Update program and system for model in accordance with output of abnormality detecting model and abnormality identifying model
The system uses anomaly detection and identification models to automate monitoring of long-term systems, addressing unpredictable abnormalities and improving detection accuracy by updating models based on discrepancies, thus enhancing abnormality resolution.
Patent Information
- Application Number
- JP2023079635
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-05-12
- Publication Date
- 2024-11-22
AI Technical Summary
Existing technologies struggle to automate monitoring of long-term operational systems effectively, particularly in environments where abnormalities occur unpredictably, and fail to address unobserved abnormalities without placing a heavy burden on personnel.
A system utilizing both anomaly detection and anomaly identification models, which learn from time-series vibration data, separately judges anomalies and updates based on discrepancies, allowing for continuous model refinement and improved abnormality identification.
Enables automated monitoring and identification of unobserved abnormalities in long-term systems, enhancing the resolution of abnormality detection and adaptation to varying operational conditions.
Smart Images

Figure 2024163764000001_ABST
Abstract
Description
[Technical field]
[0001] The present invention relates to a program and a system for updating either an anomaly detection model or an anomaly identification model in accordance with the output of the model. [Background technology]
[0002] In systems that are expected to operate over a long period of time, such as railway vehicles, machine tools, and servers (collectively referred to in this specification as "long-term operation systems"), it is important to perform maintenance inspections to prevent breakdowns by monitoring the deterioration of components over time. Many attempts have been made to automate such monitoring.
[0003] When attempting to automate monitoring using machine learning models, various cases (abnormal events) can occur in the actual field, so it is not realistic to anticipate every possible case and prepare appropriate training data in advance. For this reason, methods have been proposed that attempt to respond to unknown cases by performing unsupervised learning.
[0004] For example, Patent Document 1 discloses a technology for learning a latent variable model and a joint probability model based on training data consisting of normal data. This conventional technology discloses a system that measures the likelihood of acquired sensor data based on the latent variable model and the joint probability model, judges whether the sensor data is normal or abnormal, and learns the latent variable model and the joint probability model based on the sensor data output by the sensor. It is claimed that this enables efficient sensor selection without requiring advanced specialized knowledge or extensive experience, even in the case of new machinery or unknown failures.
[0005] The technology of Patent Document 1 is characterized by comparing restored data obtained from a model trained based on training data consisting of normal data with input data for anomaly detection, and detecting anomalies from the deviation. However, it is generally known that in long-term operating systems, even if the system is the same, the content and frequency of possible abnormal events vary greatly depending on whether the system is almost new or has been used for many years. As a result, the type of model that should be prepared in the actual field of a long-term operating system varies depending on the situation, and there is a problem that the technology of Patent Document 1 cannot fully respond to such a situation.
[0006] Patent Document 2 discloses an information processing system that includes a reconstruction means for reconstructing a normal image from an input image, an output means for outputting a reason for determining an abnormal class between the input image and the image reconstructed by the reconstruction means, and a detection means for detecting an abnormal part in the input image based on a difference between the input image and the reconstructed image and the reason for determining the abnormal class output by the output means. This is said to enable detection of an abnormal part with higher accuracy.
[0007] However, there are many sites where it is not practical to monitor long-term operating systems using images. For example, environments such as railroad bogies are not suitable for image inspection because of the intense and irregular vibrations and insufficient incident light. Furthermore, Patent Document 2 does not specifically mention how to deal with the occurrence of an abnormality that is not observed at the time of occurrence or the occurrence of an error in the detection of an abnormality. [Prior art documents] [Patent documents]
[0008] [Patent Document 1] JP 2020-119605 A [Patent Document 2] Patent Publication No. 2022-090761 Summary of the Invention [Problem to be solved by the invention]
[0009] In view of the problems with the conventional technology described above, there is a need for new technology that can automate the monitoring of long-term operating systems while also enabling responses to abnormalities that are unobserved at the time they occur, without placing an excessive personal burden on field personnel. [Means for solving the problem]
[0010] That is, the present invention can provide the following aspects.
