JP6173649B1 - Deterioration location estimation apparatus, degradation location estimation system, and degradation location estimation method - Google Patents

Abstract

The degradation point estimation device (6) acquires microphone signals (DS1, DS2, DS3) corresponding to each of the plurality of microphones (131 to 133) provided in the inspection target device, and acquires the microphone signals (DS1, DS2). , DS3), a short-time Fourier transform unit (31) for calculating a time series of short-time Fourier transform coefficients corresponding to each, and a time-series array for input to the neural network using the time series of short-time Fourier transform coefficients A time-series array generating unit (32) for generating, and a neural network unit (35) configured by a neural network, which receives an input of the time-series array and outputs a degree of deterioration corresponding to the inspection target location in the inspection target device; And a determination unit (36) for determining a deterioration part in the inspection target device using the deterioration degree.

Description

  The present invention relates to a degradation location estimation device, a degradation location estimation system, and a degradation location estimation method for estimating a degradation location in a device to be inspected.

  Conventionally, a failure location estimation device has been developed that estimates a failure location (hereinafter referred to as “failure location”) in a device to be inspected based on sound generated from the device to be inspected (hereinafter referred to as “inspection device”). Has been. Similarly, a degradation location estimation device has been developed that estimates a degradation location (hereinafter referred to as “degradation location”) in a device to be inspected based on sound generated from the device to be inspected.

  For example, the failure location estimation device disclosed in Patent Document 1 is collected in a sound collector (1) that is provided on a moving body and collects operating sounds generated from the moving body and devices located around the moving range of the moving body. Analyzing the normal operating sound and obtaining the reference sample series (104), the reference sample series analyzing means (2, 3, 4), and analyzing the collected operating sounds of the diagnostic object to obtain the target sample series (105) A target sample series analyzing means (2, 3, 4), and a mutation curve generating unit (6) for obtaining a mutation series between the reference sample series (104) and the target sample series (105) and generating a mutation curve (106); The shape feature extraction unit (7) for extracting the shape feature (107) of the mutation curve (106) and the shape feature (107) are compared with the template, and this occurs in the mobile body and the device located around the moving range of the mobile body. Collation determination unit that determines the failure location ( ) And those with a (abstract and FIG. 1 of Patent Document 1).

JP 2014-105075 A

  The failure location estimation apparatus of Patent Document 1 acquires sound generated at each location in a device to be inspected using one microphone (sound collector 1) provided on a moving body. Thereby, when a plurality of locations are arranged apart from each other along the moving direction of the moving body, it is possible to identify at which location the acquired sound is generated. However, when a plurality of locations are arranged apart from each other along a plane perpendicular to the moving direction of the moving body, it is impossible to identify at which location the acquired sound is generated. . That is, the failure location estimation apparatus of Patent Document 1 has a problem that the resolution with respect to the direction of sound arrival is low, and the failure location estimation accuracy is low.

  Further, the same problem has occurred in the degradation location estimation apparatus.

  The present invention has been made to solve the above-described problems, and includes a degradation location estimation device, a degradation location estimation system, and a degradation location estimation method that can accurately estimate degradation locations in a device to be inspected. The purpose is to provide.

The degradation point estimation apparatus of the present invention acquires microphone signals corresponding to each of a plurality of microphones provided in a device to be inspected, and calculates a time series of short-time Fourier transform coefficients corresponding to each of the microphone signals. It is composed of a short-time Fourier transform unit, a time-series array generation unit that generates a plurality of time-series arrays for input to a neural network using a time series of short-time Fourier transform coefficients, and a neural network. A neural network unit that receives an input of an array and outputs a deterioration level corresponding to the inspection target location in the inspection target device, and a determination unit that determines the deterioration location in the inspection target device using the deterioration level.

  According to this invention, since it comprised as mentioned above, the degradation location in a test object apparatus can be estimated with high precision.

It is explanatory drawing which shows the principal part of the test object apparatus which concerns on Embodiment 1 of this invention. It is explanatory drawing which shows the principal part of the degradation location estimation system which concerns on Embodiment 1 of this invention. It is a hardware block diagram which shows the principal part of the degradation location estimation apparatus which concerns on Embodiment 1 of this invention. It is a functional block diagram which shows the principal part of the degradation location estimation apparatus which concerns on Embodiment 1 of this invention. FIG. 3A is an explanatory diagram showing a cross-correlation spectrogram generated by the cross-correlation calculating unit according to Embodiment 1 of the present invention. FIG. 3B is an explanatory diagram showing an autocorrelation spectrogram generated by the autocorrelation calculating unit according to Embodiment 1 of the present invention. It is explanatory drawing which shows the structure of the neural network in the neural network part which concerns on Embodiment 1 of this invention. It is explanatory drawing which shows the three-dimensional arrangement | sequence in the input layer in the neural network shown in FIG. It is explanatory drawing which shows the three-dimensional arrangement | sequence in the convolution layer in the neural network shown in FIG. It is a flowchart which shows operation | movement of the degradation location estimation apparatus which concerns on Embodiment 1 of this invention. It is a flowchart which shows the detailed operation | movement of the short time Fourier-transform part which concerns on Embodiment 1 of this invention. It is a flowchart which shows the detailed operation | movement of the cross correlation calculating part which concerns on Embodiment 1 of this invention. It is a flowchart which shows the detailed operation | movement of the autocorrelation calculating part which concerns on Embodiment 1 of this invention. It is a flowchart which shows the detailed operation | movement of the neural network part which concerns on Embodiment 1 of this invention. It is a flowchart which shows the detailed operation | movement of the determination part which concerns on Embodiment 1 of this invention.

  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.
FIG. 1A is an explanatory diagram showing a main part of an inspection target apparatus according to Embodiment 1 of the present invention. FIG. 1B is an explanatory diagram showing a main part of the degradation point estimation system according to Embodiment 1 of the present invention. FIG. 1C is a hardware configuration diagram illustrating a main part of the degradation point estimation apparatus according to Embodiment 1 of the present invention. With reference to FIG. 1, the degradation location estimation system 100 of Embodiment 1 is demonstrated.

  In the figure, 1 is an elevator. The elevator 1 has a hoistway 2. In the hoistway 2, a basket 3 and a plurality of movable parts (not shown) are provided. The elevator 1 is provided in a building (not shown), and the basket 3 can move up and down along the hoistway 2 between the uppermost floor and the lowermost floor of the building. The elevator 1 constitutes a device to be inspected.

  A microphone array device 4, an audio interface device 5, and a degradation location estimation device 6 are provided on the ceiling of the basket 3. The microphone array device 4, the audio interface device 5, and the degradation location estimation device 6 constitute the main part of the degradation location estimation system 100.

