JP7275324B2 - Wire rope flaw detector - Google Patents

Description

The present invention relates to a wire rope flaw detector.

Conventionally, there is known a wire rope flaw detector that has a magnetizer that magnetically saturates a wire rope and a magnetic sensor that detects leakage magnetic flux that leaks from the wire rope due to a damaged portion of the wire rope (see, for example, Patent Document 1).

JP-A-09-210968

The amount of leakage magnetic flux leaking from the wire rope decreases as the wire rope becomes thinner. Therefore, in the conventional wire rope flaw detector disclosed in Patent Document 1, as the wire rope becomes thinner, the amount of leakage magnetic flux reaching the magnetic sensor decreases, and the output of the magnetic sensor decreases. As a result, the SN ratio of the conventional wire rope flaw detector is lowered.

Further, in the conventional wire rope flaw detector, even if an attempt is made to improve the SN ratio by shortening the distance between the wire rope and the magnetic sensor, there are restrictions such as assembly accuracy of the wire rope and the magnetic sensor. Therefore, the conventional wire rope flaw detector has a limit in shortening the distance between the wire rope and the magnetic sensor.

SUMMARY OF THE INVENTION An object of the present invention is to provide a wire rope flaw detector capable of more reliably improving the SN ratio.

A wire rope flaw detector according to the present invention includes a magnetizer that generates a magnetic flux that passes through a portion of a wire rope, and a magnetic sensor that generates, as a sensor signal, a signal corresponding to leakage magnetic flux leaking from the wire rope among the magnetic flux. , a control unit that processes the sensor signal, the control unit extracting a filter unit that extracts the frequency component of the sensor signal, and a plurality of feature amounts based on a plurality of values that constitute the frequency component. and a plurality of frequency components extracted by the filter unit are added to a learning model that has undergone machine learning on the correlation between the plurality of feature values and the state of the wires included in the wire rope. a learning unit that determines the presence or absence of damage to the wire by executing arithmetic processing by the learning model when a plurality of feature values extracted by the arithmetic unit are input based on the value of there is

According to the wire rope flaw detector according to the present invention, the SN ratio can be improved more reliably.

2 is an exploded perspective view showing the probe of the wire rope flaw detector according to Embodiment 1. FIG. FIG. 2 is an explanatory diagram showing the principle of flaw detection by the probe of FIG. 1; FIG. 3 is an enlarged view of part A in FIG. 2 ; FIG. 4 is a diagram more specifically explaining an example of the positional relationship between the leakage magnetic flux and the coil in FIG. 3 ; 5 is a diagram for more specifically explaining an example of the positional relationship between leakage magnetic flux leaking from a wire rope having a smaller diameter than the wire rope of FIG. 4 and a coil; FIG. FIG. 3 is a block diagram showing a functional configuration example of a control unit in FIG. 2; FIG. 7 is a diagram showing an example of frequency characteristics of the filter section of FIG. 6; 7 is a diagram showing an example of the distribution of frequency components extracted from the input signal by the filter section of FIG. 6; FIG. FIG. 7 is a diagram showing an example of a sequence yk(n) generated by a filter unit at time t1 in FIG. 6; FIG. 7 is a diagram showing an example of a sequence yk(n) generated by a filter unit at time t2 in FIG. 6; FIG. 7 is a diagram showing the concept of performing machine learning using a learning data set when the learning unit in FIG. 6 is configured as a support vector machine; FIG. 7 is a flowchart illustrating processing by a control unit when the learning unit in FIG. 6 is configured as a support vector machine; FIG. 7 is a diagram showing the concept of performing machine learning using a learning data set when the learning unit in FIG. 6 is configured as an autoencoder; FIG. 7 is a flowchart for explaining processing by the control unit when the learning unit in FIG. 6 is configured as an autoencoder; 10 is a flowchart for explaining processing by a control unit in Embodiment 2. FIG. FIG. 11 is a block diagram showing a functional configuration example of a control unit that processes a signal according to leakage magnetic flux leaking from a wire rope according to Embodiment 3; FIG. 17 is a diagram showing an example of frequency characteristics of the filter section of FIG. 16; FIG. 17 is a diagram showing an example of a time-domain waveform of a mother wavelet by the wavelet transform unit of FIG. 16; FIG. 17 is a diagram showing an example of a waveform in the frequency domain of the mother wavelet by the wavelet transform unit in FIG. 16; FIG. 17 is a conceptual diagram of the center frequency at 1/3 octave as another example of the frequency characteristics of the filter section of FIG. 16; FIG. 17 is a diagram showing an example of the distribution of frequency components extracted from the input signal by the filter section of FIG. 16; FIG. 17 is a diagram showing an example of a sequence yk(n) generated by a filter unit at time t1 in FIG. 16; FIG. 17 is a diagram showing an example of a sequence yk(n) generated by a filter unit at time t2 in FIG. 16; FIG. 17 is a flowchart for explaining processing by the control unit of FIG. 16; FIG. It is a figure explaining an example of a hardware configuration. FIG. 11 is a diagram illustrating another hardware configuration example; FIG. 27 is a diagram showing a system configuration example in which at least one of the control units of FIGS. 6 and 16 is incorporated in a terminal device as a specific example of FIG. 25 or 26; FIG. 27 is a diagram showing a system configuration example in which, as a specific example of FIG. 25 or FIG. 26, at least one of the control units of FIGS. 6 and 16 is incorporated in the determiner 401 to supply the processing content of the determiner to the data logger. FIG. 27 is a diagram showing a system configuration example in which at least one of the control units shown in FIGS. 6 and 16 is incorporated in the determiner as a specific example of FIG. 25 or 26 to supply the processing content of the determiner to the elevator control panel;

Embodiment 1.
FIG. 1 is an exploded perspective view showing a probe 1 of a wire rope flaw detector according to Embodiment 1. FIG. The probe 1 has a probe body 3 and a cover 5 .

The cover 5 is made of a non-magnetic material. A cover 5 covers the probe body 3 . Thereby, the cover 5 protects the probe body 3 . The cover 5 is provided with grooves 51 . The cross section of the groove portion 51 is formed in a U shape. The groove 51 has a first end 51_1 and a second end 51_2.

The probe body 3 has a magnetizer 11 and a magnetic sensor 13 .

The magnetizer 11 has a back yoke 111, a first permanent magnet 112_1, a second permanent magnet 112_2, a first pole piece 113_1, and a second pole piece 113_2.

The back yoke 111 is made of ferromagnetic material. The back yoke 111 has a first yoke end portion 111_1, a second yoke end portion 111_2, and a yoke center portion 111_3. One longitudinal end of the back yoke 111 is a first yoke end 111_1. The other longitudinal end of the back yoke 111 is a second yoke end 111_2. The yoke central portion 111_3 is positioned between the first yoke end portion 111_1 and the second yoke end portion 111_2.

A first pole piece 113_1 is fixed to the first yoke end portion 111_1 via a first permanent magnet 112_1. A second pole piece 113_2 is fixed to the second yoke end 111_2 via a second permanent magnet 112_2. As a result, the first permanent magnet 112_1 and the second permanent magnet 112_2 are arranged apart from each other in the longitudinal direction of the back yoke 111 . Also, the first pole piece 113_1 and the second pole piece 113_2 are arranged apart from each other in the longitudinal direction of the back yoke 111 .

The first pole piece 113_1 is made of ferromagnetic material. A first pole piece groove 113_11 is provided in the first pole piece 113_1. The cross section of the first pole piece groove portion 113_11 is formed in a U shape. The first pole piece groove portion 113_11 is fixed to the cover 5 at a position behind the first end portion 51_1.

The second pole piece 113_2 is made of ferromagnetic material. A second pole piece groove 113_21 is provided in the second pole piece 113_2. The cross section of the second pole piece groove portion 113_21 is formed in a U shape. The second pole piece groove portion 113_21 is fixed to the cover 5 at a position behind the second end portion 51_2.

The first permanent magnet 112_1 is arranged between the first pole piece 113_1 and the first yoke end 111_1. The first permanent magnet 112_1 is arranged with one magnetic pole face facing the first pole piece 113_1 and the other magnetic pole face facing the first yoke end portion 111_1. A neodymium magnet, for example, is used as the first permanent magnet 112_1. The first permanent magnet 112_1 generates magnetomotive force.

A second permanent magnet 112_2 is arranged between the second pole piece 113_2 and the second yoke end 111_2. The second permanent magnet 112_2 is arranged with one magnetic pole face facing the second yoke end portion 111_2 and the other magnetic pole face facing the second pole piece 113_2. A neodymium magnet, for example, is used as the second permanent magnet 112_2. The second permanent magnet 112_2 generates magnetomotive force.

The magnetic sensor 13 has a sensor main body 13A and a mounting portion 13B.

The attachment portion 13B is attached to the yoke central portion 111_3. The mounting portion 13B is made of a non-magnetic material.

The sensor main body 13A is arranged between the first pole piece 113_1 and the second pole piece 113_2. The sensor body 13A has a base portion 132, a coil holder 133, a first coil 131_1, and a second coil 131_2.