[0011] Aspect 1. 1. A system for monitoring an object, comprising: a sensor configured to obtain a vibration signal generated by the object; a storage means configured to acquire and store input from said sensors; A plurality of types of machine learning models configured to be stored in the storage means; Judgment method and Including, The plurality of types of machine learning models, an anomaly detection model configured to learn time series data relating to normal vibrations associated with the object; an anomaly identification model configured to learn time series data relating to abnormal vibrations associated with the object; Including, The determination means: Based on an input from the sensor that receives a vibration signal generated from the target, a first judgment using the anomaly detection model and a second judgment using the anomaly identification model are separately performed, and differences between the results of the first judgment and the results of the second judgment are identified; and and updating the anomaly detection model and / or the anomaly identification model based on a difference between a result of the first judgment and a result of the second judgment. A system characterized in that
[0012] Aspect 2. a means for acquiring data relating to the state of generated vibration that led to the judgment result when the judgment means judges that the first judgment result and the second judgment result are different; 2. The system of embodiment 1, further comprising:
[0013] Aspect 3. a display means configured to present data relating to the state of the generated vibration; 3. The system of embodiment 2, further comprising:
[0014] Aspect 4. The determination means: When the first judgment results in a judgment that the object is abnormal and the second judgment results in a judgment that the object is normal, the anomaly identification model is updated. 4. The system according to any one of aspects 1 to 3.
[0015] Aspect 5. The determination means: When the first judgment results in judging that the object is abnormal and the second judgment results in judging that the object is normal, a new anomaly identification model is created based on the vibration signal generated from the object. 5. The system according to any one of aspects 1 to 4.
[0016] Aspect 6. A program including instructions executable by a computer having a processor, the program being connectable via a network to a sensor configured to acquire vibrations generated from an object, the program including: The instructions are executed by the processor, acquiring an input related to vibration generated from the target through an input / output interface; A storage means stores a plurality of types of machine learning models, wherein the plurality of types of machine learning models include at least an anomaly detection model configured to learn time series data relating to normal vibrations associated with the object; an anomaly identification model configured to learn time series data relating to abnormal vibrations associated with the object; and performing a first determination using the anomaly detection model based on an input related to vibration generated from the target by a determination means; performing a second determination using the anomaly identification model based on an input related to vibration generated from the target by a determination means; Identifying by a judgment means whether the result of the first judgment is the same as the result of the second judgment; updating the anomaly detection model and / or the anomaly identification model based on the difference between the result of the first judgment and the result of the second judgment by a judgment means; configured to A program characterized by: Effect of the Invention
[0017] The present invention makes it possible to automate monitoring based on vibrations generated in long-term operating systems, and also makes it possible to identify the occurrence of abnormalities that were unobserved at the time of their occurrence, thereby improving the resolution for identifying abnormalities. [Brief description of the drawings]
[0018] [Figure 1] 1 illustrates an example of the configuration of a system according to an embodiment of the present invention. [Diagram 2] 2 shows an example of a display means according to an embodiment of the present invention. [Diagram 3] 2 shows an example of a display means according to an embodiment of the present invention. [Figure 4] 1 shows an example flowchart of a program process according to an embodiment of the present invention. [Diagram 5] 13 shows an example of a flowchart of a process performed by a program according to another embodiment of the present invention. [Figure 6] 1 shows an example of how the output of a machine learning model changes over time in accordance with an embodiment of the present invention. [Figure 7]1 shows an example of how the output of a machine learning model changes over time in accordance with an embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0019] The present invention can provide a system or program for monitoring a long-term operating system. Examples of such long-term operating systems include systems including vehicles such as railways, machine tools, servers, or fixed-point observation devices for houses, rivers, roads, and strategic locations, but the present invention is not limited to these and can be applied to various systems. The system according to the present invention may be a system including a wireless and / or wired network, and may be one or more pieces of hardware or a combination of hardware and software.
