Abnormality determination device, abnormality determination method, and program

Abstract

The present invention more accurately determines an abnormality. This abnormality determination device includes: a data acquisition unit that acquires abnormal time-series data, which is time-series data of a rotary machine in an abnormal state; a generation unit that corrects the abnormal time-series data and generates a plurality of pieces of simulated data, which is time-series data simulating abnormality; a learning unit that causes a learning model to learn the simulated data and a label indicating the state of the simulated data as training data and generates a trained model; a determination unit that inputs target time-series data, which is target time-series data of the rotary machine, into the trained model and performs abnormality determination of the target time-series data; and an output control unit that outputs the result of the abnormality determination of the target time-series data.

Description

Abnormality determination apparatus, abnormality determination method and program

[0001] The present disclosure relates to an abnormality determination apparatus, an abnormality determination method and a program.

[0002] Techniques for identifying abnormality causes of machines are known. For example, Patent Document 1 describes that accumulated past data of a rotating machine, simulation data is generated from the past data, machine learning is performed using the accumulated past data and the simulation data, and an abnormality cause of the rotating machine is identified using the learned data.

[0003] Japanese Unexamined Patent Publication No. 2020-64024

[0004] When determining an abnormality of a rotating machine using learned data, it is required to determine the abnormality with higher accuracy.

[0005] An object of the present disclosure is to provide an abnormality determination apparatus, an abnormality determination method and a program that can determine an abnormality with higher accuracy.

[0006] The abnormality determination apparatus according to the present disclosure includes: a data acquisition unit that acquires abnormal time-series data, which is time-series data in an abnormal state of a rotating machine; a generation unit that corrects the abnormal time-series data and generates a plurality of pieces of simulated data, which are time-series data simulating an abnormality; a learning unit that causes a learning model to perform learning using the simulated data and a label indicating a state of the simulated data as teacher data to generate a trained model; a determination unit that inputs target time-series data, which is time-series data targeted for a rotating machine, to the trained model and performs abnormality determination on the target time-series data; and an output control unit that outputs a result of the abnormality determination on the target time-series data.

[0007] The abnormality detection method relating to this disclosure includes the steps of: acquiring abnormal time series data, which is time series data of an abnormal state of a rotating machine; correcting the abnormal time series data to generate a plurality of simulated data, which is time series data that simulates an abnormality; training a learning model using the simulated data and labels indicating the state of the simulated data as training data to generate a trained model; inputting target time series data, which is time series data targeted by the rotating machine, into the trained model to perform abnormality detection on the target time series data; and outputting the result of the abnormality detection on the target time series data.

[0008] The program relating to this disclosure causes a computer to perform the following steps: acquire abnormal time series data, which is time series data of an abnormal state of a rotating machine; correct the abnormal time series data to generate multiple simulated data, which is time series data that simulates the abnormality; train a learning model using the simulated data and labels indicating the state of the simulated data as training data to generate a trained model; input target time series data, which is time series data of the rotating machine, into the trained model to perform abnormality determination of the target time series data; and output the result of the abnormality determination of the target time series data.

[0009] According to this disclosure, it is possible to determine anomalies with greater accuracy.

[0010] Figure 1 is a block diagram of the anomaly detection device of this embodiment. Figure 2 is a diagram showing an example of simulated data. Figure 3 is a diagram showing an example of simulated data. Figure 4 is a diagram showing an example of simulated data. Figure 5 is a diagram showing an example of simulated data. Figure 6 is a flowchart showing an example of the processing of the anomaly detection device.

[0011] Preferred embodiments of the present disclosure will be described in detail below with reference to the attached drawings. However, this disclosure is not limited to these embodiments, and if there are multiple embodiments, they may be combinations of these embodiments.

[0012] (Anomaly Detection Device) Figure 1 is a block diagram showing the anomaly detection device of this embodiment. The anomaly detection device 10 according to this embodiment generates multiple simulated data from time-series data of a rotating machine, trains a learning model with the generated simulated data, and generates a trained model. The anomaly detection device 10 then inputs the target time-series data of the rotating machine that is the subject of anomaly detection into the trained model, determines whether there is an anomaly in the rotating machine, and outputs the determination result. The rotating machine may be any machine that performs work by rotation, such as a compressor or a turbine. The anomaly detection device 10 may also determine an anomaly in the entire rotating machine, or it may determine an anomaly in a component mounted on the rotating machine. Examples of components mounted on a rotating machine include shafts and bearings.

