Sound inspection system and sound inspection method

Abstract

Provided is a sound inspection system for determining the state of an inspection target object on the basis of the sound of the inspection target object, the sound inspection system comprising a model generating unit for generating a normal sound model from the normal operating sound of the inspection target object, a sampling unit for sampling a plurality of feature quantities from the normal sound model, an abnormality degree calculating unit for calculating a degree of abnormality of each of the plurality of sampled feature quantities on the basis of the normal sound model, and a gamma distribution generating unit for fitting the frequency distribution of the abnormality degree with a gamma distribution to generate an abnormality degree correction model, wherein the abnormality degree and an occurrence probability are associated with one another using the abnormality degree correction model. The system can compare abnormality degree outputs of the plurality of sensor devices using the same criteria.

Description

Sound inspection system and sound inspection method

[0001] The present invention relates to a sound inspection system and a sound inspection method.

[0002] At sites such as power plants, chemical plants, and steel plants, workers often listen to the operating sounds of equipment to determine whether it is operating normally. However, experience is required to be able to distinguish between abnormal sounds. Furthermore, walking around a large site and inspecting by ear places a heavy burden on the workers. Moreover, in recent years, skilled workers have been aging, making it difficult to secure new workers. Therefore, as described in Patent Document 1, a system has been proposed in which acoustic data from a monitored object is detected by a microphone and wirelessly transmitted to a monitoring processing device located away from the monitored object. Specifically, Patent Document 1 states that "the device comprises a pressure-resistant explosion-proof housing 18 having a pressure-resistant explosion-proof structure, a microphone 20 built into the pressure-resistant explosion-proof housing 18 for detecting sounds in the equipment to be monitored, a camera 22 built into the pressure-resistant explosion-proof housing 18 for capturing images of the equipment to be monitored, an information processor 24 built into the pressure-resistant explosion-proof housing 18 for processing acoustic data from the microphone 20 and image data from the camera 22 and controlling the microphone 20 and the camera 22, a radio 26 for transmitting the acoustic data and image data processed by the information processor 24 and receiving control signals for the microphone 20 and the camera 22, and an antenna 28 built into the pressure-resistant explosion-proof housing 18 and connected to the radio 26."

[0003] JP 2009-273113 A

[0004] In the conventional technology described in Patent Document 1, a monitoring processing device calculates a frequency spectrum from acoustic data received from an on-site monitoring device and detects abnormalities in the monitored object using a neural network model. However, because the sound data is large in size as is, the measurement and analysis processes are heavy and consume a lot of power.

[0005] When installing sensor devices as retrofits on plant site equipment, it is difficult to obtain a wired power source. Therefore, because sensor devices operate using an internal battery as a power source, the battery quickly runs out when performing processes that consume a lot of power, requiring frequent battery replacement and reducing usability. To address this issue, a new sensor device has been proposed that extracts features from recorded sounds, compares them with a normal sound model generated when the sensor device is installed and stored in the sensor terminal, calculates the degree of deviation from the normal sound model as an anomaly, and then collects only the anomaly score. This reduces data communication volume and enables the sensor to operate for long periods using the internal battery.

[0006] In the sensor devices described above, a different normal sound model is generated for each sensor device, and the abnormality level is calculated based on this. This is because the normal sound model generated varies slightly depending on the installation conditions of the sensor device and subtle individual differences between the sensor devices, even when monitoring the operating sounds of equipment with the same model number. In such cases, the fact that the reference normal sound model differs for each sensor device means that the meaning of the abnormality level calculated for each sensor device will differ. This makes it difficult to establish a standard for comparing the abnormality level values ​​between different sensor devices in a system in which multiple sensor devices are in operation. As a result, it is not possible to set a common threshold for multiple sensors to monitor multiple devices, making this method less user-friendly for application to equipment monitoring.

[0007] The present invention has been made in consideration of the above-mentioned problems, and its purpose is to provide a sound inspection system and a sound inspection method that can compare the abnormality level outputs of multiple sensor devices using the same criteria.