The base portion 132 is attached to the attachment portion 13B. The coil holder 133 is attached to the base portion 132 . The coil holder 133 is made of a ferromagnetic material. The first coil 131_1 and the second coil 131_2 are attached to the coil holder 133 .

FIG. 2 is an explanatory diagram showing the principle of flaw detection by the probe 1 of FIG. The wire rope flaw detector includes a probe 1 and a controller 9 that receives signals from the probe 1 .

In FIG. 2, the outline of the cover 5 is indicated by a chain double-dashed line for convenience of illustration. In addition, in FIG. 2, the cross-sectional shape portion of the groove portion 51 is indicated by hatching for convenience of illustration. When the wire rope 2 is inspected by the wire rope flaw detector, the wire rope 2 moves relative to the probe 1 in a specific direction W_D along the longitudinal direction of the groove 51 . The probe 1 performs measurement while the wire rope 2 is in contact with the groove portion 51 .

In the example of FIG. 2, the direction of polarity of the first permanent magnet 112_1 is directed from the first yoke end portion 111_1 to the first pole piece 113_1. Also, in the example of FIG. 2, the direction of the polarity of the second permanent magnet 112_2 is the direction from the second pole piece 113_2 toward the second yoke end portion 111_2.

That is, the polarity of the first permanent magnet 112_1 is opposite to the polarity of the second permanent magnet 112_2. Therefore, in the state where the wire rope 2 is arranged in the groove 51, the magnetic flux F passing through the magnetic circuit F_C composed of a part of the wire rope 2 and the magnetizer 11 is transferred to the first permanent magnet 112_1 and the second permanent magnet 112_1. 112_2 is generated.

As a result, when the wire rope 2 is arranged in the groove 51, the section W between the portion of the wire rope 2 facing the first pole piece 113_1 and the portion facing the second pole piece 113_2 , the wire rope 2 is magnetized. In the wire rope 2, the magnetic flux F generated by the first permanent magnet 112_1 and the second permanent magnet 112_2 passes along the wire rope 2 in its longitudinal direction. That is, the magnetizer 11 generates a magnetic flux F that passes through part of the wire rope 2 .

The magnetic sensor 13 generates a signal corresponding to the leakage magnetic flux L_F leaking from the wire rope 2 among the magnetic flux F as a sensor signal. The control unit 9 processes sensor signals generated from the magnetic sensor 13 . A detailed description of the magnetic flux F and the leakage magnetic flux L_F will be given later with reference to FIGS. 3, 4 and 5. FIG.

The first permanent magnet 112_1 and the second permanent magnet 112_2 are collectively referred to as the permanent magnet 112 hereinafter. Also, the first pole piece 113_1 and the second pole piece 113_2 are collectively referred to as the pole piece 113 . Also, the first coil 131_1 and the second coil 131_2 are collectively referred to as the coil 131 .

Next, the detection principle of the leakage magnetic flux L_F by the wire rope flaw detector will be explained using FIGS. 3, 4 and 5. FIG. 3 is an enlarged view of part A in FIG. 2. FIG. As shown in FIG. 3, if there is a damaged portion B_W in a portion of the wire rope 2 through which the magnetic flux F passes, part of the magnetic flux F leaks from the wire rope 2 around the damaged portion B_W as a leakage magnetic flux L_F. .

FIG. 4 is a diagram for more specifically explaining an example of the positional relationship between the leakage magnetic flux L_F and the coil 131 in FIG. When the wire rope 2 is moved with respect to the probe 1, the first coil 131_1 and the second coil 131_2 interlink with the leakage magnetic flux L_F. Therefore, an induced voltage, which is a signal corresponding to the leakage magnetic flux L_F, is generated in the first coil 131_1 and the second coil 131_2 as a sensor signal.

By the way, the wire rope 2 is composed of a core rope and a plurality of strands 21 twisted around the core rope at a constant pitch λ. Therefore, a plurality of projections arranged in the length direction of the wire rope 2 at a constant pitch λ are formed on the outer peripheral portion of the wire rope 2 . Moreover, the strand 21 is configured by twisting a plurality of strands into a single layer or multiple layers. Therefore, if the wires included in the wire rope 2 are thin, the diameter of the wire rope 2 is reduced.

FIG. 5 is a diagram more specifically explaining an example of the positional relationship between the coil 131 and the leakage magnetic flux L_F leaking from the wire rope 2S having a diameter smaller than that of the wire rope 2 of FIG. When the wire rope 2S of FIG. 5 moves relative to the probe 1, there is more leakage interlinking the first coil 131_1 and the second coil 131_2 than when the wire rope 2 of FIG. 4 moves relative to the probe 1. The magnetic flux amount of the magnetic flux L_F is reduced. Therefore, the control unit 9 in FIG. 2 distinguishes whether the sensor signal generated from the magnetic sensor 13 is caused by the magnetic flux F caused by noise or caused by the magnetic flux F caused by the damaged portion B_W. Hard to put on. Therefore, in the present embodiment, the control unit 9 of FIG. 2 performs a mechanical analysis of the correlation between a plurality of feature quantities based on a plurality of values constituting the frequency components of the sensor signal and the state of the wires included in the wire rope. After learning, the presence or absence of damage to the wire is determined.

As described above, the wire rope 2 is constructed by twisting the strands 21 at a constant pitch λ. Therefore, the probe 1 detects noise caused by the outer circumference of the wire rope 2 at least for each pitch λ. Similarly, the wire rope 2S is also twisted at a pitch λS. Accordingly, a plurality of protrusions are similarly formed on the outer peripheral portion of the wire rope 2S and are arranged in the length direction of the wire rope 2S at a constant pitch λS. Therefore, the probe 1 detects noise caused by the outer circumference of the wire rope 2S at least for each pitch λS.

FIG. 6 is a block diagram showing a functional configuration example of the control section 9 of FIG. As shown in FIG. 6 , the control section 9 has a measuring device 91 , a synthesizer 92 , a filter section 93 and a processing section 94 .

The measuring device 91 has a first measuring device 91_1 and a second measuring device 91_2. The first measuring device 91_1 is connected to both ends of the first coil 131_1. A second measuring device 91_2 is connected to both ends of the second coil 131_2. In this example, the first coil 131_1 is located upstream of the second coil 131_2 in the specific direction W_D of the wire rope 2S.

As shown in FIG. 6, when the damaged portion B_W enters between the first pole piece 113_1 and the second pole piece 113_2, leakage magnetic flux L_F leaks from the wire rope 2 around the damaged portion B_W. The leakage magnetic flux L_F links the first coil 131_1 and then the second coil 131_2. Therefore, the time at which the induced voltage peak occurs across the first coil 131_1 is shifted by the delay time τ from the time at which the induced voltage peak occurs across the second coil 131_2. The delay time τ is represented by a value obtained by dividing the distance P between the centers of the first coil 131_1 and the second coil 131_2 by the moving speed ν of the probe 1 .

Therefore, the first measuring device 91_1 detects the induced voltage, which is a signal corresponding to the leakage magnetic flux L_F leaking from the wire rope 2S, as the sensor signal f1(t−τ). The second measuring device 91_2 also detects an induced voltage, which is a signal corresponding to the leakage magnetic flux L_F leaking from the wire rope 2S, as the sensor signal f2(t).

The combiner 92 combines the sensor signal f1(t−τ) detected by the first measuring device 91_1 and the sensor signal f2(t) detected by the second measuring device 91_2 to obtain the sensor signal x (t) is generated. Specifically, the synthesizer 92 delays the sensor signal f1(t−τ) detected by the first measuring device 91_1 by the time τ, and the sensor signal f1(t) detected by the second measuring device 91_2. is superimposed with the sensor signal f2(t). As a result, the sensor signal x(t) becomes a signal that combines the peak of the induced voltage generated across the first coil 131_1 and the peak of the induced voltage generated across the second coil 131_2.

A synthesizer 92 samples the sensor signal x(t) at a constant period Ts. Since the sampled signal becomes a signal for each cycle Ts, the time can be represented by an integer n with the cycle Ts as a unit. That is, the relationship between the analog time of the sensor signal x(t) and the time of the sampled signal is t=n·Ts. The amplitude of the sampled signal is real valued. Therefore, a discrete signal, which is a sampled signal, is expressed as a real-valued sequence {x(0), x(1), x(2), . to The sequence x(n) is supplied to the filter section 93 as the input signal x(n).

The sequence x(n) is hereinafter referred to as the input signal x(n) of the filter section 93 .

The filter unit 93 extracts the frequency component of the input signal x(n) obtained by sampling the sensor signal x(t). The filter section 93 has a plurality of FIR (Finite Impulse Response) filters 931 as a plurality of bandpass filters and a plurality of absolute value sections 932 .

FIG. 7 is a diagram showing an example of frequency characteristics of the filter section 93 of FIG. As shown in FIG. 7, each of the plurality of FIR filters 931 has a constant number of taps, gain and bandwidth b. The plurality of FIR filters 931 have a plurality of different bands as individual passbands. That is, the plurality of FIR filters 931 have individual passbands different from each other. Therefore, the filter unit 93 extracts frequency components of the input signal x(n) in each of a plurality of mutually different bands.