The microphone array device 4 includes a substantially cylindrical pedestal 11 and three microphones 13 _{1 to} 133 ₃ mounted on a substantially circular mounting surface 12 included in the pedestal 11. Each of the microphones _{131-134 3} is to obtain the sound generated by the elevator 1. The microphones 13 _{1 to} 13 ₃ output analog signals AS1, AS2, and AS3 corresponding to the acquired sound waveforms, respectively.

Here, the microphones 13 _{1 to} 13 ₃ are arranged at substantially equal intervals along the circumferential portion of the placement surface 12. That is, the microphones 13 _{1 to} 13 ₃ are arranged at positions corresponding to the vertices of an equilateral triangle (not shown) on the placement surface 12 that is substantially perpendicular to the moving direction of the cage 3. Thus, the spacing between the microphones _{131-134 3} is set to the maximum value in the mounting surface 12.

The audio interface device 5 uses, for example, an analog-digital conversion circuit that supports multi-channel input. Audio interface device 5 acquires the analog signal AS1, AS2, AS3 the microphone _{131-134 3} has outputted, and converts each analog signal AS1, AS2, AS3 into digital signals DS1, DS2, DS3. The audio interface device 5 outputs each of the digital signals DS1, DS2, and DS3 to the degradation point estimation device 6.

  Specifically, for example, the audio interface device 5 converts the analog signals AS1, AS2, and AS3 into digital signals DS1, DS2, and DS3 by a linear PCM (Pulse Code Modulation) method. At this time, the sampling frequency is set to 48 kilohertz (kHz), for example, and the number of quantization bits is set to 16 bits, for example. As a result of the conversion, each of the digital signals DS1, DS2, DS3 becomes a signal composed of a plurality of frames.

  The degradation location estimation device 6 is configured by a computer such as a personal computer (PC). The computer has an input / output interface 21, a processor 22, and a memory 23.

  The input / output interface 21 includes, for example, a USB (Universal Serial Bus) terminal and a LAN (Local Area Network) terminal. The USB terminal is communicatively connected to the audio interface device 5. The LAN terminal is communicably connected to a control device (not shown) for the elevator 1 by a LAN cable (not shown). The control device uses a signal output from the computer for operation control of the elevator 1.

  The memory 23 temporarily stores the digital signals DS1, DS2, DS3 output from the audio interface device 5. The memory 23 stores a program for causing the computer to function as the short-time Fourier transform unit 31, the time-series array generation unit 32, the neural network unit 35, and the determination unit 36 illustrated in FIG. When the processor 22 reads and executes the program stored in the memory 23, the functions of the short-time Fourier transform unit 31, the time series array generation unit 32, the neural network unit 35, and the determination unit 36 shown in FIG. 2 are realized. . In addition, the memory 23 appropriately stores data such as various values calculated by the program.

  The processor 22 is configured by, for example, a CPU (Central Processing Unit), a microprocessor, a microcontroller, or a DSP (Digital Signal Processor). The memory 23 is, for example, a RAM (Random Access Memory), a ROM (Read Only Memory), a flash memory, an EPROM (Erasable Programmable Read Only Memory), an EEPROM (Electrically Erasable Programmable Memory), or the like. An optical disk or a magneto-optical disk is used.

  FIG. 2 is a functional block diagram showing the main part of the degradation point estimation apparatus according to Embodiment 1 of the present invention. FIG. 3A is an explanatory diagram showing a cross-correlation spectrogram generated by the cross-correlation calculating unit according to Embodiment 1 of the present invention. FIG. 3B is an explanatory diagram showing an autocorrelation spectrogram generated by the autocorrelation calculating unit according to Embodiment 1 of the present invention. With reference to FIG.2 and FIG.3, the degradation location estimation apparatus 6 is demonstrated.

  The short-time Fourier transform unit 31 acquires digital signals (hereinafter referred to as “microphone signals”) DS1, DS2, DS3 output from the audio interface device 5. The short-time Fourier transform unit 31 performs a short-time Fourier transform (STFT) on each of the microphone signals DS1, DS2, and DS3, so that the short-time Fourier transform unit 31 corresponds to each of the microphone signals DS1, DS2, and DS3. A time series of time Fourier transform coefficients is calculated. That is, each time series is composed of a plurality of short-time Fourier transform coefficients, and each short-time Fourier transform coefficient indicates a complex frequency spectrum.

  The time series array generation unit 32 uses the time series of the short-time Fourier transform coefficients calculated by the short-time Fourier transform unit 31 to input a time series array for input to the neural network unit 35 (hereinafter simply referred to as “time series array”). May be generated). In the example illustrated in FIG. 2, the time series array generation unit 32 includes a cross correlation calculation unit 33 and an autocorrelation calculation unit 34.

  The cross-correlation calculating unit 33 uses the time series of the short-time Fourier transform coefficients corresponding to the microphone signal DS1 and the time series of the short-time Fourier transform coefficients corresponding to the microphone signal DS2 to cross-correlate between the microphone signals DS1 and DS2. A time-series arrangement based on a spectrum time-series is generated. The cross-correlation calculating unit 33 uses the time series of the short-time Fourier transform coefficients corresponding to the microphone signal DS2 and the time series of the short-time Fourier transform coefficients corresponding to the microphone signal DS3 to A time-series arrangement based on the time series of the cross-correlation spectrum is generated. Further, the cross-correlation calculating unit 33 uses the time series of the short-time Fourier transform coefficients corresponding to the microphone signal DS3 and the time series of the short-time Fourier transform coefficients corresponding to the microphone signal DS1 to between the microphone signals DS3 and DS1. A time-series arrangement based on the time series of the cross-correlation spectrum is generated.

  Hereinafter, a time-series arrangement of time-series cross-correlation spectra is referred to as a “cross-correlation spectrogram”. In each cross-correlation spectrogram, the first dimension is a direction corresponding to a frequency (hereinafter referred to as “frequency direction”), and the second dimension is a direction corresponding to a frame (hereinafter referred to as “frame direction”). Represented by a two-dimensional array. FIG. 3A shows an example of a cross-correlation spectrogram. The cross-correlation spectrogram shown in FIG. 3A is represented by a two-dimensional array of (W / 2 + 1) rows and T columns. Here, T is the number of frames constituting each of the microphone signals DS1, DS2, and DS3, and W is the frame length of each frame.