[0008] In order to solve the above-mentioned problems, one representative sound inspection system of the present invention is a sound inspection system that determines the state of an object to be inspected based on the sound of the object to be inspected, and is characterized by comprising: a model generation unit that generates a normal sound model from the normal operating sound of the object to be inspected; a sampling unit that samples a plurality of feature amounts from the normal sound model; an abnormality degree calculation unit that calculates the degree of abnormality of each of the sampled feature amounts based on the normal sound model; and a gamma distribution generation unit that fits the frequency distribution of the degree of abnormality with a gamma distribution to generate an abnormality degree correction model, and is characterized by using the abnormality degree correction model to link the degree of abnormality and the occurrence probability. Furthermore, one representative sound inspection method of the present invention is a sound inspection method for determining the state of an object to be inspected based on the sound of the object to be inspected, wherein a learning device generates a normal sound model from the normal operating sound of the object to be inspected, samples a plurality of feature quantities from the normal sound model, calculates the degree of abnormality of each of the sampled feature quantities based on the normal sound model, and generates an abnormality correction model by fitting the frequency distribution of the abnormality degree with a gamma distribution, and a data management device, when determining the state of the object to be inspected, links the degree of abnormality calculated based on the sound of the object to be inspected and the normal sound model to an occurrence probability using the abnormality correction model.

[0009] According to the present invention, it is possible to compare the abnormality level outputs of a plurality of sensor devices using the same criteria. Problems, configurations, and effects other than those described above will become apparent from the following description of the embodiments.

[0010] An overall diagram of the sound inspection system during training. A flowchart for creating an anomaly correction model. An example of fitting a sampled anomaly distribution with a gamma distribution. An example of the relationship between anomaly and occurrence probability. An example of setting thresholds for the anomaly time series, which differ for each site. An overall diagram of the sound inspection system during operation (Part 1). An overall diagram of the sound inspection system during operation (Part 2). An overall diagram of the sound inspection system during operation (Part 3). A table for comparing data from multiple devices.

[0011] An embodiment of the present invention will be described below with reference to the drawings. In this embodiment, as described in detail below, operation sounds generated in on-site equipment such as a plant are collected by a sensor device, and operation sounds generated when the equipment is in a normal state are learned by a learning terminal as normal sounds.

[0012] Multiple feature values ​​are randomly sampled from the normal sound model generated within the learning device as a result of learning, and the degree of anomaly is calculated for each of the obtained feature values ​​using the normal sound model. Furthermore, the distribution of the obtained degree of anomaly is fitted with a gamma distribution to obtain a gamma distribution expression that well explains the distribution of the degree of anomaly. Finally, an anomaly correction model that links any degree of anomaly with its occurrence probability is generated using the right-hand cumulative distribution of the gamma distribution, making it possible to interpret the degree of anomaly calculated using different normal sound models for each sensor device using a common index, the occurrence probability of the corresponding degree of anomaly.

[0013] An embodiment will be described with reference to Figures 1 to 9. Figure 1 shows an overall configuration diagram of a sound inspection system 1. The sound inspection system 1 of this embodiment is applied to plants such as power plants, chemical plants, and steel plants.

[0014] Plants are equipped with equipment (inspection objects 2) that generate noise, such as motors, pumps, compressors, turbines, and boilers. The noise generated here is the steady rotation noise generated by rotating machinery, and it is known that deterioration of the bearings installed as bearings is observed as a change in the operating noise. The role of the sound inspection system is to detect this change in the operating noise.

[0015] A sensor device 10 is provided near the inspection object 2. The sensor device 10 may be provided in contact with the inspection object 2 or may be provided at a distance from the inspection object 2. A plurality of sensor devices 10 may be provided on one inspection object 2.

[0016] The sensor device 10 includes, for example, a microphone 11, an analysis unit 12, a communication unit 13, and a power supply unit 14. Here, a device that integrates a sensor function and a communication function is shown, but the sensor function and the wireless communication function may be configured separately and connected to each other.

