic device 10 are displayed on the display unit 20 (step S010). If the diagnostic result is abnormal (step S011, YES), the abnormality diagnosis unit 5 stops the operation of the rotating machinery system 40. If the diagnostic result is not abnormal (step S011, YES), the abnormality diagnosis continues.

[0052] In step S006 above, the extraction unit 4A is shown to extract time-series features T1 on a time-series scale and frequency features F1 on a frequency scale based on the first database DB1. However, extraction may be performed without using the first database DB1. If at least one time-series feature T1 on a time-series scale and one or more frequency features F1 on a frequency scale are extracted, the effects described later will be obtained. Furthermore, in this embodiment, we have illustrated an operation in which the operation is immediately stopped if the diagnostic result is abnormal. However, if the diagnostic result is abnormal, the operation may also be to report it to a manager such as a worker or maintenance person and immediately stop the system.

[0053] In this embodiment, the system includes an abnormality location prediction unit 5C that has a simplified abnormality location prediction function, using the diagnosis results obtained by the abnormality diagnosis unit 5. For example, when diagnosis is performed using an autoencoder with feature quantities as input, the input / output errors are sorted in ascending order, and the abnormality location is predicted based on that order. In addition, the abnormality location prediction may also calculate the failure probability based on the faulty part. The calculated abnormality prediction data is displayed on the display unit 20.

[0054] The following describes examples of AI functions provided by the abnormality diagnosis unit 5 of this embodiment. Figure 9 is a conceptual diagram illustrating the AI ​​function in the abnormality diagnosis unit 5. Figure 10 is a conceptual diagram illustrating the AI ​​function in the abnormality diagnosis unit 5.

[0055] The anomaly diagnosis unit 5 performs diagnosis using decision trees, SVM (Support Vector Machine), local outlier factor analysis (LOF), neural networks, convolutional neural networks, and autoencoders, as shown in Figure 9. Clustering and regression analysis may also be used.

[0056] A decision tree identifies data based on the relationship between data features and threshold values, branching and identifying data through conditional branching at each node. Alternatively, a random forest, which uses multiple decision trees for diagnosis, may be used. SVM finds boundaries and interfaces that distinguish different classes or categories, constructs the best discriminant function to classify two or more classes, and then identifies them. The local outlier factorization method estimates local density from the distance to neighboring points, compares the local density of each data point, and identifies anomalies.

[0057] A neural network is a method of artificial intelligence that teaches computers to process data in a way that mimics the workings of the human brain. A neural network weights multiple features (inputs), learns the weights to minimize the error between the predicted result obtained by the weighting and the label data, and then performs identification.

[0058] Alternatively, a convolutional neural network, which adds convolutional and pooling layers to a neural network, may be used. Convolutional neural networks have a proven track record in image classification because they perform specific mathematical functions such as filtering to extract features from images that are relevant to image recognition and classification.

[0059] The convolutional layer obtains feature maps by filtering the nodes, and the pooling layer further reduces the feature maps output from the convolutional layer to create new feature maps. An autoencoder consists of an encoder and a decoder, and learns to reduce the dimensionality of input data, extract features, reconstruct them, and return them to the original data. Reducing the dimensionality of data results in the loss of some information, but it can retain highly important information so that the data can be reconstructed even without the missing parts.

[0060] Furthermore, convolutional and pooling layers can be added to the autoencoder to perform dimensionality reduction and reconstruction. This improves its adaptability in image classification applications. Also, to diagnose anomalies with an autoencoder, the diagnosis is made based on the error between the input and output. When diagnosing from the reconstruction error, which is the error between the input and output, methods such as threshold-based diagnosis and local outlier factor diagnosis can be used. Threshold-based diagnosis diagnoses an anomaly if the reconstruction error exceeds a threshold. Local outlier factor diagnosis is a clustering method that diagnoses anomalies by performing clustering using the reconstruction error output by the autoencoder as input.

[0061] In this case, when using the aforementioned convolutional neural network and autoencoder, the 2D array data of the features converted by the conversion unit 4D can be used to recognize the features as image data, thereby improving the accuracy of the diagnosis.