  The autocorrelation calculating unit 34 generates a time series array based on the time series of the autocorrelation spectrum in the microphone signal DS1 using the time series of the short-time Fourier transform coefficients corresponding to the microphone signal DS1. In addition, the autocorrelation calculation unit 34 generates a time series array based on the time series of the autocorrelation spectrum in the microphone signal DS2, using the time series of the short-time Fourier transform coefficients corresponding to the microphone signal DS2. Furthermore, the autocorrelation calculation unit 34 generates a time series array based on the time series of the autocorrelation spectrum in the microphone signal DS3 using the time series of the short-time Fourier transform coefficients corresponding to the microphone signal DS3.

  Hereinafter, the time series arrangement by the time series of the autocorrelation spectrum is referred to as “autocorrelation spectrogram”. Each autocorrelation spectrogram is represented by a two-dimensional array in which the first dimension is the frequency direction and the second dimension is the frame direction. FIG. 3B shows an example of an autocorrelation spectrogram. The autocorrelation spectrogram shown in FIG. 3B is represented by a two-dimensional array of (W / 2 + 1) rows and T columns.

  The neural network unit 35 uses a so-called “neural network”. The neural network in the neural network unit 35 (hereinafter sometimes simply referred to as “neural network”) is a hierarchical type, and has an input layer, an intermediate layer, and an output layer. Each layer in the neural network is composed of a plurality of units.

  The input layer accepts input of a plurality of time series arrays generated by the time series array generation unit 32. Specifically, for example, the input layer receives inputs of three cross-correlation spectrograms generated by the cross-correlation calculation unit 33 and three auto-correlation spectrograms generated by the auto-correlation calculation unit 34. is there.

  A plurality of units included in the output layer (hereinafter referred to as “output units”) have a one-to-one correspondence with at least some of the plurality of locations in the elevator 1 (hereinafter referred to as “inspection target locations”). doing. Each location to be inspected corresponds to, for example, one of the basket 3 provided in the hoistway 2, movable parts, and other various devices.

  In a neural network, when a time-series arrangement is input to an input layer, the output value of each output unit (so-called “activity”) is a degree of deterioration at a corresponding inspection target location (hereinafter referred to as “deterioration degree”). .) Is previously trained using dedicated training data. In other words, the neural network unit 35 outputs the degree of deterioration corresponding to each of the inspection target portions in response to the input of the time series array generated by the time series array generation unit 32.

  The determination unit 36 uses the degree of deterioration output from the neural network unit 35 to determine whether or not there are deterioration points in the plurality of inspection target points, that is, whether or not there is a deterioration point in the elevator 1. is there. If the determination unit 36 determines that there is a deteriorated part, the determination unit 36 outputs a determination result indicating the deteriorated part and the degree of deterioration corresponding to the deteriorated part. On the other hand, when it is determined that there is no deterioration portion, the determination unit 36 outputs a determination result indicating that fact. The determination result output from the determination unit 36 is input to a control device (not shown) for the elevator 1, for example.

  The short-time Fourier transform unit 31, the time series array generation unit 32, the neural network unit 35, and the determination unit 36 constitute a main part of the degradation point estimation device 6.

  FIG. 4 is an explanatory diagram showing the structure of the neural network in the neural network unit according to the first embodiment of the present invention. FIG. 5 is an explanatory diagram showing a three-dimensional array in the input layer in the neural network shown in FIG. FIG. 6 is an explanatory diagram showing a three-dimensional arrangement in the convolution layer in the neural network shown in FIG. A neural network in the neural network unit 35 will be described with reference to FIGS.

  As shown in FIG. 4, the neural network has an input layer, an intermediate layer, and an output layer. In the example shown in FIG. 4, the intermediate layer is composed of one convolution layer and two nonlinear layers. The number of layers in the neural network is L + 1, and a serial number l from 0 to L is assigned to each layer. That is, L = 4 in the example shown in FIG.

  Each layer in the neural network is composed of a plurality of units. In the figure, white circles indicate individual units. The individual units included in each layer are respectively coupled to a plurality of units that are at least part of the plurality of units included in the previous layer with respect to the layer. There is no coupling between the units in each layer. The degree of coupling between units is represented by a load factor W. Each unit also has a bias B for coupling.

  Hereinafter, in this specification, following the so-called “subscript notation”, in principle, an arbitrary object is represented by a single alphabetic capital letter, and the alphabetic capital letter following that alphabetic capital letter (ie, a subscript) It represents any element related to the object. For example, in each layer in the neural network, an index indicating a unit included in the layer is k, and an index indicating a unit included in the previous layer with respect to the layer is i. The load coefficient W indicating the degree of coupling between the unit k included in the layer l and the unit i included in the previous layer with respect to the layer l is expressed as “Wlik”. The bias B included in the unit k included in the layer l is denoted as “Blk”.

  However, as a result of substituting the value of the corresponding element for the subscript, the subscript may include uppercase letters, numbers, or symbols. For example, in FIG. 4, the load coefficient WLik indicating the degree of coupling between the unit k included in the output layer l (l = L) and the unit i included in the nonlinear layer l (l = L−1) is “WLik”. It is written. The load coefficient Wlik indicating the degree of coupling between the unit k included in the nonlinear layer l (l = L−1) and the unit i included in the nonlinear layer l (l = 2) is expressed as “WL-1ik”. Has been.

  The input layer has a structure in which a plurality of units arranged along a first dimension and a second dimension orthogonal to each other (hereinafter referred to as a “two-dimensional array surface”) are layered, that is, a two-dimensional array. It has a three-dimensional array structure in which the surface stacking direction is the third dimension. The two-dimensional array plane included in the input layer has a one-to-one correspondence with a plurality of time-series arrays input to the neural network unit 35. That is, a corresponding time series array is input to each two-dimensional array surface included in the input layer.

FIG. 5 shows an example of a three-dimensional array in the input layer. In the example shown in FIG. 5, the input layer has a three-dimensional array structure in which the (W + 2 + 1) rows and T columns two-dimensional array surfaces are (M + COMBINATIONS (M, 2)) area layers. Here, M is the number of microphones 13 _{1 to} 13 ₃ included in the microphone array device 4, and COMBINATIONS (M, 2) is a function for obtaining the number of combinations for selecting two of the M. In the first embodiment, M = 3, and the number of combinations according to COMBINATIONS (M, 2) is three. Each two-dimensional array surface has one of the three cross-correlation spectrograms generated by the cross-correlation calculating unit 33 or three auto-correlation spectrograms generated by the auto-correlation calculating unit 34. Any one of is input.

  The convolution layer performs a convolution operation on the three-dimensional region Ωgtf in the input layer. As shown in FIG. 4, the three-dimensional region Ωgtf includes a plurality of units. Here, the subscript g is an index indicating a two-dimensional array surface on which each unit in the three-dimensional region Ω is arranged, the subscript t is an index indicating a frame corresponding to each unit in the three-dimensional region Ω, and the subscript f Is an index indicating the frequency corresponding to each unit in the three-dimensional region Ω.