[0013] Furthermore, the term "time-series data" here refers to data detectable from rotating machinery, whose values ​​can change over time. Time-series data can also be described as data indicating the state of the rotating machinery, detected by sensors installed on the machinery. Examples of time-series data include data on the shaft vibration of the rotating machinery and data on the temperature of the bearings of the rotating machinery. For example, the sensor used to detect time-series data may be a vibration sensor, and the time-series data detected by that sensor will be the shaft vibration values ​​detected for each time series. Alternatively, the sensor used to detect time-series data may be a thermometer, and the time-series data detected by that sensor will be the bearing temperatures detected for each time series.

[0014] The abnormality detection device 10 according to this embodiment is, for example, a computer and, as shown in Figure 1, has an input unit 12, an output unit 14, a communication unit 16, a storage unit 18, and a control unit 20. The input unit 12 is a device that accepts user input and may be, for example, a mouse, keyboard, or touch panel. The output unit 14 is a device that outputs information and may be, for example, a display that shows images. The communication unit 16 is a module that communicates with external devices and may include, for example, an antenna. In this embodiment, the communication method used by the communication unit 16 is wireless communication, but the communication method may be arbitrary. The abnormality detection device 10 does not necessarily have an input unit 12, an output unit 14, or a communication unit 16. Furthermore, the abnormality detection device 10 may be configured as a standalone device, integrated with other devices, or as a system combining various devices such as a computing unit and a data server, and is not particularly limited.

[0015] The memory unit 18 is a memory that stores various information such as the calculation contents and programs of the control unit 20, and includes at least one of the following: a main memory device such as RAM (Random Access Memory) and ROM (Read Only Memory), and an external memory device such as an HDD (Hard Disk Drive). The program for the control unit 20 stored in the memory unit 18 may be stored on a recording medium that can be read by the abnormality determination device 10.

[0016] The control unit 20 is an arithmetic unit and includes arithmetic circuits such as a CPU (Central Processing Unit). The control unit 20 includes a data acquisition unit 30, a preprocessing unit 32, a generation unit 34, a learning unit 36, a determination unit 38, and an output control unit 40. The control unit 20 reads a program (software) from the storage unit 18 and executes it to realize the data acquisition unit 30 and the output control unit 40, and then executes their processing. The control unit 20 may execute these processes with a single CPU, or it may have multiple CPUs and execute the processing with those multiple CPUs. Furthermore, at least a part of the data acquisition unit 30, preprocessing unit 32, generation unit 34, learning unit 36, determination unit 38, and output control unit 40 may be realized in hardware.

[0017] The following describes the processing details of the abnormality detection device 10.

[0018] (Acquisition of abnormal time-series data) The data acquisition unit 30 acquires abnormal time-series data, which is time-series data of the rotating machine in an abnormal state. Abnormal time-series data refers to time-series data of the rotating machine that was detected during the period when the rotating machine was in an abnormal state. The data acquisition unit 30 may acquire abnormal time-series data by any method, for example, by acquiring abnormal time-series data from a sensor via the communication unit 16, or by having an operator input abnormal time-series data into the input unit 12. The data acquisition unit 30 may also acquire normal time-series data, which is time-series data of the rotating machine in a normal state. Normal time-series data refers to time-series data of the rotating machine that was detected during the period when the rotating machine was in a normal state.