The plurality of absolute value units 932 obtain absolute values of frequency components of the input signal x(n) supplied from the plurality of FIR filters 931, as shown in FIG. Here, the input signal x(n) is a real-valued sequence {x(0), x(1), x(2), . . . }, as described above. Therefore, the plurality of absolute value units 932 quantize the input signal x(n) by dividing it by the quantization unit and rounding it off to obtain the sequence yk(n) of integer values. However, k is a value in ascending order from 1 to N. Also, N is a natural number.

For example, when real-valued x(0) is input to the plurality of FIR filters 931, the plurality of FIR filters 931 extract the frequency component of real-valued x(0) in each of a plurality of different bands. Absolute value unit 932 performs the above operation on the frequency component of real value x(0) for each individual passband, thereby obtaining integer values y1(0), y2(0), . . . Ask for

, y1(0), y2(0), . The sequence is represented as yk(0).

The filter unit 93 performs the same processing on real-valued x(1) to obtain yk(1). The filter unit 93 performs the same processing on the real numbers after x(2), and obtains after yk(2). From the above description, the filter unit 93 obtains the sequence yk(n) from the input signal x(n).

FIG. 8 is a diagram showing an example of distribution of frequency components extracted from the input signal x(n) by the filter section 93 of FIG. As shown in FIG. 8, the moving speed ν of the wire rope 2S is, for example, trapezoidally controlled. Therefore, when the wire rope 2S is moving at a constant speed, the periodic fluctuation of the input signal x(n) due to the shape of the outer circumference of the wire rope 2S is constant. The periodic variation of the input signal x(n) caused by the shape of the outer circumference of the wire rope 2S occurs for each pitch λS of the strands 21S, as described above. A constant periodic variation of the input signal x(n) thus occurs at a particular frequency.

For example, as shown in FIG. 8, the noise frequency component f_n of the input signal x(n) appears in the bands when k=6 and k=8 when the wire rope 2S is moving at a constant speed. do.

Therefore, the constant periodic fluctuation frequency component of the input signal x(n) can be regarded as the noise frequency component f_n of the input signal x(n) among the frequency components of the input signal x(n).

Further, when the input signal x(n) is a signal synthesized by the synthesizer 92 according to the leakage magnetic flux L_F, in the time domain, the signal generated during the local minute time Δt and the input signal x(n ) are equivalent. Therefore, when the input signal x(n) is a signal synthesized by the synthesizer 92 according to the leakage magnetic flux L_F, the shorter the minute time Δt, the more the frequency component of the input signal x(n) appears. increase in number.

For example, as shown in FIG. 8, the damage frequency components f_s of the input signal x(n) are: Appears in each time zone.

As a result, the distribution of the frequency components of the input signal x(n) is distributed over a plurality of bands. Therefore, among the frequency components of the input signal x(n), the damaged frequency component f_s of the input signal x(n) also appears in bands other than the band in which the noise frequency component f_n of the input signal x(n) appears.

By the way, the damaged frequency component f_s of the input signal x(n) originally appears even when k=1 and k=2. However, we do not need to consider up to k=1 and k=2 here. Therefore, in the example of FIG. 8, when k=1 and k=2, the damage frequency component f_s of the input signal x(n) is blocked in the process of various signal processing.

Note that the frequency component distribution is composed of a sequence yk(n). The sequence yk(n) is composed of the integer values y1(n), y2(n), . . . and yN(n), as explained above. Therefore, the frequency component distribution consists of a plurality of values y1(n) to yN(n).

Also, as shown in the middle and lower parts of FIG. 8, it is possible to arrange the sequence yk(n) for k=6 and k=7 over time. The middle part of FIG. 8 shows an example in which the control unit 9 obtains the crest factor from the sequence yk(n) when k=6. The crest factor is a value obtained by calculating the ratio of the maximum value to the effective value. Further, the lower part of FIG. 8 shows an example in which the control unit 9 obtains differentiation or difference from the sequence yk(n) when k=7.

The processing unit 94 has a calculation unit 941 and a learning unit 943, as shown in FIG.

The calculation unit 941 extracts a plurality of feature amounts based on a plurality of values y1(n) to yN(n). Specifically, the calculation unit 941 extracts a plurality of feature quantities from a plurality of values y1(n) to yN(n) through a plurality of statistical calculations. The plurality of feature quantities are representative values that characterize the frequency components of the input signal x(n). A statistical operation is, for example, an operation to find a median. The median value is the value located in the center when the integer values y1(n), y2(n), . Statistical operations include, for example, maximum value, minimum value, range, average value, standard deviation, effective value, crest factor, differentiation, and difference. In addition, among statistical calculations, the range is a value obtained by calculating the difference between the maximum value and the minimum value.

FIG. 9 is a diagram showing an example of the sequence yk(n) generated by the filter unit 93 at time t1 in FIG. In one example of FIG. 9, among a plurality of values y1(n) to yN(n) forming the sequence yk(n), the value y8(n) when k=8 indicates the maximum value, and when k=9 The value y9(n) at this time indicates the minimum value. Therefore, the range is determined by the difference between the value y8(n) and the value y9(n).

FIG. 10 is a diagram showing an example of the sequence yk(n) generated by the filter unit 93 at time t2 in FIG. In the example of FIG. 10, when the values y1(n) to yN(n) forming the sequence yk(n) are arranged in ascending order, the central value is the value when k=5. In addition to median values, FIG. 10 also shows average values, standard deviations, and differentials or differences.

The learning unit 943 inputs a plurality of feature amounts based on a plurality of values y1(n) to yN(n) to the learning model 961, as shown in FIG. The learning model 961 when a plurality of feature quantities are input executes arithmetic processing. The learning unit 943 determines the presence/absence of damage to the wires included in the wire rope 2S based on the result of the arithmetic processing performed by the learning model 961 . The learning model 961 performs machine learning on the correlation between a plurality of feature amounts based on a plurality of values y1(n) to yN(n) constituting frequency components and the state of the wires included in the wire rope 2S. It is something that has already been learned.

The learning unit 943 performs learning by inputting a plurality of feature amounts into a learning model 961 in which at least one of a plurality of first feature amounts and a plurality of second feature amounts is set as a learning data set and performing arithmetic processing. The presence or absence of damage to the wires of the wire rope 2S is determined from the output of the model 961. Here, the multiple first feature amounts are a set of multiple values that characterize that the wires included in the wire rope 2S are damaged. Also, the plurality of second feature amounts is a set of values that characterize that the wires included in the wire rope 2S are not damaged.

Damage to the wires included in the wire rope 2S means physical damage occurring in at least a portion of the wire rope 2S. Physical damage is, for example, at least one of wire breakage, wire partial wire breakage, and wire abrasion marks.

Hereinafter, damage to the wires included in the wire rope 2S is appropriately referred to as damage to the wires.

<Support Vector Machine>
FIG. 11 is a diagram showing the concept of performing machine learning using a learning data set when the learning unit 943 of FIG. 6 is configured as the support vector machine 943_1. The support vector machine 943_1 sets a plurality of first feature amounts and a plurality of second feature amounts as an input as an example of supervised machine learning.

Specifically, as shown in FIG. 11, the learning unit 943 in FIG. 6 classifies the preset feature space into a first region and a second region using the identification hyperplane. It is configured as machine 943_1.

Here, the first area is an area including a plurality of first feature amounts. Specifically, the first region is a region including a first support vector SV_1 that is part of the plurality of first feature quantities. The second area is an area containing a plurality of second feature amounts. Specifically, the second region is a region including the second support vector SV_2 that is part of the plurality of second feature quantities. The discriminative hyperplane is a function that can be represented by, for example, w ^T Γ+b=0, as shown in FIG.

More specifically, the plurality of first feature quantities are represented by a sequence _Γ'q . q is 1, 2, . . . , N _e . N _e is the number of samples with damaged strands. The sequence Γ' _q is composed of a plurality of values γ'1(n), γ'2(n), . . . and γ'M(n).

For example, γ'1(n) is assigned the maximum value in the sequence yk(n). Also, γ′2(n) is assigned the smallest value in the sequence yk(n). Thus, the sequence Γ' _q is composed of values to which values calculated by a plurality of statistical operations are assigned.

Therefore, on a feature space in which there is w ^T Γ+b=0 that can be a discriminative hyperplane, when a part of the plurality of first feature quantities is the first support vector SV_1, the sequence Γ' _q is expressed by the first support vector SV_1 as An expression including w ^T Γ′ _q +b=1, for example.

That is, the support vector machine 943_1 expresses the sequence Γ' _q mapped onto the feature space by the function w ^T Γ' _q +b=1.

On the other hand, the plurality of second feature quantities are represented by a sequence Γ _p . p is 1, 2, . . . , N _o . N _o is the number of samples with no strand damage. The sequence Γ _p is composed of multiple values γ1(n), γ2(n), . . . and γM(n).

For example, γ1(n) is assigned the maximum value in the sequence yk(n). Also, γ2(n) is assigned the minimum value in the sequence yk(n). Thus, the sequence Γ _p is composed of values to which values calculated by a plurality of statistical operations are assigned.

