WO2016125256A1

WO2016125256A1 - Abnormal sound diagnosis device, abnormal sound diagnosis system, abnormal sound diagnosis method, and abnormal sound diagnosis program

Info

Publication number: WO2016125256A1
Application number: PCT/JP2015/052991
Authority: WO
Inventors: 阿部　芳春; 寛福永
Original assignee: 三菱電機株式会社; 三菱電機ビルテクノサービス株式会社
Priority date: 2015-02-03
Filing date: 2015-02-03
Publication date: 2016-08-11
Also published as: KR20170108085A; CN107209509A; JPWO2016125256A1; KR101962558B1; CN107209509B; DE112015006099T5; JP6250198B2

Abstract

An abnormal sound diagnosis device is provided with: a trajectory feature extraction unit (5) for extracting a trajectory vector (15) by converting, into a vector, a trajectory that indicates an intensity feature of an intensity time series (14) in an entire time direction obtained by an intensity time series obtaining unit (4); a discrimination parameter storage unit (6) for storing a discrimination parameter (16) learned with the vector which is the trajectory indicating the intensity feature in the entire time direction of the intensity time series of sound data generated from a reference device as an input and with information indicating a status type of a diagnosis target device as an output; a discrimination unit (7) for obtaining a K-dimensional score vector (17) with respect to each status type of the diagnosis target device from the trajectory vector (15) and the discrimination parameter (16); and a determination unit (8) for determining, by referring to the K-dimensional score vector (17), whether sound generated in the diagnosis target device is normal or abnormal and determining the kind of the abnormality.

Description

Abnormal sound diagnosis apparatus, abnormal sound diagnosis system, abnormal sound diagnosis method, and abnormal sound diagnosis program

The present invention is an abnormal sound diagnosis device that analyzes sound generated from a device to be diagnosed and diagnoses the generation of abnormal sound of the device and the type of abnormal sound that has occurred, and the device operates normally The present invention relates to an abnormal sound diagnosis apparatus, an abnormal sound diagnosis system, an abnormal sound diagnosis method, and an abnormal sound diagnosis program that do not require sound collection.

Conventionally, as an abnormal sound diagnosis device, the analysis result of the sound data collected while the device to be diagnosed is operating normally is stored as a reference value, and the analysis result of the sound data collected at the time of diagnosis is stored. A device that diagnoses that an abnormality has occurred in a device when it deviates from the reference value is known.
For example, the abnormal sound detection device disclosed in Patent Document 1 detects and stores the frequency band of the sound collected when the elevator is operating normally, and collects the sound during diagnostic operation. The presence or absence of an abnormal sound is diagnosed by excluding the sound in the frequency band stored from the sound.
Further, the abnormal sound diagnosis apparatus disclosed in Patent Document 2 acquires a normal time frequency distribution which is a reference at the time of diagnosis, and the normal time frequency distribution and the time time frequency part distribution acquired in the diagnosis mode. And the degree of abnormality is calculated, and it is determined whether an abnormality has occurred by comparing the calculated degree of abnormality with a threshold value.

JP 2012-166935 A JP 2013-200133 A

However, in the techniques of Patent Document 1 and Patent Document 2 described above, in order to diagnose a device, a sound collector is installed at the same position as that at the time of diagnosis in a device that is in a normally operating state, and is the same as at the time of diagnosis. It has been necessary to learn a reference for diagnosing the presence or absence of an abnormal sound in advance by collecting and analyzing the sound generated when the operation is performed by a sound collector.
Therefore, if the sound during normal operation cannot be collected before diagnosing the equipment, for example, in the case of an existing elevator contracted in the middle, a reference for diagnosis cannot be created and abnormal sound There was a problem that the diagnostic device could not be applied.

As mentioned above, if the sound during normal operation cannot be collected and a reference for diagnosis cannot be created, the sound during normal operation can be obtained using another device with the same specifications. It is also possible to create a reference for diagnosis by collecting sound. However, in the case of complex equipment consisting of many parts, specifications such as the location of the sound collector, the size of the parts that make up the equipment, and the placement conditions of the equipment, for example, if the equipment is an elevator, In addition, it is not practical from a cost standpoint to prepare equipment that has the same specifications such as hoistway size, hoistway material, basket loading capacity, and operating speed. There is a problem that it is difficult to create an appropriate standard using.

The present invention has been made to solve the above-described problems, and diagnoses the operating state of a device without requiring sound collection during normal operation in advance for the device to be diagnosed. For the purpose.

An abnormal sound diagnosis apparatus according to the present invention collects sound generated by a diagnosis target device and acquires sound data, and an intensity time series from time frequency distribution obtained by analyzing waveform data of the sound data. An intensity time series acquisition unit for acquiring the trajectory, a trajectory feature extraction unit for extracting a trajectory vector by converting a trajectory showing intensity features in all time directions of the intensity time series, and waveform data of sound data generated from a reference device Identification that is learned by using as input a vector that is a trajectory indicating intensity characteristics in all time directions of intensity time series obtained from the time-frequency distribution obtained by analyzing An identification parameter storage unit that stores parameters, an identification unit that acquires a score for each state type of the diagnosis target device from the trajectory vector and the identification parameter, and a diagnosis pair with reference to the score Sound generated in the device is one that includes a a determination unit or abnormal either, and abnormality type is normal.

According to the present invention, it is possible to diagnose the presence or absence of abnormal sound even for devices that cannot collect sound during normal operation in advance and create a standard for diagnosis.

