Abnormality detection system and abnormality detection method

Abstract

To accurately detect equipment abnormalities including signs of potential failures in addition to detecting equipment failures which already occurred.SOLUTION: An abnormality detection system has a processor and a storage device, the storage device holds acoustic data under normal conditions and acoustic data under abnormal conditions, the processor generates a spectrogram under normal conditions and a spectrogram under abnormal conditions from the acoustic data under normal conditions and the acoustic data under abnormal conditions, respectively, extracts normal points and abnormal points from the spectrogram under abnormal conditions, and generates a new spectrogram under abnormal conditions by adding the extracted abnormal points to the normal spectrogram, and generates a model for detecting abnormalities due to equipment failures by using the spectrogram under normal conditions and the spectrogram under abnormal conditions for learning.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention relates to a technique for detecting abnormalities in equipment based on acoustic data. [Background technology]

[0002] Known technologies for detecting abnormalities in equipment based on acoustic data include those disclosed in, for example, Japanese Patent Laid-Open No. 2022-067407 (Patent Document 1) and Japanese Patent Laid-Open No. 2023-108581 (Patent Document 2).

[0003] Patent Document 1 states that "the turnout contact degree monitoring device includes: a data acquisition unit that acquires acoustic data for each switching operation of an electric point machine that switches a turnout from one of two contact states, one of which is a fixed-side contact state in which the tongue rail is in close contact with the base rail on the fixed side and an opposite-side contact state in which the tongue rail is in close contact with the base rail on the opposite side, to the other contact state; an acoustic value calculation unit that calculates, based on the acoustic data, a separation period acoustic value that indicates the magnitude of the sound in the separation period in which one contact state is separated and a close-contact period acoustic value that indicates the magnitude of the sound in the close-contact period in which the other contact state is established; and a contact degree detection unit that detects a change in the degree of contact of the tongue rail with the base rail based on the separation period acoustic value and the close-contact period acoustic value."

[0004] Patent Document 2 states that "the abnormal sound diagnosis system of the present disclosure includes a sound acquisition unit that acquires sound data emitted from an object; a medical interview information acquisition unit that acquires medical interview information related to abnormal sounds generated in the object; a calculation processing unit that acquires a spectrogram showing the relationship between time, frequency, and sound pressure from the sound data; an extraction unit that acquires an estimated frequency range of abnormal sounds generated in the object based on the medical interview information acquired by the medical interview information acquisition unit and extracts a range corresponding to the estimated frequency range of the spectrogram acquired by the calculation processing unit; and a diagnosis unit that diagnoses the cause of abnormal sounds generated in the object based on the range extracted by the spectrogram extraction unit." [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Japanese Patent Publication No. 2022-067407 [Patent Document 2] Japanese Patent Application Publication No. 2023-108581 Summary of the Invention [Problem to be solved by the invention]

[0006] As described in Patent Documents 1 and 2 above, techniques for detecting equipment abnormalities based on acoustic data collected from the equipment are known. Generally, acoustic data collected when there are signs of an upcoming failure is less different from acoustic data collected during normal operation than when a failure, such as damage or breakdown, occurs. Therefore, when attempting to detect abnormalities, including signs of failure, it is desirable to analyze not only the volume of the sound but also the frequency components of the sound. Furthermore, since the frequency of sounds generated during operation generally differs depending on the part of the equipment, it is desirable to analyze the frequency components in order to identify the part where the abnormality is occurring.

[0007] To detect equipment anomalies, it is conceivable to create an anomaly detection model by learning based on input acoustic data. Training an anomaly detection model requires acoustic data when the equipment is normal and acoustic data when an anomaly occurs. However, because real-world equipment is managed and operated to prevent anomalies, it is rare to obtain acoustic data when an anomaly occurs, compared to acoustic data when the equipment is normal. Therefore, it has been difficult to collect a sufficient amount of acoustic data to generate an accurate anomaly detection model.

[0008] In the past, there had been insufficient consideration given to generating models that detect anomalies, including the signs of failure in each part of the system, as described above. [Means for solving the problem]

[0009] In order to solve at least one of the above problems, the present invention provides an anomaly detection system comprising a processor and a storage device, wherein the storage device holds normal-time acoustic data generated from operation sounds collected when equipment is operating normally, and abnormal-time acoustic data generated from operation sounds collected when the equipment is experiencing an abnormal fault, and the processor generates a normal-time spectrogram and an abnormal-time spectrogram from the normal-time acoustic data and the abnormal-time acoustic data, respectively, extracts normal parts and abnormal parts from the abnormal-time spectrogram, generates a new abnormal-time spectrogram by adding the extracted abnormal parts to the normal-time spectrogram, and generates a model for detecting abnormalities due to faults in the equipment by using the normal-time spectrogram and the abnormal-time spectrogram for learning. [Effects of the Invention]

[0010] According to one aspect of the present invention, it is possible to detect, with high accuracy, not only a fault that has occurred in the equipment but also an abnormality in the equipment, including a sign of a fault, based on acoustic data.