  FIG. 6 shows an example of a three-dimensional arrangement in the convolution layer. In the example shown in FIG. 6, the convolution layer has a three-dimensional array structure in which a two-dimensional array surface of ((W / 2 + 1) / DF) rows (T / DT) columns is a D area layer. That is, each two-dimensional array surface included in the convolution layer has a shape obtained by down-sampling each two-dimensional array surface included in the input layer in the frame direction and the frequency direction. Here, DT is a thinning interval in the frame direction, and DF is a thinning interval in the frequency direction. If the value of DT is 1, there is no decimation in the frame direction, and if the value of DT is D (D is an integer of 2 or more), it indicates that downsampling is performed at intervals of D-1. Similarly, if the value of DF is 1, there is no decimation in the frequency direction, and if the value of DF is D (D is an integer of 2 or more), it indicates that downsampling is performed at an interval of D-1. Yes. The two-dimensional array plane included in the convolution layer has a one-to-one correspondence with the feature d (d = 1,..., D) obtained by the convolution operation.

  During the processing by the neural network unit 35, the arrangement shape of the units in the convolution layer is converted from a three-dimensional arrangement to a one-dimensional arrangement (that is, a vector). Each layer (l = 2,..., L) after the convolution layer has a one-dimensional array structure of a plurality of units.

  Each of the two nonlinear layers has a one-dimensional arrangement of a plurality of units. Each unit included in the nonlinear layer has a nonlinear output function. As this output function, for example, a sigmoid function or a TANH function is used.

  The output layer has a one-dimensional array structure with a plurality of output units. As described above, the output unit has a one-to-one correspondence with the inspection target portion in the elevator 1. The neural network is trained in advance so that the activity of each output unit indicates the degree of deterioration at the corresponding inspection target location.

  Next, the operation of the degradation point estimation device 6 will be described with reference to the flowchart of FIG.

During operation of the elevator 1, the audio interface unit 5 acquires the analog signal AS1, AS2, AS3 the microphone _{131-134 3} has outputted, and converts the analog signal AS1, AS2, AS3 into digital signals DS1, DS2, DS3 Then, the process of outputting the digital signals DS1, DS2, DS3 to the degradation point estimation device 6 is continuously executed. Digital signals DS 1, DS 2, DS 3 output from the audio interface device 5 are temporarily stored in the memory 23. The degradation point estimation device 6 performs the following steps ST1 to ST4 on the digital signals DS1, DS2, DS3 stored in the memory 23, that is, the microphone signals DS1, DS2, DS3.

  First, in step ST1, the short-time Fourier transform unit 31 acquires the microphone signals DS1, DS2, DS3 stored in the memory 23. The short-time Fourier transform unit 31 calculates a time series of short-time Fourier transform coefficients corresponding to each of the microphone signals DS1, DS2, and DS3 by executing a short-time Fourier transform on each of the microphone signals DS1, DS2, and DS3. To do. Each time series is composed of a plurality of short-time Fourier transform coefficients, and each short-time Fourier transform coefficient indicates a complex frequency spectrum.

  Next, in step ST2, the time series array generation unit 32 generates a plurality of time series arrays using the time series of the short time Fourier transform coefficients calculated by the short time Fourier transform unit 31 in step ST1. Specifically, for example, the time series array generation unit 32 includes a cross-correlation spectrogram between the microphone signals DS1 and DS2, a cross-correlation spectrogram between the microphone signals DS2 and DS3, and a cross-correlation spectrogram between the microphone signals DS3 and DS1. The autocorrelation spectrogram in the microphone signal DS1, the autocorrelation spectrogram in the microphone signal DS2, and the autocorrelation spectrogram in the microphone signal DS3 are generated.

  Next, in step ST3, the neural network unit 35 receives an input of a plurality of time series arrays generated by the time series array generation unit 32 in step ST2. The neural network unit 35 outputs a degree of deterioration corresponding to each of the inspection target portions in response to the time-series array input.

  Next, in step ST4, the determination unit 36 determines the presence / absence of a deterioration portion in the elevator 1 using the deterioration degree output by the neural network unit 35 in step ST3. If the determination unit 36 determines that a deterioration location exists, the determination unit 36 outputs a determination result indicating the deterioration location and the degree of deterioration corresponding to the deterioration location. On the other hand, when it is determined that there is no deterioration portion, the determination unit 36 outputs a determination result indicating that fact.

  Next, the details of the process of step ST1 shown in FIG. 7, that is, the process by the short-time Fourier transform unit 31, will be described with reference to the flowchart of FIG.

  First, the short-time Fourier transform unit 31 acquires the microphone signals DS1, DS2, DS3 stored in the memory 23 (step ST11).

Next, the short-time Fourier transform unit 31 selects any _one of the microphones 13 _{1 to} 13 ₃ (step ST12). Hereinafter, the number indicating the selected microphone is m (m is an integer satisfying 0 ≦ m <M). The short-time Fourier transform unit 31 performs the following steps ST13 to ST18 on the microphone signal corresponding to the selected microphone m. Here, it is assumed that the microphone 13 ₁ corresponding to m = 0 as an example has been selected.

  The microphone signal corresponding to the microphone m is composed of T frames, and the number indicating each frame is t. In addition, a signal string in the microphone signal is expressed as g (m, i). Here, i represents a discrete time index in the signal sequence g, and is an integer that satisfies 0 ≦ i <NS (NS is the number of sound collection samples). The short-time Fourier transform unit 31 performs the following steps ST14 to T18 for each frame t (step ST13).

First, short-time Fourier transform unit 31, with respect to the microphone signal DS1 corresponding to the microphone 13 ₁ selected, each time the frame number t is increased by one, to move the frame position by LF point (step ST14), the frame number A frame is cut out at a frame position corresponding to t (step ST15). The processing of steps ST14 and ST15 is realized by the following equation (1), for example.

frame (i)
= G (m, t × LF + i) (i = 0, 1, 2,..., W−1) (1)

  Here, W is the frame length of each frame, and LF is the amount of movement of the frame. In the first embodiment, W = 1024 and LF = 512 are set.

  Next, the short-time Fourier transform unit 31 multiplies the signal indicating the extracted frame by a time window (step ST16). The process of step ST16 is realized by the following equation (2), for example.

x (i)
= Frame (i) × window (i) (i = 0, 1, 2,..., W−1)
window (i)
= 0.54-0.46 cos (2πi / (W-1)) (2)

  Here, window (i) is a time window function. In the example of the above formula (2), a Hamming window function is used.