[0062] Furthermore, the anomaly diagnosis unit 5 may use multiple algorithms for diagnosis. For example, it may use multiple algorithms to make a diagnosis and then make a decision by majority vote based on the diagnosis results, or it may use a model to set arbitrary weights and make a diagnosis based on thresholds, etc. Here, AI diagnostics require pre-training, and as shown in step S008 above, a training phase must be established, or a pre-trained model must be implemented in the AI ​​diagnostic unit. In other words, a function to set the training phase and a function to incorporate a pre-trained model trained on an external computer are required.

[0063] Furthermore, AI learning can be divided into supervised learning and unsupervised learning. Supervised learning requires pre-training that includes correct labels, whereas unsupervised learning can discover patterns and trends without requiring the aforementioned correct labels. Furthermore, if labels are necessary, there should be a function to manually set labels as needed. For example, an administrator could assign labels during the learning process, and could also re-label them after the learning period.

[0064] As described above, the anomaly diagnosis unit 5 performs anomaly diagnosis using AI. However, the condition diagnosis of the rotating machinery system 40 by the anomaly diagnosis unit 5 is not intended to be limited to AI-based diagnosis only. For example, the anomaly diagnosis unit 5 may be configured without AI functionality and diagnose the above-mentioned feature quantities using thresholds. Thus, the anomaly diagnosis by the anomaly diagnosis unit 5 is performed using one or more of the following methods: machine learning, deep learning, or threshold-based judgment.

[0065] Furthermore, the status diagnosis process flow shown in Figure 3 may be provided by an external computer. For example, operating data of the rotating machinery system 40 detected by a sensor equipped with communication functionality may be acquired by an external computer via a dedicated line or network, and abnormalities in the rotating machinery system 40 may be diagnosed according to the aforementioned flow diagram.

[0066] The following describes control methods to further improve the diagnostic accuracy of the rotating machinery system 40. In the extraction unit 4A, in order to unify the scale of the extracted features in steps S005 and S006 described above, it may perform normalization, standardization, etc., to unify the scale.

[0067]

number

[0068] In other words, the extraction unit 4A adjusts the value of the feature that exceeds the threshold range so that it becomes the upper or lower limit of the threshold range. This makes it possible to emphasize even slight differences in the feature, enabling accurate anomaly diagnosis.

[0069] Figure 11A shows an example of the accuracy of anomaly diagnosis when no threshold Kc is set. Figure 11B shows an example of the accuracy of anomaly diagnosis when a threshold Kc is set. The vertical axis represents the Mean Absolute Error (MAE), and the horizontal axis represents the belt tension. The belt tension should be within the range of 75% to 125% as the appropriate tension (normal), and anything outside this range is considered abnormal. In this example, the belt tension was defined as being within the range of 75% to 125% as the appropriate tension (normal), but this range can be changed as desired. For example, the appropriate range could be set to 80% to 120%, or to 70% or higher.

[0070] As shown in Figure 11A, when a threshold Kc is not set, even if the belt tension is at the appropriate tension of 75% to 125%, the MAE value is large and the error is large. In contrast, as shown in Figure 11B, when a threshold Kc is set, if the belt tension is at the appropriate tension of 75% to 125%, the MAE value is small and the error is small. As described above, by setting a threshold Kc that defines the threshold range, it becomes possible to diagnose the rotating machinery system 40 with high accuracy. Furthermore, the control is not limited to cutting off spectral peak values ​​exceeding the threshold Kc to the same value as the threshold Kc; values ​​below the threshold Kc may be used as long as they do not exceed the threshold Kc.

[0071] Alternatively, you may perform feature scaling in the following ways, which differ from the above. The data acquisition unit 2 acquires status signals indicating the operating status of each component constituting the rotating machinery system 40. These status signals include load information, etc. Alternatively, the operator may acquire information such as belt tension and input this as a status signal to the data acquisition unit 2.

[0072] Furthermore, the feature calculation unit 4 includes a second database DB2, shown in Figure 1, which contains correction values ​​for correcting the feature values ​​obtained at the 30% reduced belt tension to the set feature values ​​at the 100% belt tension of the first operating state, if the belt tension indicated by the status signal as the operating state is 30% less than the belt tension of the first operating state (100%). The second database DB2 may be generated, for example, by calculating the average value of multiple feature quantities for each operating state based on the status signal, thereby deriving the correction value. This enables accurate diagnosis regardless of the condition of each component, such as the belt, that makes up the rotating machinery system 40.