[0011] Other problems, configurations and effects will become apparent from the following description of the embodiments. [Brief explanation of the drawings]

[0012] [Figure 1A] 1 is a block diagram showing an example of the configuration of equipment that is a target of detection by a symptom detection system according to an embodiment of the present invention. [Figure 1B] 1 is a block diagram showing an example of the configuration of a symptom detection system according to an embodiment of the present invention. [Figure 1C] 1 is a block diagram showing an example of a hardware configuration for realizing a symptom detection system according to an embodiment of the present invention. [Figure 2] 10 is a flowchart illustrating an example of a process for creating acoustic data used by the symptom detection system according to an embodiment of the present invention to generate a learning model. [Figure 3A]FIG. 1 is a first explanatory diagram showing an example of a process in which the sign detection system according to the embodiment of the present invention creates abnormality data. [Figure 3B] FIG. 2 is a second explanatory diagram showing an example of the process in which the symptom detection system according to the embodiment of the present invention creates abnormality data. [Figure 4] FIG. 2 is an explanatory diagram showing an example of a process in which a sign detection system according to an embodiment of the present invention decomposes a spectrogram. [Figure 5] 1 is a flowchart illustrating an example of a process in which a sign detection system according to an embodiment of the present invention generates and evaluates a learning model. [Figure 6] 1 is a flowchart illustrating an example of a process in which a sign detection system according to an embodiment of the present invention performs sign detection using a learning model. DETAILED DESCRIPTION OF THE INVENTION

[0013] Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

[0014] FIG. 1A is a block diagram showing an example of the configuration of equipment that is a target of detection by a symptom detection system according to an embodiment of the present invention.

[0015] The target equipment 100 has a CPU (central processing unit) 101, a memory 102, a communication module 103, an acoustic sensor 104, and a structural part 105. Of these, the structural part 105 is the target for detecting anomalies, and the CPU 101, the memory 102, the communication module 103, and the acoustic sensor 104 have the function of collecting and transmitting data for detecting anomalies in the structural part 105. Although one structural part 105 is shown in FIG. 1A, the actual target equipment 100 may include multiple structural parts 105.

[0016] Although the type of target facility 100 is not limited, typical examples include railway facilities, factory production facilities, etc. In other words, the structural component 105 may be, for example, a component that constitutes a railway track or a component that constitutes a switch, or a component that constitutes factory production facilities, etc.

[0017] The acoustic sensor 104 converts the sound generated from the structural part 105 into acoustic data and transmits it to the CPU 101. The acoustic sensor 104 may include, for example, a microphone, an amplifier, an analog-to-digital converter, and the like (all not shown).

[0018] The CPU 101 has a data processing function 106 and a packet processing function 107. The data processing function 106 stores acoustic data acquired from the acoustic sensor 104 in the memory 102. The packet processing function 107 transmits the acoustic data stored in the memory 102 to the symptom detection system 110 (see FIG. 1B ) via the communication module 103. The communication module 103 is connected to a network (not shown) and communicates data with the symptom detection system 110.

[0019] FIG. 1B is a block diagram showing an example of the configuration of a sign detection system 110 according to an embodiment of the present invention.

[0020] The symptom detection system 110 includes a communication module 111 , a collection / storage unit 112 , an analysis algorithm 113 , and a display function 114 .

[0021] The communication module 111 communicates with the target facility 100 via a network, acquires acoustic data, and transmits it to the collection / storage unit 112. The collection / storage unit stores the acoustic data acquired from the communication module 111 in the database 115.

[0022] The analysis algorithm 113 analyzes acoustic data to perform processing to detect signs of a fault in the target equipment 100. Specifically, for example, the analysis algorithm 113 performs learning of a model for sign detection and sign detection using the model. The learning model 117 included in the analysis algorithm 113 is a learned model for sign detection, and the analysis module 116 is a function that uses the learning model 117 to detect signs.

[0023] The display function 114 is a function for displaying the result of the sign detection to the user.

[0024] In this embodiment, the function of the sign detection system 110 is described, which detects signs of upcoming failures among abnormalities in the target equipment 100, but a similar method can also be used to detect failures (such as breakdowns or damage) that have occurred in the target equipment 100. In other words, the sign detection system 110 can also be referred to as an anomaly detection system.