[0019] Furthermore, the data acquisition unit 30 may acquire a label indicating the state of the abnormal time series data (the state of the rotating machine during the period in which the abnormal time series data was detected) along with the abnormal time series data. Since the abnormal time series data is time series data in an abnormal state, the label of the abnormal time series data will indicate that the abnormal time series data (the rotating machine during the period in which the abnormal time series data was detected) is abnormal. However, the label is not limited to indicating only that it is abnormal, and in this embodiment, it is information indicating the abnormal mode of the time series data. Specifically, the label of the abnormal time series data in this embodiment may be data indicating the shape of the waveform of the abnormal time series data, and may be at least one of the following: the waveform of the time series data is spike-shaped, the time series data continues to rise as time progresses, or the time series data continues to fall as time progresses. A spike-shaped waveform means that the time series data contains waveforms in which the peak value is high (for example, above a predetermined value) and the time width of the peak is small (for example, below a predetermined time). Furthermore, when time-series data is said to be continuously increasing over time, it means that the period of time during which the time-series data continues to increase is long (for example, longer than a predetermined time). Similarly, when time-series data is said to be continuously decreasing over time, it means that the period of time during which the time-series data continues to decrease is long (for example, longer than a predetermined time). The data acquisition unit 30 may acquire labels for abnormal time-series data by any method; for example, it may acquire labels when an operator inputs them into the input unit 12.

[0020] Furthermore, if the data acquisition unit 30 also acquires normal time series data, it may also acquire labels for the normal time series data. The labels for the normal time series data may indicate that the normal time series data (the rotating machine during the period in which the normal time series data was detected) is normal.

[0021] Furthermore, the data acquisition unit 30 may also acquire correction information (described later) for correcting abnormal time-series data.

[0022] (Preprocessing) The preprocessing unit 32 performs preprocessing on the abnormal time-series data acquired by the data acquisition unit 30. Preprocessing here refers to extraction, normalization, and filtering.

[0023] The preprocessing unit 32 performs an extraction process to extract abnormal time series data from the acquired abnormal time series data within a predetermined time range. The predetermined time range is, for example, the time range in which the learning model can learn normally.

[0024] Furthermore, the preprocessing unit 32 performs a normalization process on the abnormal time series data. For example, the preprocessing unit 32 performs a normalization process on the abnormal time series data extracted in the extraction process. Normalization refers to the process of normalizing the values ​​of the abnormal time series data (for example, the values ​​of axial vibration) so that they fall within a predetermined numerical range. For example, the preprocessing unit 32 may normalize the values ​​of each time series of the abnormal time series data so that the mean is 0 and the deviation is 1. An example of a normalization process is to subtract the mean from the original data and divide by the standard deviation.

[0025] The preprocessing unit 32 performs filtering on the abnormal time series data. The preprocessing unit 32 may perform filtering on the acquired abnormal time series data before performing normalization, or it may perform filtering on the abnormal time series data after normalization. Filtering refers to the process of removing waveforms from the abnormal time series data in which the amplitude of the peak intensity is less than a threshold.

[0026] Furthermore, if the data acquisition unit 30 has also acquired normal time series data, the preprocessing unit 32 may preprocess the normal time series data in the same manner as the abnormal time series data.

[0027] In the above, extraction, normalization, and filtering were performed as preprocessing steps, but it is also acceptable to perform at least one of these steps. Furthermore, preprocessing is not mandatory, and may be omitted.

[0028] (Generating Simulated Data) Figures 2 to 5 show examples of simulated data. Line D in Figures 2 to 5 shows an example of simulated data. In Figures 2 to 5, the horizontal axis represents time, and the vertical axis represents the values ​​of the simulated data (e.g., vibration values).

[0029] The generation unit 34 corrects the acquired abnormal time series data and generates multiple simulated data sets, which are time series data that simulate the abnormality. In other words, the simulated data is not the time series data actually detected by the rotating machine's sensor, but rather time series data that simulates the time series data in an abnormal state, generated by correcting the abnormal time series data. In this embodiment, since the abnormal time series data is preprocessed, the generation unit 34 generates the simulated data based on the preprocessed abnormal time series data, but it is not limited to this, and may also generate simulated data based on abnormal time series data that has not been preprocessed.

[0030] The generation unit 34 uses abnormal time series data (for example, abnormal time series data about the shaft vibration of a rotating machine) to generate simulated data of the same type as the abnormal time series data, but with different waveforms (for example, time series data about the shaft vibration of a rotating machine, but with different waveforms). The generation unit 34 may generate the simulated data using any method with the abnormal time series data, but in this embodiment, the generation unit 34 generates the simulated data by correcting the abnormal time series data based on correction information. The correction information is information for generating the simulated data, or in other words, information that indicates which feature quantities of the abnormal time series data to change. The correction information is input (specified) by an operator, for example. The correction information, i.e., the feature quantity to be changed, may be at least one of the following: the period of the time series data, the timing when the time series data peaks, the number of peaks included in the time series data, and the intensity of the peaks included in the time series data.

