Learning device, learning method and computer readable medium, determination device, determination method and computer readable medium
By using preprocessing technology in the learning and determination device to remove drift features in the equipment monitoring data, the problem of frequent model updates in the prior art is solved, and the accuracy and stability of device status determination are improved.
Patent Information
- Application Number
- CN202110123638.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2020-01-31
- Filing Date
- 2021-01-29
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2041-01-29
AI Technical Summary
In the prior art, in equipment monitoring and diagnosis, the model is difficult to adapt to the drift phenomenon caused by the characteristics of the measurement data over time, resulting in frequent updates of the model and increasing processing load.
Through the learning device and the determination device, preprocessing technology is used to reduce the drift of the measured data, including feature quantity calculation and data conversion, remove drift features in the data, and reduce the impact on the model, thereby reducing the necessity of re-learning when the data drifts.
It effectively reduces the impact of measurement data drift on the model, reduces the frequency and processing load of model updates, and improves the accuracy and stability of equipment status judgment.
Smart Images

Figure CN113205187B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a learning device, a learning method, a computer-readable medium storing a learning program, a determination device, a determination method, and a computer-readable medium storing a determination program. Background Art
[0002] In recent years, in a control system of a factory, there is a function of monitoring and diagnosing factory equipment using a learned model. For example, Non-Patent Document 1 discloses a technique for detecting concept drift when a failure or deterioration that has not occurred before occurs in a monitoring target and updating a model at an appropriate timing.
[0003] Patent Document 1: Sakamoto et al., "Initial Experiment for Concept Drift Detection in Adaptive Monitoring System", Proceedings of the Artificial Intelligence Society Research Group SIG-KBS-B401-05, Japan, July 24, 2014
[0004] Generally, a model used in diagnosis is trained using learning data, and thus depends on the situation, conditions, and instrument settings when the learning data is obtained, and it is difficult to apply it to diagnosis under other situations. Therefore, in the above prior art, it is necessary to detect a drift phenomenon in which the characteristics of the data of the measurement object change over time, and update the model when drift is detected. Summary of the Invention
[0005] In order to solve the above problems, in a first aspect of the present invention, a learning device is provided. The learning device may include a learning data acquisition unit that acquires learning data including measurement data obtained by measuring equipment and the state of the equipment. The learning device may include a learning preprocessing unit that performs preprocessing to reduce drift in the measurement data in the learning data and outputs preprocessed learning data. The learning device may include a learning processing unit that uses the preprocessed learning data to perform processing for learning a model that determines the state of the equipment based on the preprocessed measurement data.
[0006] The learning preprocessing unit may include a feature quantity calculation unit that calculates a feature quantity indicating drift in the measurement data in the learning data. The learning preprocessing unit may include a data conversion unit that removes drift corresponding to the feature quantity from the measurement data in the learning data.
[0007] The learning device may further include a designation input unit for inputting a designation of a type of preprocessing to be performed by the preprocessing unit for learning. The preprocessing unit for learning may perform preprocessing corresponding to the input designation.
[0008] The learning preprocessing unit may perform at least one of preprocessing for reducing a center of gravity shift of the measurement data in the learning data or preprocessing for calculating a difference of the measurement data in the learning data within a sliding window in accordance with the designation.
[0009] The learning device may include a determination data acquisition unit that acquires determination data including measurement data obtained by measuring the device. The learning device may include a determination preprocessing unit that performs preprocessing to reduce drift of the determination data and outputs preprocessed determination data. The learning device may include a determination unit that uses a learned model to determine the state of the device based on the preprocessed determination data.
[0010] The learning device may further include a display processing unit that performs processing for displaying the distribution of the preprocessed learning data and the preprocessed completion determination data.
[0011] The learning data and the judgment data may include the measured values of a plurality of sensors provided in the device. The learning device may further include a selection unit that selects the measured value of at least one sensor that has a relatively high contribution to the judgment result of the model from the measured values of the plurality of sensors. The display processing unit may perform processing for displaying the distribution of the measured values of at least one sensor selected by the selection unit in the preprocessed learning data and the preprocessed judgment data.
[0012] In a second aspect of the present invention, a learning method is provided. The learning method includes the following steps: a learning device obtains learning data including measurement data obtained by measuring a device and the state of the device. The learning method includes: the learning device performs preprocessing to reduce drift of the measurement data in the learning data, and outputs the preprocessed learning data. The learning method includes: the learning device uses the preprocessed learning data to perform a process of learning a model, and the model determines the state of the device based on the preprocessed measurement data.
[0013] In a third aspect of the present invention, a computer-readable medium recording a learning program executed by a computer is provided. The learning program can cause the computer to function as a learning data acquisition unit, which acquires learning data including measurement data obtained by measuring a device and the state of the device. The learning program can cause the computer to function as a learning preprocessing unit, which performs preprocessing to reduce drift of the measurement data in the learning data and outputs the preprocessed learning data. The learning program can cause the computer to function as a learning processing unit, which uses the preprocessed learning data to perform processing to learn a model, and the model determines the state of the device based on the preprocessed measurement data.
[0014] In a fourth aspect of the present invention, a determination device is provided. The determination device may include a determination data acquisition unit that acquires determination data including measurement data obtained by measuring the device. The determination device may include a determination preprocessing unit that performs preprocessing to reduce drift of the determination data and outputs preprocessed determination data. The determination device may include a determination unit that uses a model to determine the state of the device based on the preprocessed determination data, and the model is learned using preprocessed learning data obtained by preprocessing the learning data.
[0015] In a fifth aspect of the present invention, a determination method is provided. The determination method may include: the determination device obtains determination data including measurement data obtained by measuring the device. The determination method may include: the determination device performs preprocessing to reduce drift of the determination data and outputs preprocessed determination data. The determination method may include: the determination device determines the state of the device based on the preprocessed determination data using a model, the model being learned using preprocessed learning data obtained by preprocessing the learning data.
[0016] In the sixth aspect of the present invention, a computer-readable medium is provided, which records a determination program executed by a computer. The determination program can cause the computer to function as a determination data acquisition unit, which acquires determination data including measurement data obtained by measuring the device. The determination program can cause the computer to function as a determination preprocessing unit, which performs preprocessing to reduce the drift of the determination data. The determination program can cause the computer to function as a determination unit, which determines the state of the device based on the preprocessed determination data using a model, and the model is learned using preprocessed learning data obtained by preprocessing the learning data.