1 is a block diagram illustrating a configuration of an abnormal sound diagnosis apparatus according to Embodiment 1. FIG. It is a figure which shows the structure of the identification part of the abnormal sound diagnostic apparatus which concerns on Embodiment 1. FIG. 1 is a block diagram illustrating a configuration of an identification parameter learning device according to Embodiment 1. FIG. It is a figure which shows the accumulation example of the database of the identification parameter learning apparatus which concerns on Embodiment 1. FIG. 3 is a flowchart showing an operation of the abnormal sound diagnosis apparatus according to the first embodiment. 3 is a flowchart showing an operation of the abnormal sound diagnosis apparatus according to the first embodiment. It is a figure which shows an example of the abnormal kind and the K-dimensional score vector which the determination part of the abnormal sound diagnostic apparatus concerning Embodiment 1 refers. It is explanatory drawing which shows the effect of the abnormal sound diagnosis by the abnormality diagnosis apparatus of Embodiment 1. It is explanatory drawing which shows the result of the abnormal sound diagnosis by the conventional abnormal sound diagnostic apparatus. It is explanatory drawing which shows the connection of the several intensity | strength vector by the locus | trajectory feature extraction part of the abnormality diagnosis apparatus of Embodiment 1. It is explanatory drawing which shows the effect of the abnormal sound diagnosis by the other structure of the abnormality diagnosis apparatus of Embodiment 1. FIG. It is a figure which shows the structure of the identification part of the abnormal sound diagnostic apparatus of Embodiment 2. FIG. 6 is a flowchart showing an operation of the abnormal sound diagnosis apparatus according to the second embodiment.

Hereinafter, in order to explain the present invention in more detail, modes for carrying out the present invention will be described with reference to the accompanying drawings.
Embodiment 1 FIG.
The abnormal sound diagnosis apparatus according to the first embodiment diagnoses sound generated from a device to be diagnosed (for example, an elevator), and whether the generated sound is normal sound or abnormal sound. When the sound is abnormal, the type of abnormality is determined. The device to be diagnosed is a device composed of a plurality of operating parts such as an elevator, for example, and by attaching sound collecting means for collecting the generated sound in the elevator car or outside the car. The sound generated when the car is reciprocated is collected, and the working sound of the working parts is diagnosed by determining whether the collected sound is normal or abnormal. Note that the abnormal sound diagnosis apparatus of the present invention can be applied to devices other than elevators.

In the following, a case where the abnormal sound diagnosis apparatus is implemented as software on a personal computer (hereinafter referred to as a PC) will be described as an example. The PC includes a USB terminal and a LAN terminal. A microphone is connected to the USB terminal via an audio interface circuit, and a diagnosis target device is connected to the LAN terminal via a LAN cable. The device to be diagnosed is configured to perform a predetermined driving operation according to a control instruction output from the PC. The abnormal sound diagnosis apparatus 100 is not limited to being implemented as software, and can be changed as appropriate.

FIG. 1 is a block diagram illustrating a configuration of an abnormal sound diagnosis apparatus 100 according to the first embodiment.
FIG. 1A is a diagram illustrating functional blocks of the abnormal sound diagnosis apparatus 100 according to the first embodiment. The sound collection unit 1, the waveform acquisition unit 2, the time frequency analysis unit 3, the intensity time series acquisition unit 4, the locus It comprises a feature extraction unit 5, an identification parameter storage unit 6, an identification unit 7 and a determination unit 8.
The sound collection unit 1 is configured by a sound collector such as a microphone, for example, collects sound generated from the diagnosis target device, and outputs sound data 11 in synchronization with the operation of the diagnosis target device. When the device to be diagnosed is an elevator, the sound collection unit 1 is arranged in the passenger car or outside the passenger car. The waveform acquisition unit 2 includes, for example, an amplifier and an A / D converter, samples the waveform of the sound data 11 collected by the sound collection unit 1, and outputs the waveform data 12 converted into a digital signal.

The time frequency analysis unit 3 multiplies the waveform data 12 output from the waveform acquisition unit 2 by a time window, and performs a fast Fourier transform (hereinafter referred to as FFT) operation while shifting the time window in the time direction to convert the waveform data 12 to the time frequency. Analysis is performed to obtain a time-frequency distribution 13. The intensity time series acquisition unit 4 obtains an intensity time series 14 indicating the intensity with respect to time and frequency from the time frequency distribution 13 output from the time frequency analysis unit 3. The trajectory feature extraction unit 5 smoothes the intensity time series 14 output from the intensity time series acquisition unit 4 in the time direction, and extracts a trajectory vector 15 over the entire time axis. The identification parameter storage unit 6 is a storage area for storing an identification parameter learned in advance, and an identification parameter for identifying whether the operation state of the device is normal or abnormal, and when the operation state of the device is abnormal An identification parameter for identifying the type of abnormality is stored. Details of learning of the identification parameter 16 stored in the identification parameter storage unit 6 will be described later.

The identification unit 7 collates the identification parameter 16 stored in the identification parameter storage unit 6 with the trajectory vector 15 extracted by the trajectory feature extraction unit 5, and acquires scores for a plurality of abnormal types. It is assumed that K types of abnormalities such as a normal operating state and an abnormal operating state at a specific location are set as the abnormal type. The score for the K types of abnormal types is hereinafter referred to as a K-dimensional score vector 17. The detailed configuration of the identification unit 7 will be described later. The determination unit 8 determines whether the operation state of the device is normal or abnormal based on the K-dimensional score vector 17 of the identification unit 7, and if abnormal, also determines the type of abnormality, The result 18 is output.

FIG. 1B is a block diagram illustrating a hardware configuration of the abnormal sound diagnosis apparatus 100 according to the first embodiment, and includes a processor 100a and a memory 100b. The sound collection unit 1, the waveform acquisition unit 2, the time frequency analysis unit 3, the intensity time series acquisition unit 4, the trajectory feature extraction unit 5, the identification unit 7 and the determination unit 8 execute a program stored in the memory 100b by the processor 100a. This is realized. Further, it is assumed that the identification parameter storage unit 6 is stored in the memory 100b.