[0020] In this specification, an anomaly (abnormal vibration) means a vibration that reflects the occurrence of an anomaly (or a failure) in a long-term operating system to be monitored, or reflects a sign of an anomaly occurring in the future. In this specification, an "anomaly detection model" and an "anomaly identification model" are distinguished as types of machine learning models. In this invention, the anomaly detection model and the anomaly identification model are combined, the same input is made to each, and the respective outputs are collated, and based on this, the anomaly detection model or the anomaly identification model, or both, can be updated. In this specification, "updating" a machine learning model refers to making some kind of correction to the model, and may include, for example, giving feedback data to the model to make it learn (re-learn), increasing or decreasing the number of models, and replacing a model. The number of anomaly detection models or anomaly identification models may be singular or any multiple.
[0021] The anomaly detection model is a machine learning model for detecting the occurrence of an anomaly and / or a sign of an anomaly, and may be considered as a means for analyzing an anomaly (failure) or a sign thereof in a monitored object. The anomaly detection model is configured to input and learn time series data based on vibrations generated during normal operation of a long-term operation system to be monitored or a system similar thereto. Here, normal operation refers to a state in which no malfunction or failure is found when the system is operated for a relatively long period of time. In one embodiment, the system to be monitored is operated, and input data can be created from data acquired by a sensor such as an acoustic sensor (including a sonic wave sensor and an AE sensor) during the period of normal operation. In another embodiment, a system similar to the system to be monitored (for example, a new machine of the same model) is operated, and input data can be created from data acquired by a sensor during the period of normal operation. In still another embodiment, the anomaly detection model may be in an unlearned state at the beginning of operation of the long-term operation system, and the anomaly detection model may proceed with learning as needed during operation, and may be completed following the configuration of the present invention.
[0022] Features that can be extracted when generating input data for an anomaly detection model include the frequency, amplitude, and wavelength of sound waves obtained by a sonic sensor (i.e., generally known as acoustic data), as well as the degree of vibration obtained by an acoustic emission (AE) sensor or an acceleration sensor, or the slope and differential value of the curve when these are plotted in a time series. In some embodiments, data obtained from other types of sensors (such as thermal sensors and optical sensors) may be used as such features in combination with data related to vibration.
[0023] In the system according to the present invention, a sensor (for example, a microphone or the like may be used) is used to obtain time-series data related to vibration from an object for the following reason. Image and video data is generally larger in size than vibration-related data such as audio data, and processing it inevitably requires a high-performance computer and a large-capacity communication band. In particular, in long-term operation systems related to moving objects such as railway vehicles or long-term operation systems installed in limited spaces in factories, it is often not easy to secure a sufficient power source and communication environment, and it is therefore desirable to use a vibration sensor from the viewpoint of handling.
[0024] The method for training the anomaly detection model is preferably an algorithm that can make it easier to detect abnormal output by training it to match input and output, and it is preferable to use the AutoEncoder (AE) method using a neural network or one of its derivative methods (such as the stacked AE method, the variational AE (VAE) method, the convolutional AE (CAE) method, the sparse AE method, the saturating AE method, etc.).
[0025] The anomaly identification model is a machine learning model for estimating the cause of abnormal vibrations from a monitored object, and may be considered as a means for analyzing a state in which the monitored object does not reach the level of an abnormality or failure but is unnatural. The estimation referred to here may also include outputting the probability or likelihood of each of a plurality of possible causes. The anomaly identification model is configured to input and learn data based on sounds generated when an abnormality occurs in a long-term operation system to be monitored or a system similar thereto. In one embodiment, data can be continuously acquired by sensors such as an acoustic sensor or an acceleration sensor while the monitored system is in operation, and input data (teacher data) can be created from the records when an abnormality occurs. In another embodiment, a system similar to the monitored system (for example, a machine of the same model that is known to have aged deterioration in known parts) may be operated, and input data may be created from data acquired by the sensor. In other words, data related to sounds generated when a known abnormality occurs can be set as teacher data. The feature values that can be extracted when generating the input data can be selected in the same way as those for the anomaly detection model described above.
