JP7193491B2 - Sound inspection system - Google Patents

Description

The present invention relates to a sound inspection system and control method.

At sites such as power plants, chemical plants, steel plants, and the like, workers sometimes judge whether the equipment is normal by listening to the operating sounds of the equipment. However, experience is required to be able to distinguish abnormal sounds. In addition, the workers have to walk around the large site and check by ear, which puts a heavy burden on the workers. Moreover, in recent years, the aging of skilled workers has progressed, making it difficult to secure new workers. In view of this, Patent Document 1 proposes a system in which acoustic data of an object is detected by a microphone and wirelessly transmitted to a device located away from the object.

Although it is not a technique related to sound inspection, Patent Document 2 discloses an inspection device that determines the quality of inspection results based on the score of the degree of abnormality.

Japanese Patent Application Laid-Open No. 2009-273113 JP 2019-159820 A

With the aging of field workers and the loss of know-how regarding field equipment, there is a demand for automation and labor saving in sound inspections. For this reason, attention has been paid to a system that detects malfunctions of equipment at an early stage by monitoring the operating sound of the equipment to be inspected using a microphone sensor and detecting abnormal sounds.

The monitoring processing device described in Patent Document 1 calculates a frequency spectrum from acoustic data received from a site monitoring device, and detects occurrence of an abnormality in a monitored object using a neural network model (Patent Document 1, paragraph 0066). However, in Patent Literature 1, no consideration is given to the problem that it is necessary to strengthen and change the state determination model in response to changes in environmental noise and changes in operating noise caused by changes in the operating state of equipment. This may lead to a situation in which an alarm is issued even though the driver is driving.

Although it is not a technique related to a sound inspection system, the inspection device described in Patent Document 2 has two classifiers, a first classifier before additional learning and a second classifier after additional learning. The update of the discriminator is determined from the output results of the discriminator and the second discriminator.

However, Japanese Patent Laid-Open No. 2002-200002 does not sufficiently consider at what timing additional learning is performed. Patent Document 2 describes an operation in which additional learning is performed by a learning device and the discriminator after additional learning is downloaded to an inspection device via a network. The problem of not being downloaded to the inspection device is not considered at all.

SUMMARY OF THE INVENTION The present invention has been made in view of the above problems, and an object thereof is to provide a more reliable sound inspection system and control method.
is to provide

In order to solve the above problems, a sound inspection system according to the present invention is a sound inspection system that determines the state of an inspection object based on the sound of the inspection object, and includes waveform data of the sound detected from the inspection object by a sound collector. data processing based on the lower model, and outputting the processing result; and a higher device analyzing the processing result received from the lower device based on the upper model; According to the upper learning flag issued from the side device, the lower side model is additionally learned and updated in the lower side device, and the upper side model is additionally learned and updated in the higher side device.

According to the present invention, by issuing the upper-level learning flag, the upper-level device and the lower-level device can additionally learn and update their respective models.

Explanatory drawing which shows the whole sound inspection system. The system configuration diagram of the monitoring system including the sound inspection system. Explanatory drawing which shows the detail of a visualization apparatus. FIG. 5 is an explanatory diagram of processing results sent from the abnormal noise diagnosis device to the visualization device; Explanatory drawing which shows the example of attachment of an abnormal noise diagnostic device. FIG. 5 is an explanatory diagram showing another example of attachment of the abnormal noise diagnostic device; 4 is a flowchart showing processing of the abnormal noise diagnosis device; FIG. 8 is a flowchart showing the details of measurement processing in FIG. 7; FIG. 4 is a flowchart showing processing of the visualization device; An example of a screen that presents the degree of anomaly to the user. (1) is an example of setting an additional learning flag for initial learning. (2) is an example of setting an additional learning flag associated with a change in the field state. (3) is an example of setting an additional learning flag associated with a change in ambient noise level. 4 is a flowchart of processing for performing additional learning; 4 is a flowchart showing a process of overwriting the model of the abnormal noise diagnosis device; FIG. 11 is a flowchart showing details of measurement processing according to the second embodiment; FIG. Explanatory drawing which concerns on 3rd Example and shows the whole sound inspection system. 4 is a flow chart showing the procedure for installing the sound inspection system. The schematic of the site drawing which shows arrangement|positioning of an inspection target. 4 is a graph showing intensity distribution according to angles with respect to the sound collector array; An example of an inspection correspondence table. Flowchart of measurement processing. FIG. 14 is a flowchart of processing for performing additional learning according to the fourth embodiment; FIG.

BEST MODE FOR CARRYING OUT THE INVENTION An embodiment of the present invention will be described below based on the drawings. In this embodiment, as will be described later, even if the normal operating state of the equipment to be inspected changes, or even if the acoustic environment around the object to be inspected changes, the state after the change is regarded as a new normal state. Additional learning is done with both the abnormal noise diagnosis device and the visualization device. As a result, in the sound inspection system of the present embodiment, it is possible to prevent an erroneous alarm from being output from the abnormal noise diagnosis device when the state that should be considered normal has changed.

Furthermore, in the present embodiment, the model possessed by the abnormal noise diagnosis device and the model possessed by the visualization device are controlled so as to be synchronized by a learning flag issued by the visualization device. Therefore, when the abnormal noise diagnosis device follows changes in the state to be normalized, there is no need to transmit model data from the visualization device to the abnormal noise diagnosis device. As a result, even when the visualization device and the abnormal noise diagnosis device are connected via a network with a small communication capacity (low communication speed), it is possible to quickly respond to changes in the situation (changes in the state that should be considered normal). A highly reliable sound inspection system can be provided.