Next, a detailed configuration of the identification unit 7 will be described.
FIG. 2 is an explanatory diagram showing the configuration of the identification unit 7 of the abnormal sound diagnosis apparatus 100 according to the first embodiment, and shows the configuration of the neural network in the identification unit 7.
The neural network shown in the example of FIG. 2 includes a hierarchical type, and includes a first hidden layer 72 and a second hidden layer 73 that are one input layer 71 and two hidden layers. The input layer 71, the first hidden layer 72, and the second hidden layer 73 include a unit for simulating the function of the synapse of the cranial nerve circuit. There are no connections between units in each layer, only connections between units between layers. For this reason, it is known that the neural network according to the first embodiment can stably obtain good performance by a learning method known as deep learning in the field of machine learning.

The last hidden layer also serves as the output layer. In the example of FIG. 2, the second hidden layer 73 also serves as the output layer. In general, the number M of hidden layers may be an integer of one or more layers (M ≧ 1). In the following, a case where the number of hidden layers M = 2 is described as an example based on FIG.
The input layer 71 has the same number of units as the number of dimensions (for example, L × B) of the trajectory vector 15 input from the trajectory feature extraction unit 5. Further, the second hidden layer 73, that is, the output layer has K number of nonlinear units equal to the number K of abnormal types. The number of hidden layer units excluding the output layer is set to a predetermined number in view of the discrimination performance of the neural network. Assuming that the number of units in the mth layer is U (m) (m = 0, 1, 2,..., M) with the 0th as the input layer, the number of units is restricted based on the following equation (1). is there.
U (0) = L × B
U (m) = any natural number (m = 1, 2,... M−1) (1)
U (M) = K
In the formula (1), U (m) represents the number of units in the mth layer.

Further, the load and bias necessary for calculating the response of the hidden layer are supplied from the identification parameter 16 stored in the identification parameter storage unit 6. Hereinafter, the load and bias supplied to the mth hidden layer are w (i, j, m-1) and c (j, m-1), respectively. Here, the range of i and j is i = 0, 1,. . . , U (m−1) −1 and j = 0, 1,. . . , U (m) −1.

Next, learning of the identification parameter 16 used in the identification unit 7 will be described. The identification parameter 16 stored in the identification parameter storage unit 6 is learned by the identification parameter learning device 200 shown in FIG.
FIG. 3A is a diagram illustrating functional blocks of the identification parameter learning apparatus 200 according to the first embodiment. The sound data generation unit 21, the sound database 22, the waveform acquisition unit 23, the time frequency analysis unit 24, and the intensity time series are illustrated. An acquisition unit 25, a trajectory feature extraction unit 26, a teacher vector creation unit 27, and an identification learning unit 28 are included.

The sound data generation unit 21 collects sound data using a plurality of devices with different specifications and operations as reference devices, or generates sound data by computer simulation. In the example of the first embodiment, a plurality of elevators having different specifications and operations are reference devices. The sound database 22 stores sound data 22a and abnormality type data 22b. The sound data 22a is composed of sound data generated by the sound data generation unit 21 and sound data obtained by superimposing an abnormal sound on the sound data 22a generated by the sound data generation unit 21. The abnormality type data 22b is an abnormality type of the device associated with the sound data 22a, specifically, a label indicating whether the operation state of the device is normal or abnormal, and when the operation state of the device is abnormal. A label indicating the type of abnormality is accumulated.

An example of the sound data 22a and the abnormality type data 22b stored in the sound database 22 is shown in FIG.
As shown in FIG. 4, the sound data 22a is composed of “serial number”, “solid name”, and “sound data file name”, and the abnormality type data 22b is “abnormality type C (v” corresponding to the above “serial number”. ): Example ”.
As an example of the abnormality type C (v), types such as “normal”, “top abnormality”, and “middle floor abnormality” are associated, and K abnormality types are set in total including “normal”. ing.

The waveform acquisition unit 23 samples the waveform of the sound data 22a accumulated in the sound database 22, and outputs the waveform data 31 converted into a digital signal. The time frequency analysis unit 24, the intensity time series acquisition unit (parameter intensity time series acquisition unit) 25, and the trajectory feature extraction unit (parameter trajectory feature extraction unit) 26 of the abnormal sound diagnosis apparatus 100 of FIG. The same operation as the time frequency analysis unit 3, the intensity time series acquisition unit 4 and the trajectory feature extraction unit 5 is performed, and a time frequency distribution 32, an intensity time series 33 and a trajectory vector 34 are output, respectively. The teacher vector creation unit 27 creates a teacher vector 35 using the abnormality type data 22b accumulated in the sound database 22.

The identification learning unit 28 creates learning data for learning the neural network. The learning data of a neural network generally consists of input data and output data expected to be output from the neural network when the input data is given. In the example of the block diagram shown in FIG. 3, the input data is a trajectory vector 34 input from the trajectory feature extraction unit 26, and the output data is a teacher vector 35 input from the teacher vector creation unit 27.
When the total number of sound data used for learning of the neural network is V, the input data is V trajectory vectors 34 and the output data is V teacher vectors 35.

If the trajectory vector 34 extracted from the v-th sound data in the sound data 22a is ρ (k, v), the input data x (k, v) in the identification learning unit 28 is given by the following equation (2). .
x (k, v) = ρ (k, v) (2)
That is, the input data x (k, v) is the same as the trajectory vector 34.

The V teacher vectors 35 created by the teacher vector creation unit 27 are K for the number of types of abnormal types, y (k, v) for the kth element of the vth teacher vector, and vth sound data. Is C (v), y (k, v) is given by the following equation (3) as a vector in which the element at the C (v) -th position is 1 and the other elements are 0: .

The identification learning unit 28 learns the neural network using the trajectory vector 34 that is the input data obtained as described above and the teacher vector 35 that is the output data, and uses the load and bias obtained as a result of the learning. Is stored in the identification parameter storage unit 6 as the identification parameter 36. The load and bias constituting the discrimination parameter 36 are the load w (i, j, m−1) used when calculating the response of the first hidden layer 72 and the second hidden layer 73 of the discriminator 7 described above, and This corresponds to the bias c (j, m−1).