[0025] FIG. 1C is a block diagram showing an example of a hardware configuration for implementing a symptom detection system 110 according to an embodiment of the present invention.

[0026] The symptom detection system 110 of this embodiment shown in Fig. 1B can be realized by a computer system, and Fig. 1B shows a computer system 120 as an example.

[0027] The computer system 120 includes a processor 121, a memory (main storage device) 122, an auxiliary storage device 123, an output device 124, an input device 125, and a communication interface (I / F) 126. The above components are connected to each other by a bus. The memory 122 and the auxiliary storage device 123 are storage devices that store programs and data used by the processor 121. The data in the database 115 shown in FIG. 1B may be stored in the auxiliary storage device 123, for example, and at least a portion of the data may be stored in the memory 122 as needed for processing.

[0028] The memory 122 is configured, for example, by a semiconductor memory, and is mainly used to hold programs and data that are currently being executed. For example, programs and data stored in the auxiliary storage device 123 are loaded into the memory 122 at startup or when needed. The processor 121 executes various processes in accordance with the programs stored in the memory 122. The processor 121 operates in accordance with the programs to realize various functional units (for example, the communication module 111, the collection / storage unit 112, the analysis algorithm 113, and the display function 114 shown in FIG. 1B).

[0029] The auxiliary storage device 123 is configured with a large-capacity storage device such as a hard disk drive or a solid-state drive, and is used to store programs and data for a long period of time. For example, the auxiliary storage device 123 may store the data in the collection / storage unit 112 shown in FIG. 1B, a program for implementing the analysis algorithm 113, and a program for implementing the display function 114.

[0030] Processor 121 may be comprised of a single processing unit or multiple processing units and may include single or multiple arithmetic units or multiple processing cores. Processor 121 may be implemented as one or more central processing units, microprocessors, microcomputers, microcontrollers, digital signal processors, state machines, logic circuits, graphics processing units, systems on a chip, and / or any device that manipulates signals based on control instructions.

[0031] The input device 125 is a hardware device through which a user inputs instructions, information, etc. The output device 124 is a hardware device that presents various images for input and output, such as a display device or a printing device. The communication I / F 126 is an interface for connection to a network. For example, a processor controls the output device 124 and the communication I / F 126 according to a program, thereby realizing the functions of the display function 114 and the communication module 111, respectively.

[0032] The computer system 120 may include two or more processors 121. The functions of the symptom detection system 110 may be implemented in multiple computer systems 120. In this case, the multiple computer systems 120 communicate with each other via a network. For example, some of the functions of the system of this embodiment may be implemented in one computer system 120, and other parts may be implemented in other computer systems 120.

[0033] For example, computer system 120 may be a personal computer (PC) or server device owned by a user of symptom detection system 110 (for example, a railway equipment maintenance company or a manufacturing company that manages a factory), or may be a server device accessed by the user via a communications network. In the latter case, computer system 120 may be a virtual server on the cloud. In that case, computer system 120 shown in FIG. 1C is realized by computer resources on the cloud.

[0034] FIG. 2 is a flowchart showing an example of a process for creating acoustic data used to generate the learning model 117 by the symptom detection system 110 according to the embodiment of the present invention.

[0035] To generate the learning model 117, both acoustic data collected when the target equipment 100 is operating normally and acoustic data collected when there is an abnormality in the target equipment 100 (in this embodiment, particularly when there are signs of a fault) are required. However, as described above, while acoustic data during normal times can be collected in relatively large quantities, acoustic data during abnormal times is rare. For this reason, the symptom detection system 110 executes a process to increase the amount of acoustic data during abnormal times, as shown in FIG. 2. This process is executed based on the analysis algorithm 113 of the symptom detection system 110.

[0036] First, an overview of the processing in FIG. 2 will be described. The acoustic data collected from the target facility 100 includes acoustic data of operation sounds during normal operation and acoustic data of operation sounds during abnormal operation. Even if the latter data is classified as abnormal data as a whole, not all of the data is necessarily abnormal. For example, if the target facility 100 includes multiple structural components 105, and only some of the components are abnormal while the others are normal, the collected acoustic data will include normal and abnormal parts. The same applies when an abnormality occurs temporarily or intermittently. The symptom detection system 110 extracts normal and abnormal parts from the acoustic data of operation sounds during abnormal operation and combines the abnormal parts with the acoustic data of operation sounds during normal operation to generate new acoustic data of operation sounds during abnormal operation.