The sound inspection system 3 of this embodiment is a sound inspection system that determines the state based on the sound of the inspection object 4, and the sound inspection system includes an abnormal noise diagnosis device 1 and a visualization device 2. It is The abnormal noise diagnostic apparatus 1 includes a sound collector 14 that collects operating sounds of the inspection object 4, a data processing unit 13 that analyzes the operating sounds and calculates the degree of abnormality, and a data storage unit 12 that holds analysis results. , and a communication unit 11 . The visualization device 2 includes a communication unit 23 that receives communication from the abnormal noise diagnosis device 1, a data accumulation unit 22 that accumulates data, a display unit 21 that displays analysis results, and an analysis unit 25 that analyzes data. , and a user interface unit 24 .

The visualization device 2 performs additional model learning based on the learning flag FU held in the data storage unit 22 in the analysis unit 25 , and then stores the updated model ML in the data storage unit 22 .

The abnormal noise diagnosis apparatus 1 performs additional model learning based on the learning flag FE (=learning flag FU) held in the data storage unit 12, and then holds the updated model MK. The abnormal noise diagnostic device 1 diagnoses the abnormal noise of the inspection object 4 using the updated model MK.

In this embodiment, as described above, a flag for additional learning is provided that is activated when additional learning is required for changes in operating sounds within the normal range due to changes in environmental sounds or driving conditions. there is In this embodiment, additional learning is performed based on the learning flag, thereby suppressing the output of false alarms from the abnormal noise diagnosis device 1 . As a result, in this embodiment, it is possible to realize a sound inspection system that implements robust abnormal noise inspection.

Furthermore, according to the present embodiment, additional learning is performed by both the abnormal noise diagnostic device 1 and the visualization device 2 without exchanging model data between the abnormal noise diagnostic device 1 and the visualization device 2. be able to. Therefore, even when a network with a slow communication speed is used, a situation in which the model data is downloaded from the visualization device 2 to the abnormal noise diagnostic device 1 does not fail, and a highly reliable abnormal noise inspection system is realized. can.

The present embodiment will be described below with reference to the drawings. In this embodiment, as will be described in detail below, the abnormal noise diagnostic device 1 collects operating sound data generated by an inspection object 4 such as field equipment used in a plant, and analyzes the collected sound data. Then, the degree of anomaly is calculated, and the processing results are sent to the visualization device 2 . Note that the inspection object 4 can also be called equipment or equipment.

A first embodiment will be described with reference to FIGS. 1 to 7. FIG. The sound inspection system 3 of this embodiment is a system that determines the state based on the sound of the inspection object 4 . The sound inspection system 3 collects the sound of the inspection object 4, analyzes the collected sound, transmits the processing result 5, and performs additional learning based on the learning flag FE. By storing the processing result 5 from the sound diagnosis device 1 as time-series data and using the processing result 5 to perform additional learning based on the learning flag FE, the same model as that of the abnormal sound diagnosis device 1 is generated, and a visualization device 2 having a function of visualizing the intensity.

FIG. 1 shows an overall configuration diagram of the sound inspection system 3. As shown in FIG. The sound inspection system 3 includes a visualization device 2 and an abnormal noise diagnosis device 1 . The sound inspection system 3 is applied to plants such as power plants, chemical plants, and steel plants, for example. A plurality of abnormal sound diagnosis devices 1 may be provided for one visualization device 2 . As will be described later, the sound inspection system 3 may be provided as part of a plant monitoring system that monitors the current, voltage, temperature, number of revolutions, etc. of the object 4 to be inspected.

The sound inspection system 3 of this embodiment automatically determines whether the sound (operating sound) emitted by the inspection object 4 is normal or abnormal, but the wavelength range of the sound is not limited to the human audible range. As will be described later, sounds that are higher or lower than the audible range may be included as sounds to be inspected. Furthermore, although the sound of the inspection object 4 is monitored in this embodiment, the vibration of the inspection object 4 may be monitored instead of or in addition to the sound.

Hereinafter, the abnormal noise diagnosis device 1 as an example of the "lower-order device" will be described first, and then the visualization device 2 as an example of the "upper-order device" will be described.

The abnormal sound diagnosis device 1 includes, for example, a communication section 11, a data storage section 12, a data processing section 13, and a sound collection section . The abnormal noise diagnosis device 1 includes, for example, hardware resources such as a microprocessor, memory, input/output interface, and communication interface (all not shown), and software resources such as a computer program and an operating system (all not shown). are used to implement the functions 11, 12 and 13 described above.

The plant is provided with inspection objects 4 that generate sounds, such as motors, pumps, compressors, turbines, boilers, solenoids, and valves. These operating sounds are converted into digital signals by the sound collector 14 provided in the abnormal noise diagnosis device 1 and stored as waveform data 122 in the data storage unit 12 . The waveform data 122 may be abbreviated as waveform 122 below.

The sound collecting unit 14 is composed of, for example, a microphone and an analog/digital converter (ADC). The frequency sensitivity of the microphone is preferably selected according to the inspection object 4 .