[0017] In addition, the above summary of the invention does not list all the essential features of the present invention. In addition, sub-combinations of these feature groups may also constitute inventions. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] Figure 1 The configurations of the learning device 100 and the determination device 150 according to the embodiment of the present invention are shown.
[0019] Figure 2 The following is a flow of operations of the learning device 100 according to the embodiment of the present invention.
[0020] Figure 3 The following is a flow of operations of the determination device 150 according to the embodiment of the present invention.
[0021] Figure 4 An example of a display screen according to an embodiment of the present invention is shown.
[0022] Figure 5 An example of measurement data according to the embodiment of the present invention is shown.
[0023] Figure 6 An example of measurement data after drift removal according to the embodiment of the present invention is shown.
[0024] Figure 7 An example of a computer 2200 is shown that can implement the various aspects of the present invention in whole or in part. DETAILED DESCRIPTION
[0025] The present invention will be described below by way of the embodiments of the invention, but the following embodiments do not limit the invention as defined in the claims. In addition, all combinations of features described in the embodiments are not necessarily essential to the solution of the invention.
[0026] Figure 1 The structure of the learning device 100 and the determination device 150 involved in this embodiment is shown. The learning device 100 uses the learning data whose influence of drift is reduced or removed by preprocessing to learn a model for determining the state of the equipment in the factory. The determination device 150 uses the learned model to determine the state of the equipment in the factory, etc. based on the determination data whose influence of drift is reduced or removed by preprocessing. In this way, the learning device 100 and the determination device 150 perform learning and determination by reducing or removing the influence of drift, thereby reducing the need to re-learn the model even when the measured data has drifted, and can reduce the processing load of re-learning.
[0027] Here, as factories, in addition to chemical, biological and other industrial factories, we can also cite factories that control the wellheads of gas fields, oil fields, etc. and their surroundings, factories that control power generation such as water power, thermal power, and nuclear power, factories that control environmental power generation such as sunlight and wind power, factories that control water supply and drainage, dams, etc. In addition, as equipment, we can cite tanks, pipes, valves, boilers, compressors, motors, and other devices.
[0028] In addition, the learning device 100 and the determination device 150 can also be applied to equipment installed in places other than factories. For example, the learning device 100 and the determination device 150 can also be applied to equipment that controls air conditioning in a building that is to achieve energy saving and comfort. In this embodiment, the case where the learning device 100 and the determination device 150 are applied to equipment in a factory is illustrated.
[0029] The learning device 100 includes a learning data acquisition unit 105, a learning preprocessing unit 110, and a learning processing unit 130. The learning data acquisition unit 105 acquires learning data. The learning data includes measurement data obtained by measuring the equipment in the factory and the status of the equipment corresponding to the measurement data. Here, the measurement data includes measurement values measured by one or more sensors installed in one or more equipment in the factory at each measurement timing. Such measurement data can be, for example, data indicating the temperature, pressure, flow rate, vibration or corrosion of the measurement object part of the equipment. The status of the equipment indicates whether the equipment has an abnormality or failure, etc. As an example, it can be the status after the user such as the maintenance personnel of the factory marks (annotates) the measurement data.
[0030] The learning preprocessing unit 110 is connected to the learning data acquisition unit 105, performs preprocessing to reduce the drift of the measurement data in the learning data acquired by the learning data acquisition unit 105, and outputs the preprocessed learning data obtained by the preprocessing. The learning preprocessing unit 110 includes a feature quantity calculation unit 115 and a data conversion unit 120. The feature quantity calculation unit 115 calculates the feature quantity of the measurement data in the learning data acquired by the learning data acquisition unit 105. For example, the feature quantity calculation unit 115 calculates the feature quantity representing the drift of the measurement data in the learning data. The data conversion unit 120 is connected to the feature quantity calculation unit 115, reduces (i.e., removes at least a part of) the drift of the measurement data in the learning data or removes the drift of the measurement data. Here, the data conversion unit 120 can remove the drift corresponding to the feature quantity calculated by the feature quantity calculation unit 115 from the measurement data in the learning data.
[0031] The learning processing unit 130 is connected to the learning preprocessing unit 110. The learning processing unit 130 uses the preprocessed learning data preprocessed by the learning preprocessing unit 110 to perform a process of learning a model that determines the state of the device based on the preprocessed measurement data. The learning processing unit 130 stores the learned model in the model storage unit 140.
[0032] The learning device 100 may also include a designation input unit 125. The designation input unit 125 inputs a preprocessing parameter from a user of the learning device 100, and the preprocessing parameter is a designation of the type of preprocessing to be applied to the measurement data in the learning data by the learning preprocessing unit 110. When the designation of the type of preprocessing is input from the user of the learning device 100, the feature quantity calculation unit 115 in the learning preprocessing unit 110 may calculate a feature quantity corresponding to the input designation, and the data conversion unit 120 may perform preprocessing corresponding to the input designation according to the calculated feature quantity.
[0033] The model storage unit 140 is connected to the learning device 100 , and stores the model learned by the learning processing unit 130 .
[0034] The determination device 150 includes a determination data acquisition unit 155, a determination preprocessing unit 160, and a determination unit 180. The determination data acquisition unit 155 acquires determination data. The determination data includes measurement data obtained by measuring the equipment in the factory at each measurement timing using one or more sensors installed in one or more equipment in the factory.
[0035] The determination preprocessing unit 160 is connected to the determination data acquisition unit 155, performs preprocessing to reduce the drift of the determination data acquired by the determination data acquisition unit 155, and outputs the preprocessed determination data. The determination preprocessing unit 160 includes a feature quantity calculation unit 165 and a data conversion unit 170. The feature quantity calculation unit 165 calculates the feature quantity of the measurement data acquired by the determination data acquisition unit 155. For example, the feature quantity calculation unit 165 calculates the feature quantity representing the drift of the measurement data in the determination data. The feature quantity calculation unit 165 may have the same function and structure as the feature quantity calculation unit 115 in the learning preprocessing unit 110. The data conversion unit 170 is connected to the feature quantity calculation unit 165, reduces the drift of the measurement data in the determination data (i.e., removes at least a part) or removes the drift of the measurement data. Here, the data conversion unit 170 removes the drift corresponding to the feature quantity from the measurement data in the determination data with respect to the drift corresponding to the feature quantity calculated by the feature quantity calculation unit 165. The data conversion unit 170 may have the same function and structure as the data conversion unit 120 in the learning preprocessing unit 110 .