  Next, the short-time Fourier transform unit 31 performs a fast Fourier transform (FFT) on the sequence x (i) obtained by the multiplication of the window function (step ST17). The process of step ST17 is realized by the following equation (3), for example.

  Zf = FFT (x [0: W]) (f = 0, 1, 2,..., W−1) (3)

  Here, Zf represents a complex frequency spectrum, and x [0: W] represents a sequence [x (0), x (1),..., X (W−1)]. FFT (x) represents an FFT operation function that performs a fast Fourier transform on the sequence x.

  Next, the short-time Fourier transform unit 31 separates the real part and the imaginary part of the complex frequency spectrum Zf and stores them in the memory 23 (step ST18). The process of step ST18 is realized by the following equation (4), for example.

Xmtf = Re (Zf) (0 ≦ f ≦ W / 2)
Ymtf = Im (Zf) (0 ≦ f ≦ W / 2) (4)

  Here, Xmtf represents the real part of the complex frequency spectrum Zf at the microphone m, frame t, and frequency f, and Ymtf represents the imaginary part of the complex frequency spectrum Zf at the microphone m, frame t, and frequency f. Re (c) is a function for extracting the real part a of the complex number c = a + bj, and Im (c) is a function for extracting the imaginary part b of the complex number c = a + bj. Further, j is an imaginary unit and is a square root of −1 (that is, j ** 2 = −1, where ** is a power operator).

  Since the input of the FFT operation is a real number sequence as shown in the above equation (3), the real part of Zf has even symmetry Xmtf = XmtW−f with respect to f, and the imaginary part of Zf has odd symmetry with respect to f. Ymtf = −YmtW−f. Therefore, the short-time Fourier transform unit 31 truncates the redundant portion of f> W / 2 and stores only the portion of 0 ≦ f ≦ W / 2 in the memory 23.

Short-time Fourier transform unit 31, the microphone signal DS2, DS3 For each of (step ST12) corresponding to the microphone ₁₃ 2, 13 ₃ of the residual executes processing similar to the above steps ST13～ST18. As a result, the memory 23 stores the time series of the complex frequency spectrum Zf corresponding to each of the microphone signals DS1, DS2, and DS3, that is, the time series of the short-time Fourier transform coefficients.

  Next, with reference to the flowchart of FIG. 9, the detail of the process by the cross correlation calculating part 33 among the processes of step ST2 shown in FIG. 7 is demonstrated.

First, the cross-correlation calculating unit 33 selects a combination of any two microphones among the microphones 13 _{1 to} 13 ₃ (step ST21). Hereinafter, it is assumed that a number indicating one of the two selected microphones is m, and a number indicating the other microphone is n. Each of m and n is an integer satisfying 0 ≦ m <n <M.

  The cross-correlation calculating unit 33 executes the following processes of steps ST22 to ST25 for the time series of the complex frequency spectrum corresponding to each of the selected microphones m and n. That is, the cross-correlation calculating unit 33 performs the following steps ST23 to ST25 for each frame t (step ST22).

  First, the cross-correlation calculating unit 33 acquires the real part Xmtf and the imaginary part Ymtf of the complex frequency spectrum corresponding to the microphone m and the real part Xntf and the imaginary part Yntf of the complex frequency spectrum corresponding to the microphone n from the memory 23. . The cross-correlation calculating unit 33 calculates a complex cross-correlation spectrum Cf by multiplying the complex frequency spectrum corresponding to the microphone m and the complex frequency spectrum corresponding to the microphone n (step ST23). The process of step ST23 is realized by the following equation (5), for example.

Cf = (Xmtf * Xntf + Ymtf * Yntf)
+ (Xmtf * Yntf-Ymtf * Xntf) j (5)

  Next, the cross-correlation calculating unit 33 takes the absolute value of the complex cross-correlation spectrum Cf and takes the decibel value (dB value) by converting the absolute value into a logarithm. As a result, the complex cross-correlation spectrum Cf is converted into intensity (step ST24), and the dynamic range is compressed. The process of step ST24 is implement | achieved by the following formula | equation (6), for example.

  Hf = 10 * LOG10 (ABS (Cf)) (0 ≦ f ≦ W / 2) (6)

  Here, Hf is the intensity of the cross-correlation spectrum, LOG10 (x) is a common logarithmic function, and ABS (c) is an absolute value function. The absolute value function is a function for calculating an absolute value by SQRT (a ** 2 + b ** 2) (SQRT is a square root function) for a complex number c = a + bj (a and b are real numbers and j is an imaginary unit). .

  In the memory 23, an area for storing a time-series arrangement of the cross-correlation spectrum Hf corresponding to the combination (m, n) in time series, that is, a cross-correlation spectrogram Hmntf is prepared. The cross correlation calculation unit 33 stores the calculated Hf at an address corresponding to the frame t in the region (step ST25). The process of step ST25 is implement | achieved by the following formula | equation (7), for example.

  Hmntf = Hf (0 ≦ f ≦ W / 2) (7)

  When the processing of steps ST23 to ST25 is completed for all frames t, the cross-correlation calculating unit 33 executes the processing of steps ST22 to ST25 for the next combination (m, n) (step ST21). Finally, the cross-correlation calculating unit 33 executes the processes of steps ST22 to ST25 for all combinations (m, n).

  As a result of the above processing, the memory 23 stores the time series of the cross-correlation spectrum between the microphone signals DS1 and DS2, the time series of the cross-correlation spectrum between the microphone signals DS2 and DS3, and the cross-correlation spectrum between the microphone signals DS3 and DS1. Is stored, that is, a cross-correlation spectrogram Hmntf for each time series is stored. Each cross-correlation spectrogram Hmntf satisfies the following formula (8).

Hmntf (0 ≦ m <n <M, 0 ≦ t <T, 0 ≦ f ≦ W / 2) (8)

  Next, with reference to the flowchart of FIG. 10, the details of the process by the autocorrelation calculation unit 34 in the process of step ST2 shown in FIG. 7 will be described.

First, the autocorrelation calculating unit 34 selects any _one of the microphones 13 _{1 to} 13 ₃ (step ST31). Hereinafter, the number indicating the selected microphone is m (m is an integer satisfying 0 ≦ m <M).

  The autocorrelation calculation unit 34 performs the following processes of steps ST32 to ST35 on the time series of the complex frequency spectrum corresponding to the selected microphone m. That is, the autocorrelation calculating unit 34 performs the following steps ST33 to ST35 for each frame t (step ST32).