For example, when emphasizing the human audible sound band, a microphone for the audible range should be selected. If the ultrasonic band is important, an ultrasonic microphone should be selected. Thereby, it is possible to improve the accuracy of the sound inspection. It is desirable that the sampling rate of the ADC is also determined by combining it with the microphone. It is desirable to have a sampling frequency that is twice or more than the frequency of the microphone band.

The data processing unit 13 receives the learning flag FE, the waveform 122, and the model MK stored in the data storage unit 12, and outputs the feature amount 123, the degree of abnormality 124, and the model MK. If the model is updated here, the latest model after the update is output as the model MK by additional learning. Whether or not to perform additional learning is determined based on the learning flag FE that is stored due to the learning flag FU. The learning flag FU is a flag issued by the visualization device 2, and the details will be described later.

The data storage unit 12 stores only the model MK, which is the latest model that has undergone additional learning k times. By storing only the latest model in the data storage unit 12, the storage capacity of the data storage unit 12 can be reduced. Furthermore, by storing only the latest model in the data storage unit 12, it becomes easy for the data processing unit 13 to trace which model was used to execute the process. Thereby, the hardware cost of the data storage unit 12 can be reduced.

The data processing unit 13 includes, for example, a model generating unit 131, a feature amount extracting unit 132, and an abnormality degree calculating unit 133.

The feature amount extraction unit 132 analyzes the waveform 122 and extracts the feature amount. The model generation unit 131 generates a model by additional learning (learning if the model does not exist) based on the feature amount extracted by the feature amount extraction unit 132 and the learning flag FE. The degree-of-abnormality calculator 133 calculates the degree of abnormality 124 of the waveform 122 using the latest model.

The communication unit 11 provided in the abnormal noise diagnosis device 1 transmits the processing result 5 accumulated in the data storage unit 12 to the visualization device 2 . Furthermore, the communication unit 11 also has a function of receiving the setting information 6 from the visualization device 2 .

A processing result 5 and setting information 6 are exchanged between the abnormal noise diagnosis device 1 and the visualization device 2 . A processing result 5 is transmitted from the abnormal noise diagnosis device 1 to the visualization device 2 . Setting information 6 is transmitted from the visualization device 2 to the abnormal noise diagnosis device 1 . Details of the processing result 5 and the setting information 6 will be described later.

The setting information 6 includes, for example, the learning flag FU and the model Ma generated by the visualization device 2, as shown in the lower part of FIG. The learned flag FU issued by the visualization device 2 and transmitted to the abnormal noise diagnostic device 1 is stored in the data storage unit 12 as a learned flag FE in the abnormal noise diagnostic device 2 .

The model Ma is transmitted from the visualization device 2 to the abnormal noise diagnostic device 1, for example, when it is determined that the model MK of the abnormal noise diagnostic device 1 is not suitable for abnormal noise diagnosis. That is, in order to replace the model held by the abnormal noise diagnostic device 1 with an appropriate model, the visualization device 2 transmits the model Ma to the abnormal noise diagnostic device 1 .

The communication data such as the processing result 5 and the setting information 6 shown in FIG. 1 are only data related to abnormal noise diagnosis. Signals related to system operation such as system error information and measurement instruction information are separately transmitted and received through communication units 11 and 23 as necessary.

The visualization device 2 will be explained. The visualization device 2 includes, for example, a display unit 21, a data storage unit 22, a communication unit 23, a user interface unit 24, and an analysis unit 25.

The communication unit 23 has a function of receiving the processing result 5 from the abnormal noise diagnostic device 1 and a function of transmitting setting information 6 from the visualization device 2 to the abnormal noise diagnostic device 1 . The data accumulation unit 22 accumulates various data. The analysis unit 25 uses various data accumulated in the data accumulation unit 22 to perform data analysis. The user interface unit 24 is a device operated by a user, and instructs the visualization device 2 to generate a learning flag FU and the like. The display unit 21 displays the processing result 5 as it is or after processing it. The user can refer to the information (processing result 5, trend analysis result, etc.) provided from the display unit 21 to determine whether to instruct the generation of the learning flag FU.

The data accumulation unit 22 described above accumulates the processing result 5 received from the abnormal noise diagnosis device 1 through the communication unit 23 . FIG. 4 is a storage example of the processing result 5 showing the diagnosis result at a certain time. As shown in FIG. 4, the processing result 5 can include, for example, a learning flag FE (121), a waveform 122, a feature amount 123, and an abnormality degree 124. The learning flag FU 25 and the model ML generated by the analysis unit 25 are also stored in the data storage unit 22 .

FIG. 3 is an explanatory diagram showing in detail the contents stored in the visualization device 2. As shown in FIG. In the data storage unit 22, the processing result 5, the learning flag FU, and the model ML are stored not only as the latest data but also as past data. Groups 220(1) to 220(n) of data corresponding to each abnormal noise diagnosis device 1 include sets Dt0 to Dtn of processing results 5 for each measurement time. As described above, the processing results 5 measured and diagnosed by each abnormal noise diagnosis device 1 are stored in the data storage unit 22 of the visualization device 2 . Unlike the small-sized abnormal noise diagnosis device 1 installed on site, the visualization device 2 can be equipped with a large-capacity storage device. (from the history Dt0 to the latest history Dtn), the set of the learning flag FU and the model ML can be stored for each abnormal noise diagnosis. That is, the diagnostic results of each abnormal noise diagnosis device 1 are stored in the data storage unit 22 without being overwritten. The user can also erase the stored contents of the data storage unit 22 as appropriate.