[0037] First, the symptom detection system 110 collects acoustic data of the operating sounds of the target equipment 100 under normal and abnormal conditions (step 201). Next, the symptom detection system 110 generates spectrogram image data based on the collected acoustic data of the operating sounds under normal conditions and the operating sounds under abnormal conditions (step 202). Here, a spectrogram is an image in which pixel values ​​obtained by converting signal strength for each time and frequency identified from the image data according to a predetermined conversion rule are plotted in a two-dimensional space with time on one axis and frequency on the other axis.

[0038] The conversion rule may be, for example, a rule that associates signal strength with pixel brightness, a rule that associates signal strength with pixel color, or a combination thereof. The symptom detection system 110 stores parameters that associate signal strength with pixel values ​​as conversion rules, and by changing the parameters, various spectrogram image data can be generated from the same acoustic data. Such conversion rules may be stored as part of the analysis algorithm 113. The processing from step 203 onward, which will be described later, is performed on the spectrogram image data generated in accordance with the conversion rules.

[0039] Next, the sign detection system 110 decomposes the spectrogram image data of each of the normal operating sound and the abnormal operating sound based on the frequency for each structural part 105 (step 203). For example, the sign detection system 110 may store information on the frequency range of the operating sound for each structural part 105 and decompose the spectrogram image data based on that information.

[0040] Next, the sign detection system 110 extracts frequencies for each predetermined length of time (e.g., 0.1 seconds) from the image data of each spectrogram decomposed in step 203 (step 204). For example, the sign detection system 110 extracts image data for a predetermined length of time from the image data of the spectrogram decomposed in step 203, and extracts the frequencies of the signals contained in the image data. For example, the sign detection system 110 may extract the frequency of the signal with the maximum intensity from the acoustic signals corresponding to the extracted image data, or may extract one or more frequencies where the intensity is maximum and greater than a predetermined standard, or may extract frequencies that indicate the characteristics of the acoustic signals corresponding to the extracted image data using any other method.

[0041] Next, the symptom detection system 110 compares the frequencies extracted from the operating sound under normal conditions in step 204 with the frequencies extracted from the operating sound under abnormal conditions (step 205). Specifically, the symptom detection system 110 compares the frequencies extracted from the operating sound under normal conditions and the operating sound under abnormal conditions, which correspond to the same structural part 105.

[0042] Next, the symptom detection system 110 determines whether, as a result of the comparison in step 205, the difference between the frequency extracted from the operating sound under normal conditions and the frequency extracted from the operating sound under abnormal conditions is equal to or greater than a predetermined standard (for example, within ±N hertz) and whether the difference continues for a predetermined period of time (for example, M seconds) or more (step 206).

[0043] If the difference in frequency is less than a predetermined standard or the duration of the difference is less than a predetermined time (step 206: No), the precursor detection system 110 extracts the operating sound during the abnormality that is being compared as a normal part included in the acoustic data during the abnormality (step 207).

[0044] On the other hand, if the difference in frequency is equal to or greater than a predetermined standard and the duration of the difference is equal to or greater than a predetermined time (step 206: Yes), the sign detection system 110 extracts the operation sound during abnormal operation that is the comparison target as an abnormal portion included in the sound data during abnormal operation (step 208).The sign detection system 110 then generates new sound data of the operation sound during abnormal operation by combining the extracted abnormal portion with sound data of the operation sound during normal operation (step 209).For example, new sound data of the operation sound during abnormal operation may be generated by replacing the extracted abnormal portion with a portion of the sound data of the operation sound during normal operation that has the same frequency range and duration as the extracted abnormal portion.

[0045] Next, the symptom detection system 110 checks the newly created acoustic data of the operating sound during an abnormality (step 210). This process is performed for the purpose of eliminating acoustic data that cannot actually occur. For example, the symptom detection system 110 may store conditions that actually occurring acoustic data should satisfy and determine whether the generated acoustic data satisfies those conditions, or a person may visually check the spectrogram of the generated acoustic data based on experience, etc.

[0046] FIG. 3A is a first explanatory diagram showing an example of a process in which the symptom detection system 110 according to the embodiment of the present invention creates abnormality data.

[0047] 3A are examples of spectrograms of acoustic data of operating sounds under normal and abnormal conditions, respectively. Both of these are spectrograms obtained by converting actually collected acoustic data.

[0048] Regions 301A and 302A enclosed by thick solid lines are regions of different frequency ranges in spectrogram 300A under normal conditions, and may originate from, for example, different structural parts 105. Regions 301B and 302B are regions of spectrogram 300B under abnormal conditions that are in the same frequency range as regions 301A and 302A.