[0031] In other words, the generation unit 34 of this embodiment generates a plurality of simulated data by changing at least one of the period, peak timing, number of peaks, and peak intensity of the acquired abnormal time series data. For example, if the correction information specifies a period T, time t, number of peaks n, and peak intensity H, the generation unit 34 generates n peaks with a period T, time t, and peak intensity H. The correction information may have a range set for time t. For example, if the correction information specifies a period T, time t ｘ ~t ｙ Given the number of peaks n and the peak intensity H, a time t is generated with a period T. ｘ ~t ｙ At this point in time, n peaks are generated with peak intensity H.

[0032] Furthermore, the generation unit 34 sets labels for the generated simulated data that indicate the abnormal state of the simulated data. That is, the generation unit 34 generates multiple datasets of simulated data and labels. The generation unit 34 may arbitrarily set the labels for the simulated data, but for example, it may set the labels associated with the abnormal time series data used to generate the simulated data as the labels for the simulated data. That is, for example, if there is abnormal time series data that has been labeled by the data acquisition unit 30 as having a spike-like waveform, the labels for the simulated data generated based on that abnormal time series data will be labels indicating that the waveform is spike-like.

[0033] Furthermore, if the data acquisition unit 30 has also acquired normal time series data, the generation unit 34 may generate normal simulated data, which is time series data that simulates normal conditions, based on the normal time series data. In this case, the method for generating normal simulated data based on normal time series data is the same as the method for generating simulated data based on abnormal time series data, so the explanation will be omitted. The generation unit 34 will set a label indicating that the normal simulated data is normal.

[0034] Figure 2 shows an example of normal simulated data. As shown in Figure 2, the normal simulated data is generated based on normal time series data and is labeled as normal.

[0035] Figure 3 shows an example of simulated data for an anomaly (a). An anomaly (a) refers to an abnormal state in which the time-series data continues to decline over time. As shown in Figure 3, the simulated data for an anomaly (a) shows that the value continues to decline over time and is labeled as an anomaly (a) (a label indicating that the value continues to decline over time).

[0036] Figure 4 shows an example of simulated data for an anomaly (b). An anomaly (b) refers to an abnormal state in which the time-series data continues to rise over time. As shown in Figure 4, the simulated data for an anomaly (b) shows that the value continues to rise over time and is labeled as an anomaly (b) (a label indicating that the value continues to rise over time).

[0037] Figure 5 shows an example of simulated data for anomaly (c). Anomaly (c) indicates an abnormal state in which the time-series data contains spike-like waveforms. As shown in Figure 5, the simulated data for anomaly (c) contains spike-like waveforms and is labeled as anomaly (c) (a label indicating the presence of spike-like waveforms). Figure 5(a) shows the case with one spike-like waveform, and Figure 5(b) shows the case with multiple spike-like waveforms.

[0038] The labels associated with the simulated data may be in any format indicating abnormal or normal conditions, but they may also be defined by numerical values, for example. That is, for example, the generation unit 34 may associate a label of 0 with normal simulated data generated by correcting normal time series data. The generation unit 34 may also associate a label of a value other than 0 with simulated data generated by correcting abnormal time series data. In this case, for example, different numerical values ​​may be assigned to different types of labels associated with abnormal simulated data. For example, the generation unit 34 may associate a label of 1 with simulated data showing a downward trend, a label of 2 with simulated data showing an upward trend, and a label of 3 with simulated data containing spike-like waveforms.

[0039] Furthermore, it is preferable that the generation unit 34 adds noise using random numbers to the abnormal time series data to generate simulated data. When the generation unit 34 generates normal simulated data from normal time series data, it is preferable not to add noise using random numbers.