Return to FIG. The display unit 21 has a function of visualizing the data accumulated in the data accumulation unit 22 and displaying it to the user. As described above, the data accumulation unit 22 accumulates time-series data for each measurement (for each abnormal noise diagnosis), so it is possible to check the time-series data of the degree of anomaly and the learning flag, the waveform, etc. for each measurement. becomes possible.

The analysis unit 25 analyzes the processing result 5 received from the abnormal noise diagnosis device 1 . The analysis unit 25 includes, for example, a model generation unit 251, a feature amount extraction unit 252, an anomaly degree calculation unit 253, a trend analysis unit 254, and an anomaly detection unit 255.

The analysis unit 25 has a function of generating the model ML by reading out the feature quantity 123, the waveform 122, and the past model from the data accumulation unit 22 and learning or by performing additional learning.

The correct operation is that the model ML should be the same as the model MK. By comparing the calculation result of the degree of abnormality calculated by the degree-of-abnormality calculation unit 253 using the model ML and the degree of abnormality 124 calculated by the abnormal noise diagnosis device 1, it is confirmed that the model MK is the same as the model ML. can be confirmed. As a result, it is possible to confirm on the visualization device side whether or not the model has been appropriately updated in the abnormal noise diagnosis device 1 .

In addition, the analysis unit 25 has a function of performing trend analysis using the time-series data of the degree of abnormality 124 by the trend analysis unit 254 . By analyzing the temporal change in the operating sound of the inspection object 4, it is possible to know a sign of failure. When it is determined that additional learning is necessary based on the analysis result of the trend analysis unit 254 and the calculation result of the abnormality detection unit 255, the analysis unit 25 can generate a learning flag FU25.

The anomaly detection unit 255 has a function of detecting an anomaly based on the analysis result of the trend analysis unit 254 or a threshold value set by the user. The detection result of the abnormality detection unit 22 is notified to the user from the display unit 21 .

The user interface unit 24 has a function of performing various settings by the user. The user can set values such as a threshold value and a learning period to the visualization device 2 via the user interface unit 24 . Furthermore, when the user determines the need for additional learning from the information displayed on the display unit 21, the user can input an instruction to the user interface unit 24 to cause the visualization device 2 to generate a learning flag FU.

By comparing the learning flag FU generated by the visualization device 2 and the learning flag FE generated by the abnormal noise diagnosis device 1, the visualization device can determine whether the additional learning has been correctly executed or not. can be confirmed.

FIG. 2 is a system configuration diagram of a monitoring system including a sound inspection system. As described above, various inspection objects 4 are provided in the plant. The visualization device 2 of the monitoring system remotely monitors the inspection items for each inspection object 4 and provides the monitoring results to the user.

FIG. 2 shows a case where the visualization device 2 monitors different types of inspection items. may

The visualization device 2 is connected to the wireless master device 71 via the first communication network CN1. The radio base station 71 is connected to a plurality of radio repeaters 72 via the second communication network CN2. Each wireless repeater 72 is connected to each diagnostic device 1, 73A, 73B, 73C on site via the second communication network CN2.

As described above, the abnormal noise diagnosis device 1 detects the operating sound of the inspection object 4 such as a motor by the sound collector 14, diagnoses the presence or absence of abnormality based on the learning flag, and transmits the processing result 5 to the wireless repeater. 72 and the wireless master device 71 to the visualization device 2 .

The temperature diagnosis device 73A detects the temperature of the inspection object 4 with a temperature sensor 74A, diagnoses whether there is an abnormality in the temperature, and visualizes the diagnosis result via the wireless repeater 72 and the wireless master device 71. 2.

The meter reading device 73B detects the value of the meter, which is the inspection object 4B, with the camera 74B, and transmits the result of recognizing the image of the meter value to the visualization device 2 via the wireless repeater 72 and the wireless master device 71. .

The current diagnosis device 73C detects the driving current of the object 4C to be inspected by a current sensor 74C, diagnoses whether there is an abnormality in the current value, and makes the diagnosis result visible via the wireless repeater 72 and the wireless master device 71. to the conversion device 2.

A plurality of different diagnostic devices may be provided for the same inspection object 4 , or one diagnostic device may diagnose a plurality of different inspection objects 4 . Also, the type of diagnostic device is not limited to the example shown in FIG. A pressure diagnostic device, a color diagnostic device, a flow diagnostic device, a weight diagnostic device, a voltage diagnostic device, a frequency diagnostic device, an illumination diagnostic device, etc. may be added to the monitoring system.

In this embodiment, a plurality (a large number) of diagnostic devices are connected to the wireless communication network CN2, and process results (diagnostic results) are transmitted to the visualization device 2 regularly or irregularly. In general, the surveillance wireless communication network CN2 installed in a plant or the like has stricter restrictions on communication capacity and communication speed than general mobile communication services. A plurality of packets are sent from a plurality of diagnostic devices to the visualization device 2 at any time to the monitoring wireless communication network CN2 with such communication restrictions. This embodiment enables efficient and reliable remote monitoring in such a wireless communication environment.