Therefore, on a feature space in which w ^T Γ+b=0 that can be a discriminative hyperplane, the sequence Γ _p includes the second support vector SV_2 when a part of the plurality of second feature amounts is the second support vector SV_2. As a formula, for example, w ^T Γ _p +b=−1.

That is, the support vector machine 943_1 expresses the sequence Γ _p mapped onto the feature space by the function w ^T Γ _p +b=-1.

Note that the support vector machine 943_1 sets the first support vector SV_1 to the value that is closest to one of the values that make up the sequence Γ _p among the values that make up the sequence Γ′ _q . Also, the support vector machine 943_1 sets the value closest to one of the values forming the sequence Γ' _q among the values forming the sequence Γ _p to the second support vector SV_2.

Also, the support vector machine 943_1 obtains w ^T and b that result in the distance d1 and the distance d2. Here, the distance d1 is the distance at which w ^T Γ+b=0 is farthest from the first support vector SV_1. On the other hand, the distance d2 is the distance at which w ^T Γ+b=0 is farthest from the second support vector SV_2. The support vector machine 943_1 computes the identification hyperplane by obtaining such w ^T and b.

That is, the support vector machine 943_1 generates the learning model 961_1 by mapping both the plurality of first feature amounts and the plurality of second feature amounts as the learning data set to the feature space and calculating the discrimination hyperplane.

Specifically, the support vector machine 943_1 determines that there is damage to the wire when the output of the learning model 961_1 when a plurality of feature quantities are input to the learning model 961_1 corresponds to the one classified into the first region. judge. Also, the support vector machine 943_1 determines that there is no damage to the wire when the output of the learning model 961_1 when a plurality of feature quantities are input to the learning model 961_1 corresponds to the second region.

FIG. 12 is a flowchart for explaining processing by the control unit 9 when the learning unit 943 of FIG. 6 is configured as the support vector machine 943_1. The processing of steps S11 to S13 is learning processing. The processing in steps S14 and S15 is feature amount extraction processing. The process of steps S16 to S20 is a wire determination process. The processing of steps S21 and S22 is period determination processing.

<Learning processing>
In step S11, the support vector machine 943_1 determines whether a learning data set has been input. When the support vector machine 943_1 determines that the learning data set has been input, the current process shifts to the process of step S12. When the support vector machine 943_1 determines that the learning data set has not been input, it continues the process of step S11.

In step S12, the support vector machine 943_1 determines whether the learning data set includes both the plurality of first feature amounts and the plurality of second feature amounts. When the support vector machine 943_1 determines that the learning data set includes both the plurality of first feature amounts and the plurality of second feature amounts, the current process shifts to the process of step S13. When the support vector machine 943_1 determines that the learning data set does not include both the plurality of first feature amounts and the plurality of second feature amounts, the current process returns to the process of step S11.

In step S13, the support vector machine 943_1 generates a learning model 961_1 by performing machine learning using the learning data set to calculate a discriminative hyperplane that can be classified into a first region and a second region. The support vector machine 943_1 shifts the current process to the process of step S14.

<Feature amount extraction processing>
In step S14, the calculation unit 941 determines whether or not the frequency component of the sensor signal x(t) has been extracted. When determining that the frequency component of the sensor signal x(t) has been extracted, the calculation unit 941 shifts the current process to the process of step S15. If the calculation unit 941 determines that the frequency component of the sensor signal x(t) has not been extracted, it repeats the process of step S14. In other words, the process of step S14 is a process in which the calculation unit 941 determines whether or not the filter unit 93 has extracted the frequency component of the sensor signal x(t).

In step S15, the calculation unit 941 extracts a plurality of feature quantities from a plurality of values y1(n) to yN(n) forming the frequency components by a plurality of statistical calculations. The calculation unit 941 shifts the current process to the process of step S16.

<Strand determination processing>
In step S16, the support vector machine 943_1 inputs a plurality of feature quantities to the learning model 961_1. The support vector machine 943_1 shifts the current process to the process of step S17. In step S17, the support vector machine 943_1 determines whether or not the output of the learning model 961_1 corresponds to the classification of the plurality of feature quantities into the first region. When the support vector machine 943_1 determines that the output of the learning model 961_1 corresponds to the first area classified into the plurality of feature amounts, the current process shifts to the process of step S18. When the support vector machine 943_1 determines that the output of the learning model 961_1 does not correspond to the first region classified into the plurality of feature amounts, the support vector machine 943_1 shifts the current process to the process of step S19.

In step S18, the support vector machine 943_1 determines that the wire is damaged. The support vector machine 943_1 shifts the current process to the process of step S21. In step S19, the support vector machine 943_1 determines whether or not the output of the learning model 961_1 corresponds to the classification of the plurality of feature quantities into the second region. When the support vector machine 943_1 determines that the output of the learning model 961_1 corresponds to the second region classified into the plurality of feature amounts, the support vector machine 943_1 shifts the current process to the process of step S20. When the support vector machine 943_1 determines that the output of the learning model 961_1 does not correspond to the classification of the plurality of feature quantities into the second area, the current process proceeds to the process of step S21.

In step S20, the support vector machine 943_1 determines that the wires are not damaged. The support vector machine 943_1 shifts the current process to the process of step S21.

<Period determination processing>
In step S21, the filter unit 93 determines whether or not to terminate the determination of the presence or absence of damage to the wires. When the filter unit 93 determines to terminate the determination of the presence or absence of damage to the wires, the current process is terminated. When the filter unit 93 determines not to terminate the determination of the presence or absence of damage to the wires, the current process proceeds to the process of step S22. The filter unit 93 determines whether or not the speed of the wire rope 2S corresponds to the constant speed movement period. When the filter unit 93 determines that the speed of the wire rope 2S corresponds to the constant-velocity movement period, the filter unit 93 causes the current process to return to the process of step S14. When the filter unit 93 determines that the speed of the wire rope 2S does not correspond to the constant speed movement period, it continues the process of step S22.

<Auto Encoder>
FIG. 13 is a diagram showing the concept of performing machine learning using the learning data set when the learning unit 943 of FIG. 6 is configured as the autoencoder 943_2.

Specifically, as shown in FIG. 13, the learning unit 943 in FIG. 6 is configured as an autoencoder 943_2 that suppresses the error between the input and the output to a certain amount by weighting factors wf, wg, and wh. there is The autoencoder 943_2 sets a plurality of second feature amounts as an input as an example of machine learning by unsupervised learning.

More specifically, the autoencoder 943_2 has a layered structure formed by stacking a portion of the autoencoder 962_1, a portion of the autoencoder 962_2, and the final layer 965 .

First, the autoencoder 962_1 has an encoder 963_1 and a decoder 964_1. The autoencoder 962_1 uses, for example, a sigmoid function as an activation function. The autoencoder 943_2 sets the input of the encoder 963_1 to the input of the autoencoder 962_1. The autoencoder 943_2 sets the output of the encoder 963_1 to the input of the decoder 964_1. The autoencoder 943_2 sets the output of the decoder 964_1 to the output of the autoencoder 962_1. The autoencoder 943_2 calculates weighting factors wf, wg, and wh through learning so that the same input to the autoencoder 962_1 is reconstructed on the output side of the autoencoder 962_1.

Autoencoder 943_2 then sets the output of encoder 963_1, which is part of autoencoder 962_1, to the input of autoencoder 962_2. The autoencoder 962_2 has an encoder 963_2 and a decoder 964_2. The autoencoder 962_2 uses, for example, a sigmoid function as an activation function. The autoencoder 943_2 sets the input of the encoder 963_2 to the input of the autoencoder 962_2. The autoencoder 943_2 sets the output of the encoder 963_2 to the input of the decoder 964_2. The autoencoder 943_2 sets the output of the decoder 964_2 to the output of the autoencoder 962_2. Similarly to the autoencoder 962_2, the autoencoder 943_2 learns the weighting coefficient wf, the weighting coefficient wg, and the weighting coefficient wh so that the same input to the autoencoder 962_2 is reproduced on the output side of the autoencoder 962_2. Calculate.

Autoencoder 943_2 then sets the output of encoder 963_2, which is part of autoencoder 962_2, to the input of final layer 965 . Therefore, the autoencoder 943_2 sets the output of the encoder 963_1 to the input of the encoder 963_2. Autoencoder 943_2 adds a final layer 965 to the output of encoder 963_2. The final layer 965 uses the softmax function as the activation function. Next, the auto-encoder 943_2 fine-tunes the weighting factors wf, wg, and wh using error backpropagation in the hierarchical structure of the encoder 963_1, the encoder 963_2, and the final layer 965, thereby learning the learning model 961_2. Generate.

That is, the autoencoder 943_2 learns by calculating the weighting factor wf, the weighting factor wg, and the weighting factor wh by using a plurality of second feature quantities as the learning data set for the input and output of the autoencoder 943_2. Generate model 961_2.

Also, the autoencoder 943_2 determines that there is no damage to the wire when the output of the learning model 961_2 when the plurality of feature amounts are input to the learning model 961_2 can reconstruct the plurality of feature amounts. The autoencoder 943_2 determines that there is damage to the wire when the output of the learning model 961_2 fails to reconstruct the plurality of feature amounts when the plurality of feature amounts are input to the learning model 961_2.

