Abnormality diagnosis device, abnormality diagnosis method, and abnormality diagnosis program

Abstract

This abnormality diagnosis device (100) comprises a signal processing device (10), the signal processing device (10) being provided with a signal processing unit (4), a data storage unit (5), and a discerning unit (6). The signal processing unit (4) performs STFT on waveform data of an input signal and calculates a feature amount. The data storage unit (5) stores feature amount data comprising normal, already-known waveform data. The discerning unit (6) makes a comparison between first feature amount data comprising a plurality of feature amounts calculated by the signal processing unit (4) and second feature amount data that is the feature amount data stored in the data storage unit (5), and discerns the quality of the waveform data of the input signal. The feature amount is the amount representing the degree of variation of spectrum intensity fluctuations over time in a specific frequency band in waveform data.

Description

[Technical Field]

[0001] The present disclosure relates to an abnormality diagnosis device, an abnormality diagnosis method, and an abnormality diagnosis program for diagnosing an abnormality in an electric device. [Background technology]

[0002] In a typical electrical equipment assembly process, an operational test of the electrical equipment is conducted during or after assembly is completed. Furthermore, during the operational test, a sensory test is conducted to check whether there are any abnormalities in the vibrations or operating sounds generated by the electrical equipment. Since such a sensory test is conducted using the worker's hearing or tactile sense, it is characterized by relying on the worker's senses. For this reason, in order to quantify the vibration or operating sound information used to determine whether the electrical equipment is good or bad, a diagnostic device may be used to convert the vibration or operating sound information acquired by a microphone, vibration sensor, etc. into waveform data and process the signal.

[0003] The following Patent Document 1 discloses a technology for diagnosing whether or not an abnormality exists in an electrical device by comparing a spectrum pattern obtained by performing a Fast Fourier Transform (FFT) process on waveform data of vibrations generated by the electrical device with a predetermined spectrum pattern at the time of a known abnormality, based on the comparison result. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Publication No. 62-93620 Summary of the Invention [Problem to be solved by the invention]

[0005] As described above, the technology of Patent Document 1 determines an abnormality using only the spectral pattern. Therefore, the technology of Patent Document 1 has a problem in that it is difficult to distinguish the type of abnormality when the spectral patterns of different types of abnormalities are the same, for example, when the spectral patterns are the same but the occurrence time intervals are different. Furthermore, while the technology of Patent Document 1 can distinguish known abnormalities in which the spectral pattern fluctuates over time with respect to waveform data, it has a problem in that it is difficult to distinguish unknown abnormalities that are different from normal and that an operator may sense.

[0006] The present disclosure has been made in view of the above, and aims to provide an abnormality diagnosis device that can distinguish between known and unknown abnormalities that are different from normal. [Means for solving the problem]

[0007] To solve the above-mentioned problems and achieve the object, an abnormality diagnosis device according to the present disclosure includes a microphone that converts the sound of an object to be discriminated into an analog electrical signal and a signal converter that converts the analog electrical signal into a digital signal. The signal processing device that takes in the digital signal and processes it includes a signal processing unit, a data storage unit, and a discrimination unit. The signal processing unit calculates features by performing a short-time fast Fourier transform on waveform data of an input signal. The data storage unit stores feature data based on known normal waveform data. The discrimination unit compares first feature data consisting of a plurality of features calculated by the signal processing unit with second feature data, which is feature data stored in the data storage unit, to discriminate whether the waveform data of the input signal is good or bad. The feature is a quantity that represents the degree of variation in spectral intensity fluctuation over time in a specific frequency band contained in the waveform data. [Effects of the Invention]

[0008] The abnormality diagnosis device according to the present disclosure has the effect of being able to distinguish between known and unknown abnormalities that are different from normal. [Brief explanation of the drawings]

[0009] [Figure 1] A block diagram showing a configuration example of an abnormality diagnosis device according to a first embodiment. [Figure 2] FIG. 1 is a diagram illustrating a filter bank process applicable to a signal processing device according to a first embodiment. [Figure 3] 1 is a flowchart illustrating a method for diagnosing an abnormality in an electrical device using the abnormality diagnosing device according to the first embodiment. [Figure 4] A block diagram showing a configuration example of an abnormality diagnosis device according to a second embodiment. [Figure 5] FIG. 10 is a block diagram illustrating an estimation processing algorithm used in the second embodiment. [Figure 6] A block diagram showing a configuration example of an abnormality diagnosis device according to a third embodiment. [Figure 7] A block diagram showing a configuration example of an abnormality diagnosis device according to a fourth embodiment. [Figure 8] A block diagram showing a configuration example of an abnormality diagnosis device according to a sixth embodiment. [Figure 9] FIG. 13 is a diagram illustrating a procedure for removing known noises using the abnormality diagnosis device according to the sixth embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0010] An abnormality diagnosis device, an abnormality diagnosis method, and an abnormality diagnosis program according to embodiments of the present disclosure will be described in detail below with reference to the accompanying drawings. Note that the embodiments described below are merely examples, and the scope of the present disclosure is not limited to the following embodiments.

[0011] Embodiment 1 1 is a block diagram showing an example of the configuration of an abnormality diagnosis device according to embodiment 1. The abnormality diagnosis device 100 according to embodiment 1 includes a microphone 2, a signal converter 3, and a signal processing device 10. The signal processing device 10 also includes a signal processing unit 4, a data storage unit 5, and a discrimination unit 6.

[0012] The microphone 2 converts sounds, including the operating sound of the object 1 to be discriminated, into analog electrical signals. The signal converter 3 is an analog-to-digital (AD) converter, and converts the analog electrical signals output from the microphone 2 into digital signals. The signal processing device 10 takes in the digital signals output from the signal converter 3 and performs signal processing, which will be described later. Note that the abnormality diagnosis device 100 may be provided with a plurality of signal processing devices 10.

[0013] In the signal processing device 10, the signal processing unit 4 performs necessary arithmetic processing on the waveform data of the input signal to calculate a feature amount. The feature amount is an amount that represents the degree of variation in the fluctuation of the spectral intensity over time in a specific frequency band contained in the waveform data, in other words, an amount that represents how much the fluctuation amount of the spectral intensity in a specific frequency band varies over time.