[0017] The microphone 11 collects sounds emitted by the inspection target object 2 and outputs them as an electrical signal. The electrical signal output from the microphone 11 is input to the analysis unit 12. The analysis unit 12 includes at least a normal sound model 15 and an AD conversion unit 16, and converts the sounds collected by the microphone 11 into digital data, which is then sent as a data string to the communication unit 13. The communication unit 13 transmits the collected sound data string to the learning device 20. The communication means L1 may be wired or wireless, but a high-speed communication protocol is desirable because sound data has a large data volume.

[0018] The power supply unit 14 supplies power from an internal battery to the microphone 11, the analysis unit 12, and the communication unit 13 of the sensor device 10. Any type of internal battery may be used.

[0019] The learning device 20 includes, for example, a communication unit 21, a learning unit 22, a model storage unit 23, and a network communication unit 24. For example, the learning device 20 may be configured as a general-purpose computer.

[0020] The communication unit 21 has a function of receiving a data string of sound data from the sensor device 10 and inputting it to the learning unit 22. For example, USB, Bluetooth (registered trademark), or the like is used as the serial communication means.

[0021] The learning unit 22 includes a feature extraction unit 25, a model generation unit 26, a baseline correction unit 27, a random sampling unit 29, an abnormality degree calculation unit 30, a frequency distribution generation unit 31, and a gamma distribution generation unit 32. The learning unit 22 has a function of generating a normal model by processing normal sound data acquired by the sensor device 10, and a function of generating an abnormality degree correction model using the normal sound model.

[0022] The feature extraction unit 25 calculates a vector representing the characteristics of the sound data received from the sensor device 10 as a feature. For example, this feature is a vector of about 100 dimensions that expresses the characteristics of the frequency components of the sound data and the changes in the frequency components over time. The feature extraction unit 25 may include iterative calculations such as dividing a long-duration sound source into multiple short-duration sound sources and extracting features for each short-duration sound source. In this case, multiple features are extracted depending on the length of the sound source.

[0023] The model generation unit 26 generates the part of the normal sound model excluding the baseline correction, using the feature values ​​output by the feature extraction unit 25. For example, a GMM (Gaussian Mixture Model) or the like is used as the normal sound model, but this is not limiting and any other distribution may be adopted and is also within the scope of the present patent.

[0024] The following describes the case where a normal sound model is constructed using a GMM. As shown in Equation 1, a normal sound model is constructed by a linear combination of multiple Gaussian distributions. Here, K indicates the number of Gaussian distributions, and in the kth Gaussian distribution, πk is the weighting coefficient of the Gaussian distribution, μk is the mean vector of the Gaussian distribution, and Σk is the variance-covariance matrix. Modeling may be performed after determining Equation 1 in advance, or the number of Gaussian distributions may be changed adaptively depending on the feature amounts. Furthermore, a model may be generated by incremental learning, in which multiple feature amounts are learned sequentially, or by batch learning, in which multiple feature amounts are learned at once.

[0025] The baseline correction unit 27 is a functional block that corrects the degree of anomaly using a baseline obtained by multiplying the average value of the log likelihood for M feature quantities by minus 1 for the GMM obtained by Equation 1. The baseline correction is introduced for the purpose of correcting the expected value in the normal case to zero when the degree of anomaly is calculated by applying the feature quantities to the GMM.

[0026] The baseline b calculated by the baseline correction unit 27 and the πk, μk, and Σk calculated by the model generation unit 26 are combined to form (b, πk, μk, Σk), which is stored in the model storage unit 23 as a normal sound model 28. Here, k indicates the k-th Gaussian distribution in the GMM consisting of K Gaussian distributions.

[0027] The random sampling unit 29 extracts multiple feature vectors according to the GMM shown in Equation 1. There are several known random sampling methods, such as the direct method and Gibbs sampling, and any method may be used. However, in the case of Gibbs sampling, a certain number of sampling results from the start of sampling must be discarded as a burn-in period, and the output of the random sampling unit 29 is the discarded sampling results from this burn-in period.