Specifically, the autoencoder 943_2 sets the difference between the plurality of feature amounts and the output of the learning model 961_2 when the plurality of feature amounts are input to the learning model 961_2 as the reconstruction error. The autoencoder 943_2 calculates the root mean square error by averaging the reconstruction error. The autoencoder 943_2 determines that the output of the learning model 961_2 fails to reconstruct a plurality of feature quantities when the mean square error exceeds the error tolerance. The autoencoder 943_2 determines that the output of the learning model 961_2 has reconstructed a plurality of feature quantities when the mean square error is equal to or less than the error tolerance. The error tolerance value is a value set according to a plurality of feature quantities.

FIG. 14 is a flowchart for explaining processing by the control unit 9 when the learning unit 943 in FIG. 6 is configured as the autoencoder 943_2. The processing of steps S41 to S43 is learning processing. The processing of steps S44 and S45 is feature amount extraction processing. The process of steps S46 to S49 is a wire determination process. The processing of steps S50 and S51 is period determination processing. Of the learning process, the feature amount extraction process, the wire determination process, and the period determination process, the period determination process is the same as the processes in steps S21 and S22 of FIG. Therefore, description thereof is omitted.

<Learning processing>
In step S41, the autoencoder 943_2 determines whether or not a learning data set has been input. When the autoencoder 943_2 determines that the learning data set has been input, the current process shifts to the process of step S42. When the autoencoder 943_2 determines that the learning data set has not been input, it continues the process of step S41.

In step S42, the autoencoder 943_2 determines whether or not the learning data set includes a plurality of second feature amounts. When the autoencoder 943_2 determines that the learning data set includes a plurality of second feature amounts, the autoencoder 943_2 shifts the current process to the process of step S43. When the autoencoder 943_2 determines that the learning data set does not include a plurality of second feature amounts, the autoencoder 943_2 returns the current process to the process of step S41.

In step S43, the autoencoder 943_2 performs machine learning using the learning data set to calculate a weighting factor wf, a weighting factor wg, and a weighting factor wh that suppresses the error between the input and the output to a certain amount, thereby obtaining a learning model. 961_2. The autoencoder 943_2 shifts the current process to the process of step S44.

<Feature amount extraction processing>
In step S44, the calculation unit 941 determines whether or not the frequency component of the sensor signal x(t) has been extracted. When determining that the frequency component of the sensor signal x(t) has been extracted, the calculation unit 941 shifts the current process to the process of step S45. If the calculation unit 941 determines that the frequency component of the sensor signal x(t) has not been extracted, it continues the process of step S44.

In step S45, the calculation unit 941 extracts a plurality of feature quantities from a plurality of values y1(n) to yN(n) forming the frequency components by a plurality of statistical calculations. The calculation unit 941 shifts the current process to the process of step S46.

<Strand determination processing>
In step S46, the autoencoder 943_2 inputs a plurality of feature quantities to the learning model 961_2. The autoencoder 943_2 shifts the current process to the process of step S47. In step S47, the autoencoder 943_2 determines whether or not the output of the learning model 961_2 has reconstructed a plurality of feature quantities. When the autoencoder 943_2 determines that the output of the learning model 961_2 has reconfigured a plurality of feature amounts, the autoencoder 943_2 shifts the current process to the process of step S48. When the autoencoder 943_2 determines that the output of the learning model 961_2 has not reconstructed a plurality of feature amounts, the current process proceeds to the process of step S49.

In step S48, the autoencoder 943_2 determines that the wire is not damaged. The autoencoder 943_2 shifts the current process to the process of step S50.

In step S49, the autoencoder 943_2 determines that the wire is damaged. The autoencoder 943_2 shifts the current process to the process of step S50.

As described above, the wire rope flaw detector includes the magnetizer 11 , the magnetic sensor 13 and the controller 9 . The magnetizer 11 generates a magnetic flux F that passes through a portion of the wire rope 2S. The magnetic sensor 13 generates, as a sensor signal x(t), a signal corresponding to leakage magnetic flux L_F leaking from the wire rope 2S out of the magnetic flux F. FIG. The control unit 9 processes the sensor signal x(t).

The control unit 9 has a filter unit 93 that extracts the frequency component of the sensor signal x(t), an arithmetic unit 941 and a learning unit 943 .

The learning unit 943 generates a learned learning model 961 as a pre-stage for determining whether or not the wires are damaged. The learned learning model 961 performs machine learning on the correlation between a plurality of feature amounts based on a plurality of values y1(n) to yN(n) that constitute the frequency components and the state of the wires included in the wire rope 2S. Is going.

The learning unit 943 determines the presence/absence of damage to the wire by causing the learning model 961 that has been trained to perform arithmetic processing when a plurality of feature quantities are input to the learning model 961 . Here, the plurality of feature amounts are extracted based on the plurality of values y1(n) to yN(n) that constitute the frequency components extracted at different timings from the previous stage of the determination of the presence or absence of damage to the wire. there is

That is, the processing unit 94 performs arithmetic processing of the learning model 961 when the input of the plurality of feature amounts based on the plurality of values y1(n) to yN(n) constituting the frequency component is used as the learned learning model 961. The presence or absence of damage to the wire is determined from the execution result. Here, the learning model 961 performs machine learning on the correlation between a plurality of feature quantities based on a plurality of values y1(n) to yN(n) constituting frequency components and the state of the wires included in the wire rope 2S. have learned.

Therefore, the learning model 961 can input a plurality of feature amounts based on a plurality of values y1(n) to yN(n) that constitute the frequency components, and can determine the number of wires included in the wire rope 2S without performing complicated calculations. It is possible to output the result of predicting the state.

Therefore, in the time domain, even if the S/N ratio is low because the induced voltage generated in the magnetic sensor 13 is low, the following results are obtained. The processing unit 94 can determine the presence/absence of damage to the wire while hiding the noise frequency component f_n based on the output from the learning model 961 to which the frequency component in the frequency domain is input.

From the above description, the wire rope flaw detector can more reliably improve the SN ratio.

Further, the calculation unit 941 extracts a plurality of feature quantities from a plurality of values y1(n) to yN(n) through a plurality of statistical calculations. The learning unit 943 uses at least one of a plurality of first feature quantities that characterize the presence of damage to the wire and a plurality of second feature quantities that characterize the absence of damage to the wire as a data set for learning. The presence or absence of damage to the wire is determined from the output of the learning model 961 .

That is, the learning unit 943 inputs a plurality of feature amounts to the learning model 961 in which at least one of the plurality of first feature amounts and the plurality of second feature amounts is set as a learning data set. The presence or absence of damage to the wire is determined based on the output of Here, the plurality of first feature amounts characterize that there is damage to the wire. A plurality of second feature quantities characterize that the wires are not damaged.

Therefore, the learning unit 943 can clearly determine the presence or absence of damage to the wire from a side different from the leakage magnetic flux L_F in the time domain. Therefore, the wire rope flaw detector can more reliably improve the SN ratio.

In addition, the learning unit 943 can classify the feature space set in advance into a first region including a plurality of first feature values and a second region including a plurality of second feature values using the identification hyperplane. support vector machine 943_1. The support vector machine 943_1 generates a learning model 961_1 by mapping both a plurality of first feature quantities and a plurality of second feature quantities as a learning data set to a feature space and calculating a discrimination hyperplane.

Therefore, the support vector machine 943_1 can classify, using the identification hyperplane, whether a plurality of feature quantities correspond to the case where the strand is damaged or the case where the strand is not damaged, using the learning model 961_1. Therefore, the wire rope flaw detector can improve the generalization ability for determining the presence or absence of damage to the wire.

Also, the support vector machine 943_1 determines that the wire is damaged when the output of the learning model 961_1 when a plurality of feature quantities are input to the learning model 961_1 corresponds to the one classified into the first region. The support vector machine 943_1 determines that there is no wire damage when the output of the learning model 961_1 when a plurality of feature quantities are input to the learning model 961_1 corresponds to the second region.

Therefore, the support vector machine 943_1 determines whether or not the wires are damaged depending on whether the output of the learning model 961_1 corresponds to the classification of the plurality of feature quantities into either the first region or the second region. do. Therefore, the support vector machine 943_1 can determine the presence or absence of damage to the wire depending on which region it belongs to. Therefore, the wire rope flaw detection apparatus can replace the process of determining the presence or absence of damage to the wire with a simple classification process, thereby reducing erroneous determinations.

Also, the learning unit 943 is configured as an autoencoder 943_2 that suppresses the error between the input and the output to a certain amount by weighting coefficient wf, weighting coefficient wg, and weighting coefficient wh. The autoencoder 943_2 computes the weighting factor wf, the weighting factor wg, and the weighting factor wh by using a plurality of second feature quantities as the learning data set for the input and output of the autoencoder 943_2, respectively, thereby learning the learning model 961_2. to generate

Therefore, the autoencoder 943_2 can indicate in the output which of the plurality of feature amounts corresponds to the case where the strand is damaged or the case where the strand is not damaged. Therefore, the autoencoder 943_2 can determine whether or not the wire is damaged by reducing the number of dimensions by weighting each of the plurality of feature amounts.