[0026] As a method for training the anomaly identification model, a supervised learning algorithm that uses a stochastic process model for maximum likelihood estimation is preferable, an algorithm that uses a Markov decision process model with a neural network is more preferable, and in particular an algorithm that uses a hidden Markov model (HMM).
[0027] FIG. 1 is a conceptual diagram for explaining the basic principle of the present invention. One or more sensors 110 acquire time series data on vibration from a monitored system 99, and store the data in a storage means 120 via an interface 180. The storage means 120 includes an anomaly detection model 130 and an anomaly identification model 140, and the data is input to each model. The storage means 120 may further include other machine learning models. The outputs of the anomaly detection model 130 and the anomaly identification model 140 are sent to a determination means 150. The determination means 150 makes a determination as to whether or not abnormal vibrations (e.g., abnormal sounds) are occurring based on the respective outputs.
[0028] The method of judgment based on the output of each model by the judgment means 150 is arbitrary, but may be, for example, a method using the following parameters.
[0029] True Positive Rate (TPR) or False Positive Rate (FPR): The same signal is evaluated N times, the number of times that the CMV (degree of certainty that it is that sound) exceeds the CMT (a specified threshold) is counted, and the TPR (the percentage of times that the CMV of the input sound exceeds the CMT) or FPR (the percentage of times that the CMV of the input sound does not exceed the CMT) is calculated. The presence or absence of an abnormality is judged depending on whether the TPR or FPR reaches a specified level.
[0030] Mean: The same signal is evaluated N times, the average CMV of each signal is calculated, and the presence or absence of an abnormality is determined based on whether or not the average CMV value reaches a specified level.
[0031] Standard deviation (SD): The same signal is evaluated N times, the standard deviation of each CMV is calculated, and the presence or absence of an abnormality is determined based on whether or not the value reaches a specified level.
[0032] Coefficient of variation (CV): The coefficient of variation is calculated by dividing the standard deviation by the average value, and the presence or absence of an abnormality is judged depending on whether or not this value reaches a predetermined level.
[0033] Then, the judgment means 150 checks whether the judgment based on the output of the anomaly detection model 130 and the judgment based on the output of the anomaly identification model 140 match. If they match (either abnormal vibration is occurring or abnormal vibration is not occurring), it can be determined that the system 99 is normal or abnormal, respectively. On the other hand, if the judgments do not match, the time series data on which the judgments were based can be checked, and one or both models can be updated based on the checked time series data (for example, feedback learning using the data). In the feedback, other corresponding sensor data (temperature, humidity, air pressure, etc.) may be used in combination with the time series data. In addition, situation data such as time, date, operation information, weather, operator name, etc. associated with the time series data obtained by the situation data acquisition means 160 may be used in feedback learning in combination with the time series data.
[0034] The display means 170 is a means capable of presenting the time series data on which the discrepancy in judgment occurred and other related data, if necessary, to the user. For example, the display means 170 may be hardware (display, speaker, etc.) or a combination of such hardware and software (a program for displaying a CUI, GUI, etc.).
[0035] In one embodiment, when a mismatch occurs, it may be determined which model made the wrong decision, and the feedback learning described above may be performed only on the model that made the wrong decision. Such a determination may be performed, for example, by having the sensor 110 or another monitoring means (e.g., the situation data acquisition means 160) continue to monitor the system 99 for a certain period of time when a mismatch occurs but the system 99 has not actually failed, and checking whether the system 99 actually fails during that period.