[0014] Feature data based on known normal waveform data is stored in the data storage unit 5. The discrimination unit 6 compares the feature data consisting of a plurality of feature amounts calculated by the signal processing unit 4 with the feature data based on known normal waveform data stored in the data storage unit 5, and determines whether the waveform data of the input signal input to the signal processing device 10 is good or bad.

[0015] In this paper, a data group consisting of multiple feature quantities calculated by the signal processing unit 4 may be referred to as "first feature quantity data," and feature quantity data based on normal, known waveform data stored in the data storage unit 5 or a collection of feature quantity data may be referred to as "second feature quantity data."

[0016] A general-purpose microphone can be used for the microphone 2, but it is desirable to minimize the inclusion of ambient noise and the like in the operating sounds generated by the object to be identified 1. This can be achieved by using a sound-collecting microphone with sound collection directionality, a parabolic sound-collecting microphone that combines a general microphone with a parabolic reflector, or an appropriate combination of these sound-collecting microphones.

[0017] A general audio interface can be used for the signal converter 3. Here, using an audio interface with a high sampling frequency is desirable because it allows a wide range to be set in the frequency direction when performing FFT processing in the signal processing described below. Note that if the signal processing device 10 has a function equivalent to that of the signal converter 3, that function may be used.

[0018] The signal processing device 10 may use a general computer or a programmable logic controller (PLC). A PLC is also called a sequencer. Alternatively, the signal processing device 10 may use an electronic board configured with a signal processing device and a storage medium. Examples of the signal processing device include a microcomputer and an FPGA (Field Programmable Gate Array), and examples of the storage medium include non-volatile or volatile semiconductor memory such as RAM (Random Access Memory), flash memory, EPROM (Erasable Programmable ROM), and EEPROM (Electrically EPROM). The signal processing device 10 can perform a series of operations from waveform data acquisition to classification according to procedural processing. The signal processing device 10 may also have a screen capable of instructing and displaying the classification status. In this case, the screen can be used to configure settings required for the classification process, and the status and results of the classification can be displayed on the screen to notify the operator.

[0019] Next, the algorithm of the discrimination process in the signal processing device 10 will be explained. The signal processing unit 4 performs a Short Time Fast Fourier Transform (STFT) on the waveform data converted by the signal converter 3, dividing the waveform data into an arbitrary number of parts in both the frequency direction and the time direction. In this paper, the result of the STFT process is expressed as a matrix having i frequency and j time dimensions as shown in the following equation (1): In F represent.

[0020]

number

[0021] In the above equation (1), i is the number of divisions of the waveform data in the frequency direction, and j is the number of divisions of the waveform data in the time direction.

[0022] Furthermore, based on the matrix F that is the result of the STFT processing, the signal processing unit 4 calculates a feature vector T expressed by the following equation (2), in which the standard deviation of the spectral intensity in the time direction in the matrix F is used as a feature.

[0023]

number

[0024] In the above formula (2), each element t1...t of the feature vector T i In this case, the feature value t at an arbitrary frequency i is calculated by the following equation (3).

[0025]

number

[0026] In the above formula (3), "f - " is written with " above the letter "f". - " is an alternative notation that means that the symbol "f - l " is the j elements f in the i row of the above formula (1) i,1 …f i,j is the average value of

[0027] As described above, the data storage unit 5 stores feature data based on normal known waveform data. The feature data stored in the data storage unit 5 is also calculated by STFT. Here, the number of normal known waveform data is k, and a set of feature data based on multiple normal known waveform data is called T. ’ k It is expressed as T ’1 is the feature vector obtained using the first waveform data, waveform data 1, and T ’ k is the feature vector obtained using the waveform data k, which is the kth waveform data. ’ k The data regarding the mean v and standard deviation σ of the waveform data are defined as the second feature data. The mean v and standard deviation σ can be expressed by the following equations (4) and (5). In this case, it is desirable that the number k of normal known waveform data is at least 100.

[0028]

number

number

[0029] In this paper, the mean v and standard deviation σ shown in the above equations (4) and (5) are examples of the second feature data, and statistics other than the mean v and standard deviation σ may be stored in the data storage unit 5 as the second feature data.

[0030] As described above, the discriminator 6 discriminates whether the waveform data of the input signal is good or bad. In this discrimination process, the discriminant index value Z i is used.

[0031]

number

[0032] Discrimination index value Z i is used to determine the degree of deviation between the first feature amount data and the second feature amount data. i The value of is compared with a predetermined threshold for each frequency band, and at least one or some of the discrimination index values ​​Z i If the threshold value is exceeded, the waveform data to be judged is judged to be abnormal or deviating from the normal state.

[0033] In the above STFT processing, the filter bank processing shown in FIG. 2 may be used in combination. FIG. 2 is a diagram for explaining the filter bank processing applicable to the signal processing apparatus according to Embodiment 1.

[0034] In the filter bank processing, for the matrix F which is the processing result of STFT, the number of divisions of the waveform data in the frequency direction is set to i ’ (i ’ (i), and the number of divisions of the waveform data in the time direction is set to j ’ (j ’ (j). Specifically, in FIG. 2, an example where i ’ = i / 2 and j ’ = j / 2 is shown.

[0035] In the left diagram of FIG. 2, for the four upper-left elements f 1,1 , f 1,2 , f 2,1 , f 2,2 of the matrix F, an operation of f 1,1 + f 1,2 + f 2,1 + f 2,2 , that is, an addition operation, is performed and this is set to f ’ [[ID=4​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​, f i-1,j , f i,j-1 , f i,j For these, f i-1,j-1 + f i-1,j + f i,j-1 + f i,j Perform the operation of this and make it f ’ i ’ ,j ’ Let it be. For the other elements of the matrix F, an addition operation is performed for every four elements. In this way, the newly recombined matrix F ’ shown in the right figure of FIG. 2 is generated.