[0036] The determination unit 180 is connected to the model storage unit 140 and the determination preprocessing unit 160. The determination unit 180 determines the state of the plant based on the preprocessed determination data using the learned model stored in the model storage unit 140. Here, the learned model stored in the model storage unit 140 is learned using preprocessed learning data obtained by performing learning preprocessing on the learning data.
[0037] The determination device 150 may also include a designation input unit 175. The designation input unit 175 inputs a preprocessing parameter from a user of the determination device 150, and the preprocessing parameter is a designation of the type of preprocessing to be applied to the measurement data in the determination data by the determination preprocessing unit 160. When the designation of the type of preprocessing is input from the user of the determination device 150, the feature quantity calculation unit 165 in the determination preprocessing unit 160 may calculate a feature quantity corresponding to the input designation, and the data conversion unit 170 may perform preprocessing corresponding to the input designation according to the calculated feature quantity.
[0038] The determination device 150 may include a selection unit 185 and a display processing unit 190. The selection unit 185 is connected to the model storage unit 140. When the measurement data in the learning data and the determination data include measurement values of a plurality of sensors installed in one or more devices in the factory, the selection unit 185 selects the measurement value of at least one sensor having a relatively high contribution to the determination result of the model among the measurement values of the plurality of sensors.
[0039] The display processing unit 190 is connected to the data conversion unit 120 and the data conversion unit 170, and receives the preprocessed learning data and the preprocessed determination data. In addition, the display processing unit 190 is connected to the selection unit 185, and performs processing to display the distribution of the preprocessed learning data and the preprocessed determination data received from the data conversion unit 120 and the data conversion unit 170. Here, the display processing unit 190 can perform the following processing, that is, displaying the distribution of the measured values of at least one sensor selected by the selection unit 185 in the preprocessed learning data and the preprocessed determination data received from the data conversion unit 120 and the data conversion unit 170 on the display screen. In addition, the display processing unit 190 can include a display unit such as a display for displaying the display screen. Alternatively, the display processing unit 190 can also perform processing to display the display screen on an external display unit.
[0040] According to the learning device 100 and the judgment device 150 shown above, the model is learned by reducing or removing the drift of the measurement data during learning, and the model is used for diagnosis (judgment) after reducing or removing the drift of the measurement data. This makes it possible to achieve learning and diagnosis in which the influence of the drift of the measurement data is reduced or removed.
[0041] Figure 2 The following is a flow of operations of the learning device 100 according to the present embodiment. In S200 (step 200 ), the learning data acquisition unit 105 acquires learning data.
[0042] In S210, the designation input unit 125 inputs the designation of the type of preprocessing to be applied to the measurement data in the learning data from the user of the learning device 100. Here, the preprocessing that can be applied by the learning preprocessing unit 110 may include the removal of the center of gravity shift and the removal of the drift by the difference calculation in the sliding window. The designation input unit 125 may input the designation of the preprocessing parameters such as which of the preprocessing is to be performed as the type of preprocessing, and may input the parameters accompanying the preprocessing such as the length of the sliding window.
[0043] In S220, the feature quantity calculation unit 115 calculates a feature quantity indicating a drift of the measurement data in the learning data. Here, when the type of preprocessing, etc. is input through the designation input unit 125, the feature quantity calculation unit 115 calculates a feature quantity corresponding to the designated preprocessing and parameters. When the type of preprocessing, etc. is not input, the feature quantity calculation unit 115 calculates a feature quantity corresponding to the default setting.
[0044] For example, for drift in which an offset is applied to the value of the measurement data from the sensor, the feature quantity calculation unit 115 may calculate the average value of the measurement data of each sliding window during a predetermined length or number of samples. In addition, the feature quantity calculation unit 115 may calculate the centroid of the measurement data of each sliding window during a predetermined length or number of samples for the measurement data including the measurement values from one or more sensors. Here, in the case of calculating such an average or centroid, the feature quantity calculation unit 115 may exclude outliers to calculate the average or centroid. The feature quantity calculation unit 115 may regard a measurement value that deviates from the average of all measurement values in the target period by a predetermined coefficient such as 3 times the standard deviation as an outlier. In addition, the feature quantity calculation unit 115 may also calculate the period (e.g., every week) or timing at which the average or centroid changes as a feature quantity representing the drift of the measurement data.
[0045] In S230, the data conversion unit 120 performs preprocessing to reduce the drift of the measurement data in the learning data, and outputs the preprocessed learning data. Here, the data conversion unit 120 applies the preprocessing specified by the user to the measurement data when the type of preprocessing is input through the designation input unit 125. When the type of preprocessing is not input, the data conversion unit 120 removes the drift corresponding to the feature quantity calculated in S220. For example, for the drift associated with the change of the average value in the measurement data from a certain sensor, the data conversion unit 120 subtracts the average value calculated in S220 from each sample value of the measurement data to remove the influence of the change of the average value. In addition, for the drift associated with the movement of the center of gravity of the measurement data from multiple sensors, the data conversion unit 120 subtracts the center of gravity calculated in S220 from the group of sample values of each time of the measurement data to remove the influence of the change of the center of gravity. When the type of preprocessing is not specified, the data conversion unit 120 can apply the preprocessing determined by default. Alternatively, the data conversion unit 120 may decide to use a sliding window with a period shorter than the change and to remove changes in indicators (average, center of gravity, etc.) that have changed based on the feature quantity calculated by the feature quantity calculation unit 115 indicating the periodic or timing of the change in the average or center of gravity.
[0046] In addition, the data conversion unit 120 can perform preprocessing to remove drift on the measured data without using the feature quantity calculated by the feature quantity calculation unit 115, depending on the type of drift removal. Such drift removal methods include, for example, differential calculation within a sliding window. More specifically, when using the difference of the measured values within the sliding window, such as the value obtained by subtracting the minimum value from the maximum value of the measured values within the sliding window, the initial and final values of the measured values within the sliding window, or the maximum value of the difference between adjacent measured values within the sliding window, the influence of the change in the average value is removed with the calculation. Therefore, the data conversion unit 120 does not need to use feature quantities such as the average value calculated by the feature quantity calculation unit 115 for such drift removal. In addition, even in this case, the data conversion unit 120 can determine the size of the sliding window using feature quantities that represent the period or timing of changes in the average, etc.