[0073] Alternatively, for example, the ratio of the feature quantities at the initial stage when the rotating machinery system 40 is set up to the feature quantities during operation may be derived, and the derived ratio may be used as a correction amount to correct the feature quantities. For example, the control unit 3 uses the first few hours after the start of operation of the rotating machinery system 40 as the initial stage and obtains the average values ​​F_mean1, F_mean2, etc. for each feature during this initial stage. Next, the control unit 3 obtains the feature F1, F2, etc., which are the feature during the operation in which the diagnosis is performed, derives F1 / F_mean1, F2 / F_mean2, etc., and may use this ratio as a correction amount for scaling the feature. In other words, the extraction unit 4A may cut off the extracted feature quantities at a threshold Kc so that the extracted feature quantities do not exceed the threshold range. If the cutoff threshold Kc and the feature quantity x are denoted by the following formula, the result is given by the following equation.

[0074] The following describes diagnostic devices and methods for rotating machinery systems that differ from those described above. Figure 12 is a flowchart showing another example of the flow of the condition diagnosis process performed by the diagnostic device 10 of this embodiment. The steps corresponding to each step shown in Figure 3 are numbered the same as in Figure 3. For simplification, some of the steps shown in Figure 3 have been omitted from the illustration.

[0075] In this diagnostic device 10, when the rotating machinery system 40 is driven, the sensor 1 and the data acquisition unit 2 acquire operating data (step S001). One or more time-series features are extracted from the acquired operating data, and one or more frequency features are extracted from the data converted to a frequency scale (steps S005, S006). At this time, the feature calculation unit 4 extracts each of these features as features of the rotating machinery 42, features of the power transmission mechanism 43, and features of the load equipment 44, respectively.

[0076] The abnormality diagnosis unit 5 performs an abnormality diagnosis using the extracted characteristics of the rotating machine 42, the power transmission mechanism 43, and the load equipment 44 (step S009). The result of the abnormality diagnosis is displayed on the display unit 20 (step S010). If the diagnosis result is abnormal, the abnormality diagnosis unit 5 stops the operation (step S011, YES), and if the diagnosis result is not abnormal (step S011, NO), the abnormality diagnosis continues.

[0077] Alternatively, the features of the rotating machinery 42, the power transmission mechanism 43, and the load equipment 44 may be calculated from the operating data acquired by the same sensor 1, or multiple sensors 1 may be used, with different sensors 1 used for each component.

[0078] Furthermore, the frequency analysis performed when calculating the features of the rotating machine 42, the power transmission mechanism 43, and the load equipment 44 may be different analyses. In other words, the frequency analysis performed when calculating the features of the rotating machine 42, the power transmission mechanism 43, and the load equipment 44 may be the same analysis or different analyses.

[0079] The abnormality diagnosis of the rotating machinery 42, the abnormality diagnosis of the power transmission mechanism 43, and the abnormality diagnosis of the load equipment 44 may be performed using the same algorithm, or they may be performed using different algorithms. Note that the time-series feature T1 and frequency feature F1 may be different from the features of the rotating machine 42, the power transmission mechanism 43, and the load equipment 44. Also, different numbers of features may be extracted for each. In other words, the features of the rotating machine 42, the power transmission mechanism 43, and the load equipment 44 may be the same or different. Even with this diagnostic process, abnormalities in the rotating machinery system 40, which consists of the rotating machinery 42, the power transmission mechanism 43, and the load equipment 44, can be diagnosed with high accuracy, similar to the diagnostic process described above.

[0080] Although the description describes a method for individually diagnosing abnormalities in the rotating machinery 42, the power transmission mechanism 43, and the load equipment 44, the method is not limited to this. The components of the rotating machinery system 40 can be divided into multiple parts, and abnormalities can be diagnosed for each of them. For example, the power transmission mechanism 43 can be further divided into a belt-related mechanism and a gear-related mechanism, and abnormalities can be diagnosed for each of them. Alternatively, the rotating machinery system 40 may be divided into the rotating machinery 42 and machinery other than the rotating machinery 42, and the power transmission mechanism 43 and load equipment 44 may be examined together for abnormalities. Such a diagnostic device for abnormalities in rotating machinery systems can individually diagnose abnormalities in the rotating machinery, power transmission mechanism, and load equipment, allowing for a more accurate diagnosis of abnormalities in the rotating machinery system.