[0049] In this example, the density of the pixels in the spectrogram corresponds to the signal intensity. Region 301B contains a signal intensity peak that is not seen in region 301A. This is an example of a difference between the two, which may be due to an abnormality in one of the structural components 105, for example.

[0050] FIG. 3B is a second explanatory diagram showing an example of the process in which the sign detection system 110 according to the embodiment of the present invention creates abnormality data.

[0051] 3B is a diagram showing an example of a spectrogram, shown schematically for ease of explanation. A normal spectrogram 310A includes signal intensity peaks 311A ​​and 312A, indicated by a thick solid line and a thin solid line, respectively. These are, for example, both maximum values ​​whose intensities are greater than the signals at the frequencies before and after them, and the intensity of peak 311A ​​is greater than the intensity of peak 312A.

[0052] Spectrogram 310B in the event of an abnormality contains signal intensity peaks 311B and 312B that are similar to peaks 311A ​​and 311B in the event of a normal state, respectively. Spectrogram 310B in the event of an abnormality further contains signal intensity peaks in regions 313B, 314B, and 315B indicated by dashed rectangular lines that were not seen in spectrogram 310A in the event of a normal state. Sign detection system 110 performs the comparison in step 205 and the determination in step 206 for regions 313B, 314B, and 315B.

[0053] For example, if the difference between the frequency of the peak in region 313B and the frequency of peak 311A ​​or 312A (i.e., the frequency of peak 311B or 312B) is equal to or greater than a predetermined standard and the duration is equal to or greater than a predetermined length, then symptom detection system 110 extracts region 313B as a different region in step 208. Similarly, regions 314B and 315B are also extracted as different regions when predetermined conditions are met.

[0054] Then, in step 209, the symptom detection system 110 creates new acoustic data for an abnormal state by overwriting at least one of the areas 313B, 314B, and 315B onto areas of the normal spectrogram 310A that have the same frequency range as the area 313B, 314B, and 315B.

[0055] When normal spectrogram 310A is compared with the areas of abnormal spectrogram 310B other than the areas determined to be abnormal, all data are normal but different from each other. Therefore, even if all of areas 313B, 314B, and 315B were overwritten with areas of the same frequency range as those of normal spectrogram 310A, the resulting spectrogram for abnormality would be different from spectrogram 310B for abnormality. Therefore, new acoustic data for abnormality can be created by the above process.

[0056] FIG. 4 is an explanatory diagram showing an example of a process in which the sign detection system 110 according to the embodiment of the present invention decomposes a spectrogram.

[0057] 4 is an example of a spectrogram acquired from the target equipment 100. The target equipment 100 includes a plurality of structural parts 105, and the spectrogram 400 includes a region 401 corresponding to a frequency of one of the structural parts 105 (e.g., part A) and a region 402 corresponding to a frequency of another structural part 105 (e.g., part B). In this case, in step 203, the symptom detection system 110 extracts the region 401 and the region 402.

[0058] FIG. 5 is a flowchart showing an example of a process in which the sign detection system 110 according to the embodiment of the present invention generates and evaluates the learning model 117.

[0059] First, the symptom detection system 110 converts acoustic data under normal conditions into a spectrogram (step 501). This process can be performed in the same manner as step 202 in FIG. 2. Next, the symptom detection system 110 breaks down the spectrogram into parts while checking the frequency and operating time (step 502). This process can be performed in the same manner as step 203 in FIG. 2.

[0060] Next, the sign detection system 110 adjusts a parameter (step 503). This parameter is a parameter that associates signal strength with pixel value and is held as a conversion rule by the sign detection system 110, and a predetermined parameter may be used when step 503 is executed for the first time.

[0061] Next, the symptom detection system 110 generates a learning model 117 using the spectrogram of the normal acoustic data converted using the adjusted parameters (step 504). There are no specific limitations on the specific method for generating the learning model 117. The learning model 117 can be generated using any machine learning algorithm, including known methods.

[0062] Next, the symptom detection system 110 inputs the spectrogram of the acoustic data at the time of an abnormality, converted using the above-mentioned adjusted parameters, into the generated learning model, evaluates the learning model 117 (step 505), and determines whether the degree of abnormality is equal to or greater than a threshold (step 506). Here, the degree of abnormality is an index showing how far the learning model 117 differs from the data at the time of an abnormality, and the larger this value, the greater the difference between the output when acoustic data at a normal time is input to the learning model 117 and the output when acoustic data at an abnormal time is input.