[0036] Note that FIG. 2 is an example, and i ’ <i, j ’ <Any i that satisfies the relationship of j ’ , j ’ can be selected. Also, the division of the waveform data does not have to be at equal intervals. Therefore, the number of divisions i ’ does not have to be a divisor of the number of divisions i, and the number of divisions j ’ also does not have to be a divisor of the number of divisions j. Also, there is no dependency between the number of divisions in the frequency direction and the number of divisions in the time direction, and each can be performed independently.

[0037] By appropriately setting the number of divisions i ’ , j ’ in the recombined matrix F ’ , the characteristics of the sound of the discriminant object 1 can be emphasized. Therefore, by using the matrix F ’ recombined by the filter bank processing, it becomes possible to improve the accuracy of determining the quality of the waveform data.

[0038] FIG. 3 is a flowchart for explaining an abnormal diagnosis method of an electrical device performed using the abnormal diagnosis apparatus according to Embodiment 1. The processing flow of FIG. 3 can be executed using a computer, a PLC, or a processor provided with a storage unit. Hereinafter, these computer, PLC, or processor will be simply referred to as "computer".

[0039] The computer stores second feature data, i.e., feature data representing the degree of variation in the spectral intensity fluctuation over time in a specific frequency band possessed by normal known waveform data. The computer also stores a program that realizes the functions of the microphone 2, signal converter 3, and signal processing device 10 described above. Alternatively, the computer is configured to allow the program to be loaded from an external device and executed.

[0040] The computer converts the operating sound of the electrical device into an analog electrical signal (step S11). The computer converts the analog electrical signal converted in step S11 into a digital signal (step S12). The computer performs STFT on the waveform data converted in step S12 to calculate feature quantities (step S13).

[0041] The computer calculates a discrimination index value Z based on the first feature amount data consisting of the plurality of feature amounts calculated in step S13 and the second feature amount data stored in the computer. i is compared with a threshold value (step S14). A supplementary explanation of the processing in step S14 will be given below. As described above, the feature vector T representing the first feature data is expressed by the above formula (2), and each element of the feature vector T is expressed by the above formula (3). In addition, the discrimination index value Z i An example of this is expressed by the above formula (6).

[0042] The computer detects the discriminant index value Z that exceeds the threshold. i It is determined whether or not there is a discrimination index value Z that exceeds the threshold (step S15). i If there is no discrimination index value Z that exceeds the threshold value (step S15, No), the computer determines that the electrical device is normal (step S16). i If there is at least one (Yes in step S15), the computer determines that the electrical device is not normal or that there is a possibility of an abnormality (step S17).

[0043] By using the flowchart of FIG. 3, it is possible to determine whether the electrical device being evaluated is normal or abnormal. The computer displays the discrimination index value Z i It is desirable that the device has a function to plot and display the threshold values ​​on a graph. This function allows the operator to visually and easily be notified of the pass / fail judgment result.

[0044] As explained in the section [Problems to be Solved by the Invention], the conventional technology determines anomalies using only spectral patterns, which makes it difficult to distinguish between anomalies with identical spectral patterns but different occurrence time intervals. The conventional technology also has the problem of making it difficult to distinguish between unknown anomalies that are different from the usual anomalies that an operator might sense. In contrast, the anomaly diagnosis method according to embodiment 1 determines whether or not an anomaly exists in the waveform data to be determined based on the degree of deviation between the feature values ​​of the waveform data to be determined and the feature values ​​of known normal waveform data stored in a computer, thereby making it possible to determine whether an unknown anomaly is acceptable or not.

[0045] In the above description, it is assumed that the feature quantities of normal known waveform data are stored in the computer, but this is not limiting. The feature quantities of normal known waveform data only need to be accessible to the computer and only need to be stored in a storage unit or storage area that is not a component of the computer. In other words, it is only necessary for the computer to be configured to be able to reference the feature quantities of normal known waveform data.

[0046] As described above, the abnormality diagnosis device according to the first embodiment includes a microphone that converts the sound of an object to be identified into an analog electrical signal and a signal converter that converts the analog electrical signal into a digital signal. The signal processing device that receives and processes the digital signal includes a signal processing unit, a data storage unit, and a discrimination unit. The signal processing unit calculates features by performing a short-time fast Fourier transform on the waveform data of the input signal. The data storage unit stores feature data based on normal, known waveform data. The discrimination unit compares first feature data, consisting of a plurality of features calculated by the signal processing unit, with second feature data, which is feature data stored in the data storage unit, to determine whether the waveform data of the input signal is normal or abnormal. The feature data represents the degree of variation in the spectral intensity fluctuation over time in a specific frequency band contained in the waveform data. The abnormality diagnosis device configured in this manner determines whether the waveform data to be identified has an abnormality based on the degree of deviation between the feature data of the waveform data to be identified and the feature data of normal, known waveform data stored in the data storage unit. This makes it possible to discriminate between known and unknown abnormalities that are different from normal, such as those that cause spectral patterns to fluctuate over time.

[0047] In the above process, the signal processing device may perform filter bank processing on the feature generated by the short-time fast Fourier transform, and generate first feature data based on the feature that has undergone the filter bank processing. In this way, the time width or frequency width when obtaining the first feature data can be changed arbitrarily. This makes it possible to improve the accuracy of determining whether the waveform data is good or bad.

[0048] Furthermore, the abnormality diagnosis method according to the first embodiment is a method for diagnosing an abnormality in an electrical device using a computer configured to be able to reference a feature that represents the degree of variation in the temporal fluctuation of spectral intensity in a specific frequency band possessed by normal, known waveform data. This abnormality diagnosis method includes the following first to fourth steps. The first step is a step of converting the operating sound of the electrical device into an analog electrical signal. The second step is a step of converting the analog electrical signal into a digital signal. The third step is a step of calculating a feature by performing a short-time fast Fourier transform on the waveform data converted in the second step. The fourth step is a step of comparing first feature data consisting of a plurality of feature data calculated in the third step with second feature data stored in the computer to determine whether the waveform data is good or bad. By having a computer execute these first to fourth steps, it becomes possible to determine known and unknown abnormalities that are different from normal, such as those in which the spectral pattern fluctuates over time.