An example of attachment of the abnormal noise diagnostic device 1 will be described with reference to FIGS. 5 and 6. FIG. As shown in FIG. 5 , the abnormal noise diagnostic device 1 and the sound collector 14 may be attached in contact with the inspection object 4 . Alternatively, as shown in FIG. 6, the abnormal noise diagnosis device 1 and the sound collector 14 can be provided separately from the inspection object 4. FIG. That is, the abnormal noise diagnosis device 1 may be either a contact type device or a non-contact type device. Furthermore, the contact sound collector 14 and the non-contact sound collector 14 may be connected to one abnormal sound diagnosis device 1 .

The processing of the abnormal noise diagnosis device 1 will be described with reference to the flowchart of FIG. The abnormal noise diagnosis device 1 waits until it receives a measurement instruction (S11). When the measurement instruction is received (S11: YES), the abnormal noise diagnosis device 1 receives setting information (S12).

The abnormal noise diagnosis device 1 shifts to the measurement process (S13), acquires the sound of the inspection object 4 using the sound collector 14, and analyzes the acquired sound. After the measurement process, the abnormal noise diagnosis device 1 transmits the process result 5 to the communication unit 11 (S14). The communication unit 11 transmits the processing result 5 to the visualization device 2 (S15). Thereafter, the abnormal noise diagnosis device 1 returns to step S11 and waits.

FIG. 8 is a flow chart showing the measurement process shown in step S13 of FIG. The abnormal sound diagnosis device 1 acquires the operation sound of the inspection object 4 as the waveform 122 by the sound collector 14, converts it into a digital signal, and stores it in a memory (not shown) (S21).

Using the waveform 122 stored in the memory, the abnormal sound diagnosis apparatus 1 performs noise removal processing (S22) and frequency analysis processing (S23), and then extracts feature amounts (S24).

Then, the abnormal sound diagnosis device 1 calculates the degree of abnormality from the feature amount extracted in step S24 and the model MK (S25).

The abnormal noise diagnosis device 1 determines whether the learning flag is set (S26). If the learning flag is not set (S26: NO), additional learning is not required, so this process ends. On the other hand, if the learning flag is set (S26: YES), the abnormal noise diagnostic device 1 performs additional learning on the model MK (S27), and terminates this process.

The processing of the visualization device 2 will be described using the flowchart of FIG. The visualization device 2 receives the processing result 5 from the abnormal noise diagnosis device 1 (S31). The visualization device 2 associates the received processing result 5 with time-series information and registers it in the data storage unit 22 as time-series data (S32).

The visualization device 2 uses the time-series data registered in the data accumulation unit 22 to perform trend analysis and abnormality detection processing by the analysis unit 25 (S33). The calculation result of step S33 is displayed by the display unit 21 (S34).

If the learning flag FU is not set (S35: NO), the visualization device 2 ends this process. If the learning flag FU is set (S35: YES), the visualization device 2 performs additional learning (S36) by creating a model, and then terminates this process.

FIG. 10 shows an example of data disclosed to the user from the display unit 21. As shown in FIG. The vertical axis of FIG. 10 indicates the degree of abnormality, and the horizontal axis indicates time. For example, the user can grasp that an abnormality has occurred in the inspection object 4 and its progress from the trend chart of the degree of abnormality shown in FIG. 10 .

The state of the object to be inspected 4 can also be defined based on a pre-specified threshold value of the degree of abnormality (here, for example, normal level, warning level, and warning level). For example, the user can determine that an abnormality has occurred in the inspection object 4 at time T1. Then, for example, when the degree of anomaly approaches the alert level, the user plans and prepares for maintenance. The user can perform maintenance on the inspection object 4 at time T2 when the degree of anomaly approaches the warning level. The sound inspection system of this embodiment can realize such CBM (Condition Based Maintenance).

The threshold may be defined by the user, or may be automatically calculated based on past performance managed by the sound inspection system. Other methods may be used to determine the threshold. The threshold may be flexibly determined according to the type of inspection object 4 or the environment.

A method of generating a learning flag will be described with reference to FIG. The learning flag generating method shown in FIG. 11(1) is a method in which additional learning is performed during a predetermined initial learning period, and the learning flag is turned off during the normal operation period. The initial learning period can be preset by the user. For example, if a period of one week or longer is set as the initial learning period in order to learn the normal operation sound of the inspection object 4, the difference in normal operation sound (normal sound) depending on the day of the week can be learned. In FIG. 11(1), at time T3 when the initial learning period ends, the initial learning is automatically switched to normal operation.

FIG. 11(2) shows how to create a learning flag when performing additional learning due to a site layout change during normal operation. For example, a change in the site layout may change the sound environment around the object to be inspected 4 , resulting in an increase in the degree of anomaly, and the abnormal noise diagnosis device 1 may output an alarm.

Therefore, by causing the model to additionally learn the environmental noise that has changed with the site layout change, the sound detected by the sound collector 14 after the site layout change can be recognized as a new normal sound. As a result, it is possible to suppress the output of an erroneous alarm, and when an abnormality actually occurs in the inspection object 4, the accuracy of detecting the abnormality is improved.

In the example of FIG. 11(2), the site layout is changed at time T4, so the user manually decides to perform additional learning at time T5.

FIG. 11(3) shows, for example, a change in the layout of the site, where a device that emits noise is installed near the inspection object 4, and the environmental noise level increases. The inspection object 4 is normal although the degree of abnormality is rapidly increasing. Therefore, the model is additionally trained with the changed environmental noise.