[0028] The abnormality degree calculation unit 30 calculates the degree of abnormality using the sampling results obtained by the random sampling unit 29 as input y. For a plurality of sampling results, the operation of calculating one abnormality degree for each is repeated the number of times equal to the number of samples.

[0029] The frequency distribution generation unit 31 generates a frequency distribution of the multiple degrees of anomaly generated by the degree-of-anomaly calculation unit 30. The data interval for generating the frequency distribution is adjusted to some extent freely depending on the number of samples in the random sampling unit 29. For example, when sampling 100,000 points, it is preferable to set the data interval of the degree of anomaly to about 0.5 to 1.

[0030] The gamma distribution generation unit 32 performs fitting on the frequency distribution obtained by the frequency distribution generation unit 31 using the gamma distribution function shown in Equation 2. Here, the fitting parameters are α, β, and A. In Equation 2, b is the baseline correction value included in the normal sound model 28. The fitting result (α, β, A) is stored in the model storage unit 23 as the abnormality correction model 33.

[0031] The model storage unit 23 has a function of storing the normal sound model 28 and the abnormality degree correction model 33 generated by the learning unit 22. The model storage unit 23 includes information for identifying which sensor device the normal sound model 28 and the abnormality degree correction model 33 correspond to. For example, unique ID information of the sensor device is used. The normal sound model is transferred to the sensor device via the communication unit. The abnormality degree correction model 33 is transferred to the data management device 40 via the network communication unit 24.

[0032] The network communication unit 24 connects the learning device 20 and the data management device 40 via communication means L2. When the learning device 20 is configured as, for example, a general computer, the network communication unit is, for example, a wired or wireless LAN.

[0033] The data management device 40 is composed of a network communication unit 41, an anomaly correction model 42, anomaly data 43, an anomaly threshold setting unit 44, and a time-series data generation unit 45, and has the function of generating, storing, and visualizing time-series data on anomaly levels.

[0034] The network communication unit 41 is used by the data management device 40 to receive the anomaly degree correction model 42 and the anomaly degree data 43 from the learning device 20 via the communication means L2. The anomaly degree data 43 also includes additional information such as a unique ID for identifying the sensor device 10 and a timestamp at which the anomaly degree was acquired.

[0035] The anomaly correction model 42 is generated by the learning device 20 and stored in the data management device 40. The anomaly threshold setting unit 44 sets a threshold as the degree of anomaly that results in a certain occurrence probability according to the anomaly correction model 42.

[0036] The time-series data generating unit 45 has a function of generating, displaying, and saving time-series data including thresholds using information from the abnormality threshold setting unit 44 and the abnormality data 43 .

[0037] The procedure for generating an abnormality degree correction model will be described using the flow shown in FIG.

[0038] Sound data is acquired in S100, and then features are extracted in S101. Next, a normal sound model is acquired by learning multiple features (S102). Hereinafter, GMM is assumed as the normal sound model in the explanation of Fig. 2. However, the normal sound model is not limited to GMM, and other probability distributions may be used.

[0039] The generated GMM is saved as a normal sound model (S103). Next, using the obtained GMM, multiple feature quantities according to the GMM are randomly sampled (S104). As a guideline, about 1,000 to 10,000 points are preferable in terms of speed and accuracy.

[0040] The degree of abnormality of the randomly sampled feature quantities is applied to the GMM obtained in S103 to calculate the log likelihood, and the average value of the plurality of log likelihoods is stored in the normal sound model as a baseline b (S105).

[0041] Next, random sampling of feature quantities for generating an anomaly correction model is performed from the normal sound model (S106). The number of samples at this time is preferably about 10,000 to 100,000. If the number of sampling points is too small, the shape of the frequency distribution becomes rough, reducing the accuracy of the anomaly correction model. If the number of sampling points is too large, sampling takes time, resulting in a prolonged processing time.

[0042] Next, the degree of abnormality of the sampled feature is calculated using a normal sound model as the logarithmic likelihood of the GMM (S107).