[0063] If the degree of abnormality is not equal to or greater than the threshold (step 506: No), the sign detection system 110 returns to step 503, changes the parameters, and executes the processing from step 504 onwards again. If the degree of abnormality is equal to or greater than the threshold (step 506: Yes), the processing of generating and evaluating the learning model 117 ends.

[0064] For example, the symptom detection system 110 may generate a learning model 117 for each part by executing steps 503 to 506 for each part disassembled in step 502. This allows optimal parameters to be adjusted for each part, making it possible to generate a learning model 117 with high accuracy in symptom detection for each part. Furthermore, as a result of the processing shown in FIG. 2, the amount of acoustic data during abnormal conditions used for evaluation in step 505 increases, making it possible to generate a learning model 117 with higher accuracy.

[0065] FIG. 6 is a flowchart showing an example of a process in which the sign detection system 110 according to the embodiment of the present invention performs sign detection using the learning model 117.

[0066] First, the symptom detection system 110 acquires and stores newly acquired acoustic data from the target equipment 100, which is unclear as to whether it is normal or abnormal (i.e., the data is to be used to determine whether it is normal or abnormal) (step 601).

[0067] Next, the symptom detection system 110 checks the frequency of the acquired acoustic data (step 602) and removes low frequency components (step 603).

[0068] Next, the symptom detection system creates a spectrogram of the acoustic data from which low-frequency components have been removed (step 604). This process can be performed in the same manner as step 202 in FIG.

[0069] Next, the symptom detection system decomposes the spectrogram into components (step 605). This process can be performed in the same manner as step 203 in FIG.

[0070] Next, the symptom detection system performs image analysis using the learning model 117 (step 606) and outputs the results (step 607). This completes the symptom detection process. This enables accurate detection of symptoms of failure for each component.

[0071] If a learning model 117 for each part has been generated, the sign detection system inputs the spectrogram for each part decomposed in step 605 into the learning model for each part in step 606, thereby obtaining the results for each part and outputting them in step 607. At this time, if parameters have been adjusted for each part, the parameters for each part may be used to generate a spectrogram for each part.

[0072] Furthermore, the system according to the embodiment of the present invention may be configured as follows.

[0073] (1) An anomaly detection system (e.g., precursor detection system 110), comprising a processor (e.g., processor 121) and a storage device (e.g., memory 122 and auxiliary storage device 123), wherein the storage device holds (e.g., database 115) acoustic data under normal conditions generated from operation sounds collected when equipment (e.g., target equipment 100) is operating normally, and acoustic data under abnormal conditions generated from operation sounds collected when an abnormality exists in the equipment, and the processor generates a spectrogram under normal conditions and a spectrogram under abnormal conditions from the acoustic data under normal conditions and the acoustic data under abnormal conditions, respectively (e.g., step 202), extracts normal parts and abnormal parts from the spectrogram under abnormal conditions (e.g., steps 203 to 208), generates a new spectrogram under abnormal conditions by adding the extracted abnormal parts to the spectrogram under normal conditions (e.g., step 209), and generates a model for detecting anomalies in the equipment by using the spectrogram under normal conditions and the spectrogram under abnormal conditions for learning (e.g., steps 501 to 506).

[0074] This makes it possible to detect abnormalities, including signs of equipment failure, with high accuracy based on acoustic data.

[0075] (2) In the anomaly detection system described in (1) above, the equipment includes a plurality of parts (e.g., structural part 105), and the processor decomposes the normal spectrogram and the abnormal spectrogram into spectrograms of the operating sounds of each of the plurality of parts based on frequency (e.g., step 203), extracts a plurality of locations by dividing the spectrogram of the operating sounds of each part in the abnormal state by time (e.g., step 204), and compares each of the extracted locations with the spectrogram of the operating sounds of each part in the normal state (e.g., step 205), thereby extracting the normal locations and the abnormal locations from the spectrogram in the abnormal state.

[0076] This makes it possible to detect abnormalities in each component of equipment with high accuracy based on acoustic data.

[0077] (3) In the anomaly detection system described in (2) above, when the difference between the frequency of the peak signal strength at a location extracted from the spectrogram of the operating sound of each component during the anomaly and the frequency of the peak signal strength in the spectrogram of the operating sound of each component during normal operation is equal to or greater than a predetermined standard, and the duration of the difference is equal to or greater than a predetermined length of time (e.g., step 206: Yes), the processor extracts the location as an abnormal location.

[0078] This makes it possible to detect abnormalities in each component of equipment with high accuracy based on acoustic data.

[0079] (4) In the anomaly detection system described in (2) above, the processor generates the new spectrogram for the abnormal state by adding the abnormal point extracted from the spectrogram of the operating sound of one of the parts during an abnormal state to the spectrogram of the operating sound of the same part during normal state.