[0047] In addition, the data conversion unit 120 may also perform processing to standardize the size of the measurement data. For example, the data conversion unit 120 may standardize the size of the measurement data by dividing the difference between the measurement value and the average value by the standard deviation as the pre-processed measurement data. In this way, the data conversion unit 120 can also perform appropriate pre-processing for drift such as a change in the sensitivity of the sensor.
[0048] In S240, the learning processing unit 130 uses the preprocessed learning data from the learning preprocessing unit 110 to perform a learning process on the model, and the model determines the state of the device based on the preprocessed measurement data. As an example, the model inputs one or more measurement values from one or more sensors at each timing of the determination object, and outputs a determination result of the state of the device corresponding to the one or more measurement values. Here, the model can be a support vector machine (SVM), a neural network, or other machine learning models.
[0049] For example, when using SVM, the learning processing unit 130 learns a hyperplane that separates a set of points where the device is marked as normal and a set of points where the device is marked as abnormal with the maximum margin when mapping points whose coordinates are a set of one or more measurement values at each timing in the preprocessed learning data to a multidimensional space. The hyperplane is represented by a weight vector including a set of weights corresponding to each measurement value.
[0050] In addition, for example, when a neural network is used, the learning processing unit 130 uses an algorithm such as back propagation to adjust the weights between neurons and the bias of each neuron so that when a group of one or more measurement values from one or more sensors contained in the preprocessed learning data is input into the input layer, the error between the judgment result output from the output layer and the target judgment result is minimized.
[0051] In addition, since the equipment in the factory rarely has abnormalities, most of the measurement data in the learning data are attached with labels indicating normality, while there are few measurement data attached with labels indicating abnormality. In order to solve this imbalance problem, the feature quantity calculation unit 115 can undersample the measurement data (normal data) attached with labels indicating normality in S220 and supply it to the learning processing unit 130 via the data conversion unit 120. That is, for example, the feature quantity calculation unit 115 can learn the model by using samples selected from each timed sample in the normal data at regular intervals or randomly, thereby reducing the number of samples of normal data.
[0052] In addition, the feature quantity calculation unit 115 may also oversample the measurement data (abnormal data) to which a mark indicating abnormality is attached in S220. That is, for example, the feature quantity calculation unit 115 may increase the number of samples of abnormal data by repeatedly using samples of abnormal data or randomly generating new samples of abnormal data using the average and variance of abnormal data. In addition, the feature quantity calculation unit 115 may randomly add points between samples of abnormal data as new samples using algorithms such as SMOTE (Synthetic Minority Oversampling Technique). The feature quantity calculation unit 115 may further reduce the imbalance by using undersampling of normal data and oversampling of abnormal data.
[0053] In addition, the learning processing unit 130 may also optimize the hyperparameters of the model (e.g., the number of layers and nodes of the neural network) to adjust the correct answer rate of the model to be higher. In addition, the learning processing unit 130 may also learn the model in a manner that enables each sample of normal data to be classified into one class by using a machine learning model that performs one-class classification such as a one-class SVM.
[0054] Figure 3 The following is an operation flow of the determination device 150 according to the present embodiment. In S300, the determination data acquisition unit 155 acquires determination data. The determination data acquisition unit 155 may acquire sensor data provided in real time from one or more sensors installed in one or more devices in the factory as determination data.
[0055] In S310, the designation input unit 175 receives an input from a user of the determination device 150 regarding a designation of a preprocessing parameter such as a type of preprocessing to be applied to the measurement data in the determination data. Figure 2 The same process as S210 is performed, and the preprocessing parameters such as the type of preprocessing are input.
[0056] In S320, the feature quantity calculation unit 165 calculates a feature quantity indicating a drift of the measurement data in the determination data. Figure 2 The same processing as S220 is performed to calculate the feature quantity.
[0057] In S330, the data conversion unit 170 performs preprocessing to reduce the drift of the determination data and outputs the preprocessed determination data. Figure 2The feature quantity calculation unit 165 and the data conversion unit 170 may apply preprocessing to the determination data that is the same as the preprocessing performed by the feature quantity calculation unit 115 and the data conversion unit 120. Alternatively, the feature quantity calculation unit 165 and the data conversion unit 170 may apply preprocessing to the determination data that is different from at least a part of the preprocessing performed by the feature quantity calculation unit 115 and the data conversion unit 120. For example, when the period of drift is different between the learning data and the determination data, the feature quantity calculation unit 165 and the data conversion unit 170 may apply preprocessing to reduce drift to the determination data at a period different from that of the preprocessing performed on the learning data.
[0058] In S340 , when the measurement data in the determination data includes measurement values of a plurality of sensors, the selection unit 185 selects the measurement value of at least one sensor having a relatively high contribution to the determination result of the model as a parameter to be displayed on the display screen.
[0059] In S350, the display processing unit 190 displays on the display screen the distribution of the measured values of at least one sensor selected in S340 among the measured values of each sensor included in the preprocessed learning data received from the learning preprocessing unit 110 and the preprocessed determination data received from the determination preprocessing unit 160. Thus, the display processing unit 190 can display the ranges in which the preprocessed learning data judged to be normal, the preprocessed learning data judged to be abnormal, and the preprocessed determination data are respectively distributed in the coordinate space in which the measured values of each sensor having a high contribution to the determination result are selected as coordinate axes.
[0060] In S360, the designation input unit 175 receives an instruction from the user of the determination device 150 to change the preprocessing parameters such as the type of preprocessing to be applied to the measurement data in the determination data. If there is an instruction to change the type of preprocessing, the determination device 150 advances the process to S330, applies the changed preprocessing to the determination data (S330), and performs a process of redisplaying the distribution of the measurement values of the preprocessed learning data and the preprocessed determination data (S340 to S350). If there is no instruction to change the type of preprocessing, the determination unit 180 uses the learned model stored in the model storage unit 140 to determine the state of the device based on the preprocessed determination data obtained by the determination preprocessing unit 160 (S370).
[0061] Figure 4An example of a display screen 400 according to the present embodiment is shown. The display screen 400 includes a distribution display unit 410 , a stability display unit 420 , a preprocessing parameter input unit 430 , and a preprocessing algorithm input unit 440 .