From the above description, the wire rope flaw detection apparatus can reduce the amount of calculation required for determination, and emphasizes the feature amount that contributes to the determination of the presence or absence of damage to the strand among the plurality of feature amounts. It is possible to improve the accuracy of determining the presence or absence of damage.

Also, the autoencoder 943_2 determines that there is no damage to the wire when the output of the learning model 961_2 when the plurality of feature amounts are input to the learning model 961_2 can reconstruct the plurality of feature amounts. The autoencoder 943_2 determines that there is damage to the wire when the output of the learning model 961_2 fails to reconstruct the plurality of feature amounts when the plurality of feature amounts are input to the learning model 961_2.

Therefore, the autoencoder 943_2 determines the presence/absence of damage to the wire according to whether or not the output of the learning model 961_2 has reconstructed a plurality of feature quantities. Therefore, the wire rope flaw detector can determine the presence or absence of damage to the wire even if each of the plurality of feature quantities has a non-linear relationship.

Note that the autoencoder 943_2 may perform the following processing when the learning model 961_2 determines that the wire is damaged. The autoencoder 943_2 estimates by sparse optimization which input dimension is the factor in determining that there is damage to the wire. Thereby, the autoencoder 943_2 can use the estimation result for more detailed analysis of the disconnection point of the wire rope 2 .

In addition, the filter unit 93 has a plurality of bandpass filters having a plurality of different bands as individual passbands. Therefore, the filter unit 93 can extract frequency components in a plurality of mutually different bands. Therefore, the wire rope flaw detector can analyze the induced voltage generated by the magnetic sensor 13 as the sensor signal x(t) in the frequency domain.

Further, the calculation unit 941 calculates at least one of the total value, the average value, and the median value of the plurality of values y1(n) to yN(n) that constitute the frequency components, by performing a plurality of statistical calculations. Require at least one.

Therefore, the calculation unit 941 does not compare all of the plurality of values y1(n) to yN(n) that constitute the frequency component, but uses a plurality of different representative values corresponding to the types of the plurality of statistical calculations. Extract values as multiple features.

Therefore, since the wire rope flaw detection apparatus uses a plurality of representative values extracted from various aspects as a plurality of feature values, it is possible to significantly improve the accuracy of determining the presence or absence of damage to the wire.

Embodiment 2.
In the second embodiment, descriptions of configurations and functions that are the same as or equivalent to those of the first embodiment are omitted. Embodiment 2 differs from Embodiment 1 in that one of learning model 961_1 and learning model 961_2 of Embodiment 1 is already stored in learning unit 943 . Other configurations are the same as those of the first embodiment. In other words, the rest of the configuration is the same as or equivalent to that of the first embodiment, and these portions are given the same reference numerals.

FIG. 15 is a flowchart for explaining processing by the control unit 9 in the second embodiment. As described above, either one of the learning model 961_1 and the learning model 961_2 is already stored in the learning unit 943 before the process of step S61 in FIG. 15 is started. Therefore, when inspecting the wire rope 2S, the learning process comprising the processes of steps S11 to S13 as described with reference to FIG. 12 is unnecessary. Similarly, when inspecting the wire rope 2S, the learning process comprising the processes of steps S41 to S43 as described with reference to FIG. 14 is unnecessary. Also, in FIG. 15, the period determination process composed of the processes of steps S21 and S22 as described with reference to FIG. 12 is omitted. Similarly, FIG. 15 omits the period determination process including the processes of steps S50 and S51 as described with reference to FIG.

Specifically, the processing of steps S61 and S62 is feature amount extraction processing. The processing of steps S61 and S62 is the same as the processing of steps S44 and S45. Therefore, description thereof is omitted.

Further, the processes of step S63 and steps S65 to S67 are wire determination processes. The processing of steps S63 and steps S65 to S67 is the same as the processing of steps S46 to S49. Therefore, description thereof is omitted. Here, the processing of steps S65 to S67 is executed by the learning model 961_2 generated from the autoencoder 943_2.

Further, the processes of step S63 and steps S68 to S71 are wire determination processes. The processing of steps S63 and steps S68 to S71 is the same as the processing of steps S16 to S20. Therefore, description thereof is omitted. Here, the processing of steps S68 to S71 is executed by the learning model 961_1 generated from the support vector machine 943_1.

In step S64, the control unit 9 determines whether the learning model 961 is the learning model 961_2 generated by the autoencoder 943_2 or the learning model 961_1 generated by the support vector machine 943_1. If the control unit 9 determines that the learning model 961 is the learning model 961_2 generated by the autoencoder 943_2, the process of step S64 proceeds to the process of step S65. On the other hand, when the control unit 9 determines that the learning model 961 is the learning model 961_1 generated from the support vector machine 943_1, the process of step S64 proceeds to the process of step S68.

In this way, by using either one of the learning model 961_1 and the learning model 961_2, the sensor signal x(t) is input as an input to the control unit 9, and the damage of the wire is output from the control unit 9. is output.

Therefore, either one of the learning model 961_1 and the learning model 961_2 can be utilized for fault diagnosis of the wire rope 2S.

From the above description, one of the learning model 961_1 and the learning model 961_2 is stored in the learning unit 943, so that when the sensor signal x(t) is input to the control unit 9, the strand is output. As a result, either one of the learning model 961_1 and the learning model 961_2 can be utilized for fault diagnosis of the wire rope 2S.

Embodiment 3.
In the third embodiment, descriptions of configurations and functions that are the same as or equivalent to those of the first and second embodiments are omitted. Embodiment 3 differs from Embodiments 1 and 2 in that the bandpass filters of Embodiments 1 and 2 are implemented by wavelet transform. Other configurations are similar to those of the first and second embodiments. In other words, the rest of the configuration is the same as or equivalent to that of the first embodiment, and these portions are given the same reference numerals.

FIG. 16 is a block diagram showing a functional configuration example of the control section 9 that processes a signal corresponding to the leakage magnetic flux L_F leaking from the wire rope 2S according to the third embodiment. As shown in FIG. 16, the filter unit 193 includes a wavelet transform unit 1931 as a bandpass filter that generates a distribution of frequency components of the sensor signal x(t) by performing wavelet transform on the sensor signal x(t). I have.

FIG. 17 is a diagram showing an example of frequency characteristics of the filter section 193 of FIG. The band-pass filter is implemented by a wavelet transform basis function processed by the wavelet transform unit 1931 . As shown in FIG. 17, the bandwidth b _k of each of the multiple bands becomes narrower as the center frequency ω _ck of the band becomes lower.

Next, Morlet Wavelet will be described with reference to FIG. 18 as an example of basis functions for wavelet transform. FIG. 18 is a diagram showing an example of a time-domain waveform of the mother wavelet by the wavelet transform unit 1931 in FIG. The mother wavelet is represented by Equation (1).

A daughter wavelet is represented by the following equation (2). The daughter wavelet scale is represented by the following equation (3). The daughter wavelet expressed in equation (2) can expand or contract the amplitude of the waveform shown in FIG. 18 according to the scale expressed in equation (3). Also, the daughter wavelet represented by equation (2) can translate the waveform shown in FIG. 18 in the time axis direction according to the scale represented by equation (3). where s ₀ is the scale constant. s _k is a scale function with k as argument and multiplied by s ₀ .

Next, the Fourier transforms of the mother wavelet and daughter wavelet are explained as follows. First, the equation obtained by Fourier transforming the mother wavelet is represented by the following equation (4). On the other hand, the Fourier transform of the daughter wavelet is represented by Equation (5).

FIG. 19 is a diagram showing an example of a frequency domain waveform of the mother wavelet by the wavelet transform unit 1931 in FIG. As shown in FIG. 19, the frequency characteristic of the Morlet Wavelet is the frequency of the passband specified by the bandwidth b _k and the center frequency ω ₀ of the bandwidth b _k among the frequency components of the input signal x(n). It becomes a band-pass filter to pass. The center frequency ω _ck in FIG. 19 is represented by the following equation (6). As shown in Equation (6), the center frequency ω _ck is expressed by a value obtained by dividing ω ₀ /s ₀ by the m-th root of 2. Here, as explained above, m is a natural number.

Also, the bandwidth b _k in FIG. 19 is represented by the following equation (7).

Here, equations (6) and (7) for m=1 are explained as follows. First, the following formula (8) is a formula when m=1 in formula (6). From the expression (8), the center frequency ω _ck is expressed by dividing ω ₀ /s ₀ by 2.

On the other hand, the following formula (9) is a formula when m=1 in formula (7). From the description of equation (9), the bandwidth b _k is expressed by the value obtained by dividing the square root of the natural logarithm of 2 by 2 and dividing by s ₀ .

Therefore, since Δk=1/m according to equation (7), equation (7) becomes Δk=1/m=1/1=1. Therefore, the relationship between the bandwidth b _k of one of the two adjacent bands and the bandwidth b _k+1 of the other band whose center frequency ω _ck is lower than that of the other band is b _k+ It satisfies the relationship of ₁ =2 ^-Δk · _bk =bk ₊₁ = ^2-1 · _bk . Therefore, when m=1, the relationship between the bandwidth b _k of one band and the bandwidth b _k+1 of the other band is one octave.