  First, the autocorrelation calculating unit 34 acquires the real part Xmtf and the imaginary part Ymtf of the complex frequency spectrum corresponding to the microphone m from the memory 23. The autocorrelation calculating unit 34 calculates the autocorrelation spectrum Af by multiplying the complex frequency spectrum corresponding to the microphone m and the complex conjugate of the complex frequency spectrum (step ST33). The process of step ST33 is realized by the following equation (9), for example. In addition, since the imaginary part of the autocorrelation spectrum Af is 0 according to the theory of Fourier transform, the calculation of the imaginary part is omitted in Equation (9).

Af
= Xmtf * Xmtf + Ymtf * Ymtf (0 ≦ f ≦ W / 2) (9)

  Next, the autocorrelation calculation unit 34 takes an absolute value of the autocorrelation spectrum Af, and takes the decibel value (dB value) by converting the absolute value into a logarithm. Thereby, the autocorrelation spectrum Af is converted into the intensity Gf (step ST34), and the dynamic range is compressed. The process of step ST34 is realized by the following equation (10), for example.

Gf
= 10 * LOG10 (ABS (Af)) (0 ≦ f ≦ W / 2) (10)

  The memory 23 is provided with an area for storing a time-series arrangement of the autocorrelation spectrum corresponding to the microphone m, that is, an autocorrelation spectrogram Gmtf. The autocorrelation calculating unit 34 stores the calculated Gf at an address corresponding to the frame t in the region (step ST35). The process of step ST35 is implement | achieved by the following formula | equation (11), for example.

  Gmtf = Gf (0 ≦ f ≦ W / 2) (11)

  When the processing of steps ST33 to ST35 is completed for all frames t, autocorrelation calculation unit 34 executes the processing of steps ST32 to ST35 for the next microphone m (step ST31). Finally, the autocorrelation calculating unit 34 executes the processes of steps ST32 to ST35 for all the microphones m.

  As a result of the above processing, the memory 23 stores the time series of the autocorrelation spectrum in the microphone signal DS1, the time series of the autocorrelation spectrum in the microphone signal DS2, and the time series of the autocorrelation spectrum in the microphone signal DS3. In other words, the autocorrelation spectrogram Gmtf for each time series is stored. Each autocorrelation spectrogram Gmtf satisfies the following expression (12).

  Gmtf (0 ≦ m <M, 0 ≦ t <T, 0 ≦ f ≦ W / 2) (12)

  Next, with reference to FIGS. 4 to 6 and the flowchart of FIG. 11, the details of the process of step ST3 shown in FIG.

  First, the neural network unit 35 acquires a plurality (three in the first embodiment) cross-correlation spectrograms Hmntf and a plurality (three in the first embodiment) autocorrelation spectrograms Gmtf from the memory 23. The neural network unit 35 is a three-dimensional array in which a two-dimensional array surface corresponding to each of the cross-correlation spectrograms Hmntf and a two-dimensional array surface corresponding to each of the autocorrelation spectrograms Gmtf are stacked, that is, the three-dimensional array shown in FIG. Is generated (step ST41). The process of step ST41 is realized by writing a value in the memory 23 according to the following equation (13), for example.

g ← 0
for much that 0 ≦ m <M:
Egtf = Gmtf (0 ≦ f ≦ W / 2)
g ← g + 1
for much that 0 ≦ m <M−1:
for m such that m <n <M:
Egtf = Hmntf (0 ≦ f ≦ W / 2)
g ← g + 1 (13)

  Here, Egtf is a value of a three-dimensional array corresponding to the frequency f, the frame t, and the two-dimensional array surface g. The neural network unit 35 inputs the generated three-dimensional array to the input layer of the neural network.

  Next, the neural network unit 35 calculates the activity of the corresponding unit in the convolution layer from the activity of each unit in the input layer (step ST42). At this time, the convolution layer performs a convolution operation on the three-dimensional region Ωgtf in the input layer. The process of step ST42 is realized by the following equation (14), for example.

l ← 1
Alk = OFUN (Σgtf⊂Ωgtf (k) Wltfk * Egtf + Bltk)
(14)

  Here, l is an index indicating each layer in the neural network (here, l = 1 indicating a convolution layer), k is an index indicating each unit in the convolution layer, Alk is an activity of unit k, and OFUN is an output. The function Σ is a sum operation. Further, Ωgtf (k) is a set of indexes (g, t, f) indicating each unit in the input layer coupled to the unit k. Each (g, t, f) in the set indicates a two-dimensional array surface on which the unit is arranged in the input layer, a frame corresponding to the unit in the input layer, and a frequency corresponding to the unit in the input layer. It is an index. Wltffk represents the load coefficient of unit k relative to Egtf, and Bltk represents the bias of unit k. Note that the index k in the convolution layer performs a convolution operation so that all units included in the convolution layer can be licked.

  Next, the neural network unit 35 converts the array shape of the units in the convolution layer from a three-dimensional array to a one-dimensional array (step ST43). This processing is realized by storing the above Alk in a one-dimensional array with k as an index.

  Next, the neural network unit 35 sequentially executes the processing of the following step ST45 for each of the nonlinear layers l (l = 2,..., L−1) (step ST44). That is, the neural network unit 35 calculates the activity of the corresponding unit in the layer l from the activity of each unit in the layer p (p = 1-1) (step ST45). More specifically, the neural network unit 35 multiplies the activity of each unit in the layer p by a load coefficient to obtain a sum of multiplication values related to a plurality of units, and adds a bias to the value of the sum. Further, the activity of the corresponding unit in the layer l is calculated by passing the output function, respectively. The process of step ST45 is realized by the following equation (15), for example.

p ← l-1
Alk = OFUN (Σi Wlik * Api + Blk) (15)

  Here, Alk is the activity of the unit k in the layer l, OFUN is the output function, Σi is the summation operation for the unit i in the layer p coupled to the unit k, Api is the activity of the unit i in the layer p, Blk Is a bias. As the output function OFUN, a sigmoid function or a TANH function which is a nonlinear function can be used.

  Next, the neural network unit 35 calculates the activity of the corresponding unit in the output layer L from the activity of each unit in the nonlinear layer L-1 that is the previous layer with respect to the output layer L (step ST46). The process of step ST46 is realized by the following equation (16), for example.

p ← L-1
ALk = Σi WLik * Api + BLk (0 ≦ k <K) (16)

  Here, ALk is the activity of the unit k in the layer L, Σi is the summation operation for the unit i in the layer p (p = L−1) coupled to the unit k, and Api is the layer p (p = L−1). Of the unit i, BLk is a bias, and K is the number of inspection object portions in the elevator 1. In the first embodiment, since the neural network unit 35 calculates a linear deterioration degree, an output function is not provided in the output layer, and nonlinear processing using the output function is not executed.