[0062] The distribution display unit 410 displays the distribution of the measured values of at least one sensor selected by the selection unit 185 in the preprocessed learning data and the preprocessed judgment data. In the example of this figure, the selection unit 185 selects the measured values of three sensors with relatively high contribution to the judgment result of the model. Here, the selection unit 185 can judge that the measured value of a certain sensor has a large contribution to the judgment result when the change of the measured value of the sensor has a large impact on the judgment result of the model. For example, in SVM, the selection unit 185 can also select a predetermined number of elements of the weight vector representing the hyperplane showing the boundary of the judgment result in the order of absolute value from large to small (for example, 2 in two-dimensional display and 3 in three-dimensional display), and select the measured value of the sensor corresponding to the selected element as the measured value with relatively high contribution. In addition, in the neural network, the selection unit 185 can select a predetermined number of sensors in the order of influence on the output value from large to small when the function from each input node that inputs the measured value of each sensor to the output node that outputs the judgment result is expanded.
[0063] The display processing unit 190 according to the present embodiment displays the distribution of the preprocessed learning data and the preprocessed determination data in a three-dimensional space as a three-dimensional scatter diagram on the distribution display unit 410, wherein the three-dimensional space has the measurement values of the three selected sensors as three axes. The circles in the figure represent each sample of the preprocessed learning data classified as normal, the cross represents each sample of the preprocessed learning data classified as abnormal, and the triangle represents each sample of the preprocessed determination data.
[0064] In this figure, the distribution of the preprocessed judgment data (the triangle in the figure) deviates from the distribution of the preprocessed learning data (the circle in the figure) which is classified as normal in the three-dimensional coordinate space. Here, in the case where the drift of the measured data is successfully removed, the distribution of the preprocessed judgment data should overlap with the distribution of the preprocessed learning data. Therefore, as shown in this figure, in the case where the distribution of the preprocessed judgment data deviates from the distribution of the preprocessed learning data, the display processing unit 190 can make the user realize that the drift of the judgment data has not been successfully removed sufficiently, and can take the following measures, namely, by Figure 3The drift of the determination data can be removed by changing the type of preprocessing in S360. In addition, the display processing unit 190 can allow the user to confirm whether the erroneous determination of the determination unit 180 is caused by the residual drift by displaying the distribution display unit 410 on the display screen 400.
[0065] The stability display unit 420 displays the time series data of the stability index value, which represents the judgment result of the state of the factory determined by the judgment unit 180 using the learned model and based on the preprocessed judgment data. The display processing unit 190 displays the judgment results of each timing obtained by assigning the preprocessed judgment data of each timing to the model as a time series chart. When using SVM, as the stability index value, the display processing unit 190 can display the product sum of the measured values of each sensor at each timing and the weight corresponding to each sensor as the stability index value. When using a neural network, the display processing unit 190 can also display the output value of the output node that outputs the judgment result when the measured values of each sensor are input to the input layer as the stability index value. The dotted line of the stability display unit 420 in the figure represents the normal and abnormal threshold values for the stability index value. The judgment unit 180 determines that the judgment object equipment of the factory is abnormal based on the fact that the stability index value is less than or equal to the threshold.
[0066] The preprocessing parameter input unit 430 is an input field displayed by the display processing unit 190 to receive the designation of the preprocessing parameters from the user. The preprocessing algorithm input unit 440 is an input field displayed by the display processing unit 190 to receive the designation of the type of preprocessing from the user. In this figure, the designation input unit 125 and the designation input unit 175 receive the type of preprocessing and other parameters for preprocessing through the input of the user to the preprocessing parameter input unit 430 and the preprocessing algorithm input unit 440. In the example of this figure, in the preprocessing algorithm input unit 440, the use of "center of gravity shift" is designated as the preprocessing algorithm, and in the preprocessing parameter input unit 430, "1 week" is designated as the preprocessing parameter. Receiving these, the designation input unit 125 and the designation input unit 175 set the learning preprocessing unit 110 and the judgment preprocessing unit 160 to reduce the drift of the center of gravity shift of the measurement data and set the window length to 1 week. At least one of preprocessing of reducing the center of gravity shift of the measurement data or preprocessing of calculating the difference of the measurement data within the sliding window can be performed in accordance with the designation of the learning preprocessing unit 110 and the judgment preprocessing unit 160, the preprocessing parameter input unit 430 and the preprocessing algorithm input unit 440.
[0067] According to the display screen 400 shown above, the user can confirm whether the drift of the measurement data has been successfully removed through the display of the distribution display unit 410. In addition, in the case where the drift has not been successfully removed, the designation input unit 125 and the designation input unit 175 can receive the designation of appropriate preprocessing from the user via the preprocessing parameter input unit 430 and the preprocessing algorithm input unit 440. In addition, in the case where the user who sees the display of the stability display unit 420 concludes that the determination of the determination device 150 is incorrect, the determination device 150 may also provide the group of the incorrectly determined timing measurement data and the correct state of the determination object device as additional learning data to the learning data acquisition unit 105 for relearning the model. In this case, the determination device 150 may also send a message to the learning device 100 for relearning the model. Figure 2 The learning device 100 may also perform the processing from S200 based on the instruction. Figure 2 Re-learning processing starts from S200.
[0068] Furthermore, when the feature quantity calculation unit 115 has a function of performing at least one of undersampling or oversampling of the learning data, the display processing unit 190 may also display at least one of an input field for specifying whether to apply undersampling / oversampling, an input field for specifying the rate of undersampling / oversampling, or an input field for specifying other parameters of undersampling / oversampling. In this case, the specification input unit 125 may input a specification related to undersampling / oversampling, and the feature quantity calculation unit 115 may perform undersampling / oversampling of the learning data based on the specification.
[0069] Figure 5 An example of measurement data according to this embodiment is shown. Figure 6 An example of measurement data after drift removal according to the present embodiment is shown. The horizontal axis of these graphs represents the passage of time, and the vertical axis represents the measurement value.
[0070] In the example of this figure, the measurement data is time series data of the measurement value of the temperature sensor for measuring the temperature of the measurement part. Figure 5 As shown, in the measurement data, drift occurs in which the center of gravity moves per rotation.