[0040] (Generation of a Learning Model) The learning unit 36 ​​trains the pre-training learning model to generate a learning model that has been trained to recognize the correspondence between time-series data and the presence or absence of abnormalities in the rotating machinery (time-series data). The learning unit 36 ​​uses simulated data and the labels associated with that simulated data as training data to train the learning model and generate a trained model. The learning model may be any AI model, but it is a learning model trained by deep learning, and consists of a model that defines the neural network constituting the classifier trained by deep learning (neural network configuration information) and variables. The learning model N may be capable of determining the label of the data based on the input data. In the example of this embodiment, the learning model N is an RNN (Recurrent Neural Network) model, but it is not limited to an RNN model and may be any type of learning model. An RNN is a neural network that has an input layer, a hidden layer, and an output layer, and has a structure in which the output of the hidden layer (forward propagation) is recursively input to the hidden layer (backpropagation).

[0041] The learning unit 36 ​​takes simulated data as input values ​​and outputs the labels of the input simulated data, and inputs this dataset as training data into the learning model. The learning unit 36 ​​prepares multiple datasets of simulated data and labels, and inputs each of the multiple datasets into the learning model. In other words, in this embodiment, since multiple sets of simulated data with different waveforms are generated, the learning model can be trained with high accuracy by using multiple sets of simulated data with different waveforms as training data.

[0042] Furthermore, the learning unit 36 uses at least a plurality of pieces of simulated data generated from abnormal time-series data as teacher data, but normal simulated data generated from normal time-series data may also be used as teacher data. In this case, the learning unit 36 uses the normal simulated data and the label (normal label) of the normal simulated data as teacher data. Further, the learning unit 36 may also use the abnormal time-series data itself as teacher data. In this case, the learning unit 36 uses the abnormal time-series data and the label of the abnormal time-series data as teacher data. Further, the learning unit 36 may also use the normal time-series data itself as teacher data. In this case, the learning unit 36 uses the normal time-series data and the label of the normal time-series data as teacher data.

[0043] The learning unit 36 may set the order in which a dataset of teacher data is input to a learning model based on label types. For example, after inputting a dataset of normal simulated data (simulated data with a normal label assigned thereto) to a learning model to cause the learning model to perform learning, the learning unit 36 may input a dataset of simulated data (simulated data with an abnormal label assigned thereto) to the learning model to cause the learning model to perform learning. For example, in the present embodiment, it is preferable that the generating unit 34 inputs the dataset of teacher data to the learning model in ascending order of the labels associated with the simulated data. That is, for example, first, the learning unit 36 causes the learning model to learn normal simulated data assigned a label of 0 and the label 0. Next, the learning unit 36 causes the learning model to learn downward-trending simulated data associated with a label of 1 and the label 1. Next, the learning unit 36 causes the learning model to learn upward-trending simulated data associated with a label of 2 and the label 2. Finally, the learning unit 36 causes the learning model to learn simulated data in which a spike-shaped waveform associated with a label of 3 is generated and the label 3. By causing the learning model to perform learning based on the order of labels, the learning unit 36 improves the output accuracy of the trained model. Note that the order in which the learning unit 36 learns the simulated data and labels is not limited to ascending order. For example, the order in which the learning unit 36 learns the simulated data and labels may be descending order, and is not limited to ascending or descending order, and learning may be performed with the order changed.

[0044] (Abnormality Determination) The abnormality determination apparatus 10 determines (monitors) whether there is an abnormality in a rotating machine using time-series data to be subjected to abnormality determination, based on the learning model that has undergone machine learning as described above. Hereinafter, a method for determining whether there is an abnormality in a rotating machine will be described. Note that in the present embodiment, the abnormality determination apparatus 10 performs both a process of causing the learning model to perform machine learning and an abnormality determination process using the learning model described later. However, the present invention is not limited thereto, and the abnormality determination apparatus 10 does not have to perform the process of causing the learning model to perform machine learning, and may perform only the abnormality determination process using the learning model described later. In this case, the abnormality determination apparatus 10 acquires, for example via communication, the learning model that has undergone machine learning, stores the learning model in the storage unit 18, and determines whether there is an abnormality in the rotating machine using the learning model read from the storage unit 18.

[0045] (Acquisition of Target Time-Series Data) The data acquisition unit 30 acquires target time-series data that is time-series data to be subjected to abnormality determination. The data acquisition unit 30 acquires the target time-series data from a sensor of the rotating machine, for example via the communication unit 16.