FIG. 3B is a block diagram showing a hardware configuration of the identification parameter learning apparatus 200 according to Embodiment 1, and includes a processor 200a and a memory 200b. In the sound data generation unit 21, the waveform acquisition unit 23, the time frequency analysis unit 24, the intensity time series acquisition unit 25, the trajectory feature extraction unit 26, the teacher vector generation unit 27, and the identification learning unit 28, the processor 200a is stored in the memory 200b. This is realized by executing the program. The sound database 22 is stored in the memory 200b.

Next, the operation of the abnormal sound diagnosis apparatus 100 will be described with reference to FIGS.
5 and 6 are flowcharts showing the operation of the abnormal sound diagnosis apparatus 100 according to Embodiment 1, FIG. 5 shows the operation of the sound collection unit 1 and the waveform acquisition unit 2, and FIG. 6 shows the time-frequency analysis unit 3 The operation of each component thereafter is shown. In the following description, a device to be diagnosed by the abnormal sound diagnosis apparatus 100 is simply referred to as a device.
When the abnormal sound diagnosis apparatus 100 detects the start of operation of the device (step ST1), the sound collecting unit 1 collects sound generated from the device (step ST2). The waveform acquisition unit 2 samples the sound waveform 11 by acquiring and amplifying the sound data 11 collected in step ST2 and performing A / D conversion (step ST3). For example, the waveform acquisition unit 2 is a 16-bit linear having a sampling frequency of 48 kHz. PCM (pulse code modulation) digital signal waveform data is converted (step ST4).

Next, the abnormal sound diagnosis apparatus 100 determines whether or not the operation of the device is finished (step ST5). If the operation of the device has not ended (step ST5; NO), the process returns to step ST2 and the above-described process is repeated. On the other hand, when the operation of the device is completed (step ST5; YES), the waveform acquisition unit 2 connects the waveform data acquired in step ST4 and outputs it as a series of waveform data 12 (step ST6). This completes the sound collection and waveform data acquisition process. Next, proceeding to the flowchart of FIG. 6, an abnormal sound diagnosis process using the acquired waveform data 12 is performed.

The time-frequency analysis unit 3 acquires the waveform data 12 output from the waveform acquisition unit 2, and extracts a frame from the waveform data 12 while shifting a time window of, for example, 1024 points in the time direction at intervals of 16 milliseconds. Then, a time frequency distribution g (t, f), which is a frequency spectrum series, is obtained by FFT calculation for each frame to obtain a time frequency distribution 13 (step ST11).
Here, t is a time index corresponding to the shift interval for shifting the time window, and f is an index indicating the frequency of the result of the FFT operation. The time t and the frequency f are integers that satisfy 0 ≦ t ≦ T and 0 ≦ f ≦ F, respectively. T is the number of frames in the time direction of the time frequency distribution 13, and F is an index corresponding to a Nyquist frequency that is ½ of the sampling frequency fs of the waveform data 12 (F = fs / 2).

Next, the intensity time series acquisition unit 4 has a band of 1 octave width in the time frequency distribution 13 obtained in step ST11, for example, with five frequencies of 0.5 kHz, 1 kHz, 2 kHz, 4 kHz, and 8 kHz as center frequencies. The sum of the frequency components included in the five frequency bands is obtained, and the intensity time series 14 of each band is obtained (step ST12). If the intensity time series 14 of each band is G (t, b), G (t, b) is given by the following equation (4).

In Equation (4), b is an integer satisfying 0 ≦ b <B as a band index (B is the number of bands, and B = 5 in this example). Further, Ω (b) represents a set of frequencies f for which the sum is obtained in the time frequency distribution g (t, f) with respect to the band b.

The trajectory feature extraction unit 5 smoothes the intensity time series 14 in the time direction for each band (step ST13), obtains a smoothing intensity at a point that equally divides the entire time axis into L, and creates an L-dimensional intensity vector ( Step ST14). In this example, the intensity time series 14 is smoothed in the time direction in five bands. Further, the intensity of the created L-dimensional intensity vector is normalized (step ST15), and the L-dimensional intensity vector of each normalized band is connected to create an L × B-dimensional trajectory vector 15 (step ST16).

The intensity time series G to (t, b) (t = 0, 1,..., T−1, b = 0, 1,..., B−1) after the smoothing in step ST13 are as follows. Calculated based on equation (5).

In Expression (5), smooth_t (x (t)) is a function that outputs a new time series obtained by smoothing the series x (t) relating to t in the subscript t direction.

Further, the smoothing strength H (l, b) of the L equally divided point obtained in step ST14 (l = 0, 1,..., L−1, b = 0, 1,..., B−1). Is calculated based on the following equation (6).

In Equation (6), τ (l) is a real function representing the interpolation position with respect to the subscript t in G to (t, b), and w (l) is a function that gives a weighting coefficient for interpolation. It is given by equations (7) and (8).

Int (x) in Expression (8) is a function for obtaining the integer part of the argument x.

In the generation of the trajectory vector 15 in step ST16, the smoothing strength H (l, b) of the L equally divided points of each band is connected as an L-dimensional vector to obtain the trajectory vector 15, and k of the trajectory vector 15 Assuming that the th element is ρ (k) (k = 0, 1,..., L × B−1), ρ (k) is given by the following equation (9).

Next, the identification unit 7 inputs the trajectory vector 15 input from the trajectory feature extraction unit 5 to the input layer 71 of the neural network, and uses the identification parameter stored in the identification parameter storage unit 6 to activate the output unit. And a K-dimensional score vector 17 is generated (step ST17).
The processing in step ST17 will be described with reference to a specific configuration example of the identification unit 7 in FIG. First, the i-th element in the trajectory vector 15 is copied to the i-th unit of the input layer. If the value of the i-th unit in the input layer is x (i, 0), x (i, 0) is given by the following equation (10).
x (i, 0) = ρ (i) (10)
In equation (10), ρ (i) represents the value of the i-th element of the trajectory vector 15.