[0036] There are roughly two cases where it can be confirmed that the anomaly detection model 130 made a correct judgment and the anomaly identification model 140 made an incorrect judgment: A1. The system 99 actually has a tendency to cause an abnormality, and the anomaly detection model 130 correctly detects vibrations based on the abnormality, but the anomaly identification model 140 is unable to identify the abnormality because the abnormality is unknown to the anomaly identification model 140. A2. System 99 does not have a tendency to cause an abnormality, and anomaly detection model 130 correctly determines that system 99 is normal, but anomaly identification model 140 erroneously determines that system 99 will cause an abnormality.
[0037] In either case of A1 or A2, the anomaly identification model 140 can be updated based on the data in which the anomaly identification model 140 made an erroneous judgment (such an update can be performed by some update means not shown). For example, the data and, if necessary, data (situation data) on the situation in which the data was obtained (date and time, weather, temperature, etc.) may be fed back to the anomaly identification model 140 to allow learning to be performed. Alternatively, a new anomaly identification model 140 trained using such data may be generated and used to replace the original anomaly identification model 140 or to operate in parallel with the original anomaly identification model 140.
[0038] There are roughly two cases in which it can be confirmed that the anomaly detection model 130 makes an incorrect judgment and the anomaly identification model 140 makes a correct judgment: B1. The system 99 actually has a tendency to cause an abnormality, and the anomaly detection model 130 cannot detect the vibration based on the abnormality, but the anomaly identification model 140 can correctly identify the abnormal vibration. B2. System 99 does not tend to cause anomalies, and anomaly detection model 130 mistakenly determines that system 99 is abnormal, but anomaly identification model 140 correctly determines that system 99 is normal.
[0039] In either case of B1 or B2, the anomaly detection model 130 may be updated based on the data in which the anomaly detection model 130 made an erroneous judgment. For example, as in the above, feedback may be provided to the anomaly detection model 130 to allow it to learn, or the model may be replaced or added.
[0040] In another embodiment, both models may be updated, the details of such decisions being discussed further below.
[0041] FIG. 2 shows an example of software implementation (GUI) of the display means 170. When the date and time for which an anomaly is to be confirmed are input, the relevant audio files around that time are listed. When the Sound icon is pressed, the audio file can be output from the computer as sound, and when the Spectrum icon is pressed, the audio file can be output from the computer as a visual spectrum display. The Model checkbox is for generating a new machine learning model based on the checked audio file, and can be generated by pressing the Modeling button.
[0042] FIG. 3 is an example of a GUI for confirming a newly generated machine learning model as described above. A simulation result in which an arbitrary audio file is input as an evaluation file to the new machine learning model can be presented. Based on the result, the user can further include the new machine learning model in the system according to the present invention (displayed as "abnormal sound 3 model" in the case of FIG. 3). That is, the storage means 120 of the present system can further have machine learning models other than the anomaly detection model 130 and the anomaly identification model 140. It will be understood that the above description can be expanded to set the operation of the determination means 150 even when there are three or more machine learning models (or when there are a plurality of any of the above machine learning models).
[0043] In one embodiment, the user interface such as the GUI may further include a means for receiving input from a user. Such input may include, for example, data on the long-term operating system of interest (for example, in the case of a system related to railroad vehicles, the type of track on which the vehicle runs and weather information for the area in which the vehicle runs). Based on such input, the system may suggest one or more models that are considered appropriate from among multiple models that are candidates for the anomaly detection model 130 and the anomaly identification model 140.
[0044] In an embodiment of the present invention, a server connectable to a network can also be provided. The server includes a storage means (which may be, for example, the above-mentioned storage means 120) for storing a plurality of types of machine learning models, and a communication means for connecting to an external sensor (which may be, for example, the above-mentioned sensor 110) via the network. The server can obtain input data related to generated vibrations from an external long-term operation system via the sensor and the communication means.
[0045] In some embodiments, a system may be provided that includes a detection device and a server as described above, and a network configured to connect the detection device and the server.