[0043] The obtained abnormality degree is subjected to frequency analysis to obtain a frequency distribution (S108).

[0044] The frequency distribution for abnormality degree x>0 is fitted with a gamma distribution using Equation 2 to obtain fitting parameters A, α, and β (S109). These fitting parameters are the abnormality degree correction model.

[0045] Finally, the obtained anomaly correction model is used to perform a calculation to link the anomaly degree x with its occurrence probability P(x) (S110), where P(x) is given by the right-sided cumulative distribution of the gamma distribution.

[0046] The fitting shown in S109 in Fig. 2 will be described with reference to Fig. 3. Fig. 3 shows an example of a result 311 obtained by fitting a frequency distribution 310 using a gamma distribution. As shown in Fig. 3, the frequency distribution of the degree of anomaly can be well fitted using an appropriate set of parameters A, α, β and a baseline b.

[0047] FIG. 4 shows the relationship between the degree of anomaly and the occurrence probability. 412 is a plot of the right-hand cumulative distribution of the gamma distribution calculated using the anomaly correction model against the degree of anomaly. This relationship is generated individually when multiple sensor devices are present. In other words, the relationship between the degree of anomaly and its occurrence probability is unique to each sensor device. Therefore, the relationship shown in FIG. 4 (anomaly correction model 42) is calculated and stored for each sensor device.

[0048] Using FIG. 5, it will be explained how the sound inspection system provided in this embodiment enables time series of various abnormalities in different sensor devices A to C to be interpreted in a unified manner.

[0049] Sensor devices A to C are different sensor devices installed in different locations, and the anomaly level values ​​obtained are calculated for each normal sound model. Therefore, the anomaly level values ​​obtained from each sensor device cannot be compared with those of other sensor devices. However, because the relationship between the anomaly level and the occurrence probability can be calculated using the anomaly level correction model in each sensor device, it is possible to set an anomaly level with an occurrence probability of, for example, 1E-10 for each sensor device. If this is set as a threshold, for example, an alert can be issued if an anomaly level exceeding the occurrence probability of 1E-10 is observed in all sensor devices.

[0050] A first example of equipment monitoring performed by the sound inspection system of Example 1 will be shown using Figure 6. Note that Figure 1 shows a system for generating a normal sound model and an abnormality degree correction model, and that there are new function blocks that are used when equipment monitoring is performed.

[0051] The sensor device 10 includes, for example, a microphone 11, an analysis unit 12, a communication unit 13, and a power supply unit 14. Here, a device that integrates a sensor function and a wireless communication function is shown, but the sensor function and the wireless communication function may be configured separately and connected to each other.

[0052] The microphone 11 collects sounds emitted by the inspection target object 2 and outputs them as electrical signals. The electrical signals output from the microphone 11 are input to the analysis unit 12. The analysis unit 12 includes a normal sound model 15, an AD conversion unit 16, a feature extraction unit 17, and an abnormality degree calculation unit 18, analyzes the sounds collected by the microphone 11, and sends the analysis results to the communication unit 13.

[0053] The AD conversion unit 16 samples and quantizes the amplitude of the analog signal, converts it into a digital value, and sends it to the feature extraction unit 17. The feature extraction unit 17 has the same function as the feature extraction unit 25 in Figure 1. However, unlike Figure 1, it operates in the sensor device 10. The abnormality degree calculation unit 18 is a functional block that calculates the degree of abnormality using the normal sound model 15, just like the abnormality degree calculation unit 30 in Figure 1. However, unlike Figure 1, it operates in the sensor device 10. The normal sound model 15 is generated by the learning device 20 in Figure 1 and transferred to and stored in the sensor device 10.

[0054] The communication unit 13 communicates with the communication unit 21 of the wireless base station 60 to transmit a packet of the analysis result generated by the analysis unit 12 to the data management device 40. The analysis result packet is transmitted to the wireless base station 60 as shown by communication means L1, which is a wireless communication path, and is further transmitted from the wireless base station 60 to the data management device 40 as shown by communication means L2, which is a wireless communication path.