In addition to layout changes, it is also possible to respond to changes in other site conditions. For example, even if the operation mode of the inspection object 4 is changed or if the operation of devices around the inspection object 4 is stopped, the above-described method can be used.

The method of generating the learning flag is not limited to the above example. A learning flag for the sound inspection system may be generated by cooperating with other systems such as other sensor systems provided in the plant such as cameras, production control systems, and factory operation systems.

The additional learning process will be described using the flowchart of FIG. 12 . The visualization device 2 acquires a trigger for additional learning by the method shown in FIG. 11 (S41). As noted above, additional learning triggers may be manually generated or automatically generated.

When the visualization device 2 acquires a trigger for additional learning, it generates a learning flag FU (S42) and causes the communication unit 23 to transmit it to the abnormal noise diagnosis device 1 (S43).

When the abnormal noise diagnosis device 1 receives the learning flag FU from the visualization device 2 (S51), it generates a learning flag FE corresponding to the learning flag FU (S52). The abnormal noise diagnosis device 1 performs additional learning in the model generation unit 131 (S53). The abnormal noise diagnosis device 1 causes the communication unit 11 to transmit the processing result 5 to the visualization device 2 (S54). Then, the abnormal noise diagnosis device 1 erases the learning flag FE (S55).

When the visualization device 2 receives the processing result 5 from the abnormal noise diagnosis device 1 (S44), the learning flag FU generated by the visualization device 2 matches the learning flag FE generated by the abnormal noise diagnosis device 1. (S45).

When the learning flag FU and the learning flag FE match, the visualization device 2 performs additional learning using the waveform 122 included in the processing result 5 (S46) to generate the latest model ML (S47). The visualization device 2 associates the latest model ML, the learning flag FE, the learning flag FU, the degree of anomaly 124, the waveform 122, the feature quantity 123, and the time stamp, and stores them in the data storage unit 22. (S48).

Processing for overwriting a model will be described with reference to the flowchart of FIG. 13 . When the performance of the model MK of the abnormal noise diagnostic device 1 deteriorates for some reason, the model Ma is transmitted from the visualization device 2 to the abnormal noise diagnostic device 1, and the model MK being used in the abnormal noise diagnostic device 1 is converted to the model Ma overwrite with

For example, when the user judges that the diagnostic performance of the abnormal noise diagnosis device 1 has deteriorated due to an increase in the occurrence rate of false alarms, a trigger for overwriting the model MK being used in the abnormal noise diagnosis device 1 can be visualized. 2 manually (S61). A trigger for overwriting the model may be automatically generated and input to the visualization device 2 after obtaining user approval.

When a trigger for overwriting the model is input, the visualization device 2 selects one past model Ma from the model group stored in the data storage unit 22 (S62), and transmits it to the abnormal noise diagnosis device 1. Send (S63). The visualization device 2 can select the newest model Ma before performance degradation from among the model group stored in the data storage unit 22 . The model Ma may be designated by a trigger input in step S61, or the visualization device 2 may automatically extract candidate models and wait for selection by the user.

When the abnormal noise diagnosis device 1 receives the model Ma from the visualization device 2 (S71), it replaces the model MK currently in use with the model Ma (S72). At this point, only the model Ma received from the visualization device 2 is stored in the data storage unit 12 .

The abnormal sound diagnosis device 1 calculates the degree of abnormality 124 of the waveform 122 using the waveform 122 detected by the sound collector 14 and the model Ma (S73), and visualizes the processing result 5 described with reference to FIG. 2 (S74).

According to this embodiment configured as described above, the abnormal noise diagnosis device 1 and the visualization device 2 are connected to the model MK of the abnormal noise diagnosis device 1 and the model MK of the visualization device 2 via the learning flag FU. ML can be additionally learned and synchronized. In other words, the abnormal noise diagnosis device 1 and the visualization device 2 can basically match each other's models without transmitting model data. Therefore, in a so-called poor communication environment, the visualization device 2 can update and manage the performance of the abnormal noise diagnosis device 1, and the reliability of the sound inspection system 3 and the usability of the sound inspection system 3 for the user can be improved. and can be improved.

Furthermore, in this embodiment, the model MK of the abnormal noise diagnostic device 1 can be overwritten with the latest model Ma among the past model groups of the visualization device 2 . Therefore, even if the model MK performs erroneous learning due to a temporary change in the environmental sound, the abnormal noise diagnosis apparatus 1 can be restored to the model Ma before the temporary change in the environmental sound. performance can be restored.

A second embodiment will be described with reference to FIG. In each of the following embodiments, including the present embodiment, differences from the first embodiment will be mainly described.

The flowchart of FIG. 14 shows the measurement process of the abnormal noise diagnosis device 1 according to this embodiment. The difference between the measurement processing of this embodiment and the measurement processing of the first embodiment described in FIG. is inserted.

The frequency band of the operating sound differs depending on the type of inspection object 4 to be inspected. Therefore, in this embodiment, before extracting the feature amount 123, the device-specific filtering process (S28) is performed to extract and process the frequency band in which the feature is likely to appear. As a result, the feature quantity 123 can be extracted more accurately, and the diagnostic accuracy of the abnormal noise diagnosis device 1 is improved.