[0081] The hardware configuration of the control unit 3 will be described below. The control unit 3, as an example of its hardware, consists of a processor 3A and a storage device 3B, which functions as memory, as described above. The storage device 3B includes a volatile storage device such as random access memory (not shown) and a non-volatile auxiliary storage device such as flash memory. Alternatively, a hard disk may be provided as an auxiliary storage device instead of flash memory. Processor 3A executes a program input from storage device 3B. In this case, the program is input to processor 3A from the auxiliary storage device via volatile storage. Processor 3A may also output data such as calculation results to the volatile storage device of storage device 3B, or it may save the data to the auxiliary storage device via the volatile storage device.

[0082] The diagnostic device of this embodiment, configured as described above, A diagnostic device comprising: an acquisition unit for acquiring operating data of components constituting a rotating machinery system; and a control unit for diagnosing the state of the components constituting the rotating machinery system based on characteristic quantities obtained from the acquired operating data, The control unit, One or more first features are selected from the time-series data obtained from the operating data, and one or more second features are selected from the frequency data obtained by frequency analysis of the operating data. Based on the selected first and second features, the state of the rotating machinery system is diagnosed. It is.

[0083] The physical phenomena vary depending on the type of anomaly occurring in the rotating machinery system. Therefore, diagnosis using only one type of signal, such as time-series features or frequency features, cannot fully reflect the abnormal state of the rotating machinery system, and diagnostic accuracy tends to decrease. For example, if the tension of a belt decreases and an abnormal condition occurs, the rotational motion of the rotating machine is not directly transmitted to the belt, and the belt vibrates like a string. In this case, when the torque is large, the force on the belt is large, and when the torque is small, the force on the belt is small, and the string vibration of the belt changes in accordance with the change in torque. In this way, when the intensity of the spectral value indicating an abnormality changes with the change in torque, diagnosis using only thresholds in frequency features may result in misdiagnosis where a normal value is detected as abnormal, so diagnosis in combination with time series features is necessary. By performing diagnosis using both time series features and frequency features in this way, the accuracy of diagnosing abnormalities in, for example, power transmission mechanisms that transmit the rotational energy of rotating machines and loads, which are susceptible to the effects of torque changes, can be improved.

[0084] The diagnostic device of this embodiment selects one or more first features, which are time-series features, and one or more second features, which are frequency features. This enables highly accurate and efficient condition diagnosis of rotating machinery systems.

[0085] Furthermore, in the diagnostic device of this embodiment configured as described above, The control unit, The system includes a first database that shows the correlation between the time-series data and the frequency data, Based on the aforementioned database, one or more of the aforementioned first and second features are selected, and the state of the rotating machine system is diagnosed based on the selected first and second features. It is.

[0086] Thus, a first database showing the correlation between time-series data and frequency data may be provided, and control may be performed to extract first and second features based on this first database. The degree to which each feature obtained from each component constituting a rotating mechanical system reflects the abnormality differs depending on the type and degree of the abnormality occurring in the rotating mechanical system. Therefore, for example, if the time-series data is a torque waveform showing the temporal fluctuation of torque, since this torque fluctuation is related to the motor current, the first database may associate current spectrum data as frequency data with the torque waveform as time-series data. In this way, compared to randomly extracting features from time-series data and frequency data, it becomes possible to diagnose the state of rotating machinery with higher diagnostic accuracy and efficiency.

[0087] Furthermore, in the diagnostic device of this embodiment configured as described above, The control unit, A trained model for diagnosing the state of the rotating machine system is constructed by machine learning using two-dimensional array data obtained by arranging the first and second features. The state of the rotating machine system is diagnosed based on the trained model and the array data generated using the acquired first and second features. It is. This necessitates a broad range of knowledge regarding rotating machinery, power transmission mechanisms, load equipment, etc., and involves numerous condition changes to consider during condition diagnosis. Furthermore, even when a vast amount of diverse operating data is acquired, it enables highly accurate and efficient condition diagnosis of rotating machinery systems. Furthermore, even in cases where changes in the state of a rotating machinery system occur, such as fluctuations in load conditions, which can easily affect the features, if the array data is configured to include features related to the load data, accurate state diagnosis of the rotating machinery system becomes possible regardless of such load fluctuations.