[0055] The power supply unit 14 supplies power from an internal battery to the microphone 11, the analysis unit 12, and the communication unit 13 of the sensor device 10, which is a wireless slave station. Any type of internal battery may be used.

[0056] The wireless master station 60 is a relay device that constitutes part of a sensor network that is spread throughout the plant. A sound sensor network may be included as part of the sensor network. In this case, the sensor network may include a sound sensor network that detects and diagnoses abnormal sounds, as well as a sensor network that detects temperature, humidity, pressure, voltage, current, frequency, resistance, flow rate, flow velocity, color, images, etc. Alternatively, all of the sensor networks in the plant may be sound sensor networks.

[0057] The wireless master station 60 can wirelessly communicate with a plurality of sensor devices 10 and transmit packets (analysis result data) from each sensor device 10 to the data management device 40. Packets can also be transferred between adjacent sensor devices 10 using a so-called bucket brigade method.

[0058] The data management device 40 may have the functions of the wireless master station 60. In this case, the sensor device 10 can directly communicate with the data management device 40 by wireless communication.

[0059] The data management device 40 holds the analysis results received from the sensor device 10. The data management device 40 may be configured as a computer including, for example, a microprocessor, a main memory device, an auxiliary memory device, an input / output circuit, a communication circuit, a user interface device (none of which are shown), and the like.

[0060] The data management device 40 includes, for example, a network communication unit 41 , an abnormality degree correction model 42 , abnormality degree data 43 , an abnormality degree threshold setting unit 44 , and a time-series data generation unit 45 .

[0061] The network communication unit 41 has a function of communicating with the wireless master station 60. The abnormality degree data 43 is data extracted from a packet of analysis results received from the sensor device 10, which is a wireless slave station, and stored in association with the time of collection. The functions of the abnormality degree correction model 42 and the abnormality degree threshold setting unit 44 are the same as those in FIG. 1.

[0062] The time-series data generation unit 45 has a function of providing time-series data and threshold information from the anomaly threshold setting unit 44 and the anomaly data 43 in response to a request from the monitoring terminal 50. The monitoring terminal 50 outputs the time-series data to a display, a printer, or the like in the form of a graph. It is also possible to analyze the pattern of anomaly fluctuations by providing a data analysis process such as a clustering process in the data management device for the time-series data held by the time-series data generation unit 45.

[0063] 7 shows a second example of equipment monitoring performed by the sound inspection system of the first embodiment. The configuration of the analysis unit of the sensor device 10 is different from that of the first example shown in FIG.

[0064] 6 in that the analysis unit 12 of the sensor device 10 has an anomaly degree conversion unit 19 and an anomaly degree correction model 42. The anomaly degree correction model 42 is the same as the anomaly degree correction model 42 in Fig. 1 during learning, and is transferred to the sensor device 10 from the learning device 20 and stored therein.

[0065] The abnormality degree conversion unit 19 has a function of converting the abnormality degree into an occurrence probability using the abnormality degree correction model 42. For this conversion, for example, the relationship shown in FIG.

[0066] 8 shows a third example of the case where the sound inspection system of the first embodiment performs facility monitoring. The configuration of the data management device is different from that of the first example shown in FIG.

[0067] The data management device 40 differs from that shown in Fig. 6 in that it is newly provided with an anomaly degree conversion unit 19. This anomaly degree conversion unit 19 has the function of converting the degree of anomaly into the probability of occurrence of that degree of anomaly using an anomaly degree correction model 42 and anomaly degree data 46. For this conversion, the relationship shown in Fig. 4, for example, is used.

[0068] Figure 9 shows a table comparing data from multiple devices, for example, to determine repair priorities. By applying the method of the present invention, the anomaly probability is calculated using the same criteria for different devices, allowing accurate comparison of the device conditions. This enables condition-based maintenance, whereby devices with a high anomaly probability are given priority for maintenance.