  Hereinafter, the subscript k indicates each unit in the output layer L, and indicates an inspection target portion corresponding to the unit. In the memory 23, an area for storing the deterioration degree Rk corresponding to each of the inspection target locations k is prepared. The neural network unit 35 stores the activity degree ALk calculated by the above equation (16) in the area of the memory 23 as the deterioration degree Rk (step ST47). The process of step ST47 is realized by the following equation (17), for example.

  Rk = ALk (0 ≦ k <K) (17)

  Next, the details of the process of step ST4 shown in FIG. 7, that is, the process by the determination unit 36 will be described with reference to the flowchart of FIG.

  First, the determination unit 36 acquires the deterioration degree Rk corresponding to each of the inspection target locations k from the memory 23 (step ST51).

  Next, the determination unit 36 obtains the maximum deterioration degree R * of the acquired deterioration degrees Rk and the index k * indicating the inspection target location k corresponding to the maximum deterioration degree R * (step ST52). The process of step ST52 is realized by the following equation (18), for example.

k * = ARGMAXk Rk
R * = Rk * (18)

  Here, ARGMAXk Rk is a function that returns an index of Rk that takes the maximum value with respect to the subscript k (where k * satisfies Rk * ≧ Rk).

  Next, the determination unit 36 compares the maximum deterioration degree R * with a predetermined threshold value θ (step ST53).

  When the maximum deterioration degree R * is equal to or less than the threshold value θ (step ST53 “NO”), the determination unit 36 determines that there is no deterioration portion (step ST54). The determination unit 36 outputs a determination result in which the index indicating the degradation location is −1 and the degradation level corresponding to the degradation location is 0 (step ST55). In other words, the determination result in this case indicates that there is no deterioration portion in the elevator 1.

  On the other hand, when the maximum deterioration degree R * exceeds the threshold θ (step ST53 “YES”), the determination unit 36 determines that a deterioration portion exists (step ST56). The determination unit 36 outputs a determination result in which the index indicating the degradation location is k * and the degradation level corresponding to the degradation location is R * (step ST57). In other words, the determination result in this case indicates the degradation location in the elevator 1 and the degradation level corresponding to the degradation location.

  The time series array generated by the time series array generation unit 32 is not limited to the cross-correlation spectrogram Hmntf and the autocorrelation spectrogram Gmtf. For example, the time series array generation unit 32 may generate only the autocorrelation spectrogram Gmtf without generating the cross correlation spectrogram Hmntf. Thereby, since only the autocorrelation spectrogram Gmtf is input to the neural network, the number of parameters in the neural network can be reduced.

  Alternatively, for example, the time series array generation unit 32 replaces the cross-correlation spectrogram Hmntf and the autocorrelation spectrogram Gmtf with a time series based on the real part of the autocorrelation spectrum Af corresponding to each of the microphone signals DS1, DS2, and DS3. An array and a time-series array based on a time series of the imaginary part of the autocorrelation spectrum Af corresponding to each of the microphone signals DS1, DS2, and DS3 may be generated. Similar to the cross-correlation spectrogram Hmntf and the autocorrelation spectrogram Gmtf, these time-series arrays can be represented by a two-dimensional array in which the first dimension is the frequency direction and the second dimension is the frame direction. In this case, it is required to execute processing corresponding to intensity conversion and cross-correlation calculation in the neural network. For this reason, it is preferable to increase the number of layers in the neural network from the viewpoint of maintaining the accuracy of output by the neural network.

  Alternatively, for example, the time series array generation unit 32 replaces the cross-correlation spectrogram Hmntf and the autocorrelation spectrogram Gmtf with a time series based on the real part time series of the complex frequency spectrum Zf corresponding to each of the microphone signals DS1, DS2, DS3. An array and a time series array by a time series of the imaginary part of the complex frequency spectrum Zf corresponding to each of the microphone signals DS1, DS2, and DS3 may be generated. Similar to the cross-correlation spectrogram Hmntf and the autocorrelation spectrogram Gmtf, these time-series arrays can be represented by a two-dimensional array in which the first dimension is the frequency direction and the second dimension is the frame direction. In this case, it is required to execute processing corresponding to intensity conversion, autocorrelation calculation, cross-correlation calculation, and the like in the neural network. For this reason, it is preferable to increase the number of layers in the neural network from the viewpoint of maintaining the accuracy of output by the neural network.

  In addition, the determination result output from the determination unit 36 when it is determined that there is a deteriorated portion may be at least indicating the deteriorated portion, and may not indicate the degree of deterioration corresponding to the deteriorated portion.

  Moreover, the determination part 36 should just perform some determination regarding a degradation location using the degradation degree Rk which the neural network part 35 output, and the said determination is not limited to determination of the presence or absence of a degradation location. .

  Further, the microphone array device 4 may be any one having a plurality of microphones, and the number of the microphones is not limited to three. For example, the microphone array device 4 may have four microphones. In this case, these microphones may be arranged along the circumferential portion of the placement surface 12 at positions corresponding to the vertices of the regular square. By increasing the number of microphones, the resolution with respect to the direction of arrival of the sound is improved, so that it is possible to identify inspection target locations arranged closer to each other. As the number of microphones increases, it is required to increase the number of parameters in the neural network.

  Further, the microphone array device 4 may be provided on the floor portion of the cage 3 instead of the ceiling portion of the cage 3. In addition, the microphone array device 4 may be provided at any location in the hoistway 2. However, from the viewpoint of reducing the number of microphones while improving the resolution with respect to the direction of sound arrival, the microphone array device 4 is preferably provided on a moving body such as the basket 3.

  Further, the inspection target device is not limited to the elevator 1. The inspection target device may be, for example, a railway vehicle. In this case, the microphone array device 4, the audio interface device 5, and the degradation location estimation device 6 may be provided in the railway vehicle. In addition, the inspection target device may be any device as long as it is a device including a moving body or a device configured by a moving body.

As described above, the degradation point estimation device 6 according to the first embodiment acquires the microphone signals DS1, DS2, DS3 corresponding to each of the plurality of microphones 13 _{1 to} 13 ₃ provided in the inspection target device, A short-time Fourier transform unit 31 that calculates a time series of short-time Fourier transform coefficients corresponding to each of the microphone signals DS1, DS2, and DS3, and a time series for input to the neural network using the time series of short-time Fourier transform coefficients A time-series array generating unit 32 that generates an array, and a neural network, a neural network unit 35 that receives an input of the time-series array and outputs a degree of deterioration corresponding to the inspection target location in the inspection target device; And a determination unit 36 that determines the deterioration location in the inspection target device using the degree. As a result, it is possible to consider the direction of arrival of the sound and improve the estimation accuracy of the deteriorated part.