For example, the device-specific filters can consist of bandpass filter banks. In the setting information 6 transmitted from the visualization device 2 to each abnormal noise diagnosis device 1, the pass frequency band for each inspection object 4 may be specified. In this case, the setting information 6 shown in the lower part of FIG. 3 includes the value of the pass frequency band.

Furthermore, the device-by-device filter may be updated so as to follow changes in the environmental sound of the inspection object 4 . For example, similar to the model update mechanism, an update flag for updating the setting value of the device-specific filter is used to update the device-specific filter processing of the abnormal noise diagnosis device 1 according to the condition of the inspection object 4. may

The present embodiment configured in this way also has the same effect as the first embodiment. Furthermore, in the present embodiment, since the device-specific filtering process is executed, it is possible to improve the determination accuracy of the sound inspection system.

A third embodiment will be described with reference to FIGS. 15 to 20. FIG. In the sound inspection system 3A of the present embodiment, a sound collector array 14A is used as the sound collector of the abnormal noise diagnosis device 1A.

FIG. 15 is an explanatory diagram showing the entire sound inspection system 3A. A sound inspection system 3A of this embodiment includes an abnormal sound diagnosis device 1A and a visualization device 2A, and the abnormal sound diagnosis device 1A is provided with a sound collector array 14A.

In the sound inspection system 3A of the present embodiment, since the abnormal noise diagnosis device 1A uses the sound collector array 14A, a plurality of devices such as the inspection objects 4(1) to 4(6) can be diagnosed as one abnormal noise diagnosis. It can be inspected by the device 1A. An inspection object correspondence table 126 is stored in the data storage unit 12 of the abnormal noise diagnostic apparatus 1A. The inspection object correspondence table 126 manages position information of a plurality of devices to be inspected. Six inspection objects 4, ie, inspection objects 4(1) to 4(6) will be described as an example, but the number of inspection objects 4 may be one or more.

An example of the inspection object correspondence table 126 will be described with reference to FIG. The inspection object correspondence table 126 includes, for example, management numbers, inspection object names, types, angles, distances, and peak intensities extracted from intensity distribution calculation results as items.

Return to FIG. In the abnormal noise diagnosis apparatus 1A of the present embodiment, waveforms 122, models MK, degrees of abnormality 124, feature quantities 123, and models Ma (not shown) are generated corresponding to the number of inspection objects 4. Alternatively, it is analyzed and stored in the data storage unit 12 .

The sound collector array 14A can be configured as a microphone array. By using a microphone array, the direction of the sound source can be specified. The data processing unit 13 of the abnormal sound diagnosis device 1A has a sound source separation processing function (not shown) that separates sounds detected from each inspection object 4 .

The installation operation will be described with reference to FIGS. 16 to 19. FIG. FIG. 16 is a flow chart showing the procedure of the installation operation.

In the installation operation, the sound inspection system 3A first finds the positional relationship between the abnormal noise diagnostic device 1A and each inspection object 4 in order to distinguish the sound of each inspection object 4 using the sound collector array 14A. A site drawing is created (S81).

FIG. 17 shows an example of the site drawing 8 . The site drawing 8 shows the relative positions of the abnormal noise diagnosis device 1A and the inspection objects 4(1) to 4(6). A point indicated as "0" in FIG. 17 is a reference direction.

Return to FIG. Next, the sound inspection system 3A extracts the distances and angles between the abnormal noise diagnosis device 1A and each inspection object 4 from the site drawing 8 (S82). The angle can be defined, for example, by determining a reference direction and defining the angle formed by the reference direction and the inspection object 4 .

The sound inspection system 3A uses the sound collector array 14A of the abnormal noise diagnosis device 1A to measure the sound of each inspection object 4 (S83).

FIG. 18 calculates the intensity distribution for each angle of the sound from each inspection object 4 . The vertical axis in FIG. 18 indicates the sound intensity, and the horizontal axis indicates the angle from the reference direction.

The sound source directions can be separated by the number obtained by subtracting "1" from the number of sensors that the sound collector array 14 has. FIG. 18 shows an example measured by a sound collector array 14A having seven or more sensors.

The abnormal sound diagnosis apparatus 1A is placed at a position where the sound intensity changes according to the distance between the inspection object 4 and the sound collector array 14A and where the sound peaks are separated according to the angle from the reference direction. should be installed. However, this description merely describes a preferable example, and does not limit the installation position of the abnormal noise diagnosis device 1A.

Return to FIG. Finally, the sound inspection system 3A creates an inspection object correspondence table 126 shown in FIG. 19 based on the site drawing 8 and the intensity distribution calculation result by angle (FIG. 18).

The measurement process performed by the abnormal noise diagnostic apparatus 1A of this embodiment will be described with reference to the flowchart of FIG.

The abnormal sound diagnosis apparatus 1A acquires the same number of sound waveform data as the number of sensors of the sound collector array 14A by the sound collector array 14A (S91). Thereafter, all processing is performed using the waveform data detected in step S91.

The abnormal noise diagnosis device 1A sets the inspection object 4 to be processed (S92). Normally, the abnormal noise diagnostic apparatus 1A performs calculations in order from the top number of the inspection object correspondence table 126 . The abnormal noise diagnosis apparatus 1A sets the angle of the inspection object 4 to be processed from the inspection object correspondence table 126 (S93), and executes sound source separation and noise elimination using the angle information (S94).