[0069] As described above, the disclosed system is a sound inspection system 1 that determines the condition of an inspection object based on the sound of the inspection object, and includes a model generation unit 26 that generates a normal sound model from the normal operating sound of the inspection object, a sampling unit (random sampling unit 29) that samples multiple feature quantities from the normal sound model, an abnormality degree calculation unit 30 that calculates the abnormality degree of each of the sampled feature quantities based on the normal sound model, and a gamma distribution generation unit 32 that fits the frequency distribution of the abnormality degrees with a gamma distribution to generate an abnormality degree correction model, and is characterized in that the abnormality degree and the occurrence probability are linked using the abnormality degree correction model.This system can indicate the degree of abnormality of the condition of the inspection object by its occurrence probability, eliminating the effects of installation conditions and individual differences and making it possible to compare the abnormality degree outputs of multiple sensor devices using the same standard.

[0070] The disclosed system also includes a learning device 20 having the model generation unit 26, the sampling unit, the abnormality degree calculation unit 30, and the gamma distribution generation unit 32, a sensor device 10 that collects sounds of the inspection target, and a data management device 40, wherein the model generation unit 26 generates a normal sound model from the operation sounds collected by the sensor device 10, and the data management device 40 links the degree of abnormality with the occurrence probability using the abnormality degree correction model generated by the learning device 20. In this system, the learning device that generates the normal sound model and the abnormality degree correction model can be excluded from the configuration during inspection.

[0071] Furthermore, the data management device 40 calculates the right-hand cumulative distribution of the gamma distribution using the anomaly correction model in order to link the degree of anomaly with the occurrence probability, thereby making it possible to easily obtain the occurrence probability from the gamma distribution.

[0072] As an example, when determining the state of the inspection object, the sensor device 10 collects sounds from the inspection object, calculates the abnormality level of the collected sounds based on the normal sound model, and transmits the calculated abnormality level, and the data management device 40 outputs data indicating the state of the inspection object based on the abnormality level transmitted by the sensor device. This configuration reduces the processing load on the sensor device 10, and enables the data management device 40 to output the abnormality level of the sensor device 10 according to a unified standard.

[0073] As an example, when determining the state of the inspection object, the sensor device 10 collects sounds from the inspection object, calculates the degree of abnormality of the collected sounds based on the normal sound model, and converts the degree of abnormality into an occurrence probability based on the abnormality correction model. In this configuration, the sensor device 10 converts the degree of abnormality into an occurrence probability, so that the outputs of the sensor devices 10 can be directly compared.

[0074] As an example, when determining the state of the inspection object, the sensor device collects sounds from the inspection object, calculates the degree of abnormality of the collected sounds based on the normal sound model, transmits the calculated degree of abnormality, and the data management device converts the degree of abnormality into an occurrence probability based on the abnormality correction model. This configuration reduces the processing load on the sensor device 10, and allows the data management device 40 to output the occurrence probability of the state of the sensor device 10.

[0075] Furthermore, in the disclosed system, the learning device 20 generates a plurality of abnormality degree correction models corresponding to the plurality of sensor devices 10. Furthermore, the data management device 40 displays and outputs a plurality of occurrence probabilities for the plurality of sensor devices 10 in a comparable manner. Therefore, for the abnormality degree calculated based on a different normal sound model for each sensor device, by calculating the relationship between the abnormality degree and the probability of occurrence of that abnormality, the abnormality degree and the occurrence probability are linked, and thus monitoring can be performed by setting a threshold based on the occurrence probability value.

[0076] The present invention is not limited to the above-described embodiments, and various modifications are possible. For example, the above-described embodiments have been described in detail to clearly explain the present invention, and the present invention is not necessarily limited to those including all of the described components. Furthermore, not only can such components be removed, but also replacement or addition of components is possible. For example, the object of inspection is not limited to rotating machines, as long as it is a device that continuously generates a certain amount of sound. The abnormality to be detected will appear as a change in the continuously generated sound. Furthermore, there may be multiple output thresholds.