[0046] (Determination Unit) The determination unit 38 inputs the target time-series data to the trained model and performs abnormality determination on the target time-series data. The determination unit 38 inputs the target time-series data to the learning model that has undergone machine learning of the correspondence relationship between time-series data of the same type as the target time-series data and the presence or absence of an abnormality in the rotating machine, and performs abnormality determination on the target time-series data. That is, for example, when the target time-series data is shaft vibration, the determination unit 38 inputs the target time-series data to the learning model that has undergone machine learning of the correspondence relationship between shaft vibration time-series data and the presence or absence of an abnormality.

[0047] The determination unit 38 may input the target time-series data (raw data) itself acquired by the data acquisition unit 30 to the trained model. Alternatively, for example, the preprocessing unit 32 may preprocess the target time-series data, and the preprocessed target time-series data may be input to the trained model. The preprocessing method for the target time-series data is the same as the preprocessing for the abnormal time-series data described above, and thus description thereof is omitted.

[0048] When the trained model receives the target time series data as input, it outputs a determination result indicating whether or not there are anomalies in the target time series data. The determination unit 38 acquires this determination result output from the trained model. In this embodiment, the determination unit 38 acquires information indicating the label associated with the target time series data, calculated by the trained model, as the determination result. For example, the determination unit 38 may acquire the numerical value of the label associated with the target time series data from the trained model as the determination result, or it may acquire the probability that the target time series data is assigned to each label as the determination result.

[0049] Furthermore, when the determination unit 38 determines whether the waveform of the target time series data is abnormal, it may make the abnormal determination after a predetermined time (for example, after a few minutes). For example, the determination unit 38 may determine that the waveform of the target time series data includes a spike-like waveform several minutes after it reaches a peak (for example, after detecting a downward trend in the waveform). Also, the determination unit 38 may have the learning unit 36 ​​retrain the determination result. For example, if the determination unit 38 incorrectly determines that the target time series data is abnormal, it will have the learning unit 36 ​​retrain the correct determination result. For example, if the determination unit 38 determines that the target time series data includes a spike-like waveform and is showing an upward trend, it will have the learning unit 36 ​​retrain that the target time series data includes a spike-like waveform.

[0050] (Output of Results) The output control unit 40 outputs the determination result acquired by the determination unit 38. The output method by the output control unit 40 is arbitrary. For example, the output control unit 40 may output the determination result to the output unit 14 (for example, to display it on a display). Alternatively, the output control unit 40 may transmit (output) the determination result to another device via the communication unit 16.

[0051] (Processing Flow) The processing flow of the anomaly detection device 10 described above will now be explained. Figure 6 is a flowchart showing an example of the processing of the anomaly detection device. As shown in Figure 6, the anomaly detection device 10 acquires anomaly time series data using the data acquisition unit 30 (step S10). The anomaly detection device 10 performs preprocessing on the anomaly time series data using the preprocessing unit 32 (step S12). The anomaly detection device 10 generates multiple simulated data by correcting the anomaly time series data using the generation unit 34 (step S14). The anomaly detection device 10 prepares multiple datasets consisting of simulated data and labels indicating the state of the simulated data using the learning unit 36, and trains a learning model using the multiple datasets as training data to generate a trained model (step S16). The anomaly detection device 10 acquires the target time series data to be subject to anomaly detection using the data acquisition unit 30, inputs the target time series data into the trained model using the determination unit 38, performs anomaly detection on the target time series data (step S18), and outputs the determination result using the output control unit 40 (step S20).

[0052] As described above, the anomaly detection device 10 according to this embodiment acquires anomaly time-series data, corrects the anomaly time-series data, and generates multiple simulated data. According to this embodiment, by generating multiple simulated data, it is possible to generate simulated data that more closely resembles the anomaly state. Therefore, according to this embodiment, anomalies can be detected with greater accuracy.