[0049] Furthermore, the abnormality diagnosis program according to the first embodiment is a program for causing a computer configured to be able to refer to features that represent the degree of variation in spectral intensity fluctuations over time in a specific frequency band possessed by normal, known waveform data to diagnose an abnormality in an electrical device. This abnormality diagnosis program includes the following first to fourth steps. The first step is a step of converting the operating sound of the electrical device into an analog electrical signal. The second step is a step of converting the analog electrical signal into a digital signal. The third step is a step of calculating features by performing a short-time fast Fourier transform on the waveform data converted in the second step. The fourth step is a step of comparing first feature data consisting of a plurality of features calculated in the third step with second feature data stored in the computer to determine whether the waveform data is good or bad. By having a computer execute a program including these first to fourth steps, it becomes possible to determine known and unknown abnormalities that are different from normal, such as those in which spectral patterns fluctuate over time.

[0050] Embodiment 2 In the configuration of the abnormality diagnosis device according to the first embodiment shown in Fig. 1, there is a risk that noise from the surroundings may be mixed in, leading to a mistaken discrimination between noise and abnormal sound. In the second embodiment, a configuration that improves this issue is proposed.

[0051] FIG. 4 is a block diagram showing an example of the configuration of an abnormality diagnosis device according to embodiment 2. In FIG. 4, an abnormality diagnosis device 100A according to embodiment 2 adds a noise collecting microphone 11 for collecting ambient noise to the configuration of the abnormality diagnosis device 100 according to embodiment 1 shown in FIG. 1. The other configuration is the same as or equivalent to that in FIG. 1, and the same or equivalent components are given the same reference numerals, and redundant explanations will be omitted. Note that FIG. 4 illustrates an example in which there are three noise collecting microphones 11, but this example is not limiting. The number of noise collecting microphones 11 may be one, or may be two or more than four.

[0052] Next, the operation of the abnormality diagnosis device according to the second embodiment will be described. As mentioned above, in the configuration of the first embodiment, there is a risk that ambient noise may be mixed in, resulting in an erroneous discrimination between noise and abnormal sound. Therefore, in the second embodiment, the time period in which ambient noise is mixed in is estimated using waveform data collected using the noise collection microphone 11. Then, the method of the first embodiment is applied to the waveform data excluding the estimated time period. In this way, it is possible to provide a discrimination method that reduces the influence of ambient noise.

[0053] Next, an estimation processing algorithm used in the second embodiment will be described with reference to Fig. 5. Fig. 5 is a block diagram for explaining the estimation processing algorithm used in the second embodiment. Fig. 5 shows adders 21 and 23 and a filter 22 as control elements that realize the technique of the second embodiment.

[0054] First, let us assume that the matrix F shown in the above equation (1) and the feature vector T shown in the above equation (2) are obtained from the waveform data of the audio signals collected by microphone 2 and noise collecting microphone 11. In this case, let us denote the feature vector obtained from microphone 2 at time t as s(t), and the feature vectors obtained from the three noise collecting microphones 11 as n1(t), n2(t), and n3(t), respectively.

[0055] 5, the filter 22 is only required to reproduce the degree of attenuation of the value of each element of the matrix F. Here, the constant vector representing the characteristics of the filter 22 is represented by w, and the estimated value of the noise included in the feature vector s(t) observed by the microphone 2 is represented by d. ~ (t) is expressed as "d ~ " is a "d" with " above it. ~ This is an alternative notation that means that the symbol " is attached. In this case, the estimated noise value d ~ (t) can be expressed by the following equation (7).

[0056]

number

[0057] In the above formula (7), the symbol " ○ " is an operator representing a multiplication process for each element of a vector. Note that the constant vector w representing the characteristics of the filter 22 is a vector whose elements are constants obtained by appropriately calculating the attenuation rate of each element. Any method may be used to calculate the attenuation rate of each element.

[0058] According to the above equation (7), the noise feature amount is calculated by adding up the sound pressure levels of the three noise collecting microphones 11 and multiplying the result by a constant. Therefore, the period during which the noise feature amount calculated by the above equation (7) exceeds a predetermined threshold can be estimated as the period during which noise is present. Therefore, by excluding the estimated period, i.e., the period during which noise is present, and performing judgment based on the above algorithm, it becomes possible to judge the quality of waveform data with the influence of the mixed noise reduced.

[0059] Furthermore, let us consider the characteristics of the filter 22. First, the characteristics of the filter 22 are good when the actual noise d(t) observed by the microphone 2 and its estimated value d ~ This is considered to be equivalent to a reduction in the error with (t). To consider this error, we define the observed signal matrix N expressed by the following equation (8).

[0060]

number

[0061] In the above equation (8), the symbol "T" represents transposition, and L represents the number of frames in the observed signal matrix N. The number of frames L corresponds to the number of short-time frames in the STFT.

[0062] Also, a noise matrix D is defined, which represents the actual noise observed by the microphone 2, and is expressed by the following equation (9).

[0063]

number

[0064] Here, the operation of extracting the i-th column of the observed signal matrix N shown in the above equation (8) and generating a column vector is called N i The operation of extracting the i-th column of the noise matrix D shown in equation (9) above and generating a column vector is denoted as D i The operation to extract the i-th value of the constant vector w is expressed as w i In this case, if the squared error when focusing on the i-th frequency in the matrix F representing the STFT processing result is represented as e, this squared error e can be expressed as in the following equation (10).

[0065]

number

[0066] Minimize the above equation (10) w i This w i is expressed as w i Differentiated by (de / w i ) becomes zero i , i.e. de / w i = 0 satisfies w i and is expressed by the following equation (11).

[0067]

number

[0068] By performing the calculation of equation (11) above for all i, a constant vector w having optimal parameters can be obtained. The calculation process described above requires information on the noise d(t) collected by microphone 2 and information on the feature vectors n1(t), n2(t), and n3(t) collected by noise-collecting microphone 11. However, alternative methods can be used. An example of an alternative method is to use data recorded in a situation where only noise is present and no operational noise of the measurement target is present. This method may use data recorded from only noise in an actual noisy environment, or data recorded in a laboratory by playing a signal equivalent to noise through a speaker. An example of the latter method is a test signal that has energy in all frequency bands and fluctuates over time. An example of this test signal is the TSP signal used in the TSP (Time-Stretched-Pulse) method.