[0077] 1: Sound inspection system, 2: Inspection object, 10: Sensor device, 11: Microphone, 12: Analysis unit, 13: Communication unit, 14: Power supply unit, 15: Normal sound model, 16: AD conversion unit, 17: Feature extraction unit, 18: Abnormality degree calculation unit, 19: Abnormality degree conversion unit, 20: Learning device, 21: Communication unit, 22: Learning unit, 23: Model storage unit, 24: Network communication unit, 25: Feature extraction unit, 26: Model generation unit, 27: Base Line correction unit, 28: normal sound model, 29: random sampling unit, 30: abnormality degree calculation unit, 31: frequency distribution generation unit, 32: gamma distribution generation unit, 33: abnormality degree correction model, 40: data management device, 41: network communication unit, 42: abnormality degree correction model, 43: abnormality degree data, 44: abnormality degree threshold setting unit, 45: time series data generation unit, 46: abnormality degree data, 50: monitoring terminal, 60: wireless master station

Claims

1. A sound inspection system that determines the condition of an object to be inspected based on the sound of the object to be inspected, comprising: a model generation unit that generates a normal sound model from the normal operating sounds of the object to be inspected; a sampling unit that samples a number of features from the normal sound model; an abnormality degree calculation unit that calculates the degree of abnormality of each of the sampled features based on the normal sound model; and a gamma distribution generation unit that fits a frequency distribution of the degree of abnormality with a gamma distribution to generate an abnormality degree correction model, wherein the sound inspection system links the degree of abnormality to a probability of occurrence using the abnormality degree correction model.

2. A sound inspection system as described in claim 1, comprising: a learning device having the model generation unit, the sampling unit, the abnormality calculation unit and the gamma distribution generation unit; a sensor device that collects sounds of the object to be inspected; and a data management device, wherein the model generation unit generates a normal sound model from the operating sounds collected by the sensor device, and the data management device links the degree of abnormality to the occurrence probability using the abnormality correction model generated by the learning device.

3. A sound inspection system as described in claim 2, characterized in that the data management device calculates the right-sided cumulative distribution of a gamma distribution using the anomaly correction model to link the degree of anomaly with the occurrence probability.

4. A sound inspection system as described in claim 2, characterized in that, when determining the condition of the object to be inspected, the sensor device collects the sound of the object to be inspected, calculates the degree of abnormality of the collected sound based on the normal sound model, and transmits the calculated degree of abnormality, and the data management device outputs data indicating the condition of the object to be inspected based on the degree of abnormality transmitted by the sensor device.

5. A sound inspection system as described in claim 2, characterized in that, when determining the state of the inspection object, the sensor device collects the sound of the inspection object, calculates the degree of abnormality of the collected sound based on the normal sound model, and converts the degree of abnormality into an occurrence probability based on the abnormality correction model.

6. A sound inspection system as described in claim 2, characterized in that, when determining the condition of the inspection object, the sensor device collects the sound of the inspection object, calculates the degree of abnormality of the collected sound based on the normal sound model, and transmits the calculated degree of abnormality, and the data management device converts the degree of abnormality into an occurrence probability based on the abnormality correction model.

7. A sound inspection system according to claim 2, characterized in that the learning device generates a plurality of anomaly correction models corresponding to a plurality of the sensor devices.

8. A sound inspection system according to claim 7, characterized in that the data management device displays and outputs a plurality of occurrence probabilities for a plurality of the sensor devices in a comparative manner.

9. A sound inspection method for judging the condition of an object to be inspected based on the sound of the object to be inspected, comprising: a learning device that generates a normal sound model from the normal operating sound of the object to be inspected; samples a plurality of features from the normal sound model; calculates the degree of abnormality of each of the sampled features based on the normal sound model; fits a frequency distribution of the degree of abnormality with a gamma distribution to generate an abnormality correction model; and a data management device that, when judging the condition of the object to be inspected, links the degree of abnormality calculated based on the sound of the object to be inspected and the normal sound model to an occurrence probability using the abnormality correction model.

Images