[0080] This makes it possible to detect abnormalities in each component of equipment with high accuracy based on acoustic data.

[0081] (5) The anomaly detection system described in (2) above, wherein the storage device holds parameters for generating the spectrogram from the acoustic data, and the processor generates a normal spectrogram and an abnormal spectrogram from the normal acoustic data and the abnormal acoustic data, respectively, based on the parameters (e.g., step 501), generates the model for each of the components by learning the normal spectrogram (e.g., steps 502 and 504), evaluates the degree of anomaly for each of the components by inputting the abnormal spectrogram into the model (e.g., steps 502 and 505), and if the degree of anomaly is smaller than a predetermined standard (e.g., step 506: No), changes the parameters (e.g., step 503).

[0082] This makes it possible to detect abnormalities in each component of equipment with high accuracy based on acoustic data.

[0083] (6) In the anomaly detection system described in (5) above, the processor acquires acoustic data of the anomaly detection target (e.g., step 601), generates a spectrogram of the anomaly detection target from the acoustic data of the anomaly detection target based on the parameters (e.g., step 604), decomposes the spectrogram of the anomaly detection target into spectrograms of the operating sounds of each of the multiple components based on frequency (e.g., step 605), inputs the spectrogram of the anomaly detection target for each component into the model (e.g., step 606), and outputs the results (e.g., step 607).

[0084] This makes it possible to detect abnormalities in each component of equipment with high accuracy based on acoustic data.

[0085] The present invention is not limited to the above-described embodiments, but includes various modifications. For example, the above-described embodiments have been described in detail to facilitate a better understanding of the present invention, and the present invention is not necessarily limited to those including all of the described configurations. Furthermore, it is possible to replace part of the configuration of one embodiment with the configuration of another embodiment, or to add the configuration of another embodiment to the configuration of one embodiment. Furthermore, it is possible to add, delete, or replace part of the configuration of each embodiment with other configurations.

[0086] Furthermore, the above-described configurations, functions, processing units, processing means, etc. may be partially or entirely implemented in hardware, for example, by designing them as integrated circuits. The above-described configurations, functions, etc. may also be implemented in software, with a processor interpreting and executing a program that implements each function. Information such as the programs, tables, and files that implement each function can be stored in storage devices such as nonvolatile semiconductor memory, hard disk drives, and solid-state drives (SSDs), or in computer-readable, non-transitory data storage media such as IC cards, SD cards, and DVDs.

[0087] In addition, the control lines and information lines shown are those that are considered necessary for the explanation, and not all control lines and information lines in the product are necessarily shown. In reality, it can be considered that almost all components are interconnected. [Explanation of symbols]

[0088] 100 Target Equipment 101 CPU 102 memory 103, 111 Communication module 104 Acoustic Sensor 105 Structural parts 106 Data Management Department 107 Packet Processing Unit 110 Predictive Detection System 112 Collection / Storage Unit 113 Analysis Algorithm 114 Display function 115 databases 116 Analysis Module 117 Learning Model 120 Computer Systems 121 processors 122 memory 123 Auxiliary storage 124 Output Device 125 Input Devices 126 Communication Interface

Claims

1. An anomaly detection system, comprising: a processor and a storage device, The storage device stores normal acoustic data generated from operation sounds collected when the equipment is operating normally, and abnormal acoustic data generated from operation sounds collected when an abnormality occurs in the equipment, The processor: generating a normal state spectrogram and an abnormal state spectrogram from the normal state acoustic data and the abnormal state acoustic data, respectively; extracting normal and abnormal portions from the spectrogram when the abnormality occurs; generating a new spectrogram during an abnormal state by adding the extracted abnormal portion to the spectrogram during a normal state; An anomaly detection system characterized by generating a model for detecting anomalies in the equipment by using the spectrograms in normal times and the spectrograms in abnormal times for learning.

2. The anomaly detection system according to claim 1, The equipment includes a plurality of components; The processor: decomposing the normal spectrogram and the abnormal spectrogram into spectrograms of operation sounds of each of the plurality of components based on frequency; An anomaly detection system characterized by extracting multiple locations by dividing the spectrogram of the operating sounds of each component during an abnormality by time, and comparing each of the extracted multiple locations with the spectrogram of the operating sounds of each component during normal times, thereby extracting the normal locations and the abnormal locations from the spectrogram during the abnormal time.