[0046] In one embodiment, a program can be provided that includes instructions executable by a computer (which may be local or remote) having a processor that can be coupled via a network to a sensor configured to acquire sounds generated from a target system. The instructions included in the program can be executed by the processor to acquire input related to vibrations generated from the target system via an input / output interface, and a determination process can be performed using the machine learning model as described above.
[0047] FIG. 4 is an example of a flowchart of processing by a program according to an embodiment of the present invention. First, in step 402, time series data related to vibrations generated from an object is acquired (from an external sensor connected via a network, etc.). Then, in step 404, an anomaly detection model configured to learn time series data related to normal vibrations related to the object is used to determine whether the object is abnormal. In addition, in step 406, an anomaly identification model configured to learn time series data related to abnormal vibrations related to the object is used to determine whether abnormal vibrations are generated from the object. Note that in this example, steps 404 and 406 are described in parallel, but the present invention is not limited to this, and the processing may be performed in series.
[0048] Then, in step 408, the system checks whether the judgments of the anomaly detection model and the anomaly identification model are the same. In this example, if the judgments match and there is no anomaly, nothing is done and the system returns to the step of acquiring input of generated sound 402. In another example, some kind of update of the model may be performed as described later. If the judgments do not match, the system proceeds to step 410.
[0049] In step 410, it is confirmed whether the judgments of the anomaly detection model and the anomaly identification model match and indicate that there is an anomaly. If they match and indicate that there is an anomaly, the system can notify the user to actually check the target (for example, by stopping the operation and checking the inside) in step 412. If the judgments do not match, the system can proceed to step 414 and end the series of processes.
[0050] In step 414, one of the models can be updated and it can be checked again whether the judgments match. Alternatively, after updating the model, a notification can be issued to prompt the user to proceed to step 412 and check the object. In another example, the process can proceed from step 412 to step 414 to end the series of processes. If it is confirmed that there is no abnormality in the object (if it is determined that one of the models has recognized it incorrectly), the process can return to input acquisition 402 of the generated vibration.
[0051] FIG. 5 is an example of a flowchart of a process by a program according to another embodiment of the present invention. First, as in the example of FIG. 4, in step 502, time series data related to vibrations generated from an object is acquired, and in steps 504 and 506, a determination is made as to whether or not an abnormal vibration has occurred from the object using an anomaly detection model or an anomaly identification model configured to learn time series data related to normal or abnormal vibrations related to the object, respectively. Thereafter, in step 508, it is confirmed whether or not the anomaly detection model has determined that the object is abnormal and whether or not the anomaly identification model has determined that the object is normal. If not, the process may return to step 502. If yes, the process may proceed to step 510, update the anomaly identification model, and then return to step 502. Such a configuration is useful when, in the case of proceeding to step 510, there is a high probability that an anomaly that is unobserved at the time of occurrence is estimated to have occurred in the long-term operation system of the object.
[0052] As a further modification of the flow shown in Fig. 5, in step 510, a new anomaly identification model may be generated using the vibration signal (e.g., time-series data over a certain period related to acoustic data or acceleration data) that was used when the judgment was based. The new anomaly identification model thus generated may coexist with the previously existing anomaly identification model (processing is performed using the increased number of anomaly identification models). Alternatively, the new anomaly identification model thus generated may replace the previously existing anomaly identification model (if there were originally multiple anomaly identification models, one of them may be replaced, or all of them may be replaced).
[0053] It should be noted that the looping process in the above-described flowchart is merely an example, and in another example, the process may be ended instead of looping back to the first step.
[0054] In an embodiment, a computer readable medium (SSD, HDD, optical media, etc.) may be provided that stores the above-mentioned program. EXAMPLES
[0055] Hereinafter, an embodiment in which a damper of a bogie of a railway vehicle is set as a system to be monitored will be described. As an example, consider a case where the system is equipped with an abnormality (abnormal sound) identification model α that has been trained using sound data related to an abnormality of "damper deformation" as training data, and an abnormality (abnormal sound) identification model β that has been trained using sound data related to an abnormality of "rubber pieces getting stuck in a moving part" as training data. FIG. 6 shows an example of the transition of the output (score) of each model when a known abnormality related to β occurs. FIG. 7 shows an example of the transition of the output of each model when an unknown abnormality occurs. In FIGS. 6 and 7, the time axis is set from the front of the screen to the back (diagonal on the paper), and the vertical axis indicates the anomaly score.