[0088] Furthermore, in the diagnostic device of this embodiment configured as described above, The array data is composed of a set number of rows and columns and is divided into a plurality of cells, comprising one or more cells into which the first feature quantity is input and one or more cells into which the second feature quantity is input. The control unit, The first and second features input into the cell are adjusted to values ​​within a threshold range derived according to the variance of their respective values. It is. Thus, by providing array data composed of multiple cells with a set number of rows and columns, adjusting the feature quantities used in each cell to values ​​within a threshold range, and excluding outliers in the feature quantities, it becomes possible to perform highly accurate and efficient condition diagnosis of rotating machinery systems, even when using a vast amount of diverse operating data.

[0089] Furthermore, in the diagnostic device of this embodiment configured as described above, The control unit, A threshold range is set for the spectral values ​​of the frequency components in each frequency band of the frequency data, and the average of the spectral values ​​of the frequency components adjusted to values ​​within the threshold range in each frequency band is used as the second feature. It is. This enables condition diagnosis of rotating machinery systems while taking floor level into consideration, resulting in more accurate and efficient condition diagnosis of rotating machinery systems.

[0090] Furthermore, in the diagnostic device of this embodiment configured as described above, The control unit, The rotation frequency of the component is derived, and the frequency or spectral peak value of the sideband component that occurs around a frequency that is n times the power supply frequency (where n is a natural number), determined by the difference or sum of the power supply frequency that drives the rotating machine as the component and the rotation frequency, is used as the second feature quantity. It is. By using the frequency or spectral peak value of sideband wave formation caused by anomalies as a feature, it becomes possible to diagnose the condition of rotating machinery systems with even higher diagnostic accuracy and efficiency.

[0091] Furthermore, in the diagnostic device of this embodiment configured as described above, The control unit, Based on the status signals indicating the operating state of each component constituting the rotating machinery system, the average value of the feature quantities over multiple cycles in each operating state is derived. The system includes a second database containing correction values ​​that, based on the average value of the derived feature quantities, correct the acquired feature quantity values ​​to the value of the feature quantity in a first operating state, which is a preset operating state, from the operating state at the time the feature quantity was acquired. Based on the second database, the value of the feature obtained is corrected to the value of the feature in the first operating state. It is. This enables highly accurate and efficient diagnosis of the condition of rotating machinery systems, regardless of changes in the state of the rotating machinery system, such as changes in belt tension or load.

[0092] While this disclosure describes exemplary embodiments, the various features, aspects, and functions described in the embodiments are not limited to the application of any particular embodiment, but can be applied individually or in various combinations to the embodiments. Accordingly, countless variations not illustrated are conceivable within the scope of the art disclosed in this specification. These include, for example, modifications, additions, or omissions of at least one component. [Explanation of Symbols]

[0093] 2 Data acquisition unit (acquisition unit), 3 Control unit, 10 Diagnostic device, 20 Display unit (notification unit), 30 Diagnostic system, 40 Rotating machinery system, 42 Rotating machinery (components), 43 Power transmission mechanism (components), 44 Load equipment (components), T1 Time series feature (first feature), F1 Frequency feature (second feature), DB1 First database, DB2 Second database, M Map data (array data).

Claims

1. A diagnostic device comprising: an acquisition unit for acquiring operating data of components constituting a rotating machinery system; and a control unit for diagnosing the state of the components constituting the rotating machinery system based on characteristic quantities obtained from the acquired operating data, The control unit, The system includes a first database that shows the correlation between time-series data obtained from the aforementioned operating data and frequency data obtained by frequency analysis of the aforementioned operating data, regarding abnormalities occurring in the rotating machinery system. Based on the aforementioned first database, multiple types of first features are selected from the time-series data associated by the aforementioned first database, and multiple types of second features are selected from the frequency data. The selected plurality of first features include at least two first features that are distinct from each other, and the selected plurality of second features include at least two second features that are distinct from each other. A trained model for diagnosing the state of the rotating mechanical system is constructed by machine learning using array data that constitutes a two-dimensional feature space formed by arranging the selected multiple types of first features and the multiple types of second features. The control unit, The state of the rotating machine system is diagnosed based on the trained model and the distribution characteristics of multiple types of features, namely the first and second features, in the array data constituting the feature space. Diagnostic equipment.