[0053] Here, in order to identify the cause of the anomaly, simulation data is generated by simulating the anomaly data, and this data is used to train a learning model, which then identifies the cause of the anomaly. This simulation data is generated based on the machine and is used to identify the cause of the anomaly at the fault location. However, the anomaly state data differs depending on the cause, etc., and generating multiple simulated data of possible scenarios makes it easier to determine the anomaly. In contrast, the anomaly determination device 10 of this embodiment makes it easier to determine the anomaly by performing the above processing, and can determine the anomaly with greater accuracy.

[0054] (Effects) As described above, the abnormality determination device 10 according to the first aspect of this disclosure includes: a data acquisition unit 30 that acquires abnormal time series data, which is time series data of an abnormal state of a rotating machine; a generation unit 34 that corrects the abnormal time series data and generates a plurality of simulated data, which is time series data that simulates an abnormality; a learning unit 36 ​​that uses the simulated data and labels indicating the state of the simulated data as training data to train a learning model and generate a trained model; a determination unit 38 that inputs target time series data of the rotating machine to the trained model and performs abnormality determination of the target time series data; and an output control unit 40 that outputs the result of the abnormality determination of the target time series data. According to this disclosure, abnormalities can be determined with greater accuracy.

[0055] An anomaly detection device 10 according to a second aspect of this disclosure is the anomaly detection device 10 according to the first aspect, wherein the generation unit 34 generates simulated data by changing at least one of the following: the period of the anomaly time series data, the timing at which the anomaly time series data peaks, the number of peaks in the anomaly time series data, and the intensity of the peaks in the anomaly time series data. According to this disclosure, anomalies can be detected with greater accuracy.

[0056] An anomaly detection device 10 according to a third aspect of this disclosure is an anomaly detection device 10 according to a second aspect, wherein the learning unit 36 ​​generates a trained model using the fact that the waveform of the simulated data is spike-shaped as a label, and the determination unit 38 detects that the waveform of the target time series data is spike-shaped as an anomaly detection result. According to this disclosure, anomalies can be detected with greater accuracy.

[0057] An anomaly detection device 10 according to the fourth aspect of this disclosure is an anomaly detection device 10 according to any of the first to third aspects, wherein the learning unit 36 ​​generates a trained model using as labels whether the simulated data is continuously increasing or continuously decreasing over time, and the determination unit 38 detects as an anomaly determination that the target time series data is continuously increasing or continuously decreasing over time. According to this disclosure, anomalies can be determined with greater accuracy.

[0058] An anomaly detection device 10 according to the fifth aspect of this disclosure is an anomaly detection device 10 according to any of the first to fourth aspects, further comprising a preprocessing unit 32 that normalizes anomaly time series data, and a learning unit 36 ​​generates a trained model by training a learning model with simulated data generated based on the normalized anomaly time series data. According to this disclosure, by normalizing the data, unnecessary data such as noise can be removed.

[0059] An anomaly detection device 10 according to the sixth aspect of this disclosure is an anomaly detection device 10 according to the fifth aspect, wherein the preprocessing unit 32 performs filtering processing to remove waveforms in which the amplitude of the peak intensity of the target time series data is less than a threshold. According to this disclosure, it is possible to detect the necessary data that has passed the filtering processing.

[0060] An anomaly detection device 10 according to the seventh aspect of this disclosure is an anomaly detection device 10 according to any of the first to sixth aspects, wherein the anomaly time-series data includes at least one of the shaft vibration data of the rotating machine and the bearing temperature data. According to this disclosure, anomalies can be determined with greater accuracy based on continuous data in a time series.

[0061] An abnormality detection device 10 according to the eighth aspect of this disclosure is an abnormality detection device 10 according to any of the first to seventh aspects, wherein the rotating machine includes at least one of a compressor and a turbine. According to this disclosure, abnormalities in the rotating machine can be determined with greater accuracy.

[0062] An anomaly detection method according to the ninth aspect of this disclosure includes the steps of: acquiring anomaly time series data, which is time series data of an abnormal state of a rotating machine; correcting the anomaly time series data to generate a number of simulated data, which are time series data that simulate the anomaly; training a learning model using the simulated data and labels indicating the state of the simulated data as training data to generate a trained model; inputting target time series data, which is time series data of the rotating machine, into the trained model to perform anomaly detection on the target time series data; and outputting the result of the anomaly detection on the target time series data. According to this disclosure, anomalies can be detected with greater accuracy.