Next, the output of each unit is calculated in order from the first hidden layer 72 to the second hidden layer 73. The output of each unit is obtained by applying a load to the output from all units in the previous layer to obtain the sum, subtracting the bias, and performing non-linear conversion using a sigmoid function. If the output of the jth unit of the mth hidden layer is x (j, m), x (j, m) is calculated by the following equation (11).

In Expression (11), σ (x) is a sigmoid function having a nonlinear input / output characteristic indicating a soft threshold characteristic, and is given by the following Expression (12).

In the above equation (11), when m = 1, x (i, 0) is required. This is because the i-th element ρ () of the trajectory vector 15 as shown in the above equation (10). equal to i).

The calculation based on equation (11) is m = 1,. . . , M, the output x (k, M) of the last hidden layer is obtained. In the example of FIG. 2, the output x (k, 2) of the second hidden layer 73 is obtained. The output is regarded as the output of the output layer. If the k-th output of the output layer is o (k), o (k) is given by the following equation (13).
o (k) = x (k, M) (13)
Finally, normalize the K outputs of the output layer. By normalizing, the sum of K outputs becomes 1. If the result of normalization is a score vector value s (k), the score vector value s (k) is given by the following equation (14) known as a softmax operation.

The K-dimensional score vector 17 obtained by the above process is output to the determination unit 8.

Returning to the flowchart of FIG.
The determination unit 8 compares the elements of the K-dimensional score vector 17 generated in step ST17, determines a possible abnormality type based on the index of the largest element (step ST18), and outputs a determination result (step ST18). Step ST19), the process is terminated. If the possible abnormality type is k *, k * is given by the following equation (15).

In addition, although the structure which outputs one element with the largest score of the K-dimensional score vector 17 was shown, you may comprise so that a some element may be output with those scores.

FIG. 7 is a diagram illustrating an example of an abnormality type and a K-dimensional score vector referred to by the determination unit 8 of the abnormal sound diagnosis apparatus 100 according to the first embodiment.
As shown in FIG. 7, “K-dimensional score vectors” are associated with K “abnormality types”, respectively. The K-dimensional score vector becomes “1” when all the values of the K score vectors constituting the K-dimensional score vector are added. In the example of FIG. 7, since the score vector of the abnormality type “top abnormality” is “0.64” and takes the maximum value, the determination unit 8 determines that the possible abnormality type is “top abnormality”.

Next, the effect when the abnormal sound diagnosis apparatus 100 configured as described above is applied to an elevator will be described with reference to FIGS. 8 and 9.
FIG. 8 is an explanatory diagram showing the effect of abnormal sound diagnosis performed by the abnormal sound diagnosis apparatus 100 according to the first embodiment. For comparison, FIG. 9 shows a result of abnormal sound diagnosis by a conventional abnormal sound diagnosis apparatus.
First, an abnormal sound diagnosis method using a conventional abnormal sound diagnosis apparatus and results obtained will be described with reference to FIG. In the conventional abnormal sound diagnosis apparatus, the traveling section 301 of the car 300 is divided, and the signal intensity of the sound generated at normal time for each divided section is stored as a reference value. In the example of FIG. 9A, the travel section is divided into six, and the first reference value, the second reference value,..., The sixth reference value are acquired and stored.

The abnormality was detected in each section by comparing the stored reference value with the intensity time series of the sound data acquired at the time of diagnosis. Since the signal strength of normal sound in each section differs depending on the use and operating environment of each elevator, the reference value acquired for a certain elevator cannot be applied to the abnormal sound diagnosis of different elevators, or Even if it can be applied, there is a problem that the accuracy of the abnormal sound diagnosis is lowered. For this reason, in the conventional abnormal sound diagnosis apparatus, it is necessary to perform a learning operation in advance for each elevator and store a reference value.

FIG. 9B shows the result of comparison with the sound signal intensity at the time of diagnosis when a reference value created by a different elevator is applied to another elevator, and based on the signal intensity of the sound at the time of diagnosis. In the plotted diagram, there is a normal operation solid 303 whose signal strength exceeds the reference value 305, or there is an abnormal operation solid 304 whose signal strength does not exceed the reference value 305. As described above, no matter how the reference value 305 is set, there is a problem that the normal operation state and the abnormal operation state of the device cannot be clearly separated based on the sound signal intensity at the time of diagnosis. It was.

Next, the effect of the abnormal sound diagnosis performed by the abnormal sound diagnosis apparatus 100 according to Embodiment 1 will be described with reference to FIG.
In the abnormal sound diagnosis apparatus 100 according to the first embodiment, as shown in FIG. 8A, sounds generated while the car 300 reciprocates between the lowermost floor and the uppermost floor are collected, and the obtained sound data is collected. Then, the time frequency is analyzed, an intensity time series is obtained, and the trajectory over the entire length in the time direction of the intensity time series is vector-converted as an integral to extract a trajectory vector. In the example of FIG. 8A, for simplification of explanation, two types of abnormality, “normal” and “abnormal” (K = 0 to 1), the number of bands is set to 1 band (B = 1), and L A case where x 1-

dimensional trajectory vectors

306 and 307 are extracted is shown. The trajectory vector 306 indicates a vector when the abnormality type is “1: abnormal”, and the trajectory vector 307 indicates a vector when the abnormal type is “0: normal”. FIG. 8B shows a result of plotting the positions of the trajectory vector 306 and the trajectory vector 307 in the space when the trajectory vector 306 and the trajectory vector 307 are input to the discriminating unit 7.

FIG. 8B shows a first feature axis (main axis) and a second feature axis (axis orthogonal to the main axis) by performing, for example, principal component analysis using a vector of normal individuals and a vector of abnormal individuals as a set. It is a diagram showing the arrangement of each vector in the L × 1D space where these feature axes are stretched.
The principal component analysis is a process for displaying the mutual positional relationship of vectors in a multidimensional space, and is not a process constituting the present invention. The first feature axis and the second feature axis are not calculated by the configuration of the present invention, but are described to show that the trajectory vectors are classified in space.