The abnormal sound diagnostic apparatus 1A performs frequency analysis on the waveform data (S95), and further extracts feature amounts (S96). The abnormal sound diagnosis device 1A calculates the degree of abnormality using the feature amount extracted in step S96 (S97).

The abnormal noise diagnosis device 1A checks whether the learning flag is set (S98). If the learning flag is set (S98: YES), the abnormal noise diagnosis device 1A performs additional learning (S99) accompanied by model creation. On the other hand, if the learning flag is not set (S98: NO), the abnormal noise diagnosis device 1A skips step S99.

Only one learning flag exists regardless of the number of inspection objects 4 . That is, when the learning flag is set, models are created for all the target inspection objects 4 .

The abnormal noise diagnostic apparatus 1A confirms whether the calculations in the above steps have been performed for all inspection objects 4 listed in the inspection object correspondence table 126 (S100). If there is an unprocessed inspection object 4 (S100: NO), the abnormal noise diagnosis device 1A returns to step S92. When the abnormal noise diagnostic apparatus 1A completes the calculations in the above-described steps for all the inspection objects 4 listed in the inspection object correspondence table 126 (S100: YES), this processing ends.

Similar to the measurement process described in FIG. 14, also in the measurement process of this embodiment, filter processing for each inspection object is performed between the step S95 of analyzing the frequency and the step of extracting the feature amount (S96). You may insert a step to The per-inspection filter can consist of a bank of bandpass filters. It is desirable to use an individual filter for each inspection object for all the inspection objects 4 listed in the inspection object correspondence table 126 .

The present embodiment configured in this way also has the same effect as the first embodiment. Furthermore, in this embodiment, since the sound collector array 14A is used, the overall configuration can be simplified compared to the case where a sound collector is installed for each inspection object.

A fourth embodiment will be described with reference to FIG. In this embodiment, the waveform 122 is not included in the processing result 5 sent from the abnormal noise diagnosis device 1 to the visualization device 2 . Therefore, the waveform 122 is not stored in the data storage unit 22 of the visualization device 2 .

The flowchart of FIG. 21 shows the additional learning process according to this embodiment. The additional learning process of this embodiment differs from the additional learning process described with reference to FIG. 12 in steps S54B and S46B.

In step S54B of this embodiment, the processing result 5 includes the learning flag FE (121), the feature amount 123, and the degree of abnormality 124, but does not include the waveform 122. FIG. In step S<b>46</b>B of this embodiment, the visualization device 2 performs additional learning based on the feature quantity received from the abnormal noise diagnosis device 1 .

The present embodiment configured in this way also has the same effect as the first embodiment. Furthermore, in the present embodiment, waveform data is not included in the processing result 5 sent from the abnormal noise diagnosis device 1 to the visualization device 2, so the data size of the processing result 5 can be reduced. As a result, the visualization device 2 can manage many abnormal sound diagnostic devices 1 even in a congested wireless communication network or a wireless communication network with a low communication speed.

In addition, this invention is not limited to embodiment mentioned above. Those skilled in the art can make various additions, modifications, etc. within the scope of the present invention. The above-described embodiments are not limited to the configuration examples illustrated in the attached drawings. The configuration and processing method of the embodiment can be changed as appropriate within the scope of achieving the object of the present invention.

In addition, each component of the present invention can be selected arbitrarily, and the present invention includes an invention having a selected configuration. Furthermore, the configurations described in the claims can be combined in addition to the combinations specified in the claims.

1, 1A: abnormal noise diagnosis device, 2, 2A: visualization device, 3, 3A: sound inspection system, 4, 4B, 4C: inspection object, 5: processing result, 6: setting information, 11: communication unit, 12: data storage unit, 13: data processing unit, 14: sound collector, 14A: sound collector array, 21: display unit, 22: data storage unit, 23: communication unit, 24: user interface unit, 25: analysis Part 71: Wireless base station, 72: Wireless repeater

Claims (7)

A sound inspection system that determines an abnormal state of the inspection object based on the sound of the inspection object,
a lower-side device that processes the waveform data of the sound detected from the inspection object by the sound collecting unit based on the lower-side model and outputs the processing result;
a higher-level device that analyzes the processing result received from the lower-level device based on a higher-level model;
with
According to the upper learning flag issued from the upper device, the lower model is additionally learned and updated in the lower device, and the upper model is additionally learned and updated in the higher device. is,
The processing result output by the lower device includes at least a lower learning flag corresponding to the higher learning flag, a feature amount extracted from the waveform data, and a feature calculated using the lower model. and the degree of anomaly for the waveform data,
The upper device saves the processing result output from the lower device as time-series data, additionally learns the upper model based on the processing result, and the inspection target object based on the degree of abnormality determine the abnormal state of
sound inspection system. the upper learning flag is generated at a predetermined timing and transmitted to the lower device;
The sound inspection system according to claim 1. The predetermined timing is based on a user's operation input from a user interface unit of the host device,
The sound inspection system according to claim 2. The host device has a trend analysis unit that analyzes trends from the processing results saved as time-series data,
The predetermined timing is based on the trend analysis result by the trend analysis unit,
The sound inspection system according to claim 2. The higher-level device stores the old higher-level model before additional learning when additionally learning the higher-level model.
The sound inspection system according to any one of claims 1-4. The processing result further includes the waveform data,
The sound inspection system according to claim 5. The sound collecting unit includes a microphone array,
The sound inspection system according to claim 1.

Images