The inspection target device is intended to include mobile, microphone _{131-134 3} is provided on the moving body. As a result, the number of microphones can be reduced while allowing the direction of sound arrival to be taken into account.

Further, the degradation point estimation system 100 of the first embodiment includes a plurality of microphones 13 _{1 to} 13 ₃ provided in the inspection target device. Deterioration location estimation system 100 acquires microphone signals DS1, DS2, DS3 corresponding to each of microphones 13 _{1 to} 13 ₃ and time-series Fourier transform coefficients corresponding to each of microphone signals DS1, DS2, DS3. A short-time Fourier transform unit 31 that calculates the time series, a time-series array generation unit 32 that generates a time-series array for input to the neural network using the time series of the short-time Fourier transform coefficients, and a neural network, A deterioration having a neural network unit 35 that receives an input of a time-series arrangement and outputs a deterioration degree corresponding to the inspection target part in the inspection target device, and a determination part 36 that determines the deterioration point in the inspection target device using the deterioration degree. A location estimation device 6 is provided. As a result, it is possible to consider the direction of arrival of the sound and improve the estimation accuracy of the deteriorated part.

Further, in the failure location estimation method of the first embodiment, the short-time Fourier transform unit 31 uses the microphone signals DS1, DS2, and DS3 corresponding to the plurality of microphones 13 _{1 to} 13 ₃ provided in the inspection target device. A step of acquiring and calculating a time series of short-time Fourier transform coefficients corresponding to each of the microphone signals DS1, DS2, and DS3 (step ST1), and the time-series array generation unit 32 generates a time series of short-time Fourier transform coefficients. Generating a time series array for input to the neural network using the step (step ST2), and the neural network unit 35 constituted by the neural network accepts the input of the time series array and places it in the inspection target location in the inspection target device. The step of outputting the corresponding deterioration degree (step ST3) and the determination unit 36 use the deterioration degree. And a step of determining the deterioration point in the inspection target device (step ST4) Te. As a result, it is possible to consider the direction of arrival of the sound and improve the estimation accuracy of the deteriorated part.

  In the present invention, any constituent element of the embodiment can be modified or any constituent element of the embodiment can be omitted within the scope of the invention.

  The degradation point estimation apparatus of the present invention can be used for estimation of a degradation point in an inspection target device such as an elevator or a railway vehicle, for example.

1 elevator, 2 hoistway, 3 car, 4 microphone array system, 5 audio interface device, 6 deterioration location estimating apparatus, 11 pedestal, 12 mounting _surface, 13 _1, 13 2, 13 ₃ microphones, 21 input-output interface, 22 Processor, 23 Memory, 31 Short-time Fourier transform unit, 32 Time series array generation unit, 33 Cross-correlation calculation unit, 34 Autocorrelation calculation unit, 35 Neural network unit, 36 Judgment unit, 100 Deterioration location estimation system

Claims (12)

A short-time Fourier transform unit that acquires a microphone signal corresponding to each of a plurality of microphones provided in the inspection target device and calculates a time series of short-time Fourier transform coefficients corresponding to each of the microphone signals;
A time series array generation unit for generating a plurality of time series arrays for input to a neural network using the time series of the short-time Fourier transform coefficients;
A neural network unit configured by the neural network, which receives an input of the time series array and outputs a degree of deterioration corresponding to the inspection target portion in the inspection target device;
A determination unit that determines a deterioration location in the inspection target device using the deterioration degree,
A degradation point estimation device comprising:   The time series array generation unit uses the time series of the short-time Fourier transform coefficients to generate the time series array based on the time series of the autocorrelation spectrum in the microphone signal and the time based on the cross correlation spectrum between the microphone signals. The degradation location estimation apparatus according to claim 1, wherein a sequence array is generated.   The time series array generation unit generates the time series array based on the time series of the autocorrelation spectrum in the microphone signal using the time series of the short-time Fourier transform coefficients. Deterioration location estimation device.   The time series array generation unit uses the time series of the short-time Fourier transform coefficients, and uses the time series array of the real part time series of the short-time Fourier transform coefficients and the short-time Fourier transform coefficients. The deterioration location estimation apparatus according to claim 1, wherein the time series array is generated by a time series of the imaginary part.   2. The degradation point estimation apparatus according to claim 1, wherein the neural network includes an input layer that receives a plurality of inputs of the time series array, and a convolution layer that performs a convolution operation on the input layer. . The input layer has a three-dimensional array structure in which a two-dimensional array surface by a plurality of units is stacked, and the time-series array corresponding to each of the two-dimensional array surfaces is input.
The degradation point estimation device according to claim 5, wherein the convolution layer performs the convolution operation on a three-dimensional region in the input layer.   The degradation point estimation apparatus according to claim 1, wherein the microphone is arranged at a position corresponding to each vertex of a regular polygon.   The degradation location estimation apparatus according to claim 1, wherein the inspection target device includes a moving body, and the microphone is provided in the moving body.   The said inspection object apparatus is comprised by the elevator, the said mobile body is comprised by the cage | basket | car of the said elevator, The said microphone is provided in the ceiling part or floor | floor part of the said basket, It is characterized by the above-mentioned. Deterioration location estimation device as described.   The determination unit determines the presence or absence of the deterioration portion, and outputs a determination result indicating the deterioration portion and the degree of deterioration corresponding to the deterioration portion when it is determined that the deterioration portion exists. The degradation location estimation apparatus according to claim 1. A plurality of microphones provided in the device to be inspected;
A short-time Fourier transform unit that obtains a microphone signal corresponding to each of the microphones and calculates a time series of short-time Fourier transform coefficients corresponding to each of the microphone signals, and a time series of the short-time Fourier transform coefficients A time-series array generating unit that generates a plurality of time-series arrays for input to the neural network, and the neural network, and accepts the input of the time-series array, and the inspection target location in the inspection target device A deterioration location estimation apparatus comprising: a neural network unit that outputs a deterioration level corresponding to the determination unit; and a determination unit that determines a deterioration location in the inspection target device using the deterioration level;
Deterioration location estimation system with A step of obtaining a microphone signal corresponding to each of a plurality of microphones provided in the device to be inspected by the short-time Fourier transform unit and calculating a time series of short-time Fourier transform coefficients corresponding to each of the microphone signals; When,
A time series array generating unit generating a plurality of time series arrays for input to the neural network using the time series of the short-time Fourier transform coefficients;
A step of receiving a time-series array input and outputting a degree of deterioration corresponding to an inspection target location in the inspection target device, wherein the neural network unit configured by the neural network;
A step of determining a deterioration portion in the inspection target device using the deterioration degree;
A degradation location estimation method comprising:

Images