Specifically, in Equation (8), when k=0, the center frequency ω _ck becomes the frequency ω ₀ /s ₀ . Also, in Equation (9), when k=0, the bandwidth b _k is represented by Equation (10) below. Therefore, from equations (8) and (9), the center frequency ω _ck and the bandwidth b _k are halved each time k is increased by one.

Also, since Δk=1/m, when m is a natural number other than 1, the following relationship holds. Of two bands adjacent to each other, the relationship between the bandwidth _bk of one band and the bandwidth _bk+1 of the other band whose center frequency _ωck is lower than that of the other band is: bk ₊₁ = It satisfies the relationship of 2 ^−Δk ·b _k =b _k+1 =2 ^−1/m ·b _k . Therefore, when m is other than 1, the relationship between the bandwidth b _k of one band and the bandwidth b _k+1 of the other band is 1/m octave. The center frequency ω _ck when m=3 is explained as follows.

FIG. 20 is a conceptual diagram of the center frequency ω _ck at ⅓ octave as another example of the frequency characteristics of the filter section 193 of FIG. 16 . As shown in FIG. 20, the center frequency ω _ck can be expressed by a value obtained by dividing ω ₀ /s ₀ by the cube root of 2.

From the above description, the center frequency ω _ck and the center frequency ω _ck+1 of two bands adjacent to each other are represented by the following equation (11). Equation (11) expresses the difference in magnitude between the center frequency ω _ck and the center frequency ω _ck+1 of two adjacent bands based on Equation (8). From equation (11), the center frequency ω _ck becomes 2 ^−1/m each time k increases by 1.

Also, the bandwidth b _k and the bandwidth b _k+1 in each of the two bands adjacent to each other are represented by the following equation (12). Equation (12) expresses the difference in magnitude between the bandwidth b _k and the bandwidth b _k+1 of two adjacent bands based on Equation (9). From equation (12), the bandwidth b _k becomes 2 ^−1/m for each increment of k.

FIG. 21 is a diagram showing an example of distribution of frequency components extracted from the signal according to the leakage magnetic flux L_F by the filter unit 193 of FIG. The example of FIG. 21 is similar to the example of FIG. 8 except for the bandwidth b _k and the center frequency ω _ck of two bands adjacent to each other. Therefore, description of FIG. 21 is omitted.

FIG. 22 is a diagram showing an example of the sequence yk(n) generated by the filter unit 193 at time t1 in FIG. In the example of FIG. 22, among a plurality of values y1(n) to yN(n) forming the sequence yk(n), the value y6(n) when k=6 is the maximum value, k=1, The values y1(n), y2(n), y3(n), y5(n) and y7(n) at 2, 3, 5 and 7 show the minimum values. Therefore, the range is determined by the difference between the value y6(n) and the values y1(n), y2(n), y3(n), y5(n) and y7(n).

FIG. 23 is a diagram showing an example of the sequence yk(n) generated by the filter unit 193 at time t2 in FIG. In the example of FIG. 23, when a plurality of values y1(n) to yN(n) forming the sequence yk(n) are arranged in ascending order, the central value is the value when k=5. Therefore, the value when k=5 is adopted as the median value.

FIG. 24 is a flow chart for explaining the processing by the controller 9 of FIG. Note that the process of step S81 is a learning process. The processing from step S82 to step S84 is feature amount extraction processing. The process of step S85 is a wire determination process. The process of step S86 is a period determination process. Of the learning process, feature quantity extraction process, wire determination process, and period determination process, the learning process, wire determination process, and period determination process are as follows. That is, the learning process, wire determination process, and period determination process are a series of processes including steps S11 to S13, steps S16 to S20, and steps S21 and S22 in FIG. Also, the learning process, wire determination process, and period determination process are a series of processes consisting of steps S41 to S43, steps S46 to S49, and steps S50 and S51 in FIG. Therefore, the learning process, the wire determination process, and the period determination process are one of the series of processes described above. Therefore, their description is omitted.

<Feature amount extraction processing>
In step S<b>82 , the synthesizer 92 inputs the input signal x(n) corresponding to the sensor signal x(t) to the filter section 193 . The combiner 92 shifts the current process to the process of step S83.

In step S<b>83 , the wavelet transform unit 1931 of the filter unit 193 extracts the frequency component of the input signal x(n). The filter unit 193 shifts the current process to the process of step S84. The description of the processing executed by the absolute value unit 932 between the processing of step S83 and the processing of step S84 is omitted.

In step S84, the computing unit 941 extracts a plurality of feature quantities from the multiple values y1(n) to yN(n) composed of the frequency components extracted by the wavelet transforming unit 1931 through a plurality of statistical operations. The calculation unit 941 shifts the current process to the process of step S85.

From the above explanation, in the wire rope flaw detector, the bandwidth b _k of each of the plurality of bands becomes narrower as the center frequency ω _ck of the band becomes lower. Therefore, the lower the center frequency ω _ck of the band, the higher the frequency resolution and the lower the time resolution. The higher the center frequency ω _ck of the band, the lower the frequency resolution and the higher the time resolution. Therefore, the wire rope flaw detector can more accurately detect where on the time axis abrupt fluctuations occur, and more accurately determine the frequency of slow fluctuations, thus enabling efficient analysis.

The relationship between the bandwidth b _k of one of the two adjacent bands and the bandwidth b _k+1 of the other band having a center frequency lower than that of the other band is Δk=1/m , the relationship b _k+1 =2 ^−Δk ·b _k is satisfied. Therefore, the band can be changed by the mth root of 2. Therefore, the temporal resolution is significantly improved in the high frequency region, and the spatial resolution is particularly significantly improved in the low frequency region.

Further, the filter unit 193 extracts frequency components from the sensor signal x(t) by performing wavelet transform on the sensor signal x(t). Since the wavelet is a local function, there is a high correlation between the wavelet and the detection of the damaged portion B_W of the wire that occurs locally. Therefore, the filter unit 193 can emphasize the damage frequency component f_s among the frequency components. Therefore, the wire rope flaw detector can emphasize the frequency component of the induced voltage generated when the wire is damaged, so that the SN ratio can be significantly improved.

Also, in each embodiment, the function of each part of the wire rope flaw detector is implemented by a processing circuit. That is, the wire rope flaw detection apparatus includes a processing circuit for executing the synthesizer 92 , the filter section 93 , the filter section 193 , the calculation section 941 and the learning section 943 . The processing circuit, even if it is dedicated hardware, is a CPU that executes programs stored in memory (Central Processing Unit, also referred to as a central processing unit, processing unit, arithmetic unit, microprocessor, microcomputer, processor, DSP) may be

FIG. 25 is a diagram illustrating a hardware configuration example. In FIG. 25, processing circuitry 201 is connected to bus 202 . If processing circuitry 201 is dedicated hardware, processing circuitry 201 may be, for example, a single circuit, multiple circuits, a programmed processor, an ASIC, an FPGA, or a combination thereof. Each function of each part of the wire rope flaw detector may be realized by the processing circuit 201 , or the functions of each part may be collectively realized by the processing circuit 201 .

FIG. 26 is a diagram illustrating another hardware configuration example. In FIG. 26, processor 203 and memory 204 are connected to bus 202 . When the processing circuit is a CPU, the function of each part of the wire rope flaw detector is implemented by software, firmware, or a combination of software and firmware. Software or firmware is written as a program and stored in memory 204 . The processing circuit reads out and executes a program stored in the memory 204 to implement the function of each unit. That is, when the wire rope flaw detector is executed by the processing circuit, the step of controlling the combiner 92, the filter unit 93, the filter unit 193, the calculation unit 941, and the learning unit 943 is executed as a result. A memory 204 is provided for storing programs. Further, it can be said that these programs cause a computer to execute a procedure or method for executing the synthesizer 92, the filter section 93, the filter section 193, the calculation section 941, and the learning section 943. Here, the memory 204 is, for example, non-volatile or volatile semiconductor memory such as RAM, ROM, flash memory, EPROM, EEPROM, magnetic disk, flexible disk, optical disk, compact disk, mini disk, DVD, etc. Applicable.

A part of the function of each part of the wire rope flaw detector may be realized by dedicated hardware, and another part may be realized by software or firmware. For example, the filter unit 93 and the filter unit 193 can realize their functions by a processing circuit as dedicated hardware. The functions of the arithmetic unit 941 and the learning unit 943 can be realized by the processing circuit reading and executing a program stored in the memory 204 .

As such, the processing circuitry may implement each of the above functions through hardware, software, firmware, or a combination thereof. Next, an example of realizing each of the above functions will be specifically explained as follows.

FIG. 27 is a diagram showing a system configuration example in which the controller 9 of at least one of FIGS. 6 and 16 is incorporated in a terminal device 501 as a specific example of FIG. 25 or 26 and used. As shown in FIG. 27, the wire rope flaw detector detects damage to the wire rope 2S with a probe 1. FIG. The wire rope 2S suspends, for example, an elevator car. In addition, the wire rope 2S may be used for a crane.