2. The array data is composed of a plurality of cells having a set number of rows and columns, comprising a plurality of cells into which the first feature quantity is input and a plurality of cells into which the second feature quantity is input. The control unit, The first and second feature quantities input into the cell are adjusted to values ​​within a threshold range derived according to the variance of their respective values. The diagnostic device according to claim 1.

3. The control unit, A threshold range is set for the spectral values ​​of the frequency components in each frequency band of the frequency data, and the average of the spectral values ​​of the frequency components adjusted to values ​​within the threshold range in each frequency band is used as the second feature quantity. The diagnostic device according to claim 2.

4. The control unit, The rotation frequency of the component is derived, and the frequency or spectral peak value of the sideband component that occurs around a frequency that is n times the power supply frequency (where n is a natural number), determined by the difference or sum of the power supply frequency that drives the rotating machine as the component and the rotation frequency, is used as the second feature quantity. The diagnostic device according to claim 1.

5. The control unit, The rotation frequency of the component is derived, and the frequency or spectral peak value of the sideband component that occurs around a frequency that is n times the power supply frequency (where n is a natural number), determined by the difference or sum of the power supply frequency that drives the rotating machine as the component and the rotation frequency, is used as the second feature quantity. The diagnostic device according to claim 2.

6. The control unit, The rotation frequency of the component is derived, and the frequency or spectral peak value of the sideband component that occurs around a frequency that is n times the power supply frequency (where n is a natural number), determined by the difference or sum of the power supply frequency that drives the rotating machine as the component and the rotation frequency, is used as the second feature quantity. The diagnostic device according to claim 3.

7. The control unit, Based on the status signals indicating the operating state of each component constituting the rotating machinery system, the average value of the feature quantities over multiple cycles in each operating state is derived. The system includes a second database containing correction values ​​that, based on the average value of the derived feature quantities, correct the acquired feature quantity values ​​to the value of the feature quantity in a first operating state, which is a preset operating state, from the operating state at the time the feature quantity was acquired. Based on the second database, the value of the feature obtained is corrected to the value of the feature in the first operating state. A diagnostic device according to any one of claims 1 to 6.

8. The control unit, As the aforementioned machine learning method, an autoencoder, which is an unsupervised machine learning method, is used to derive the reconstruction error between the input value of the array data and the output value of the array data output from the autoencoder, and the state of the rotating mechanical system is diagnosed based on the reconstruction error. A diagnostic device according to any one of claims 1 to 6.

9. The control unit, Extract the first and second feature quantities that indicate at least the load state of the component, A diagnostic device according to any one of claims 1 to 6.

10. The control unit, Based on the first and second feature quantities selected for each component, the state of the rotating mechanical system is diagnosed for each component. A diagnostic device according to any one of claims 1 to 6.

11. The control unit, As the second feature quantity, the operating data to be subjected to frequency analysis is: At least one of the following is used for the maximum value of each of the operating data in a set period: variance, standard deviation, kurtosis, skewness, current, voltage, vibration velocity, vibration acceleration, sound, strain, temperature, pressure, power, AE. A diagnostic device according to any one of claims 1 to 6.

12. The control unit, From the aforementioned time-series data, the first feature is selected such that it includes at least one of the following: variance, standard deviation, kurtosis, skewness, maximum value, mean, sum, root mean square value, crest factor, clearance coefficient, total harmonic distortion (THD) of the maximum value of each operating data in a set period, variance of the RMS value in a set period, and variance of the maximum value in a set period. A diagnostic device according to any one of claims 1 to 6.

13. A diagnostic device according to any one of claims 1 to 6, The system includes a notification unit that connects to the diagnostic device via a network and notifies the administrator of the diagnostic results obtained by the diagnostic device, Diagnostic system.

14. A diagnostic method using the diagnostic device described in any one of claims 1 to 6, Based on the first database, multiple types of first features are selected from the time-series data of the operation data associated by the first database, and multiple types of second features are selected from the frequency data obtained by frequency analysis of the operation data. The selected plurality of first features include at least two first features that are distinct from each other, and the selected plurality of second features include at least two second features that are distinct from each other. A trained model for diagnosing the state of the rotating mechanical system is constructed by machine learning using array data that constitutes a two-dimensional feature space formed by arranging the selected multiple types of first features and the multiple types of second features. The state of the rotating machine system is diagnosed based on the trained model and the distribution characteristics of multiple types of features, namely the first and second features, in the array data constituting the feature space. Diagnostic methods.

Images