3. The anomaly detection system according to claim 2, The processor extracts a location as an abnormal location when the difference between the frequency of the peak signal strength at the location extracted from the spectrogram of the operating sound of each component during the abnormal state and the frequency of the peak signal strength at the location extracted from the spectrogram of the operating sound of each component during the normal state is equal to or greater than a predetermined standard, and the duration of the difference is equal to or greater than a predetermined length of time.

4. The anomaly detection system according to claim 2, The anomaly detection system is characterized in that the processor generates the new spectrogram for the abnormal time by adding the abnormal point extracted from the spectrogram of the operating sound of one of the parts when an abnormality occurs to the spectrogram of the operating sound of the same part when it is normal.

5. The anomaly detection system according to claim 2, the storage device holds parameters for generating the spectrogram from the acoustic data; The processor: generating a normal state spectrogram and an abnormal state spectrogram from the normal state acoustic data and the abnormal state acoustic data, respectively, based on the parameters; generating the model by learning the normal spectrogram for each of the components; Evaluating the degree of anomaly for each of the components by inputting the spectrogram at the time of anomaly into the model; An anomaly detection system characterized in that, when the degree of anomaly is smaller than a predetermined standard, the parameter is changed.

6. The anomaly detection system according to claim 5, The processor: Acquire acoustic data of the target for anomaly detection, generating a spectrogram of the abnormality detection target from the acoustic data of the abnormality detection target based on the parameters; decomposing the spectrogram of the anomaly detection target into spectrograms of operation sounds of the plurality of components based on frequency; an anomaly detection system, comprising: inputting a spectrogram of the anomaly detection target for each of the components into the model; and outputting the result.

7. An anomaly detection method executed by a computer system, comprising: the computer system includes a processor and a storage device; The storage device stores normal acoustic data generated from operation sounds collected when the equipment is operating normally, and abnormal acoustic data generated from operation sounds collected when an abnormality occurs in the equipment, The anomaly detection method includes: a first step in which the processor generates a spectrogram in a normal state and a spectrogram in an abnormal state from the acoustic data in a normal state and the acoustic data in an abnormal state, respectively; a second step in which the processor extracts normal portions and abnormal portions from the spectrogram at the time of the abnormality; a third step in which the processor generates a new spectrogram in an abnormal state by adding the extracted abnormal portion to the spectrogram in a normal state; a fourth step in which the processor generates a model for detecting an abnormality in the equipment by using the normal spectrogram and the abnormal spectrogram for learning.

8. The anomaly detection method according to claim 7, The equipment includes a plurality of components; In the second step, the processor decomposing the normal spectrogram and the abnormal spectrogram into spectrograms of operation sounds of each of the plurality of components based on frequency; An anomaly detection method comprising: extracting a plurality of locations by dividing a spectrogram of the operating sounds of each component during an abnormal state by time; comparing each of the extracted locations with a spectrogram of the operating sounds of each component during normal states; and extracting the normal locations and the abnormal locations from the spectrogram during the abnormal state.

9. The anomaly detection method according to claim 8, In the second step, the processor extracts a location as an abnormal location if the difference between the frequency of the peak signal strength at the location extracted from the spectrogram of the operating sound of each component during the abnormal state and the frequency of the peak signal strength at the location extracted from the spectrogram of the operating sound of each component during the normal state is equal to or greater than a predetermined standard and the duration of the difference is equal to or greater than a predetermined length of time.

10. The anomaly detection method according to claim 8, In the third step, the processor generates the new spectrogram for the abnormal state by adding the abnormal point extracted from the spectrogram of the operation sound of one of the parts when an abnormality occurs to the spectrogram of the operation sound of the same part when it is normal.

11. The anomaly detection method according to claim 8, the storage device holds parameters for generating the spectrogram from the acoustic data; In the fourth step, the processor generating a normal state spectrogram and an abnormal state spectrogram from the normal state acoustic data and the abnormal state acoustic data, respectively, based on the parameters; generating the model by learning the normal spectrogram for each of the components; Evaluating the degree of anomaly for each of the components by inputting the spectrogram at the time of anomaly into the model; An anomaly detection method characterized in that, when the degree of anomaly is smaller than a predetermined standard, the parameter is changed.

12. The anomaly detection method according to claim 11, A sixth step further comprises: In the sixth step, the processor: Acquire acoustic data of the target for anomaly detection, generating a spectrogram of the abnormality detection target from the acoustic data of the abnormality detection target based on the parameters; decomposing the spectrogram of the anomaly detection target into spectrograms of operation sounds of the plurality of components based on frequency; an anomaly detection method comprising inputting a spectrogram of the anomaly detection target for each of the components into the model and outputting the result.

Images