[0071] The feature quantity calculation unit 115 calculates the center of gravity of each sliding window and the cycle of the center of gravity change (one cycle in the example of this figure) as a feature quantity representing the drift of the measurement data in response to receiving the learning data including such measurement data. Figure 4 The preprocessing algorithm ( Figure 4"Center of gravity shift" in ) and preprocessing parameters ( Figure 4 In this case, the feature amount calculation unit 115 calculates the feature amount according to the designated preprocessing algorithm and preprocessing parameters.
[0072] The data conversion unit 120 may perform preprocessing according to the feature quantity calculated by the feature quantity calculation unit 115, that is, in each interval obtained by dividing the measurement data into the above-mentioned period (one week in the example of this figure), the value of the centroid of the interval is subtracted from the measurement data. Figure 6 As shown in FIG. 1 , the data conversion unit 120 can obtain measurement data from which the influence of periodic drift has been removed. The learning processing unit 130 can use the preprocessed learning data including the measurement data from which the influence of drift has been removed to learn the model. In addition, if the drift of the measurement data is removed by using the average or centroid of each interval obtained by dividing the measurement data into a certain period, there may be a difference between the boundaries of the intervals. Figure 6 The discontinuous points are shown around timings 24 to 29 and around timings 55 to 57 in FIG. Therefore, the learning processing unit 130 may not use the measured values between the predetermined margin periods after the section switching for learning.
[0073] In addition, in the determination phase based on the determination data, the determination device 150 can process the measurement data in the same manner as in the above-mentioned learning phase. Figure 5 The determination data including such measurement data is calculated accordingly, and the center of gravity of each sliding window and the period of center of gravity change are calculated as feature quantities representing the drift of the measurement data. Figure 4 The preprocessing algorithm and preprocessing parameters specified by the user are received through the preprocessing parameter input unit 430 and the preprocessing algorithm input unit 440. In this case, the feature quantity calculation unit 165 calculates the feature quantity according to the specified preprocessing algorithm and preprocessing parameter.
[0074] The data conversion unit 170 may perform preprocessing according to the feature quantity calculated by the feature quantity calculation unit 165, that is, in each interval obtained by dividing the measurement data into the above-mentioned period, the value of the centroid of the interval is subtracted from the measurement data. Figure 6 As shown in FIG. 1 , the data conversion unit 170 can obtain measurement data from which the influence of periodic drift has been removed. The determination unit 180 can use the preprocessed determination data including the measurement data from which the influence of drift has been removed to determine the state of the device. Figure 6For the discontinuous points shown around timings 24 to 29 and around timings 55 to 57 in FIG. 1 , the determination unit 180 may not use the measured values during the predetermined margin period after the section switching for determination.
[0075] According to the learning device 100 and the determination device 150 shown above, the model can be learned in a manner not affected by the drift of the measurement data, and the factory diagnosis can be performed using such a model. Therefore, even if the characteristics of the measurement data change over time, the learning device 100 and the determination device 150 can reduce the need to relearn the model.
[0076] In addition, the learning device 100 and the determination device 150 can also receive specifications of the type of preprocessing from the user and perform the specified preprocessing, so that the preprocessing can be changed in accordance with the following events known to the user, such as the shutdown of a factory or part of the equipment for periodic maintenance, the implementation of sensor calibration, or the setting changes of the sensor or the instrument that reads the sensor data.
[0077] Furthermore, according to the learning device 100 and the determination device 150 , since a model from which the influence of drift of measurement data is eliminated is used, the model can be applied to equipment having the same or similar structure in the same factory or in different factories without relearning.
[0078] In addition, in the above description, for the sake of convenience, an example is given of a case where one preprocessing is applied to the measurement data. Alternatively, the learning device 100 and the determination device 150 may perform different preprocessing on the measurement data from different sensors. For example, the learning device 100 and the determination device 150 may perform a gravity center shifting process with a cycle of one week on the measurement data a from sensors A1 to A3, perform a gravity center shifting process with a cycle of 30 minutes on the measurement data b from sensors B1 to B2, and perform a differential calculation process using a sliding window of one day on the measurement data c from sensor C.
[0079] In the above description, the case where the selection unit 185 and the display processing unit 190 are provided in the determination device 150 is exemplified. Alternatively, the selection unit 185 and the display processing unit 190 may be provided in the learning device 100. In addition, the learning device 100 may be provided as a device in which the determination device 150 is integrated, and conversely, the determination device 150 may be provided as a device in which the learning device 100 is integrated. In addition, at least one of the learning device 100 or the determination device 150 may be implemented by cloud computing.
[0080] Various embodiments of the present invention may be described with reference to flowcharts and block diagrams, where a block may represent (1) a stage of a process for performing an operation or (2) a portion of a device having the function of performing an operation. Specific stages and portions may be implemented by dedicated circuits, programmable circuits, and / or processors, the programmable circuits being supplied together with computer-readable instructions stored on a computer-readable medium, and the processors being supplied together with computer-readable instructions stored on a computer-readable medium. Dedicated circuits may include digital and / or analog hardware circuits, and may also include integrated circuits (ICs) and / or discrete circuits. Programmable circuits may include reconfigurable hardware circuits, the reconfigurable hardware circuits including logical AND, logical OR, logical XOR, logical NAND, logical NOR, and other logical operations, as well as memory elements such as flip-flops, registers, field programmable gate arrays (FPGAs), and programmable logic arrays (PLAs).
[0081] The computer-readable medium may include any tangible device capable of storing commands executed by an appropriate device, and as a result, the computer-readable medium having the commands stored therein has a product that includes commands that can be executed in order to create a means for performing the operations specified in the flowchart or block diagram. As examples of computer-readable media, electronic storage media, magnetic storage media, optical storage media, electromagnetic storage media, semiconductor storage media, etc. may be included. As more specific examples of computer-readable media, floppy (registered trademark) disks, magnetic disks, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), electrically erasable programmable read-only memories (EEPROM), static random access memories (SRAM), compact disk read-only memories (CD-ROM), digital versatile disks (DVD), Blu-ray (RTM) disks, memory sticks, integrated circuit cards, etc. may be included.
[0082] Computer-readable commands may include assembly commands, command set architecture (ISA) commands, machine commands, machine-dependent commands, microcode, firmware commands, state setting data, or any source code or object code written in any combination of one or more programming languages, including object-oriented programming languages such as Smalltalk, JAVA (registered trademark), C++, and existing procedural programming languages such as the "C" programming language or similar programming languages.