[0063] A program according to the tenth aspect of this disclosure causes a computer to perform the following steps: acquire abnormal time series data, which is time series data of an abnormal state of a rotating machine; correct the abnormal time series data to generate a number of simulated data, which is time series data that simulates the abnormality; train a learning model using the simulated data and labels indicating the state of the simulated data as training data to generate a trained model; input target time series data, which is time series data of the rotating machine, into the trained model to perform abnormality determination of the target time series data; and output the result of the abnormality determination of the target time series data. According to this disclosure, abnormalities can be determined with greater accuracy.

[0064] Although embodiments of the present disclosure have been described above, the embodiments are not limited to those described herein. Furthermore, the aforementioned components include those that can be easily conceived by those skilled in the art, those that are substantially the same, and those that fall within the so-called equivalent range. Moreover, the aforementioned components can be combined as appropriate. Furthermore, various omissions, substitutions, or modifications of the components can be made without departing from the gist of the embodiments described above.

[0065] 10 Anomaly detection device 30 Data acquisition unit 32 Preprocessing unit 34 Generation unit 36 ​​Learning unit 38 Determination unit 40 Output control unit

Claims

1. An anomaly detection device comprising: a data acquisition unit that acquires anomaly time series data, which is time series data of an abnormal state of a rotating machine; a generation unit that corrects the anomaly time series data to generate multiple simulated data, which is time series data that simulates an anomaly; a learning unit that uses the simulated data and labels indicating the state of the simulated data as training data to train a learning model and generate a trained model; a determination unit that inputs target time series data, which is time series data targeted by the rotating machine, into the trained model and performs an anomaly determination of the target time series data; and an output control unit that outputs the result of the anomaly determination of the target time series data.

2. The anomaly detection device according to claim 1, wherein the generation unit generates the simulated data by changing at least one of the following: the period of the anomaly time series data, the timing at which the anomaly time series data reaches a peak, the number of peaks in the anomaly time series data, and the intensity of the peaks in the anomaly time series data.

3. The anomaly detection device according to claim 1 or 2, wherein the learning unit generates the trained model using the fact that the waveform of the simulated data is spike-shaped as the label, and the determination unit detects that the waveform of the target time series data is spike-shaped as the result of the anomaly determination.

4. The anomaly determination device according to claim 1 or 2, wherein the learning unit generates the trained model using as a label whether the simulated data is continuously increasing or continuously decreasing over time, and the determination unit detects as a result of the anomaly determination whether the target time series data is continuously increasing or continuously decreasing over time.

5. An anomaly detection device according to claim 1 or 2, further comprising a preprocessing unit for normalizing the abnormal time series data, wherein the learning unit generates a trained model by training the learning model with the simulated data generated based on the normalized abnormal time series data.

6. The anomaly determination device according to claim 5, wherein the preprocessing unit performs filtering to remove waveforms in which the amplitude of fluctuations in the peak intensity of the target time series data is less than a threshold.

7. The abnormality determination device according to claim 1 or 2, wherein the abnormality time series data includes at least one of the shaft vibration data of the rotating machine and the bearing temperature data.

8. The abnormality determination device according to claim 1 or 2, wherein the rotating machine includes at least one of a compressor and a turbine.

9. An anomaly detection method comprising: acquiring anomaly time series data, which is time series data of an abnormal state of a rotating machine; correcting the anomaly time series data to generate multiple simulated data, which is time series data that simulates an anomaly; training a learning model using the simulated data and labels indicating the state of the simulated data as training data to generate a trained model; inputting target time series data, which is time series data targeted by the rotating machine, into the trained model to perform anomaly detection on the target time series data; and outputting the result of the anomaly detection on the target time series data.

10. A program that causes a computer to perform the following steps: acquire abnormal time series data, which is time series data of an abnormal state of a rotating machine; correct the abnormal time series data to generate multiple simulated data, which is time series data that simulates the abnormality; train a learning model using the simulated data and labels indicating the state of the simulated data as training data to generate a trained model; input target time series data, which is time series data of the rotating machine, into the trained model and perform abnormality determination of the target time series data; and output the result of the abnormality determination of the target time series data.

Images