[0069] As described above, the abnormality diagnosis device according to the second embodiment further includes a noise collection microphone for collecting ambient noise. The signal processing device uses waveform data collected by the noise collection microphone to estimate a time period in which ambient noise is present, and calculates features by performing a short-time fast Fourier transform on the waveform data excluding the estimated time period. This makes it possible to avoid erroneously determining that noise contained in the sound of the object to be identified is abnormal.

[0070] Furthermore, in the anomaly diagnosis method according to the second embodiment, the anomaly diagnosis method described in the first embodiment includes, between the second and third steps, an estimation step for estimating a time period in which ambient noise is present, using waveform data collected using a noise collection microphone for collecting ambient noise. Then, in the third step, a feature quantity is calculated by performing a short-time fast Fourier transform on the waveform data excluding the time period estimated in the estimation step. This provides a discrimination method that reduces the influence of ambient noise. This makes it possible to avoid erroneously determining that noise contained in the sound of the object being discriminated is an anomaly.

[0071] Embodiment 3 Fig. 6 is a block diagram showing an example of the configuration of an abnormality diagnosis device according to embodiment 3. In Fig. 6, abnormality diagnosis device 100B according to embodiment 3 is configured by adding a soundproof wall 12 for reducing ambient noise to the configuration of abnormality diagnosis device 100 according to embodiment 1 shown in Fig. 1. Soundproof wall 12 is arranged so as to surround object 1 to be discriminated. The other configuration is the same as or equivalent to that in Fig. 1, and the same or equivalent components are designated by the same reference numerals, and redundant explanations will be omitted.

[0072] 6, a general sound-absorbing material, a metal plate, or a combination thereof can be used as the soundproof wall 12. It is desirable to change the thickness of the soundproof wall 12 according to the level of the surrounding noise.

[0073] In the third embodiment, the method of determining whether or not there is an abnormality in the waveform data is basically the same as in the first embodiment. However, with the configuration of the first embodiment, there is a risk of noise being mixed in from the surroundings. Furthermore, with the configuration of the second embodiment, there is a risk of misidentifying noise as an abnormal sound in an environment where the surrounding noise cannot be completely removed or where noise is constantly mixed in. In contrast, in the third embodiment, the surrounding noise can be attenuated by the soundproof wall 12. Although the method of the third embodiment is a simple method, it can provide a determination method that reduces the influence of surrounding noise.

[0074] As described above, the abnormality diagnosis device according to the third embodiment includes a soundproof wall arranged to surround the object to be discriminated. The soundproof wall attenuates surrounding noise, thereby providing a discrimination method that reduces the influence of surrounding noise. This makes it possible to avoid misjudging noise contained in the sound of the object to be discriminated as an abnormality.

[0075] Embodiment 4 Fig. 7 is a block diagram showing an example of the configuration of an abnormality diagnosis device according to embodiment 4. In Fig. 7, abnormality diagnosis device 100C according to embodiment 4 adds a movable unit 13 for moving microphone 2 and a movable control unit 14 for controlling the operation of movable unit 13 to the configuration of abnormality diagnosis device 100 according to embodiment 1 shown in Fig. 1. The other configuration is the same as or equivalent to that in Fig. 1, and the same or equivalent components are designated by the same reference numerals, and redundant explanations will be omitted.

[0076] Here, an articulated robot, a linear robot, an air cylinder, a hydraulic cylinder, or the like can be used for the movable part 13. Also, a computer, a sequencer, or the like can be used for the movement control part 14. The movement control part 14 may be equipped with a position detection sensor, a camera, or the like for determining the movement of the movable part 13.

[0077] In the fourth embodiment, the method of determining whether or not there is an abnormality in the waveform data is basically the same as in the first embodiment. However, in the first embodiment, if the object 1 to be determined moves, for example, and the microphone 2 is fixed, the distance between the object 1 to be determined and the microphone 2 changes, making it difficult to stably measure the sound generated from the object 1 to be determined. An example of a case in which the object 1 to be determined moves is when the object 1 to be determined is placed on a conveyor belt and transported along a production line.

[0078] In contrast to this, in the fourth embodiment, the movable control unit 14 can operate the movable unit 13 in accordance with the movement of the object to be distinguished 1. This can reduce or eliminate the change in the distance between the object to be distinguished 1 and the microphone 2, thereby enabling stable measurement of the sound generated from the object to be distinguished 1.

[0079] Furthermore, if the structure of the object to be distinguished 1 is complex, the distance between the object to be distinguished 1 and the microphone 2 increases, which may result in insufficient volume. Insufficient volume means that waveform data with sufficient amplitude cannot be obtained. If waveform data with sufficient amplitude cannot be obtained, the waveform data may not contain sounds necessary for distinction, leading to incorrect distinction. When a worker distinguishes between objects, he or she must place his or her ear close to the operation sound generating portion of the object to be distinguished 1 to check. However, with the configuration of embodiment 1, it is difficult to bring the microphone 2 close to the operation sound generating portion of the object to be distinguished 1. In contrast, in embodiment 4, the movable control unit 14 can operate the movable unit 13 so that the microphone 2 is close enough to the operation sound generating portion of the object to be distinguished 1. This allows the microphone 2 to be brought close enough to a position where sufficient amplitude can be obtained for measurement. This allows waveform data with sufficient amplitude to be obtained, enabling accurate determination of the quality of the waveform data.

[0080] As described above, the abnormality diagnosis device according to the fourth embodiment includes a movable unit for moving the microphone and a movable control unit for controlling the operation of the movable unit. This configuration allows the microphone to be brought closer to the operation sound generating unit in the object to be discriminated. This makes it possible to obtain waveform data with sufficient amplitude, and to accurately determine whether the waveform data is good or bad.