[0083] Computer readable instructions can be provided locally or via a local area network (LAN), a wide area network (WAN) such as the Internet to a processor or programmable circuit of a general-purpose computer, a special computer, or other programmable data processing device, and the computer readable instructions can be executed to create a means for performing the operations specified in the flowchart or block diagram. Examples of processors include computer processors, processing units, microprocessors, digital signal processors, controllers, microcontrollers, etc.
[0084] Figure 7 The computer 2200 is an example of a computer that can implement multiple embodiments of the present invention in whole or in part. The program installed in the computer 2200 can cause the computer 2200 to function as an operation associated with the device involved in the embodiment of the present invention or one or more parts of the device, or can cause the computer 2200 to perform the operation or the one or more parts, and / or can cause the computer 2200 to perform a process involved in the embodiment of the present invention or a stage of the process. Such a program can be executed by the CPU 2212 in order to cause the computer 2200 to perform specific operations associated with some or all blocks of the flowcharts and block diagrams described in this specification.
[0085] The computer 2200 according to the present embodiment includes a CPU 2212, a RAM 2214, a graphics controller 2216, and a display device 2218, which are connected to each other via a host controller 2210. The computer 2200 further includes input / output units such as a communication interface 2222, a hard disk drive 2224, a DVD-ROM drive 2226, and an IC card drive, which are connected to the host controller 2210 via an input / output controller 2220. The computer further includes a conventional input / output unit such as a ROM 2230 and a keyboard 2242, which are connected to the input / output controller 2220 via an input / output chip 2240.
[0086] The CPU 2212 controls each unit by operating according to the programs stored in the ROM 2230 and the RAM 2214. The graphic controller 2216 obtains image data generated by the CPU 2212 in a frame buffer or the like provided in the RAM 2214 or in itself, and displays the image data on the display device 2218.
[0087] The communication interface 2222 communicates with other electronic devices via a network. The hard disk drive 2224 stores programs and data used by the CPU 2212 in the computer 2200. The DVD-ROM drive 2226 reads programs or data from the DVD-ROM 2201 and provides the programs or data to the hard disk drive 2224 via the RAM 2214. The IC card drive reads programs and data from an IC card and / or writes programs and data to an IC card.
[0088] The ROM 2230 stores therein a startup program or the like executed by the computer 2200 upon activation, and / or a program that depends on the hardware of the computer 2200. The input / output chip 2240 can also connect various input / output units to the input / output controller 2220 via a parallel port, a serial port, a keyboard port, a mouse port, and the like.
[0089] The program is provided by a computer-readable medium such as DVD-ROM 2201 or IC card. The program is read from the computer-readable medium, installed in the hard disk drive 2224, RAM 2214 or ROM 2230, which are also examples of computer-readable media, and executed by CPU 2212. The information processing described in these programs is read by computer 2200, and cooperation between the program and the various types of hardware resources described above is established. The device or method can be constructed by implementing the operation or processing of information according to the use of computer 2200.
[0090] For example, when communication is performed between the computer 2200 and an external device, the CPU 2212 can execute the communication program loaded into the RAM 2214, and based on the processing described in the communication program, issue a command to perform communication processing to the communication interface 2222. Under the control of the CPU 2212, the communication interface 2222 reads the transmission data stored in the transmission buffer processing area provided in the RAM 2214, the hard disk drive 2224, the DVD-ROM 2201, or a recording medium such as an IC card, and transmits the read transmission data to the network, or writes the reception data received from the network to the reception buffer processing area provided on the recording medium.
[0091] In addition, the CPU 2212 can read all or a necessary part of a file or database stored in an external recording medium such as the hard disk drive 2224, the DVD-ROM drive 2226 (DVD-ROM 2201), or an IC card into the RAM 2214, and perform various types of processing on the data in the RAM 2214. Then, the CPU 2212 writes the processed data back to the external recording medium.
[0092] Various types of information such as various types of programs, data, tables, and databases can be stored in the recording medium and subjected to information processing. The CPU 2212 can perform various types of processing, including various types of operations specified by the command sequence of the program, information processing, conditional judgment, conditional branching, unconditional branching, information retrieval / replacement, etc., which are recorded everywhere in the present invention on the data read from the RAM 2214, and write the results back to the RAM 2214. In addition, the CPU 2212 can retrieve information in files, databases, etc. in the recording medium. For example, in the case where a plurality of entries each having an attribute value of the first attribute associated with an attribute value of the second attribute is stored in the recording medium, the CPU 2212 retrieves an entry that specifies an attribute value of the first attribute and is consistent with the condition from the plurality of entries, reads the attribute value of the second attribute stored in the entry, and thereby obtains the attribute value of the second attribute associated with the first attribute that satisfies the predetermined condition.
[0093] The program or software module described above may be stored in a computer-readable medium on or near the computer 2200. In addition, a recording medium such as a hard disk or RAM provided in a server system connected to a dedicated communication network or the Internet can be used as a computer-readable medium, thereby providing the program to the computer 2200 via the network.
[0094] The present invention has been described above using the embodiments, but the technical scope of the present invention is not limited to the scope described in the above embodiments. It should be clear to those skilled in the art that various changes or improvements can be made to the above embodiments. According to the description of the claims, the method of making such changes or improvements is also included in the technical scope of the present invention.
[0095] It should be noted that the execution order of each process such as actions, processes, steps and stages in the devices, systems, programs and methods shown in the claims, specifications and drawings can be implemented in any order unless it is specifically and clearly indicated as "earlier" or "in advance" and the output of the previous process is not used for the subsequent process. Even if the action flow in the claims, specifications and drawings is described using "first" or "next" for convenience, it does not mean that it must be implemented in this order.
[0096] Description of the label
[0097] 100 learning device, 105 learning data acquisition unit, 110 learning preprocessing unit, 115 feature value calculation unit, 120 data conversion unit, 125 designation input unit, 130 learning processing unit, 140 model storage unit, 150 determination device, 155 determination data acquisition unit, 160 determination preprocessing unit, 165 feature value calculation unit, 170 data conversion unit, 175 designation input unit, 180 determination unit, 185 selection unit, 190 display processing unit, 400 display screen, 410 distribution display Department, 420 stability display department, 430 preprocessing parameter input department, 440 preprocessing algorithm input department, 2200 computer, 2201 DVD-ROM, 2210 host controller, 2212 CPU, 2214 RAM, 2216 graphic controller, 2218 display device, 2220 input / output controller, 2222 communication interface, 2224 hard disk drive, 2226 DVD-ROM drive, 2230 ROM, 2240 input / output chip, 2242 keyboard.