[0056] Case 1: The output of the abnormal sound detection model exceeds the threshold. The output of the allophone discrimination model does not exceed the threshold. An actual abnormality was found in the damper. In this case, two possible cases are assumed: the damper malfunctions, producing a sound that deviates from normal sounds and is observed by the abnormal sound detection model, but because this is an abnormal sound that is unknown to the abnormal sound identification model, the abnormal sound identification model is unable to react (1a); and the abnormal sound identification model is unable to react appropriately even though a known abnormal sound has occurred (1b).
[0057] In case (1a), if it is confirmed that a new abnormality, such as "insufficient oil in the sliding parts," has occurred in the damper, a new abnormal sound identification model γ can be generated using the data related to that abnormal sound as training data and incorporated into the system.In case (1b), if no new abnormality has been confirmed, the abnormal sound identification models α and β can be discarded, and a re-trained abnormal sound identification model can be generated and inserted into the system.
[0058] Case 2: The output of the abnormal sound detection model exceeds the threshold. The output of the allophone discrimination model does not exceed the threshold. No abnormalities were found in the damper. In this case, there are two possible cases: the damper develops an unknown abnormality, producing a sound that deviates from normal sounds and is observed by the abnormal sound detection model; however, because this is completely unknown, neither the abnormal sound identification models α nor β are able to react (2a); or the abnormal sound detection model makes a mistake (2b).
[0059] In case (2a), the damper may have a previously unknown anomaly, prompting further investigation (e.g., a probe using a different sensor), or in case (2b), the system may be reconfigured by retraining the abnormal noise detection model.
[0060] Case 3: The output of the abnormal sound detection model exceeds the threshold. The output of the allophone discrimination model exceeds the threshold. An abnormality was found in the damper. In this example, two cases are assumed: (3a) when the damper has a known abnormality and it is correctly detected, and (3b) when an abnormality is detected but it is not the type of abnormality that was actually estimated. This system can notify the user to prompt them to check the type of abnormality, and may prompt the user to check by outputting a message or an alert sound on the GUI, for example (the same applies to the example below).
[0061] In case (3a), if the damper is actually deformed and the abnormal sound identification model α reacts, it is considered preferable to continue operating the system as is. In case (3b), if the damper is actually not deformed because rubber pieces are stuck in the moving part, it can be determined that the estimation of the cause of the abnormal sound should be optimized. In this case, it is possible to encourage re-learning of each abnormal sound identification model by taking into account the difference in features between the abnormal sound to which the abnormal sound identification model α should react and the abnormal sound to which the abnormal sound identification model β should react, or the difference in features between the abnormal sound to which the abnormal sound identification model α should react and normal sounds.
[0062] Case 4: The output of the abnormal sound detection model does not exceed the threshold. The output of the allophone discrimination model does not exceed the threshold. An abnormality was found in the damper. For example, if a model trained on subway train data is applied to above-ground trains, it may underestimate the effects of weather and temperature. In this case, it can be determined that both models need to be retrained, and if necessary, data from other sensors can be used to encourage retraining.
[0063] Case 5: The output of the abnormal sound detection model exceeds the threshold. The output of the allophone discrimination model exceeds the threshold. No abnormalities were found in the damper. As with Case 4 above, it is possible to encourage retraining of each model. This can occur when a model that was trained based on data from vehicles on routes with few curves is applied as is to vehicles on routes with many curves. Compared to Case 4, it is assumed that there is a high probability of overfitting, so it is possible to suggest retraining and model reconfiguration that takes this into consideration.