As shown in the plot result of FIG. 8B, based on the abnormality type of the trajectory vector and the position in the space, the group 308 indicating the normality of the device and the group 309 indicating the abnormality of the device are arranged. In this example, a hyperplane (straight line) orthogonal to a straight line connecting the center of gravity of the group 308 and the center of gravity of the group 309 is obtained as the boundary 310. With this arrangement, general features appearing in the intensity time series can be captured.
Although FIG. 8B shows an example in which a straight line is obtained as the boundary 310, it is assumed that a hypercurve (curve) having a complicated shape is obtained in actual diagnosis processing.
In this way, it is possible to capture general characteristics that appear in the intensity time series regardless of the elevator specifications and operating environment, and there is no need to learn reference values for each individual in advance, and there are differences in elevator specifications and operating environments. Can also make a robust diagnosis.

As described above, according to the first embodiment, the sound collection unit 1 that collects the sound generated from the device, and the waveform acquisition that acquires the waveform data obtained by sampling and converting the waveform of the collected sound data. Unit 2, time frequency analysis unit 3 that performs time frequency analysis of the acquired waveform data, intensity time series acquisition unit 4 that obtains an intensity time series indicating the intensity with respect to time and frequency from the time frequency distribution, and acquired intensity The trajectory feature extracting unit 5 that smoothes the time series in the time direction and extracts trajectory vectors over the entire time axis, and stores the identification parameters learned using the extracted trajectory vectors as input data and abnormal types as output data. An identification parameter storage unit 6; an identification unit 7 that acquires a score vector corresponding to the abnormality type based on the trajectory vector and the stored identification parameter; and an acquired score level And a determination unit 8 that determines the type of abnormality of the device based on the torque, so that the device is normal by capturing generalized features appearing in the intensity time series that do not depend on the specification or operating environment of the device. Whether or not there is an abnormality, and if it is abnormal, the type of abnormality can be determined. Thereby, it is not necessary to learn a reference for diagnosis for each device having different specifications and operations in advance, and a robust diagnosis can be performed even for differences in device specifications and operating environments. In addition, it is possible to provide an abnormal sound diagnosis apparatus that suppresses a decrease in diagnosis accuracy due to differences in device specifications and operating environments.

In the above description of the first embodiment, the case where the sound collecting unit 1 is configured by one sound collector and arranged in a device to be diagnosed has been described. However, the sound collecting unit 1 includes a plurality of sound collectors. And may be arranged at a plurality of locations of the diagnosis target device. In this case, multi-channel sound collection is performed simultaneously with the operation of the diagnosis target device, and multi-channel sound data 11 is obtained. The waveform acquisition unit 2, the time frequency analysis unit 3, and the intensity time series acquisition unit 4 acquire the waveform data 12, the time frequency distribution 13, and the intensity time series 14 for the multichannel signals, respectively. The trajectory feature extraction unit 5 acquires a multi-channel intensity vector from the multi-channel intensity time series 14 input from the intensity time-series acquisition unit 4. Further, the intensity vectors of the respective channels are connected in the time axis direction.

FIG. 10 is an explanatory diagram illustrating connection of multi-channel intensity vectors in the trajectory feature extraction unit 5 of the abnormal sound diagnosis apparatus 100 according to the first embodiment.
FIG. 10 shows a case where the intensity vectors of the three channels are connected, and the first channel vector 15a, the second channel vector 15b, and the third channel vector 15c are connected in the time axis direction of the vector to obtain L × B. A trajectory vector 15 of three dimensions (“× 3” is obtained by connecting intensity vectors of three channels) is generated. Since there is a connection between the channels in the intermediate layer of the neural network, the synchronicity between the channels can be learned. In the description up to the previous paragraph, the dimension number of the trajectory vector is L × B, but here, the dimension number of the trajectory vector is read as L × B × 3.
Thus, by using sound data collected by a plurality of sound collectors, the degree of separation in the identification space between vectors of different types of abnormalities can be improved, and diagnostic accuracy can be improved.

FIG. 11 is an explanatory diagram showing an effect when an abnormal sound diagnosis is performed based on a trajectory vector obtained by connecting multi-channel intensity vectors.
In FIG. 11A,

intensity time series

311, 312, and 313 indicate intensity time series obtained in the first frequency band, the second frequency band, and the third frequency band, respectively. , 312, and 313 are shown as a L × 1 × 3D trajectory vector 314 and a trajectory vector 315 that are connected in the time axis direction. The trajectory vector 314 indicates a vector when the abnormality type is “1: abnormal”, and the trajectory vector 315 indicates a vector when the abnormal type is “0: normal”. When the trajectory vector 314 and the trajectory vector 315 are input to the discriminating unit 7, the result of plotting the spatial positions of the trajectory vector 314 and the trajectory vector 315 in the discriminating unit 7 is shown in FIG. In FIG. 11B, a result equivalent to the result shown in FIG. 8B can be obtained.

Embodiment 2. FIG.
In the first embodiment, the case where the identification unit 7 has a neural network configuration has been described. In the second embodiment, a case where a support vector machine (hereinafter referred to as SVM) is applied as the identification unit will be described. I do.
Since the entire configuration of the abnormal sound diagnosis apparatus 100 of the second embodiment is the same as that of the first embodiment, description of the block diagram is omitted, and an identification unit having a different configuration will be described in detail below.

FIG. 12 is a diagram illustrating a configuration of the identification unit 7a of the abnormal sound diagnosis apparatus 100 according to the second embodiment.
The identification unit 7a has (K-1) K / 2 SVMs as a whole, where K is the number of abnormal types. Here, each SVM is configured to learn to classify and identify any two abnormal types of vectors among K abnormal types including normal. Each SVM has, as parameters, the number of support vectors n, n support vectors x _i (i = 0, 1, 2,..., N−1), and n coefficients α _i (i = 0, 1). , 2,..., N−1), bias b, and kernel function definition k (x ₁ , x ₂ ) described later. Hereinafter, the SVM that identifies normal or abnormal type i and abnormal type j (where i <j) is referred to as SVM [i, j] (0 ≦ i <j <K).

