Failure prediction system

Abstract

To provide a failure prediction system capable of predicting failures with high accuracy. [Solution] The failure prediction system includes a plurality of devices, a plurality of sensors, an anomaly detection device that detects anomalies in the sensor output signals, a hidden layer output storage device that stores the hidden layer output when the anomaly detection device determines an anomaly, a clustering device that groups the stored hidden layer outputs into a plurality of clusters, a clustering result storage device that stores the grouped clusters, and a counting and determination device that notifies when an anomaly is predicted. The counting and determination device performs a first determination process to determine whether the sum of the number of elements in one cluster is equal to or greater than a threshold, a second determination process to determine whether there is a bias in the number of elements in each device if the sum is equal to or greater than the threshold, and a third determination process to determine that there is an anomaly in the device with the bias in the number of elements if the bias is determined to exist.

Description

[Technical Field]

[0001] The present disclosure relates to a failure prediction system. [Background technology]

[0002] In plants such as waste treatment facilities and chemical plants, a large number of devices work together to perform their intended tasks. However, if one of the devices breaks down, the plant must be shut down unplanned. To address this issue, technologies have been developed that install corresponding sensors (e.g., vibration sensors, pressure sensors, voltage sensors, current sensors, etc.) on each piece of equipment in the plant and analyze information obtained from each sensor using machine learning or other methods to predict failures and avoid unplanned plant shutdowns. For example, Patent Document 1 discloses a system for predicting failures in rotating machinery, which clusters features in the intermediate layer (hidden layer) of a deep learning model and estimates the presence and severity of an abnormality based on the distance from the center of the cluster. When clustering is performed on multiple pieces of equipment of the same type, for example, multiple rotating machines installed in a specific plant, using machine learning, for example, the k-means method, a type of unsupervised learning, on features in the hidden layer, multiple features of each rotating machine may be included in the same cluster.The probability of an abnormality increases as the distance from the center of the cluster to a certain feature increases, making it possible to predict which rotating machines among the multiple rotating machines have an abnormality.

[0003] However, just because an abnormality is far from the center point of the cluster does not necessarily mean that it is an abnormality that is highly likely to cause the rotating machinery to fail, that is, an abnormality that is a precursor to a failure (hereinafter referred to as a "major abnormality").It is possible that an abnormality that is "different in some way from a state without failure" but is not a precursor to a failure of the rotating machinery (hereinafter referred to as a "unimportant abnormality") may occur at a location that is far away due to "variation" in probability. In other words, from the perspective of equipment failure prediction, a "non-critical abnormality" is an abnormality that can occur normally even if there is no failure in the equipment. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Publication No. 2024-72124 Summary of the Invention [Problem to be solved by the invention]

[0005] Conventional technology only predicted a malfunction and notified workers even in the case of "non-critical abnormalities," so the accuracy of predicting equipment malfunctions was insufficient, and further improvements in accuracy were desired. Therefore, the present disclosure aims to provide a failure prediction system that identifies devices that exhibit "serious abnormalities" from a mixture of information regarding "unimportant abnormalities" that can occur even when the devices are not malfunctioning and information regarding "serious abnormalities" that are precursors to failure, and accurately predicts failures in relation to multiple devices of the same type. [Means for solving the problem]

[0006] The failure prediction system of the present invention comprises: Multiple devices of the same type; a plurality of sensors of the same type installed corresponding to the plurality of devices; an anomaly detection device including at least an autoencoder, which detects an anomaly in the output signal by inputting feature amounts of the output signal of the sensor installed in the device in a fault-free state and learning the feature amounts in advance; a hidden layer output storage device that stores a hidden layer output when the anomaly detection device determines that an anomaly has occurred; a clustering device that groups the hidden layer outputs stored in the hidden layer output storage device into a plurality of clusters; a clustering result storage device that stores the grouped clusters; The system further includes a counting and determining device that, when it predicts an abnormality that indicates a precursor to a failure in any of the plurality of devices, notifies the user of the abnormality. The aggregation and determination device executes a first determination process to determine whether the sum of the number of elements in one of the multiple clusters stored in the clustering result storage device is equal to or greater than a preset threshold; a second determination process to determine whether there is a bias in the number of elements corresponding to each of the multiple devices when it is determined that the sum is equal to or greater than the threshold; and a third determination process to determine, when it is determined that there is a bias, that there is an abnormality indicating a precursor to a failure in the device among the multiple devices that has a bias toward a larger number of elements and to consider this as having been predicted. [Effects of the Invention]