Claims
1. A learning device, comprising: a learning data acquisition unit that acquires learning data including measurement data obtained by measuring a device and a state of the device; a learning preprocessing unit that performs preprocessing to reduce drift of the measurement data in the learning data and outputs the preprocessed learning data; a learning processing unit that uses the preprocessed learning data to perform a process of learning a model that determines the state of the device based on the preprocessed measurement data; a determination data acquisition unit that acquires determination data including measurement data obtained by measuring the device; a determination preprocessing unit that performs preprocessing to reduce drift of the determination data and outputs the preprocessed determination data; as well as a determination unit that uses the learned model to determine the state of the device based on the preprocessed determination data, The learning preprocessing unit includes a feature quantity calculation unit that calculates a feature quantity indicating a drift of the measurement data in the learning data. The feature amount calculation unit further reduces imbalance in the number of samples by under-sampling normal data in the measurement data in the learning data and / or over-sampling abnormal data in the measurement data in the learning data.
2. The learning device according to claim 1, wherein: The learning preprocessing unit comprises: A data conversion unit removes a drift corresponding to the feature amount from the measurement data in the learning data.
3. The learning device according to claim 1 or 2, wherein: The learning device further includes a designation input unit for inputting a designation of a type of preprocessing to be performed by the learning preprocessing unit. The learning preprocessing unit performs preprocessing corresponding to the input designation.
4. The learning device according to claim 3, wherein: The learning preprocessing unit performs at least one of preprocessing for reducing a center of gravity shift of the measurement data in the learning data or preprocessing for calculating a difference of the measurement data in the learning data within a sliding window in accordance with the designation.
5. The learning device according to claim 1, wherein: The learning preprocessing unit performs preprocessing of calculating the difference of the measurement data in the learning data within a sliding window. When using the difference of the measured values within the sliding window, such as the value obtained by subtracting the minimum value from the maximum value of the measured values within the sliding window, the initial and final values of the measured values within the sliding window, or the maximum value of the difference between adjacent measured values within the sliding window, the influence of the change in the average value is removed along with the calculation.
6. The learning device according to claim 1, wherein: The device further includes a display processing unit that performs processing for displaying the distribution of the preprocessed learning data and the preprocessed completion determination data.
7. The learning device according to claim 6, wherein: The learning data and the determination data include measurement values of a plurality of sensors installed in the device. The learning device further includes a selection unit that selects, among the measurement values of the plurality of sensors, at least one measurement value of a sensor that has a relatively high contribution to the determination result of the model. The display processing unit performs processing for displaying a distribution of the measurement values of the at least one sensor selected by the selection unit in the preprocessed learning data and the preprocessed completion determination data.
8. A learning method comprising the following steps: The learning device acquires learning data including measurement data obtained by measuring the device and the state of the device; The learning device performs preprocessing to reduce drift of the measurement data in the learning data, and outputs the preprocessed learning data; The learning device uses the preprocessed learning data to learn a model, and the model determines the state of the device based on the preprocessed measurement data; The learning device acquires determination data including measurement data obtained by measuring the device; The learning device performs preprocessing to reduce drift of the determination data and outputs the preprocessed determination data; as well as The learning device uses the learned model to determine the state of the device based on the preprocessing completion determination data. The learning process includes: calculating a feature quantity indicating a drift of the measurement data in the learning data; Calculating the feature value further includes under-sampling normal data in the measurement data in the learning data and / or over-sampling abnormal data in the measurement data in the learning data to reduce an imbalance in the number of samples.
9. A computer-readable medium recording a learning program, the learning program being executed by a computer and for causing the computer to function as a learning device, The learning device has: a learning data acquisition unit that acquires learning data including measurement data obtained by measuring a device and a state of the device; a learning preprocessing unit that performs preprocessing to reduce drift of the measurement data in the learning data and outputs the preprocessed learning data; a learning processing unit that uses the preprocessed learning data to perform a process of learning a model that determines the state of the device based on the preprocessed measurement data; a determination data acquisition unit that acquires determination data including measurement data obtained by measuring the device; a determination preprocessing unit that performs preprocessing to reduce drift of the determination data and outputs the preprocessed determination data; as well as a determination unit that uses the learned model to determine the state of the device based on the preprocessed determination data, The learning preprocessing unit includes a feature quantity calculation unit that calculates a feature quantity indicating a drift of the measurement data in the learning data. The feature amount calculation unit further reduces imbalance in the number of samples by under-sampling normal data in the measurement data in the learning data and / or over-sampling abnormal data in the measurement data in the learning data.
10. A determination device for use in the learning device according to claim 1, comprising: a determination data acquisition unit that acquires determination data including measurement data obtained by measuring the device; a determination preprocessing unit that performs preprocessing to reduce drift of the determination data and outputs the preprocessed determination data; and A determination unit determines the state of the device based on the preprocessed determination data using a model learned using preprocessed learning data obtained by preprocessing learning data.
11. A determination method using the determination device according to claim 10, comprising: The determination device obtains determination data including measurement data obtained by measuring the device; The determination device performs preprocessing to reduce drift of the determination data and outputs the preprocessed determination data; and The determination device determines the state of the device based on the preprocessed determination data using a model learned using preprocessed learning data obtained by preprocessing learning data.
12. A computer-readable medium recording a determination program, the determination program being executed by a computer and for causing the computer to function as the determination device according to claim 10, the determination device comprising a determination data acquisition unit, a determination preprocessing unit, and a determination unit, wherein: The determination data acquisition unit acquires determination data including measurement data obtained by measuring the device. The determination preprocessing unit performs preprocessing to reduce drift of the determination data and outputs the preprocessed determination data. The determination unit determines the state of the device based on the preprocessed determination data using a model learned using preprocessed learning data obtained by preprocessing learning data.
Citation Information
Patent Citations
Systems and methods for continuously modeling industrial asset performance
CN110337616A
Double-window concept drift detection method based on sample distribution statistical test
CN110717543A
Instrument calibration support system
JP2005043121A
Data analysis device, method and program
JP2015170281A
Machine learning device and machine learning method for learning fault condition and fault prediction device and fault prediction system equipped with machine learning device
JP2017033526A