Next, the operation of the identification unit 7a will be described.
FIG. 13 is a flowchart showing the operation of the abnormal sound diagnosis apparatus according to the second embodiment. In the following, the same steps as those in the abnormal sound diagnosis apparatus according to the first embodiment are denoted by the same reference numerals as those used in FIG. 6, and the description thereof is omitted or simplified. The operations of the sound collection unit 1 and the waveform acquisition unit 2 are the same as those in the flowchart shown in FIG.
When the trajectory vector 15 created by the trajectory feature extraction unit 5 is input in step ST16, the identification unit 7a inputs the trajectory vector 15 to each SVM, and uses the identification parameter stored in the identification parameter storage unit 6. The output value y (ρ) of the discrimination function of each SVM is calculated based on the following equation (16) (step ST21).

_{_{Here, k (x 1, x 2}} ) is the inner product between the mapping to a multi-dimensional space of vectors x1 φ _{(x 1)} and mapping φ _{(x 2)} to a multi-dimensional space vector _{x 2 <φ (x} ₁ ), φ (x ₂ )> (note that φ (x) is a nonlinear function of the vector x that cannot be expressed by an explicit expression). As the kernel function, for example, a Gauss kernel represented by the following equation (17) can be used. Note that σ is a Gaussian kernel parameter.

Next, the identification unit 7a calculates the classification output of each class from the output value of the identification function of each SVM calculated in step ST21, and the score vector value s indicating the score of 1 to K corresponding to the abnormality type (K) is calculated, and the calculated score vector value s (k) is output to the determination unit 8 as the K-dimensional score vector 17 (step ST22). The determination unit 8 compares the elements of the K-dimensional score vector 17 generated in step ST17, determines a possible abnormality type based on the index of the largest element (step ST18), and outputs a determination result (step ST18). Step ST19), the process is terminated.

As described above, even when the support vector machine is applied to the identification unit 7a as in the second embodiment, the trajectory feature extraction unit 5 creates the trajectory vector 15 over the entire length in the time direction of the intensity time series 14. It is possible to capture generalized features that appear in the intensity time series that do not depend on device specifications or operating environments. Thereby, it is not necessary to learn the criteria at the time of diagnosis for each individual, and robust diagnosis can be performed even with respect to differences in device specifications and operating environments. Moreover, the abnormal sound diagnostic apparatus which suppressed the fall of the diagnostic precision by the difference in an apparatus can be provided.

In the first and second embodiments described above, the trajectory vector 15 output from the trajectory feature extraction unit 5 uses the trajectory features over the entire length in the time direction of the intensity time series 14 as an L-dimensional vector by linear interpolation. Although the structure to extract was shown, you may obtain | require an L-dimensional vector using the other conversion which preserve | saves the characteristic of the locus | trajectory over the full length of the time direction of the intensity | strength time series 14 irrespective of the presence or absence of loss of information. As another conversion, for example, a trajectory of the intensity time series 14 over the entire length in the time direction may be Fourier-transformed to form an L-dimensional vector from low-order Fourier coefficients. As another conversion, a compressed feature may be output as an L-dimensional vector by principal component analysis.

It should be noted that the above-described conversion without loss indicates that the features as they are are used as the vectors without performing processing on the vectors indicating the features over the entire length in the time direction of the intensity time series 14. On the other hand, the conversion that allows loss is a process of reducing the number of dimensions by multiplying a vector indicating the characteristics of the intensity time series 14 over the entire length in the time direction by, for example, a matrix obtained by principal component analysis. It shows that the feature is used as a vector. It is considered that a part of the information included in the raw feature vector is lost due to the above-described reduction of the number of dimensions.

In Embodiment 1 and Embodiment 2 described above, when the device to be diagnosed is an elevator, the trajectory over the entire length in the time direction of the intensity time series in a route in which the elevator makes a round trip through the operating section. The trajectory vector is extracted by converting the vector into a vector, but in the reciprocating operation of the working section, it is divided into one-way sections such as the ascending section and the descending section, and the total length of the section in the time direction for each divided section It is also possible to convert the trajectory extending over to a vector and extract the trajectory vector, and to prepare the identification unit 7 for each divided section and perform the identification processing.
Thereby, in the case of an elevator, there is no abnormality when ascending, and even if there is an abnormality when descending, diagnosis for each section is possible.
The sections to be divided are not limited to the ascending section and the descending section. For example, the ascending section may be further divided into smaller sections such as a lower section, a middle section, and a higher section.

Since the abnormal sound diagnosis apparatus according to the present invention can perform abnormal sound diagnosis with high accuracy with respect to differences in equipment specifications and operations, a reference value for determining abnormal sound is created for each individual. Applicable to devices that cannot be used, and suitable for diagnosing abnormal sounds in devices.

1 sound collection unit, 2 waveform acquisition unit, 3 time frequency analysis unit, 4 intensity time series acquisition unit, 5 trajectory feature extraction unit, 6 identification parameter storage unit, 7, 7a identification unit, 8 determination unit, 21 sound data generation unit , 22 sound database, 22a sound data, 22b abnormality type data, 23 waveform acquisition unit, 24 time frequency analysis unit, 25 intensity time series acquisition unit, 26 trajectory feature extraction unit, 27 teacher vector creation unit, 28 identification learning unit, 71 Input layer, 72 1st hidden layer, 73 2nd hidden layer, 100 abnormal sound diagnosis device, 100a, 200a processor, 100b, 200b memory, 200 identification parameter learning device.