[0007] According to the failure prediction system disclosed herein, for multiple devices of the same type, from among a mixture of information on "unimportant abnormalities" that can occur even when the device is not faulty and information on "major abnormalities" that are precursors to failure, it is possible to easily identify devices that exhibit "major abnormalities" through the first, second, and third judgment processes, and to accurately predict failures. [Brief explanation of the drawings]

[0008] [Figure 1] 1 is a diagram illustrating a failure prediction system according to an embodiment. [Figure 2] FIG. 1 illustrates an example of an abnormality detection device. [Figure 3] FIG. 10 is a diagram illustrating an example of clusters grouped by the clustering device. [Figure 4] 10A and 10B are diagrams illustrating an example of a counting result and a determination result obtained by the counting and determination device. DETAILED DESCRIPTION OF THE INVENTION

[0009] A failure prediction system according to an embodiment of the present invention will be described. The embodiment is merely an example, and is not intended to exclude various modifications or applications of techniques not explicitly stated. Except for the essential components of the present invention, each component of the embodiment can be selected or modified as necessary.

[0010] 1 is a diagram showing the system configuration of a failure prediction system 100. For ease of explanation, the explanation will be given here using three identical devices: a first device 110, a second device 120, and a third device 130, but the failure prediction system 100 can be applied to a plant equipped with two or more identical devices. When the failure prediction system 100 is installed in a waste incinerator plant, the first device 110, the second device 120, and the third device 130 may be, for example, rotary machines, specifically, induced draft fans. The first device 110, the second device 120, and the third device 130 need only be of the same type, and do not have to have the same specifications. For example, as long as they are the same type of induced draft fans, they may be of different sizes or models. The rotary machines may also be motors, etc. Furthermore, the first device 110, the second device 120, and the third device 130 are not limited to rotary machines, and may be any other type of similar equipment.

[0011] The sensor output memory device 140, feature memory device 155, hidden layer output memory device 165, clustering result memory device 175, and aggregation result and judgment result memory device 185 provided in the failure prediction system 100 may each be a separate memory device, for example, a separate server or storage. These storage devices may be replaced with different storage areas, in which case all of these storage areas may be located on the same server, or all of these storage areas may be located in a cloud-based or clustered database that connects multiple servers.

[0012] Now, each component of the failure prediction system 100 in FIG. 1 will be described in detail in turn. A first sensor 111 is installed in the first device 110, a second sensor 121 in the second device 120, and a third sensor 131 in the third device 130. Here, an example in which one sensor (e.g., a vibration sensor) is installed in each device will be described, but two or more sensors (e.g., a vibration sensor, a temperature sensor, a voltage sensor, a current sensor, etc.) may be installed. Each sensor transmits time-series information (e.g., time-series vibration information) related to the performance and status of each device to the sensor output storage device 140 as a sensor output, i.e., an output signal.

[0013] The sensor output storage device 140 receives the sensor outputs from the first sensor 111, the second sensor 121, and the third sensor 131, and stores the received sensor outputs in association with the respective devices. Each sensor output is so-called "raw data" corresponding to the respective device.

[0014] The feature extraction device 150 periodically accesses the sensor output storage device 140, reads the sensor output of each device, and extracts features associated with each of the first device 110, the second device 120, and the third device 130 using a technique such as Fast Fourier Transform (FFT), and transmits them to the feature storage device 155.

[0015] The feature storage device 155 receives features corresponding to the first device 110, the second device 120, and the third device 130 from the feature extraction device 150, and stores each of the received features in association with the corresponding device.

[0016] The abnormality detection device 160 periodically accesses the feature storage device 155, reads out the feature for each device, and performs abnormality detection using the read out feature as an input. 2, the anomaly detection device 160 includes an encoder 161a that reduces the dimension of an input feature, a decoder 161b that restores the reduced-dimensional feature, a comparator 162 that compares the input feature with the feature restored by the decoder 161b and issues a mismatch signal if the two do not match, and a buffer 163 that temporarily stores the reduced-dimensional feature included in a hidden layer (intermediate layer) and, upon receiving a mismatch signal, transmits the stored feature as a hidden layer output to a hidden layer output storage device 165. The autoencoder 161 includes at least the encoder 161a and the decoder 161b. Before the failure prediction system 100 officially starts operation, the first device 110, the second device 120, and the third device 130 are operated in advance in a state without any failures, and during this operation, the autoencoder 161 is made to learn the feature amounts of the sensor outputs output by the first sensor 111, the second sensor 121, and the third sensor 131. Through this learning, the autoencoder 161 can match the feature amount input to the encoder 161a with the feature amount decoded by the decoder 161b. Therefore, after the failure prediction system 100 officially begins operation, if the feature input to the encoder 161a of the autoencoder 161 does not match the feature decoded by the decoder 161b of the autoencoder 161, it can be determined to be an "abnormality" (a concept that includes both the aforementioned "serious abnormality" and "unimportant abnormality").