[0081] Embodiment 5. In the configuration of the abnormality diagnosis device according to the first embodiment shown in FIG. 1, there is a risk that ambient noise may be mixed in, resulting in a misidentification between noise and abnormal sound. Furthermore, in the configuration of the second embodiment, there is a risk that ambient noise may not be completely removed or that abnormal sound may be misidentified in an environment where noise is constantly present. Furthermore, in the method of the second embodiment, if the operating sound generated by the object to be discriminated 1 is loud, the operating sound may also be input to the noise collecting microphone 11, making it impossible to estimate the time period in which noise is mixed in. Therefore, in the fifth embodiment, a method that improves on these issues is proposed. The method according to the fifth embodiment can be implemented using the configuration of the abnormality diagnosis device 100 according to the first embodiment shown in FIG. 1.

[0082] Next, the algorithm of the estimation process used in the fifth embodiment will be described. First, the signal processing unit 4 performs STFT on the waveform data converted by the signal converter 3, dividing the waveform data into an arbitrary number of parts in both the frequency direction and the time direction. This process is the same as in the first embodiment. The matrix F, which is the result of the STFT process, is expressed by the following equation (12) again.

[0083]

number

[0084] In the above equation (12), i is the number of divisions of the waveform data in the frequency direction, and j is the number of divisions of the waveform data in the time direction.

[0085] The signal processing unit 4 calculates a matrix R expressed by the following equation (13) for the matrix F that is the result of the STFT processing.

[0086]

number

[0087] In the above equation (13), ε is a matrix including elements of the vector of the average ν shown in the above equation (4), and is expressed by the following equation (14).

[0088]

number

[0089] As shown in the above equation (14), the matrix ε is a diagonal matrix, and its diagonal elements are expressed by the following equation (15).

[0090]

number

[0091] As shown in the above equation (15), the diagonal element e of the matrix ε i,i is the average ν i It is expressed as the reciprocal of

[0092] The matrix R is calculated as information related to the frequency of noise of a known normal product. As shown in the above equations (13) to (15), the matrix R is expressed as the product of the matrix F and the matrix ε having the reciprocal of the average ν, and is calculated as an estimate of the non-stationary noise. Note that if a sufficient number of averages ν are not available, a unit matrix may be used instead of the matrix ε.

[0093] The signal processing unit 4 calculates the sum per time for each element of the matrix R that represents the estimated value of the non-stationary noise. ’ The sum vector R ’ can be expressed as the following equation (16).

[0094]

number

[0095] The signal processor 4 calculates the sum vector R ’ Calculate the overall average value and sum vector R ’Among the elements, elements that exceed the overall average value are identified. Then, the signal processing unit 4 estimates the interval to which the element that exceeds the overall average value belongs as a time interval containing non-stationary noise, i.e., a time interval containing noise other than operation noise. Then, the methods of the first and second embodiments are applied to the waveform data excluding the estimated time interval. In this way, a discrimination method that reduces the influence of ambient noise can be provided.

[0096] In the above, the sum vector R ’ The overall average value is used as the reference value for comparison, i.e., the threshold value, but is not limited to this. ’ The threshold value may be the overall average value multiplied by an arbitrary coefficient.

[0097] As described above, according to the anomaly diagnosis device of embodiment 5, the signal processing device uses waveform data of the operation sound of the object to be distinguished to estimate a time period in which noise other than the operation sound is mixed in, and calculates feature quantities by performing a short-time fast Fourier transform on the waveform data excluding the estimated time period. This makes it possible to avoid erroneously determining that noise included in the sound of the object to be distinguished is abnormal.

[0098] Furthermore, in the anomaly diagnosis method according to the fifth embodiment, an estimation step is added between the second and third steps of the anomaly diagnosis method described in the first embodiment, in which waveform data of the operating sound of an electrical device is used to estimate a time period in which noise other than the operating sound is present. Then, in the third step, a feature quantity is calculated by performing a short-time fast Fourier transform on the waveform data excluding the time period estimated in the estimation step. This provides a discrimination method that reduces the influence of ambient noise. This makes it possible to avoid erroneously determining that noise contained in the sound of the object being discriminated is an anomaly.

[0099] Embodiment 6 In the configuration of the abnormality diagnosis device 100 according to the first embodiment shown in FIG. 1, there is a risk of misidentifying noise from an abnormal sound due to the presence of known ambient noise, such as chimes announcing the time or melodies generated by an AGV (Automated Guided Vehicle) traveling within a factory. Furthermore, in the configuration of the second embodiment, there is a risk of misidentifying noise from an abnormal sound in an environment where ambient noise cannot be completely removed or where noise is constantly present. Furthermore, in the method of the second embodiment, if the operating noise generated by the object to be discriminated 1 is loud, the operating noise may also be input to the noise collection microphone 11, making it impossible to estimate the time period in which noise is present. Therefore, in the sixth embodiment, a method that improves these points is proposed. The method according to the sixth embodiment can be implemented by using an abnormality diagnosis device 100D, as shown in FIG. 8, which includes a signal processing device 10D, in addition to the configuration of the abnormality diagnosis device 100 according to the first embodiment shown in FIG. 1. FIG. 8 is a block diagram showing an example of the configuration of the abnormality diagnosis device 100D according to the sixth embodiment. The signal processing device 10D included in the abnormality diagnosis device 100D further includes a preprocessing unit 30 in addition to the configuration of the signal processing device 10 shown in Fig. 1. Furthermore, the data storage unit 5 of the signal processing device 10D stores feature amount data based on known normal waveform data and known noise data.

[0100] Next, the algorithm of the estimation process used in the sixth embodiment will be described. First, the preprocessing unit 30 performs STFT on the waveform data converted by the signal converter 3 and the known noise data, dividing the waveform data and the noise data into an arbitrary number of parts in both the frequency direction and the time direction. The preprocessing unit 30 also performs pattern matching between matrices resulting from the respective STFT processes. A general algorithm can be used for pattern matching.

[0101] Here, the matrix representing the STFT result of the waveform data converted by the signal converter 3 is represented by I and is called the "first signal matrix." Also, the matrix representing the STFT result of the known noise data is represented by T and is called the "first noise matrix." When the magnitudes of the first noise matrix T in the time direction and frequency direction are M and N, respectively, pattern matching uses R shown in the following equation (17): NCC (Normalized Cross-Correlation) can be used.