Claims

In the abnormal sound diagnosis device that diagnoses whether the sound generated in the diagnosis target device is abnormal,
A sound collecting unit that collects sound generated by the diagnosis target device and obtains sound data;
An intensity time series acquisition unit for acquiring an intensity time series from a time frequency distribution obtained by analyzing waveform data of sound data acquired by the sound collection unit;
A trajectory feature extracting unit that extracts a trajectory vector by converting a trajectory indicating intensity characteristics in all time directions of the intensity time series acquired by the intensity time series acquiring unit;
An input is a vector that is a trajectory showing intensity characteristics in all time directions of intensity time series obtained from a time-frequency distribution obtained by analyzing waveform data of sound data generated from a reference device, and the state of the diagnosis target device An identification parameter storage unit for storing an identification parameter learned by using information indicating the type as an output;
An identification unit that acquires a score for each state type of the diagnosis target device from the locus vector extracted by the locus feature extraction unit and the identification parameter stored in the identification parameter storage unit;
An abnormal sound diagnosis, comprising: a determination unit that refers to the score acquired by the identification unit and determines whether the sound generated in the diagnosis target device is normal or abnormal and the type of abnormality apparatus.
The intensity time series acquisition unit acquires the intensity with respect to time and frequency from the time frequency distribution as the intensity time series,
The trajectory feature extraction unit converts the trajectory indicated by the intensity time series acquired by the intensity time series acquisition unit into a vector in a two-dimensional space of time and intensity with respect to frequency, and connects the converted vectors to generate the trajectory vector. The abnormal sound diagnosis apparatus according to claim 1, wherein:
2. The abnormal sound according to claim 1, wherein the trajectory feature extraction unit performs lossless vector conversion or lossy vector conversion on the intensity time series acquired by the intensity time series acquisition unit. Diagnostic device.
The abnormal sound diagnosis apparatus according to claim 1, wherein the identification unit acquires the score using a neural network technique.
The abnormal sound diagnosis apparatus according to claim 1, wherein the identification unit acquires the score using a support vector machine technique.
A plurality of the sound collection units are arranged in the diagnosis target device, collect sound generated in the diagnosis target device and collect sound data of a plurality of channels,
The intensity time series acquisition unit acquires the intensity time series of the plurality of channels from the time frequency distribution obtained by analyzing the waveform data of each sound data of the plurality of channels collected by the sound collection unit,
The trajectory feature extraction unit converts a trajectory indicating intensity characteristics in all time directions of the intensity time series of a plurality of channels acquired by the intensity time series acquisition unit into a vector, and converts the converted vectors of the plurality of channels in the time direction. The abnormal sound diagnosis apparatus according to claim 1, wherein the abnormal sound diagnosis apparatus is connected to extract the trajectory vector.
The intensity time series acquisition unit acquires the intensity time series in each operating section from the time frequency distribution divided corresponding to the operating section of the diagnosis target device,
The trajectory feature extraction unit divides the intensity time series acquired by the intensity time series acquisition unit so as to correspond to the operation section of the diagnosis target device, and shows the trajectory indicating the intensity characteristics in all time directions of the divided intensity time series. To the vector to extract the trajectory vector,
The identification unit, for each state type for each operation section of the diagnosis target device, from the trajectory vector corresponding to each operation section extracted by the trajectory feature extraction unit and the identification parameter stored in the identification parameter storage unit. The abnormal sound diagnosis apparatus according to claim 1, wherein a score is acquired.
A sound database that stores sound data generated from the reference device, abnormal sound superimposition data obtained by superimposing abnormal sound on the sound data, and abnormal type information of the device associated with the sound data and the abnormal sound superimposition data When,
A parameter intensity time series acquisition unit for acquiring an intensity time series from a time frequency distribution obtained by analyzing the waveform data of the sound data and the abnormal sound superimposed signal accumulated in the sound database;
A parameter trajectory feature extraction unit that converts a trajectory indicating intensity features in all time directions of the intensity time series into vectors from the intensity time series acquired by the parameter intensity time series acquisition unit;
A teacher vector creation unit that creates a teacher vector from the abnormality type information accumulated in the sound database;
The trajectory vector converted by the parameter trajectory feature extraction unit is input, learning is performed so that the teacher vector generated by the teacher vector generation unit is output, and the learning result is stored in the identification parameter storage unit as the identification parameter. An identification parameter learning device comprising: an identification learning unit for storing;
An abnormal sound diagnosis system comprising the abnormal sound diagnosis apparatus according to claim 1.
In the abnormal sound diagnosis method for diagnosing whether the sound generated in the diagnosis target device is abnormal,
A sound collection unit that collects sound generated by the diagnosis target device and obtains sound data; and
An intensity time series obtaining unit obtaining an intensity time series from a time-frequency distribution obtained by analyzing the waveform data of the sound data;
A trajectory feature extraction unit converts a trajectory indicating intensity features in all time directions of the intensity time series into a vector and extracts a trajectory vector;
The identification unit receives as input a vector that is a trajectory showing intensity characteristics in all time directions of intensity time series obtained from a time-frequency distribution obtained by analyzing waveform data of sound data generated from a reference device, and Obtaining a score for each state type of the diagnosis target device from the identification parameter learned as an output indicating information indicating the state type of the target device, and the trajectory vector;
An abnormal sound diagnosis method comprising: a step of determining whether a sound generated in the diagnosis target device is normal or abnormal and a type of abnormality with reference to the score.
Sound collection processing procedure for collecting sound generated by a diagnosis target device and acquiring sound data, and intensity time series acquisition processing for acquiring an intensity time series from a time-frequency distribution obtained by analyzing waveform data of the sound data Obtained by analyzing the procedure, the trajectory feature extraction processing procedure for extracting the trajectory vector by converting the trajectory showing the intensity features in the whole time direction of the intensity time series, and the waveform data of the sound data generated from the reference device. An identification parameter learned by using as input a vector that is a trajectory indicating intensity characteristics in all time directions of intensity time series acquired from the obtained time-frequency distribution, and outputting information indicating a state type of the diagnosis target device, and An identification processing procedure for obtaining a score for each state type of the diagnosis target device from a trajectory vector, and a sound generated in the diagnosis target device with reference to the score Whether it is a is or abnormal normal, determining the type of abnormality when an abnormality processing procedure and the abnormal sound diagnostic program for causing a computer to execute.