[0064] Case 6: The output of the abnormal sound detection model does not exceed the threshold. The output of the allophone discrimination model does not exceed the threshold. No abnormalities were found in the damper. In this case, the system is considered to be operating without any problems.
[0065] Case 7: The output of the abnormal sound detection model does not exceed the threshold. The output of the allophone discrimination model exceeds the threshold. An abnormality was found in the damper. In this case, it is believed that the abnormal sound detection model is missing a detection. For example, this can occur when there is a sudden malfunction or breakdown based on a momentary event, such as a foreign object flying into the vehicle from the outside colliding with it. In this case, the system can suggest that a separate abnormal sound detection model be created that learns by weighting acute abnormal sounds and then incorporated into the system.
[0066] Case 8: The output of the abnormal sound detection model does not exceed the threshold. The output of the allophone discrimination model exceeds the threshold. No abnormalities were found in the damper. In this case, it is considered that the abnormal sound identification model is making a mistake. For example, this can happen when the feature values of normal sound and abnormal sound β are similar. In this case, measures can be taken to make the feature values of normal sound and abnormal sound β different and re-learn the abnormal sound identification model.
[0067] As described above, the system of the present invention can flexibly respond to various situations and continuously update (e.g., optimize, re-learn) the machine learning model, thereby achieving the remarkable effect of always striving for improved accuracy.
Claims
1. 1. A system for monitoring an object, comprising: a sensor configured to obtain a vibration signal generated by the object; a storage means configured to acquire and store input from said sensors; A plurality of types of machine learning models configured to be stored in the storage means; Judgment method and Including, The plurality of types of machine learning models are an anomaly detection model configured to learn time series data relating to normal vibrations associated with the object; an anomaly identification model configured to learn time series data relating to abnormal vibrations associated with the object; Including, The determination means: Based on an input from the sensor that receives a vibration signal generated from the target, a first judgment using the anomaly detection model and a second judgment using the anomaly identification model are separately performed, and differences between the results of the first judgment and the results of the second judgment are identified; and and updating the anomaly detection model and / or the anomaly identification model based on a difference between a result of the first judgment and a result of the second judgment. A system characterized in that
2. a means for acquiring data relating to the state of generated vibration that led to the judgment result when the judgment means judges that the first judgment result and the second judgment result are different; The system of claim 1 further comprising:
3. a display means configured to present data relating to the state of the generated vibration; The system of claim 2 , further comprising:
4. The determination means: When the first judgment results in a judgment that the object is abnormal and the second judgment results in a judgment that the object is normal, the anomaly identification model is updated.
2. The system according to claim 1, characterized in that
5. The determination means: When the first judgment results in judging that the object is abnormal and the second judgment results in judging that the object is normal, a new anomaly identification model is created based on the vibration signal generated from the object.
5. The system according to claim 4, characterized in that
6. A program including instructions executable by a computer having a processor, the program being capable of being coupled via a network to a sensor configured to acquire vibrations generated from an object, the program comprising: The instructions are executed by the processor, acquiring an input related to vibration generated from the target through an input / output interface; A storage means stores a plurality of types of machine learning models, wherein the plurality of types of machine learning models include at least an anomaly detection model configured to learn time series data relating to normal vibrations associated with the object; an anomaly identification model configured to learn time series data relating to abnormal vibrations associated with the object; and performing a first determination using the anomaly detection model based on an input related to vibration generated from the target by a determination means; performing a second determination using the anomaly identification model based on an input related to vibration generated from the target by a determination means; Identifying by a judgment means whether the result of the first judgment is the same as the result of the second judgment; updating the anomaly detection model and / or the anomaly identification model based on the difference between the result of the first judgment and the result of the second judgment by a judgment means; configured to A program characterized by:
Citation Information
Patent Citations
Abnormality detection system, abnormality detection method, abnormality detection program, and method for generating learned model
JP2020119605A
Information processing device, information processing method and program
JP2022090761A