[0017] The hidden layer output storage device 165 receives the hidden layer output sent by the buffer 163 of the anomaly detection device 160, and stores the received hidden layer output in association with the corresponding device.

[0018] The clustering device 170 periodically accesses the hidden layer output storage device 165 and reads out the hidden layer outputs that have been determined to be "abnormal" (a concept that includes both the aforementioned "significant abnormality" and "unimportant abnormality") for all of the first device 110, the second device 120, and the third device 130. Then, the clustering device 170 clusters the read hidden layer outputs using a known clustering technique (for example, the k-Means method) and transmits the clustering results to the clustering result storage device 175. The clustering results include clusters and hidden layer outputs associated with each device classified into the clusters.

[0019] The clustering result storage device 175 receives the clustering results sent by the clustering device 170 and stores the received clustering results. An example of the clustering result is shown in Figure 3. The hidden layer outputs associated with the first device 110, the second device 120, and the third device 130 are classified, that is, grouped, into a number of clusters, for example, clusters A to H. In Figure 3, the outputs are grouped into eight clusters A to H as an example, but for ease of explanation, the following description will be continued assuming that the outputs are grouped into only four clusters A to D.

[0020] The tallying and determining device 180 refers to the clustering results and performs the first determination process, the second determination process, and the third determination process described in detail below to determine whether or not there is a "significant abnormality." First, the counting and determination device 180 periodically accesses the clustering result storage device 175, reads out the hidden layer output corresponding to each cluster, and performs a process (hereinafter referred to as the "counting process") of counting the number of hidden layer outputs (hereinafter referred to as the "number of elements") of the first device 110, the second device 120, and the third device 130 classified into each cluster. An example of the results of the aggregation process is shown in Figure 4. The first device 110 included in cluster A has 5 elements, the second device 120 included in cluster A has 5 elements, and the third device 130 included in cluster A has 5 elements. The first device 110 included in cluster B has 15 elements, the second device 120 included in cluster B has 0 elements, and the third device 130 included in cluster B has 1 element. The first device 110 included in cluster C has 9 elements, the second device 120 included in cluster C has 8 elements, and the third device 130 included in cluster C has 0 elements. The first device 110 included in cluster D has 3 elements, the second device 120 included in cluster D has 4 elements, and the third device 130 included in cluster D has 2 elements.

[0021] (First determination process) The tallying and determining device 180 then determines whether the total number of elements included in one cluster is equal to or greater than a preset number, i.e., a threshold value, determined in advance according to the design. Here, as an example, the explanation will continue assuming that the threshold value is 15. If the number of elements included in one cluster is small, proper failure prediction will not be possible, so it is desirable that the threshold value be at least 5 or more. In Figure 4, the total number of elements in cluster A is (5 + 5 + 5) = 15, the total number of elements in cluster B is (15 + 0 + 1) = 16, the total number of elements in cluster C is (9 + 8 + 0) = 17, and the total number of elements in cluster D is (3 + 4 + 2) = 9. Therefore, in the first determination process, the aggregation and determination device 180 determines that cluster D is below the threshold (indicated by an X in the determination result of the first determination process in Figure 4), and determines that clusters A, B, and C are all above the threshold (indicated by a O in the determination result of the first determination process in Figure 4).

[0022] (Second Determination Process) Next, the aggregation and determination device 180 performs a second determination process on the number of elements of the first device 110, the second device 120, and the third device 130 of clusters A, B, and C, whose number of elements is determined to be greater than or equal to the threshold in the first determination process. In the second determination process, it is determined whether there is a bias in the number of elements included in a certain cluster among the first device 110, the second device 120, and the third device 130. This determination method may use a known technique, such as a "test of difference in population ratios." When performing the determination using a "test of difference in population ratios," the number of elements associated with each device included in one cluster is tested brute-force for each pair of devices, and devices with a bias in the number of elements compared to other devices are identified. For example, if the difference in the number of elements between two devices is less than three, it can be determined that there is no bias. Here, as an example, the description will be continued assuming that there is no bias if the difference in the number of elements corresponding to two devices is less than three, and that there is a bias if the difference is three or more. However, the setting of three can be changed as appropriate depending on the design.