[0102]

number

[0103] Normalized cross-correlation R obtained by pattern matching NCC At a certain index (i,j), a certain threshold Th Near If the difference exceeds the normalized cross-correlation R NCC The index at which the value of is maximum is set as the start time, and the data from this start time to the magnitude of the first noise matrix T in the time direction is determined to be a section containing known noise.

[0104] Next, a procedure for removing a known noise will be described with reference to Fig. 9. Fig. 9 is a diagram for explaining a procedure for removing a known noise that is performed using an abnormality check device 100D according to the sixth embodiment.

[0105] The preprocessing unit 30 determines whether or not known noise data is included in the waveform data converted by the signal converter 3. In the example of Fig. 9, the portion indicated by the thick solid line frame is determined to be a section containing known noise.

[0106] If it is determined that there is a section containing a known noise, the preprocessing unit 30 calculates the value of the first noise matrix T by subtracting it from the replacement threshold value Th pow The value of the element that exceeds the noise removal variable P is calculated by pattern matching. rep, and a second noise matrix T' is generated by replacing some elements of the STFT result of the known noise data. In the example of FIG. 9, the threshold value Th pow = 100, and the values ​​of "150", "200", and "450" ​​that exceed this value of 100 in the first noise matrix T are the noise elimination variables P rep Here, the noise removal variable P rep , the average value or an uncalculated value of the part excluding the data of the section where known noise is mixed in the first signal matrix I can be used. In Fig. 9, the average value for each frequency of the part excluding the data of the section where known noise is mixed in the first signal matrix I is used as the noise removal variable P rep An example is given in which:

[0107] Finally, the pre-processing unit 30 generates a second signal matrix I' from which the influence of the known noise has been removed by replacing the data of the portion of the first signal matrix I where the known noise is present with the second noise matrix T'. The subsequent processing uses this second signal matrix I' to perform any of the processing of the first to fifth embodiments described above.

[0108] In the sixth embodiment, the configuration in which the pre-processing unit 30 is added to the front stage of the signal processing unit 4 is applied to the configuration of the abnormality diagnosis device 100 according to the first embodiment shown in FIG. 1, but it is also possible to apply this to the configuration of the second embodiment shown in FIG. 4, the configuration of the third embodiment shown in FIG. 6, and the configuration of the fourth embodiment shown in FIG. 7.

[0109] As described above, according to the anomaly diagnosis device of embodiment 6, the signal processing device includes a preprocessing unit that uses known noise data to estimate a time period in which the known noise data is mixed in from the result of a short-time fast Fourier transform, and the signal processing unit performs processing to calculate features from data excluding the noise data. This makes it possible to avoid erroneously determining that the known noise included in the sound of the object to be discriminated is abnormal.

[0110] In the abnormality diagnosis device according to the sixth embodiment, the preprocessing unit performs a short-time fast Fourier transform on the waveform data converted by the signal converter and known noise data, dividing the waveform data and the noise data into an arbitrary number of parts in both the frequency and time directions, and performs pattern matching between the matrices resulting from the short-time fast Fourier transforms. Furthermore, the signal processing unit calculates first feature data using data generated based on the pattern matching results of the preprocessing unit. This makes it possible to provide an abnormality diagnosis device that reduces the influence of ambient noise.

[0111] Furthermore, in the anomaly diagnosis method according to the sixth embodiment, the anomaly diagnosis methods described in the first to fifth embodiments include an estimation step between the second and third steps, in which a time period in which known noise data is mixed in is estimated from the results of a short-time fast Fourier transform. Then, in the third step, the data for the time period estimated in the estimation step is replaced with other data, and a short-time fast Fourier transform is performed on the waveform data to calculate feature quantities. This provides a discrimination method that reduces the influence of ambient noise. This makes it possible to avoid erroneously determining that noise contained in the sound of the object being discriminated is abnormal.

[0112] The configurations shown in the above embodiments are merely examples, and may be combined with other known technologies, or different embodiments may be combined with each other. It is also possible to omit or modify parts of the configurations as long as they do not deviate from the gist of the invention. [Explanation of symbols]

[0113] 1 Object to be discriminated, 2 Microphone, 3 Signal converter, 4 Signal processing unit, 5 Data storage unit, 6 Discrimination unit, 10, 10D Signal processing device, 11 Noise collection microphone, 12 Soundproof wall, 13 Movable unit, 14 Movable control unit, 21, 23 Adder, 22 Filter, 30 Pre-processing unit, 100, 100A, 100B, 100C, 100D Abnormality diagnosis device.

Claims

1. A microphone that converts the sound of the object to be identified into an analog electrical signal, a signal converter for converting the analog electrical signal into a digital signal; a signal processing device that receives the digital signal and processes the signal; Equipped with The signal processing device includes: a signal processing unit that performs a short-time fast Fourier transform on waveform data of an input signal to calculate a feature quantity; a data storage unit in which feature data based on normal known waveform data is stored; a discrimination unit that compares first feature data, which is calculated by the signal processing unit and comprises a plurality of feature amounts representing the degree of variation in spectral intensity fluctuation over time in a specific frequency band contained in the waveform data, with second feature data, which is feature data stored in the data storage unit, to determine whether the waveform data of the input signal is good or bad; Equipped with the signal processing unit performs a short-time fast Fourier transform on the waveform data converted by the signal converter to calculate a matrix F having i frequency dimensions and j time dimensions expressed by the following equation (1), and calculates a feature vector T expressed by the following equation (2) as the first feature data, the feature vector T having a standard deviation of spectral intensity in the time direction in the matrix F as a feature; The data storage unit stores data on a mean v and a standard deviation σ calculated based on T′ k , which is a collection of features based on a plurality of normal known waveform data, and which are expressed by the following equations (3) and (4): The determination unit determines whether or not there is an abnormality in the waveform data of the input signal based on a degree of deviation between the first feature amount data and the second feature amount data. An abnormality diagnosis device characterized by: [Equation 1] [Equation 2] [Equation 3] [Equation 4]

2. Equipped with a noise collection microphone for collecting ambient noise, The signal processing device estimates a time period in which ambient noise is present using waveform data collected by the noise collecting microphone, and calculates a feature amount by performing a short-time fast Fourier transform on the waveform data excluding the estimated time period.