The probe 1 detects damage to the wires while moving along, for example, a specific direction W_D with respect to the wire rope 2S. The probe 1 supplies, for example, a sensor signal x(t), which is an analog signal, to the AD converter 301 via a cable. AD converter 301 converts an analog signal into a digital signal. A digital signal converted by the AD converter 301 is input to the terminal device 501 . As the terminal device 501, for example, a personal computer is used. The terminal device 501 performs various signal processing on the digital signal input from the AD converter 301 to determine whether or not the wires are damaged. In addition, the terminal device 501 displays the determination result of the presence/absence of damage to the wires.

FIG. 28 is a system configuration example of supplying the processing contents of the determiner 401 to the data logger 601 by incorporating at least one of the control units 9 of FIGS. 6 and 16 into the determiner 401 as a specific example of FIG. It is a figure which shows. The probe 1 supplies the sensor signal x(t), for example composed of an analog signal, to the determiner 401 via a cable. The decision device 401 is equipped with a microcomputer. The determiner 401 is dedicated hardware. A decision device 401 converts an analog signal into a digital signal. The determiner 401 determines whether or not the wire is damaged by performing various signal processing on the converted digital signal. Also, the determiner 401 notifies the determination result of the presence/absence of damage to the wires.

Note that the determiner 401 can supply various internally processed signals to an external device as analog signals or digital signals. For example, a data logger 601 is used as the external device. The data logger 601 can display a waveform by inputting an analog signal or a digital signal from the determination device 401 . Also, the data logger 601 can record the processing content of the determination device 401 .

FIG. 29 shows a system configuration in which at least one of the controllers 9 shown in FIGS. 6 and 16 is incorporated in a judgment device 401 as a specific example of FIG. FIG. 4 is a diagram showing an example; The elevator control panel 701 receives a digital signal from the determiner 401 and can transmit monitoring information such as which wire rope 2 of which property is broken to the central monitoring center.

Although the wire rope flaw detector has been described above based on the first and second embodiments, the present invention is not limited to this.

In Embodiments 1 and 2, an example in which the learning unit 943 calculates the learning model 961 that can be applied to other data and then determines whether or not there is a wire in the target data has been described. not to be For example, the learning unit 943 may determine the presence/absence of a wire using the Mahalanobis distance MD after defining the unit space by the calculation of the MT method.

<MT method>
Specifically, the learning unit 943 defines a unit space when using the MT method. The learning unit 943 calculates the distance from the center of the unit space to the target data as the Mahalanobis distance MD. The learning unit 943 determines that those with a close Mahalanobis distance MD are normal. On the other hand, the learning unit 943 determines an abnormality when the Mahalanobis distance MD is long.

More specifically, the learning unit 943 acquires the sensor signal x(t) and a plurality of values y1(n) to yN ( n) is used to determine the presence or absence of damage to the wire.

First, the computing unit 941 computes normalized values obtained by normalizing the sensor signal x(t) and the plurality of values y1(n) to yN(n). The calculation unit 941 calculates the normalized value by (raw data-average value)/standard deviation. Next, the computing unit 941 computes the correlation matrix R using the set of normalized values as a unit space. Next, the computing section 941 computes the inverse matrix R ⁻¹ of the correlation matrix R. The calculated inverse matrix R ⁻¹ is held by the learning unit 943 .

Next, the calculation unit 941 calculates the sensor signal x(t) of the target data for the presence or absence of damage to the wire, and a plurality of values y1(n) to A matrix Y is generated by normalizing each of yN(n) in the same manner as described above. Next, the learning unit 943 multiplies the matrix Y obtained by normalizing the target data, the inverse matrix R ⁻¹ calculated in advance, and the transposed matrix Y ^T of the matrix Y obtained by normalizing the target data, and obtains the number of items. By dividing, the Mahalanobis distance MD is calculated. The learning unit 943 determines whether or not the wires are damaged by comparing a preset threshold distance with the Mahalanobis distance MD. Here, the number of items is the number of parameters used for normalization calculation, which is 1+N in the above example.

In other words, the control unit 9 derives the correspondence relationship between the input data and the presence or absence of damage to the wire using the correlation between various data. As a result, even when other input data is input, the control unit 9 can reliably determine the presence or absence of damage to the wires based on the derived correspondence relationship. Therefore, the wire rope flaw detector can realize high generalization ability.

2, 2S wire rope, 11 magnetizer, 13 magnetic sensor, 9 control section, 93, 193 filter section, 94 processing section, 941 calculation section, 943 learning section, 943_1 support vector machine, 943_2 autoencoder, 961, 961_1, 961_2 learning model.

Claims (9)

a magnetizer that generates magnetic flux through a portion of the wire rope;
a magnetic sensor that generates, as a sensor signal, a signal corresponding to leakage magnetic flux leaking from the wire rope among the magnetic flux;
a control unit that processes the sensor signal;
with
The control unit
a filter unit for extracting frequency components of the sensor signal;
a calculation unit that extracts a plurality of feature amounts based on a plurality of values that constitute the frequency component;
Based on a plurality of values constituting frequency components extracted by the filter unit, in a learned learning model that has performed machine learning on the correlation between the plurality of feature values and the state of the wires included in the wire rope. a learning unit that determines the presence or absence of damage to the wire by executing arithmetic processing by the learning model when a plurality of feature values extracted by the arithmetic unit are input;
has
The learning unit divides a preset feature space into a first region including a plurality of first feature quantities that characterize the presence of damage to the wire, and a plurality of first regions that characterize the absence of damage to the wire. Classify into a second region containing a second feature amount ,
The computing unit provides a total value of the plurality of values, a maximum value, a minimum value, a range that is a difference between the maximum value and the minimum value, an average value, a standard deviation, an effective value, a crest factor, a differential value, a difference, and a wire rope flaw detector that obtains at least one median value as at least one of the plurality of feature quantities . a magnetizer that generates magnetic flux through a portion of the wire rope;
a magnetic sensor that generates, as a sensor signal, a signal corresponding to leakage magnetic flux leaking from the wire rope among the magnetic flux;
a control unit that processes the sensor signal;
with
The control unit
a filter unit for extracting frequency components of the sensor signal;
a calculation unit that extracts a plurality of feature amounts based on a plurality of values that constitute the frequency component;
Based on a plurality of values composing the frequency components extracted by the filter unit, in a learning model that has undergone machine learning on the correlation between the plurality of feature values and the state of the wires included in the wire rope. a learning unit that determines the presence or absence of damage to the wire by executing arithmetic processing by the learning model when a plurality of feature values extracted by the arithmetic unit are input;
has
The learning unit divides a preset feature space into a first region including a plurality of first feature quantities that characterize the presence of damage to the wire, and a plurality of first regions that characterize the absence of damage to the wire. Classify into a second region containing a second feature amount,
The filter unit has a plurality of bandpass filters each having a plurality of different bands as individual passbands,
The wire rope flaw detector, wherein the bandwidth of each of the plurality of bands is narrower as the center frequency of the band is lower. The learning unit determines the presence or absence of damage to the wire from the output of the learning model that uses at least one of the plurality of first feature amounts and the plurality of second feature amounts as a learning data set. The wire rope flaw detector according to claim 1 or 2 . The learning unit is configured as a support vector machine that can be classified into the first region and the second region by the identification hyperplane,
The support vector machine maps both the plurality of first feature quantities and the plurality of second feature quantities as the learning data set to the feature space and computes the discrimination hyperplane to create the learning model. 4. The wire rope flaw detector according to claim 3 . The support vector machine determines that there is damage to the wire when the output of the learning model when the plurality of feature quantities are input to the learning model corresponds to the one classified in the first region. 5. The method according to claim 4 , wherein it is determined that there is no damage to the wire when the output of the learning model when the plurality of feature quantities are input to the learning model corresponds to the one classified into the second region. wire rope flaw detector. The learning unit is configured as an auto-encoder that suppresses an error between the input and the output to a certain amount by a weighting factor,
4. The autoencoder generates the learning model by calculating the weighting factor by using the plurality of second feature quantities as the learning data set for the input and output of the autoencoder, respectively. The wire rope flaw detector described in . The autoencoder determines that there is no damage to the wire when the output of the learning model when the plurality of feature amounts are input to the learning model can reconstruct the plurality of feature amounts, and the plurality of 7. The wire rope flaw detection according to claim 6 , wherein it is determined that there is damage to the strand if the output of the learning model when the feature amount of is input to the learning model cannot reconstruct the plurality of feature amounts. Device. Of the two adjacent bands, the relationship between the bandwidth _bk of one band and the bandwidth bk ₊₁ of the other band whose center frequency is lower than that of the one band is given by m being a natural number and Δk= 1/m, then
3. The wire rope flaw detector according to claim 2 , which satisfies the relationship b _{k+1 =2−Δk·b} ^k _. The wire rope flaw detector according to claim 2 or 8 , wherein the filter unit extracts the frequency component from the sensor signal by performing a wavelet transform on the sensor signal.

Images