[0023] In Figure 4, in cluster A, the number of elements linked to the first device 110, the second device 120, and the third device 130 is all five, so the aggregation and determination device 180 determines that there is no bias in the number of elements of the first device 110, the second device 120, and the third device 130 (an X is shown in the determination result of the second determination process in Figure 4).

[0024] In cluster B, when the number of elements of the second device 120 and the third device 130 is compared, the number of elements of the second device 120 is 0 and the number of elements of the third device 130 is 1, the difference being 1, which is less than 3, so there is no bias in the number of elements of the second device 120 and the third device 130. However, when comparing the number of elements of the first device 110 and the second device 120, the number of elements of the first device is 15 and the number of elements of the second device 120 is 0, a difference of 15, which is more than 3, and therefore there is a bias in the number of elements between the first device 110 and the second device 120. Furthermore, when comparing the number of elements of the first device 110 and the third device 130, the number of elements of the first device is 15 and the number of elements of the third device 130 is 1, a difference of 14, which is more than 3, and therefore there is a bias in the number of elements between the first device 110 and the third device 130. Therefore, the tallying and determination device 180 determines that there is a bias in the number of elements of the first device 110 relative to the second device 120 and the third device 130 (determination result of the second determination process in FIG. 4 is indicated by ◯).

[0025] In cluster C, when comparing the number of elements of the first device 110 and the second device 120, the number of elements of the first device 110 is 9 and the number of elements of the second device 120 is 8, the difference being 1, which is less than 3, so there is no bias in the number of elements of the first device 110 and the second device 120. However, when comparing the number of elements of the first device 110 and the third device 130, the number of elements of the first device is 9 and the number of elements of the third device 130 is 0, the difference being 9, which is 3 or more, so there is a bias in the number of elements between the first device 110 and the third device 130. Furthermore, when comparing the number of elements of the second device 120 and the third device 130, the number of elements of the second device is 8 and the number of elements of the third device 130 is 0, the difference being 8, which is 3 or more, so there is a bias in the number of elements between the second device 120 and the third device 130. Therefore, the tallying and determination device 180 determines that there is a bias in the number of elements of the third device 130 relative to the first device 110 and the second device 120 (determination result of the second determination process in FIG. 4 is indicated by ◯).

[0026] (Third Determination Process) Then, the counting and determination device 180 performs a third determination process on the first device 110, the second device 120, and the third device 130 of cluster B and cluster C that have been determined to have a bias in the number of elements. The third determination process is a process that determines whether or not any of the devices has a "serious abnormality" that is a precursor to a failure, and if it is determined that there is a "serious abnormality," it instructs notification of the abnormality, assuming that an abnormality that is a precursor to a failure has been predicted. As mentioned above, the number of elements, i.e., the number of hidden layer outputs, is the number of features determined to be abnormal (a concept that includes both the above-mentioned "important abnormality" and "unimportant abnormality"), so from the perspective of failure prediction, the greater the number of elements, the higher the probability of failure. Therefore, when the number of elements for one specific device is biased toward being larger than that for other devices within a specific cluster determined to have a bias, the aggregation and determination device 180 determines that the device has a "important abnormality" that is a precursor to failure.

[0027] In Figure 4, for cluster B, the first device 110 has a larger number of elements than the second device 120 and the third device 130, so the aggregation and judgment device 180 judges that the first device 110 has a ``serious abnormality.'' On the other hand, for cluster C, the number of elements in the third device 130 is biased toward being smaller than those in the first device 110 and the second device 120, so the aggregation and judgment device 180 judges that although the third device 130 shows a different trend compared to the first device 110 and the second device 120, there is no ``significant abnormality.''

[0028] The counting and determination device 180 transmits the counting results of the number of elements of the first device 110, the second device 120, and the third device 130, and the determination results for each cluster in the first determination process, the second determination process, and the third determination process to the counting result and determination result storage device 185. If the tallying and determination device 180 determines in the first determination process that the total number of elements in all clusters is not equal to or greater than the threshold, it transmits the result of the determination to the tallying result and determination result storage device 185 and ends the first determination process. Then, without performing the second and third determination processes, it accesses the clustering result storage device 175 again, performs the tallying process, and then performs the first determination process. Furthermore, if it is determined in the second determination process that there is no bias in all of the target clusters, the result of the determination is sent to the counting result and determination result storage device 185, and the second determination process ends. Then, without performing the third determination process, the clustering result storage device 175 is accessed again, and after performing the counting process, the first determination process is performed.