2. The abnormality diagnosis device according to claim 1.

3. The signal processing device estimates a time period in which noise other than the operation sound is mixed in using waveform data of the operation sound of the object to be distinguished, and calculates a feature amount by performing a short-time fast Fourier transform on the waveform data excluding the estimated time period.

2. The abnormality diagnosis device according to claim 1.

4. The signal processing device includes a pre-processing unit that uses known noise data to estimate a time period in which the known noise data is mixed from a result of a short-time fast Fourier transform, and the signal processing unit performs a short-time fast Fourier transform on waveform data in which data in the time period in which the noise data is mixed is replaced with other data, and calculates a feature amount.

2. The abnormality diagnosis device according to claim 1.

5. the preprocessing unit performs a short-time fast Fourier transform on the waveform data converted by the signal converter and known noise data, dividing the waveform data and the noise data into an arbitrary number of parts in both the frequency direction and the time direction, and performs pattern matching between matrices resulting from the respective short-time fast Fourier transform processes; The signal processing unit calculates the first feature amount data using data generated based on the pattern matching processing result by the preprocessing unit.

5. The abnormality diagnosis device according to claim 4.

6. When a matrix representing the result of a short-time fast Fourier transform of the waveform data is defined as a first signal matrix, and a matrix representing the result of a short-time fast Fourier transform of the known noise data is defined as a first noise matrix, the preprocessing unit determines that there is a section containing known noise when the waveform data contains the known noise data, generates a second noise matrix by replacing element values ​​of the first noise matrix whose values ​​exceed a first threshold with noise removal variables obtained by pattern matching, and generates a second signal matrix by replacing data of a section containing known noise in the first signal matrix with data of the second noise matrix; The signal processing unit calculates the first feature amount data using the second signal matrix.

6. The abnormality diagnosis device according to claim 5.

7. The preprocessing unit uses normalized cross-correlation for pattern matching between the first signal matrix and the first noise matrix, and when the normalized cross-correlation exceeds a second threshold at a certain index, determines that the index in the first signal matrix at which the value of the normalized cross-correlation is maximum is a start time, and determines that the data from this start time to the magnitude of the first noise matrix in the time direction is a section containing known noise.

7. The abnormality diagnosis device according to claim 6.

8. a soundproof wall disposed so as to surround the object to be discriminated; 2. The abnormality diagnosis device according to claim 1.

9. The signal processing device performs filter bank processing on the feature quantity generated by the short-time fast Fourier transform, and generates the first feature quantity data based on the feature quantity subjected to the filter bank processing.

9. The abnormality diagnosis device according to claim 2, wherein the abnormality diagnosis device is a device for detecting an abnormality.

10. a movable part for moving the microphone; a movement control unit that controls the operation of the moving unit; 9. The abnormality diagnosis device according to claim 2, further comprising:

11. 1. An abnormality diagnosis method for diagnosing an abnormality in an electrical device using a computer configured to be able to refer to a feature quantity that represents a degree of variation in spectral intensity fluctuation over time in a specific frequency band that is contained in normal known waveform data, the method comprising: a first step of converting the operating sound of the electrical device into an analog electrical signal; a second step of converting the analog electrical signal into a digital signal; a third step of performing a short-time fast Fourier transform on the waveform data converted in the second step to calculate a matrix F having i frequency and j time dimensions expressed by the following formula (1), and calculating a feature vector T expressed by the following formula (2) as first feature data, the feature vector T being the standard deviation of the spectral intensity in the time direction in the matrix F; a fourth step of comparing first feature data consisting of a plurality of feature values ​​calculated in the third step with second feature data, which is data relating to a mean v and a standard deviation σ shown in the following equations (3) and (4) and which is calculated based on T' k , a collection of feature values ​​based on a plurality of known normal waveform data, and which is stored in the computer, to determine whether the waveform data is good or bad; An abnormality diagnosis method comprising: [Equation 5] [0060] [Equation 7] [Equation 8]

12. an estimation step between the second step and the third step of estimating a time period in which ambient noise is present using waveform data collected by a noise collection microphone for collecting ambient noise; In the third step, a short-time fast Fourier transform is performed on the waveform data excluding the time interval estimated in the estimation step. The abnormality diagnosis method according to claim 11 .

13. an estimation step is included between the second step and the third step, in which a time period in which noise other than the operation sound is mixed in is estimated using waveform data of the operation sound of the electrical device; In the third step, a short-time fast Fourier transform is performed on the waveform data excluding the time interval estimated in the estimation step. The abnormality diagnosis method according to claim 11 .

14. an estimation step is included between the second step and the third step, in which a time interval in which known noise data is mixed is estimated from a result of a short-time fast Fourier transform; In the third step, a short-time fast Fourier transform is performed on waveform data obtained by replacing the data in the time interval estimated in the estimation step with other data. The abnormality diagnosis method according to claim 11 .

15. 1. An abnormality diagnosis program for causing a computer configured to be able to refer to a feature quantity representing a degree of variation in spectral intensity fluctuation over time in a specific frequency band contained in normal known waveform data to diagnose an abnormality in an electrical device, the program comprising: a first step of converting the operating sound of the electrical device into an analog electrical signal; a second step of converting the analog electrical signal into a digital signal; a third step of performing a short-time fast Fourier transform on the waveform data converted in the second step to calculate a matrix F having i frequency and j time dimensions expressed by the following formula (1), and calculating a feature vector T expressed by the following formula (2) as first feature data, the feature vector T being the standard deviation of the spectral intensity in the time direction in the matrix F; a fourth step of comparing first feature data consisting of a plurality of feature values ​​calculated in the third step with second feature data, which is data relating to a mean v and a standard deviation σ shown in the following equations (3) and (4) and which is calculated based on T' k , a collection of feature values ​​based on a plurality of known normal waveform data, and which is stored in the computer, to determine whether the waveform data is good or bad; causing the computer to execute a process including An abnormality diagnosis program characterized by: [Equation 9] [Equation 10] [0011] [0012]

Images