[0029] The tally result and determination result storage device 185 receives the determination results of the first determination process, the second determination process, and the third determination process, and stores the determination results.

[0030] If the tallying and determination device 180 determines through the third determination process that a specific device, for example, the first device 110, has a “serious abnormality,” it transmits warning information regarding the first device 110 to the alarm device / monitor 190. The alarm device / monitor 190 includes at least one of an alarm device and a monitor. When an alarm device is included, upon receiving warning information from the counting and determination device 180, the alarm device notifies the worker of the relevant device by sound information such as a voice saying, for example, "A serious abnormality has occurred in the first device." When a monitor is included, the monitor notifies the worker of the relevant device by displaying visual information such as text saying, "A serious abnormality has occurred in the first device." In this way, workers are notified of abnormalities in specific equipment and are urged to carry out early checks or repairs.

[0031] As described above, the failure prediction system of the present invention can easily identify devices that exhibit a "major abnormality" from a mixture of information on "non-major abnormalities" that can occur even when the devices are not malfunctioning and information on "major abnormalities" that are precursors to failure, and can accurately predict failures for multiple devices of the same type. [Explanation of symbols]

[0032] 100 Failure prediction system 110...First equipment 111 First sensor 120...Second device 121 Second sensor 130...Third equipment 131 Third sensor 140 Sensor output storage device 150 Feature Extraction Device 155...Feature Memory Device 160 Abnormality detection device 161 Autoencoder 161a···Encoder 161b decoder 162 Comparator 163... Buffer 165···Hidden layer output storage device 170 Clustering device 175···Clustering result storage device 180. Counting and Judging Device 185....Tallying result and judgment result storage device 190 Warning Device or Monitor

Claims

1. Multiple devices of the same type; a plurality of sensors of the same type installed corresponding to the plurality of devices; an anomaly detection device that includes at least an autoencoder and that detects an anomaly in the output signal by inputting and learning in advance features of the output signal of the sensor installed in the device in a fault-free state; a hidden layer output storage device that stores a hidden layer output when the anomaly detection device determines that an anomaly has occurred; a clustering device that groups the hidden layer outputs stored in the hidden layer output storage device into a plurality of clusters; a clustering result storage device that stores the grouped clusters; a counting and determination device that, when detecting an abnormality that indicates a sign of a failure in any of the plurality of devices, notifies the abnormality; and the aggregation and determination device performs a first determination process of determining whether or not the sum of the number of elements in one of the plurality of clusters stored in the clustering result storage device is equal to or greater than a preset threshold; a second determination process for determining whether or not there is a bias in the number of elements corresponding to each of the plurality of devices when it is determined that the sum is equal to or greater than the threshold value; a third determination process for determining, when it is determined that the bias exists, that an abnormality indicating a precursor to the failure exists for a device among the plurality of devices that has the bias on the side with a larger number of elements, and determining that the failure has been predicted; A failure prediction system that performs the following.

2. Further provided with a storage device for counting results and determination results, The tallying and determining device If it is determined in the first determination process that the sum of the number of elements is not equal to or greater than the threshold value, the result of the determination is stored in the counting result and determination result storage device, and the first determination process is terminated. The failure prediction system of claim 1, wherein if it is determined in the second judgment process that there is no bias, the judgment result is stored in the aggregation result and judgment result storage device, and the second judgment process is terminated.

3. the plurality of sensors are vibration sensors; further comprising an alarm or monitor; The failure prediction system according to claim 2, wherein the aggregation and judgment device notifies the equipment among the plurality of devices for which the abnormality is predicted by sounding the alarm device or by displaying on the monitor.

4. a sensor output storage device that stores the time-series output signals of the plurality of sensors in association with the corresponding plurality of devices; a feature extraction device that extracts a feature of the output signal stored in the sensor output storage device; a feature storage device that stores the extracted feature amounts in association with the corresponding plurality of devices; Furthermore, The failure prediction system according to claim 3 , wherein the abnormality detection device detects an abnormality in the output signal using the feature stored in the feature storage device.

5. The abnormality detection device the autoencoder including an encoder that reduces the dimension of the input feature quantity, and a decoder that restores the reduced dimension feature quantity; a comparator that compares the input feature with the feature restored by the decoder and outputs a mismatch signal if the two do not match; a buffer that temporarily stores the dimensionally reduced hidden layer output and transmits the stored hidden layer output to the hidden layer output storage device when the encoder receives the mismatch signal; The failure prediction system according to any one of claims 1 to 4, comprising:

Images